1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/WinEHFuncInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 static cl::opt<int> ReciprocalEstimateRefinementSteps(
71 "x86-recip-refinement-steps", cl::init(1),
72 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
73 "result of the hardware reciprocal estimate instruction."),
76 // Forward declarations.
77 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
80 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
81 const X86Subtarget &STI)
82 : TargetLowering(TM), Subtarget(&STI) {
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
87 // Set up the TargetLowering object.
88 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
90 // X86 is weird. It always uses i8 for shift amounts and setcc results.
91 setBooleanContents(ZeroOrOneBooleanContent);
92 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
93 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
95 // For 64-bit, since we have so many registers, use the ILP scheduler.
96 // For 32-bit, use the register pressure specific scheduling.
97 // For Atom, always use ILP scheduling.
98 if (Subtarget->isAtom())
99 setSchedulingPreference(Sched::ILP);
100 else if (Subtarget->is64Bit())
101 setSchedulingPreference(Sched::ILP);
103 setSchedulingPreference(Sched::RegPressure);
104 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
105 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
107 // Bypass expensive divides on Atom when compiling with O2.
108 if (TM.getOptLevel() >= CodeGenOpt::Default) {
109 if (Subtarget->hasSlowDivide32())
110 addBypassSlowDiv(32, 8);
111 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
112 addBypassSlowDiv(64, 16);
115 if (Subtarget->isTargetKnownWindowsMSVC()) {
116 // Setup Windows compiler runtime calls.
117 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
118 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
119 setLibcallName(RTLIB::SREM_I64, "_allrem");
120 setLibcallName(RTLIB::UREM_I64, "_aullrem");
121 setLibcallName(RTLIB::MUL_I64, "_allmul");
122 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
123 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
124 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
125 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
126 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
128 // The _ftol2 runtime function has an unusual calling conv, which
129 // is modeled by a special pseudo-instruction.
130 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
131 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
132 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
133 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
136 if (Subtarget->isTargetDarwin()) {
137 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
138 setUseUnderscoreSetJmp(false);
139 setUseUnderscoreLongJmp(false);
140 } else if (Subtarget->isTargetWindowsGNU()) {
141 // MS runtime is weird: it exports _setjmp, but longjmp!
142 setUseUnderscoreSetJmp(true);
143 setUseUnderscoreLongJmp(false);
145 setUseUnderscoreSetJmp(true);
146 setUseUnderscoreLongJmp(true);
149 // Set up the register classes.
150 addRegisterClass(MVT::i8, &X86::GR8RegClass);
151 addRegisterClass(MVT::i16, &X86::GR16RegClass);
152 addRegisterClass(MVT::i32, &X86::GR32RegClass);
153 if (Subtarget->is64Bit())
154 addRegisterClass(MVT::i64, &X86::GR64RegClass);
156 for (MVT VT : MVT::integer_valuetypes())
157 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
159 // We don't accept any truncstore of integer registers.
160 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
161 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
162 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
163 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
164 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
165 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
167 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
169 // SETOEQ and SETUNE require checking two conditions.
170 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
171 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
172 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
173 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
174 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
175 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
177 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
179 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
180 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
181 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
183 if (Subtarget->is64Bit()) {
184 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
185 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
186 } else if (!Subtarget->useSoftFloat()) {
187 // We have an algorithm for SSE2->double, and we turn this into a
188 // 64-bit FILD followed by conditional FADD for other targets.
189 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
190 // We have an algorithm for SSE2, and we turn this into a 64-bit
191 // FILD for other targets.
192 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
195 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
197 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
200 if (!Subtarget->useSoftFloat()) {
201 // SSE has no i16 to fp conversion, only i32
202 if (X86ScalarSSEf32) {
203 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
204 // f32 and f64 cases are Legal, f80 case is not
205 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
207 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
208 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
211 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
212 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
215 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
216 // are Legal, f80 is custom lowered.
217 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
218 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
220 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
222 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
225 if (X86ScalarSSEf32) {
226 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
227 // f32 and f64 cases are Legal, f80 case is not
228 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
230 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
231 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
234 // Handle FP_TO_UINT by promoting the destination to a larger signed
236 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
237 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
238 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
240 if (Subtarget->is64Bit()) {
241 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
242 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
243 } else if (!Subtarget->useSoftFloat()) {
244 // Since AVX is a superset of SSE3, only check for SSE here.
245 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
246 // Expand FP_TO_UINT into a select.
247 // FIXME: We would like to use a Custom expander here eventually to do
248 // the optimal thing for SSE vs. the default expansion in the legalizer.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
251 // With SSE3 we can use fisttpll to convert to a signed i64; without
252 // SSE, we're stuck with a fistpll.
253 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
256 if (isTargetFTOL()) {
257 // Use the _ftol2 runtime function, which has a pseudo-instruction
258 // to handle its weird calling convention.
259 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
262 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
263 if (!X86ScalarSSEf64) {
264 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
265 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
266 if (Subtarget->is64Bit()) {
267 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
268 // Without SSE, i64->f64 goes through memory.
269 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
273 // Scalar integer divide and remainder are lowered to use operations that
274 // produce two results, to match the available instructions. This exposes
275 // the two-result form to trivial CSE, which is able to combine x/y and x%y
276 // into a single instruction.
278 // Scalar integer multiply-high is also lowered to use two-result
279 // operations, to match the available instructions. However, plain multiply
280 // (low) operations are left as Legal, as there are single-result
281 // instructions for this in x86. Using the two-result multiply instructions
282 // when both high and low results are needed must be arranged by dagcombine.
283 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
285 setOperationAction(ISD::MULHS, VT, Expand);
286 setOperationAction(ISD::MULHU, VT, Expand);
287 setOperationAction(ISD::SDIV, VT, Expand);
288 setOperationAction(ISD::UDIV, VT, Expand);
289 setOperationAction(ISD::SREM, VT, Expand);
290 setOperationAction(ISD::UREM, VT, Expand);
292 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
293 setOperationAction(ISD::ADDC, VT, Custom);
294 setOperationAction(ISD::ADDE, VT, Custom);
295 setOperationAction(ISD::SUBC, VT, Custom);
296 setOperationAction(ISD::SUBE, VT, Custom);
299 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
300 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
301 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
302 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
303 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
304 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
305 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
306 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
307 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
312 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
313 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
314 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
315 if (Subtarget->is64Bit())
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
318 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
319 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
320 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
321 setOperationAction(ISD::FREM , MVT::f32 , Expand);
322 setOperationAction(ISD::FREM , MVT::f64 , Expand);
323 setOperationAction(ISD::FREM , MVT::f80 , Expand);
324 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
326 // Promote the i8 variants and force them on up to i32 which has a shorter
328 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
329 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
331 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
332 if (Subtarget->hasBMI()) {
333 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
335 if (Subtarget->is64Bit())
336 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
338 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
339 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
340 if (Subtarget->is64Bit())
341 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
344 if (Subtarget->hasLZCNT()) {
345 // When promoting the i8 variants, force them to i32 for a shorter
347 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
348 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
350 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
352 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
353 if (Subtarget->is64Bit())
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
356 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
357 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
358 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
362 if (Subtarget->is64Bit()) {
363 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
364 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
368 // Special handling for half-precision floating point conversions.
369 // If we don't have F16C support, then lower half float conversions
370 // into library calls.
371 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
372 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
373 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
376 // There's never any support for operations beyond MVT::f32.
377 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
378 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
379 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
380 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
382 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
383 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
384 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
385 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
386 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
387 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
389 if (Subtarget->hasPOPCNT()) {
390 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
392 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
393 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
395 if (Subtarget->is64Bit())
396 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
399 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
401 if (!Subtarget->hasMOVBE())
402 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
404 // These should be promoted to a larger select which is supported.
405 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
406 // X86 wants to expand cmov itself.
407 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
408 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
409 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
411 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
412 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
414 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
415 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
417 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
418 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
419 if (Subtarget->is64Bit()) {
420 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
421 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
423 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
424 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
425 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
426 // support continuation, user-level threading, and etc.. As a result, no
427 // other SjLj exception interfaces are implemented and please don't build
428 // your own exception handling based on them.
429 // LLVM/Clang supports zero-cost DWARF exception handling.
430 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
431 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
434 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
435 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
436 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
438 if (Subtarget->is64Bit())
439 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
440 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
441 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
442 if (Subtarget->is64Bit()) {
443 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
444 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
445 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
446 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
447 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
449 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
450 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
451 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
452 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
453 if (Subtarget->is64Bit()) {
454 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
455 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
456 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
459 if (Subtarget->hasSSE1())
460 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
462 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
464 // Expand certain atomics
465 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
467 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
468 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
469 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
472 if (Subtarget->hasCmpxchg16b()) {
473 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
476 // FIXME - use subtarget debug flags
477 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
478 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
479 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
482 if (Subtarget->is64Bit()) {
483 setExceptionPointerRegister(X86::RAX);
484 setExceptionSelectorRegister(X86::RDX);
486 setExceptionPointerRegister(X86::EAX);
487 setExceptionSelectorRegister(X86::EDX);
489 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
490 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
492 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
493 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
495 setOperationAction(ISD::TRAP, MVT::Other, Legal);
496 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
498 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
499 setOperationAction(ISD::VASTART , MVT::Other, Custom);
500 setOperationAction(ISD::VAEND , MVT::Other, Expand);
501 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
502 // TargetInfo::X86_64ABIBuiltinVaList
503 setOperationAction(ISD::VAARG , MVT::Other, Custom);
504 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
506 // TargetInfo::CharPtrBuiltinVaList
507 setOperationAction(ISD::VAARG , MVT::Other, Expand);
508 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
511 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
512 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
514 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
516 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
517 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
518 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
520 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
521 // f32 and f64 use SSE.
522 // Set up the FP register classes.
523 addRegisterClass(MVT::f32, &X86::FR32RegClass);
524 addRegisterClass(MVT::f64, &X86::FR64RegClass);
526 // Use ANDPD to simulate FABS.
527 setOperationAction(ISD::FABS , MVT::f64, Custom);
528 setOperationAction(ISD::FABS , MVT::f32, Custom);
530 // Use XORP to simulate FNEG.
531 setOperationAction(ISD::FNEG , MVT::f64, Custom);
532 setOperationAction(ISD::FNEG , MVT::f32, Custom);
534 // Use ANDPD and ORPD to simulate FCOPYSIGN.
535 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
536 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
538 // Lower this to FGETSIGNx86 plus an AND.
539 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
540 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
542 // We don't support sin/cos/fmod
543 setOperationAction(ISD::FSIN , MVT::f64, Expand);
544 setOperationAction(ISD::FCOS , MVT::f64, Expand);
545 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
546 setOperationAction(ISD::FSIN , MVT::f32, Expand);
547 setOperationAction(ISD::FCOS , MVT::f32, Expand);
548 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
550 // Expand FP immediates into loads from the stack, except for the special
552 addLegalFPImmediate(APFloat(+0.0)); // xorpd
553 addLegalFPImmediate(APFloat(+0.0f)); // xorps
554 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
555 // Use SSE for f32, x87 for f64.
556 // Set up the FP register classes.
557 addRegisterClass(MVT::f32, &X86::FR32RegClass);
558 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
560 // Use ANDPS to simulate FABS.
561 setOperationAction(ISD::FABS , MVT::f32, Custom);
563 // Use XORP to simulate FNEG.
564 setOperationAction(ISD::FNEG , MVT::f32, Custom);
566 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
568 // Use ANDPS and ORPS to simulate FCOPYSIGN.
569 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
570 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
572 // We don't support sin/cos/fmod
573 setOperationAction(ISD::FSIN , MVT::f32, Expand);
574 setOperationAction(ISD::FCOS , MVT::f32, Expand);
575 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
577 // Special cases we handle for FP constants.
578 addLegalFPImmediate(APFloat(+0.0f)); // xorps
579 addLegalFPImmediate(APFloat(+0.0)); // FLD0
580 addLegalFPImmediate(APFloat(+1.0)); // FLD1
581 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
582 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
584 if (!TM.Options.UnsafeFPMath) {
585 setOperationAction(ISD::FSIN , MVT::f64, Expand);
586 setOperationAction(ISD::FCOS , MVT::f64, Expand);
587 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
589 } else if (!Subtarget->useSoftFloat()) {
590 // f32 and f64 in x87.
591 // Set up the FP register classes.
592 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
593 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
595 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
596 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
600 if (!TM.Options.UnsafeFPMath) {
601 setOperationAction(ISD::FSIN , MVT::f64, Expand);
602 setOperationAction(ISD::FSIN , MVT::f32, Expand);
603 setOperationAction(ISD::FCOS , MVT::f64, Expand);
604 setOperationAction(ISD::FCOS , MVT::f32, Expand);
605 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
606 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
608 addLegalFPImmediate(APFloat(+0.0)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
612 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
613 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
614 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
615 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
618 // We don't support FMA.
619 setOperationAction(ISD::FMA, MVT::f64, Expand);
620 setOperationAction(ISD::FMA, MVT::f32, Expand);
622 // Long double always uses X87.
623 if (!Subtarget->useSoftFloat()) {
624 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
625 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
626 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
628 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
629 addLegalFPImmediate(TmpFlt); // FLD0
631 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
634 APFloat TmpFlt2(+1.0);
635 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
637 addLegalFPImmediate(TmpFlt2); // FLD1
638 TmpFlt2.changeSign();
639 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
642 if (!TM.Options.UnsafeFPMath) {
643 setOperationAction(ISD::FSIN , MVT::f80, Expand);
644 setOperationAction(ISD::FCOS , MVT::f80, Expand);
645 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
648 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
649 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
650 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
651 setOperationAction(ISD::FRINT, MVT::f80, Expand);
652 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
653 setOperationAction(ISD::FMA, MVT::f80, Expand);
656 // Always use a library call for pow.
657 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
658 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
659 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
661 setOperationAction(ISD::FLOG, MVT::f80, Expand);
662 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
663 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
664 setOperationAction(ISD::FEXP, MVT::f80, Expand);
665 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
666 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
667 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
669 // First set operation action for all vector types to either promote
670 // (for widening) or expand (for scalarization). Then we will selectively
671 // turn on ones that can be effectively codegen'd.
672 for (MVT VT : MVT::vector_valuetypes()) {
673 setOperationAction(ISD::ADD , VT, Expand);
674 setOperationAction(ISD::SUB , VT, Expand);
675 setOperationAction(ISD::FADD, VT, Expand);
676 setOperationAction(ISD::FNEG, VT, Expand);
677 setOperationAction(ISD::FSUB, VT, Expand);
678 setOperationAction(ISD::MUL , VT, Expand);
679 setOperationAction(ISD::FMUL, VT, Expand);
680 setOperationAction(ISD::SDIV, VT, Expand);
681 setOperationAction(ISD::UDIV, VT, Expand);
682 setOperationAction(ISD::FDIV, VT, Expand);
683 setOperationAction(ISD::SREM, VT, Expand);
684 setOperationAction(ISD::UREM, VT, Expand);
685 setOperationAction(ISD::LOAD, VT, Expand);
686 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
687 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
688 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
689 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
690 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
691 setOperationAction(ISD::FABS, VT, Expand);
692 setOperationAction(ISD::FSIN, VT, Expand);
693 setOperationAction(ISD::FSINCOS, VT, Expand);
694 setOperationAction(ISD::FCOS, VT, Expand);
695 setOperationAction(ISD::FSINCOS, VT, Expand);
696 setOperationAction(ISD::FREM, VT, Expand);
697 setOperationAction(ISD::FMA, VT, Expand);
698 setOperationAction(ISD::FPOWI, VT, Expand);
699 setOperationAction(ISD::FSQRT, VT, Expand);
700 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
701 setOperationAction(ISD::FFLOOR, VT, Expand);
702 setOperationAction(ISD::FCEIL, VT, Expand);
703 setOperationAction(ISD::FTRUNC, VT, Expand);
704 setOperationAction(ISD::FRINT, VT, Expand);
705 setOperationAction(ISD::FNEARBYINT, VT, Expand);
706 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
707 setOperationAction(ISD::MULHS, VT, Expand);
708 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
709 setOperationAction(ISD::MULHU, VT, Expand);
710 setOperationAction(ISD::SDIVREM, VT, Expand);
711 setOperationAction(ISD::UDIVREM, VT, Expand);
712 setOperationAction(ISD::FPOW, VT, Expand);
713 setOperationAction(ISD::CTPOP, VT, Expand);
714 setOperationAction(ISD::CTTZ, VT, Expand);
715 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
716 setOperationAction(ISD::CTLZ, VT, Expand);
717 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
718 setOperationAction(ISD::SHL, VT, Expand);
719 setOperationAction(ISD::SRA, VT, Expand);
720 setOperationAction(ISD::SRL, VT, Expand);
721 setOperationAction(ISD::ROTL, VT, Expand);
722 setOperationAction(ISD::ROTR, VT, Expand);
723 setOperationAction(ISD::BSWAP, VT, Expand);
724 setOperationAction(ISD::SETCC, VT, Expand);
725 setOperationAction(ISD::FLOG, VT, Expand);
726 setOperationAction(ISD::FLOG2, VT, Expand);
727 setOperationAction(ISD::FLOG10, VT, Expand);
728 setOperationAction(ISD::FEXP, VT, Expand);
729 setOperationAction(ISD::FEXP2, VT, Expand);
730 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
731 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
732 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
733 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
734 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
735 setOperationAction(ISD::TRUNCATE, VT, Expand);
736 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
737 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
738 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
739 setOperationAction(ISD::VSELECT, VT, Expand);
740 setOperationAction(ISD::SELECT_CC, VT, Expand);
741 for (MVT InnerVT : MVT::vector_valuetypes()) {
742 setTruncStoreAction(InnerVT, VT, Expand);
744 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
745 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
747 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
748 // types, we have to deal with them whether we ask for Expansion or not.
749 // Setting Expand causes its own optimisation problems though, so leave
751 if (VT.getVectorElementType() == MVT::i1)
752 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
754 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
755 // split/scalarized right now.
756 if (VT.getVectorElementType() == MVT::f16)
757 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
761 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
762 // with -msoft-float, disable use of MMX as well.
763 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
764 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
765 // No operations on x86mmx supported, everything uses intrinsics.
768 // MMX-sized vectors (other than x86mmx) are expected to be expanded
769 // into smaller operations.
770 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
771 setOperationAction(ISD::MULHS, MMXTy, Expand);
772 setOperationAction(ISD::AND, MMXTy, Expand);
773 setOperationAction(ISD::OR, MMXTy, Expand);
774 setOperationAction(ISD::XOR, MMXTy, Expand);
775 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
776 setOperationAction(ISD::SELECT, MMXTy, Expand);
777 setOperationAction(ISD::BITCAST, MMXTy, Expand);
779 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
781 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
782 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
784 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
785 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
786 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
787 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
788 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
789 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
790 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
791 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
792 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
793 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
794 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
796 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
797 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
800 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
801 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
803 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
804 // registers cannot be used even for integer operations.
805 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
806 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
807 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
808 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
810 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
811 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
812 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
813 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
814 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
815 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
816 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
817 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
818 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
819 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
820 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
821 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
822 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
823 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
824 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
825 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
826 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
827 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
828 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
829 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
830 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
831 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
832 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
834 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
835 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
836 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
837 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
839 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
840 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
845 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
846 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
847 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
848 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
850 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
851 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
852 MVT VT = (MVT::SimpleValueType)i;
853 // Do not attempt to custom lower non-power-of-2 vectors
854 if (!isPowerOf2_32(VT.getVectorNumElements()))
856 // Do not attempt to custom lower non-128-bit vectors
857 if (!VT.is128BitVector())
859 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
860 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
861 setOperationAction(ISD::VSELECT, VT, Custom);
862 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
865 // We support custom legalizing of sext and anyext loads for specific
866 // memory vector types which we can load as a scalar (or sequence of
867 // scalars) and extend in-register to a legal 128-bit vector type. For sext
868 // loads these must work with a single scalar load.
869 for (MVT VT : MVT::integer_vector_valuetypes()) {
870 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
871 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
872 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
873 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
874 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
875 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
881 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
882 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
883 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
884 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
885 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
886 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
887 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
888 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
890 if (Subtarget->is64Bit()) {
891 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
892 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
895 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
896 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
897 MVT VT = (MVT::SimpleValueType)i;
899 // Do not attempt to promote non-128-bit vectors
900 if (!VT.is128BitVector())
903 setOperationAction(ISD::AND, VT, Promote);
904 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
905 setOperationAction(ISD::OR, VT, Promote);
906 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
907 setOperationAction(ISD::XOR, VT, Promote);
908 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
909 setOperationAction(ISD::LOAD, VT, Promote);
910 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
911 setOperationAction(ISD::SELECT, VT, Promote);
912 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
915 // Custom lower v2i64 and v2f64 selects.
916 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
917 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
918 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
919 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
921 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
922 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
924 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
925 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
926 // As there is no 64-bit GPR available, we need build a special custom
927 // sequence to convert from v2i32 to v2f32.
928 if (!Subtarget->is64Bit())
929 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
931 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
932 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
934 for (MVT VT : MVT::fp_vector_valuetypes())
935 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
937 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
938 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
939 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
942 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
943 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
944 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
945 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
946 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
947 setOperationAction(ISD::FRINT, RoundedTy, Legal);
948 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
951 // FIXME: Do we need to handle scalar-to-vector here?
952 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
954 // We directly match byte blends in the backend as they match the VSELECT
956 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
958 // SSE41 brings specific instructions for doing vector sign extend even in
959 // cases where we don't have SRA.
960 for (MVT VT : MVT::integer_vector_valuetypes()) {
961 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
962 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
963 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
966 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
967 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
968 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
969 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
970 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
971 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
972 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
974 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
975 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
976 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
977 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
978 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
979 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
981 // i8 and i16 vectors are custom because the source register and source
982 // source memory operand types are not the same width. f32 vectors are
983 // custom since the immediate controlling the insert encodes additional
985 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
986 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
987 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
988 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
990 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
991 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
992 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
993 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
995 // FIXME: these should be Legal, but that's only for the case where
996 // the index is constant. For now custom expand to deal with that.
997 if (Subtarget->is64Bit()) {
998 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1003 if (Subtarget->hasSSE2()) {
1004 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1005 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1006 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1008 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1009 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1011 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1012 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1014 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1015 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1017 // In the customized shift lowering, the legal cases in AVX2 will be
1019 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1020 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1022 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1023 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1025 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1028 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1029 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1030 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1031 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1032 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1033 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1034 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1036 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1037 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1038 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1040 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1041 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1042 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1043 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1044 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1045 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1046 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1047 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1048 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1049 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1050 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1051 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1053 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1054 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1055 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1056 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1057 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1058 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1059 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1060 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1063 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1064 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1066 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1067 // even though v8i16 is a legal type.
1068 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1069 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1070 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1072 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1073 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1074 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1076 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1077 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1079 for (MVT VT : MVT::fp_vector_valuetypes())
1080 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1082 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1083 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1085 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1086 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1088 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1089 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1091 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1092 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1093 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1094 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1096 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1097 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1098 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1100 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1101 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1102 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1103 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1104 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1105 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1106 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1107 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1108 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1109 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1110 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1111 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1113 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1114 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1115 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1116 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1118 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1119 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1120 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1121 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1122 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1123 setOperationAction(ISD::FMA, MVT::f32, Legal);
1124 setOperationAction(ISD::FMA, MVT::f64, Legal);
1127 if (Subtarget->hasInt256()) {
1128 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1129 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1130 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1131 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1133 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1134 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1135 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1136 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1138 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1139 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1140 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1141 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1143 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1144 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1145 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1146 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1148 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1149 // when we have a 256bit-wide blend with immediate.
1150 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1152 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1153 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1154 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1155 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1156 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1157 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1158 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1160 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1161 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1162 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1163 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1164 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1165 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1167 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1168 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1169 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1170 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1172 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1173 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1174 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1175 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1177 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1178 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1179 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1180 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1183 // In the customized shift lowering, the legal cases in AVX2 will be
1185 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1186 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1188 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1189 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1191 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1193 // Custom lower several nodes for 256-bit types.
1194 for (MVT VT : MVT::vector_valuetypes()) {
1195 if (VT.getScalarSizeInBits() >= 32) {
1196 setOperationAction(ISD::MLOAD, VT, Legal);
1197 setOperationAction(ISD::MSTORE, VT, Legal);
1199 // Extract subvector is special because the value type
1200 // (result) is 128-bit but the source is 256-bit wide.
1201 if (VT.is128BitVector()) {
1202 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1204 // Do not attempt to custom lower other non-256-bit vectors
1205 if (!VT.is256BitVector())
1208 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1209 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1210 setOperationAction(ISD::VSELECT, VT, Custom);
1211 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1212 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1213 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1214 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1215 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1218 if (Subtarget->hasInt256())
1219 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1222 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1223 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1224 MVT VT = (MVT::SimpleValueType)i;
1226 // Do not attempt to promote non-256-bit vectors
1227 if (!VT.is256BitVector())
1230 setOperationAction(ISD::AND, VT, Promote);
1231 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1232 setOperationAction(ISD::OR, VT, Promote);
1233 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1234 setOperationAction(ISD::XOR, VT, Promote);
1235 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1236 setOperationAction(ISD::LOAD, VT, Promote);
1237 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1238 setOperationAction(ISD::SELECT, VT, Promote);
1239 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1243 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1244 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1245 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1246 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1247 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1249 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1250 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1251 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1253 for (MVT VT : MVT::fp_vector_valuetypes())
1254 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1256 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1257 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1258 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1259 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1260 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1261 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1262 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1263 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1264 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1265 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1266 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1267 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1269 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1270 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1271 setOperationAction(ISD::XOR, MVT::i1, Legal);
1272 setOperationAction(ISD::OR, MVT::i1, Legal);
1273 setOperationAction(ISD::AND, MVT::i1, Legal);
1274 setOperationAction(ISD::SUB, MVT::i1, Custom);
1275 setOperationAction(ISD::ADD, MVT::i1, Custom);
1276 setOperationAction(ISD::MUL, MVT::i1, Custom);
1277 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1278 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1279 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1280 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1281 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1283 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1284 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1285 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1286 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1287 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1288 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1290 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1291 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1292 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1293 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1294 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1295 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1296 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1297 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1299 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1300 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1301 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1302 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1303 if (Subtarget->is64Bit()) {
1304 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1305 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1306 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1307 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1309 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1310 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1311 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1312 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1313 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1314 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1315 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1316 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1317 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1318 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1319 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1320 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1321 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1322 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1323 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1324 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1326 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1327 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1328 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1329 if (Subtarget->hasDQI()) {
1330 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1331 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1333 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1334 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1335 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1336 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1337 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1338 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1339 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1340 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1341 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1342 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1343 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1344 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1345 if (Subtarget->hasDQI()) {
1346 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1347 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1349 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1350 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1351 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1353 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1355 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1356 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1357 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1358 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1360 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1361 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1362 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1363 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1364 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1366 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1367 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1369 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1371 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1372 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1373 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1374 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1375 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1376 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1377 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1378 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1379 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1380 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1381 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1383 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1384 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1386 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1387 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1389 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1391 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1392 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1394 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1395 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1397 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1398 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1400 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1401 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1402 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1403 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1404 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1405 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1407 if (Subtarget->hasCDI()) {
1408 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1409 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1411 if (Subtarget->hasDQI()) {
1412 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1413 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1414 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1416 // Custom lower several nodes.
1417 for (MVT VT : MVT::vector_valuetypes()) {
1418 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1420 setOperationAction(ISD::AND, VT, Legal);
1421 setOperationAction(ISD::OR, VT, Legal);
1422 setOperationAction(ISD::XOR, VT, Legal);
1424 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1425 setOperationAction(ISD::MGATHER, VT, Custom);
1426 setOperationAction(ISD::MSCATTER, VT, Custom);
1428 // Extract subvector is special because the value type
1429 // (result) is 256/128-bit but the source is 512-bit wide.
1430 if (VT.is128BitVector() || VT.is256BitVector()) {
1431 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1433 if (VT.getVectorElementType() == MVT::i1)
1434 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1436 // Do not attempt to custom lower other non-512-bit vectors
1437 if (!VT.is512BitVector())
1440 if (EltSize >= 32) {
1441 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1442 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1443 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1444 setOperationAction(ISD::VSELECT, VT, Legal);
1445 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1446 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1447 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1448 setOperationAction(ISD::MLOAD, VT, Legal);
1449 setOperationAction(ISD::MSTORE, VT, Legal);
1452 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1453 MVT VT = (MVT::SimpleValueType)i;
1455 // Do not attempt to promote non-512-bit vectors.
1456 if (!VT.is512BitVector())
1459 setOperationAction(ISD::SELECT, VT, Promote);
1460 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1464 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1465 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1466 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1468 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1469 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1471 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1472 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1473 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1474 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1475 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1476 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1477 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1478 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1479 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1480 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1481 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1482 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1483 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1484 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1485 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1486 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1487 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1488 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1489 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1490 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1491 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1492 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1493 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1494 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1495 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1496 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1497 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1499 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1500 const MVT VT = (MVT::SimpleValueType)i;
1502 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1504 // Do not attempt to promote non-512-bit vectors.
1505 if (!VT.is512BitVector())
1509 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1510 setOperationAction(ISD::VSELECT, VT, Legal);
1515 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1516 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1517 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1519 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1520 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1521 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1522 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1523 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1524 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1525 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1526 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1527 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1528 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1530 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1531 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1532 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1533 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1534 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1535 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1536 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1537 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1540 // We want to custom lower some of our intrinsics.
1541 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1542 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1543 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1544 if (!Subtarget->is64Bit())
1545 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1547 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1548 // handle type legalization for these operations here.
1550 // FIXME: We really should do custom legalization for addition and
1551 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1552 // than generic legalization for 64-bit multiplication-with-overflow, though.
1553 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1554 // Add/Sub/Mul with overflow operations are custom lowered.
1556 setOperationAction(ISD::SADDO, VT, Custom);
1557 setOperationAction(ISD::UADDO, VT, Custom);
1558 setOperationAction(ISD::SSUBO, VT, Custom);
1559 setOperationAction(ISD::USUBO, VT, Custom);
1560 setOperationAction(ISD::SMULO, VT, Custom);
1561 setOperationAction(ISD::UMULO, VT, Custom);
1565 if (!Subtarget->is64Bit()) {
1566 // These libcalls are not available in 32-bit.
1567 setLibcallName(RTLIB::SHL_I128, nullptr);
1568 setLibcallName(RTLIB::SRL_I128, nullptr);
1569 setLibcallName(RTLIB::SRA_I128, nullptr);
1572 // Combine sin / cos into one node or libcall if possible.
1573 if (Subtarget->hasSinCos()) {
1574 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1575 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1576 if (Subtarget->isTargetDarwin()) {
1577 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1578 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1579 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1580 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1584 if (Subtarget->isTargetWin64()) {
1585 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1586 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1587 setOperationAction(ISD::SREM, MVT::i128, Custom);
1588 setOperationAction(ISD::UREM, MVT::i128, Custom);
1589 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1590 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1593 // We have target-specific dag combine patterns for the following nodes:
1594 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1595 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1596 setTargetDAGCombine(ISD::BITCAST);
1597 setTargetDAGCombine(ISD::VSELECT);
1598 setTargetDAGCombine(ISD::SELECT);
1599 setTargetDAGCombine(ISD::SHL);
1600 setTargetDAGCombine(ISD::SRA);
1601 setTargetDAGCombine(ISD::SRL);
1602 setTargetDAGCombine(ISD::OR);
1603 setTargetDAGCombine(ISD::AND);
1604 setTargetDAGCombine(ISD::ADD);
1605 setTargetDAGCombine(ISD::FADD);
1606 setTargetDAGCombine(ISD::FSUB);
1607 setTargetDAGCombine(ISD::FMA);
1608 setTargetDAGCombine(ISD::SUB);
1609 setTargetDAGCombine(ISD::LOAD);
1610 setTargetDAGCombine(ISD::MLOAD);
1611 setTargetDAGCombine(ISD::STORE);
1612 setTargetDAGCombine(ISD::MSTORE);
1613 setTargetDAGCombine(ISD::ZERO_EXTEND);
1614 setTargetDAGCombine(ISD::ANY_EXTEND);
1615 setTargetDAGCombine(ISD::SIGN_EXTEND);
1616 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1617 setTargetDAGCombine(ISD::SINT_TO_FP);
1618 setTargetDAGCombine(ISD::SETCC);
1619 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1620 setTargetDAGCombine(ISD::BUILD_VECTOR);
1621 setTargetDAGCombine(ISD::MUL);
1622 setTargetDAGCombine(ISD::XOR);
1624 computeRegisterProperties(Subtarget->getRegisterInfo());
1626 // On Darwin, -Os means optimize for size without hurting performance,
1627 // do not reduce the limit.
1628 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1629 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1630 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1631 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1632 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1633 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1634 setPrefLoopAlignment(4); // 2^4 bytes.
1636 // Predictable cmov don't hurt on atom because it's in-order.
1637 PredictableSelectIsExpensive = !Subtarget->isAtom();
1638 EnableExtLdPromotion = true;
1639 setPrefFunctionAlignment(4); // 2^4 bytes.
1641 verifyIntrinsicTables();
1644 // This has so far only been implemented for 64-bit MachO.
1645 bool X86TargetLowering::useLoadStackGuardNode() const {
1646 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1649 TargetLoweringBase::LegalizeTypeAction
1650 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1651 if (ExperimentalVectorWideningLegalization &&
1652 VT.getVectorNumElements() != 1 &&
1653 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1654 return TypeWidenVector;
1656 return TargetLoweringBase::getPreferredVectorAction(VT);
1659 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1661 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1663 const unsigned NumElts = VT.getVectorNumElements();
1664 const EVT EltVT = VT.getVectorElementType();
1665 if (VT.is512BitVector()) {
1666 if (Subtarget->hasAVX512())
1667 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1668 EltVT == MVT::f32 || EltVT == MVT::f64)
1670 case 8: return MVT::v8i1;
1671 case 16: return MVT::v16i1;
1673 if (Subtarget->hasBWI())
1674 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1676 case 32: return MVT::v32i1;
1677 case 64: return MVT::v64i1;
1681 if (VT.is256BitVector() || VT.is128BitVector()) {
1682 if (Subtarget->hasVLX())
1683 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1684 EltVT == MVT::f32 || EltVT == MVT::f64)
1686 case 2: return MVT::v2i1;
1687 case 4: return MVT::v4i1;
1688 case 8: return MVT::v8i1;
1690 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1691 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1693 case 8: return MVT::v8i1;
1694 case 16: return MVT::v16i1;
1695 case 32: return MVT::v32i1;
1699 return VT.changeVectorElementTypeToInteger();
1702 /// Helper for getByValTypeAlignment to determine
1703 /// the desired ByVal argument alignment.
1704 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1707 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1708 if (VTy->getBitWidth() == 128)
1710 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1711 unsigned EltAlign = 0;
1712 getMaxByValAlign(ATy->getElementType(), EltAlign);
1713 if (EltAlign > MaxAlign)
1714 MaxAlign = EltAlign;
1715 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1716 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1717 unsigned EltAlign = 0;
1718 getMaxByValAlign(STy->getElementType(i), EltAlign);
1719 if (EltAlign > MaxAlign)
1720 MaxAlign = EltAlign;
1727 /// Return the desired alignment for ByVal aggregate
1728 /// function arguments in the caller parameter area. For X86, aggregates
1729 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1730 /// are at 4-byte boundaries.
1731 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1732 if (Subtarget->is64Bit()) {
1733 // Max of 8 and alignment of type.
1734 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1741 if (Subtarget->hasSSE1())
1742 getMaxByValAlign(Ty, Align);
1746 /// Returns the target specific optimal type for load
1747 /// and store operations as a result of memset, memcpy, and memmove
1748 /// lowering. If DstAlign is zero that means it's safe to destination
1749 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1750 /// means there isn't a need to check it against alignment requirement,
1751 /// probably because the source does not need to be loaded. If 'IsMemset' is
1752 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1753 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1754 /// source is constant so it does not need to be loaded.
1755 /// It returns EVT::Other if the type should be determined using generic
1756 /// target-independent logic.
1758 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1759 unsigned DstAlign, unsigned SrcAlign,
1760 bool IsMemset, bool ZeroMemset,
1762 MachineFunction &MF) const {
1763 const Function *F = MF.getFunction();
1764 if ((!IsMemset || ZeroMemset) &&
1765 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1767 (Subtarget->isUnalignedMemAccessFast() ||
1768 ((DstAlign == 0 || DstAlign >= 16) &&
1769 (SrcAlign == 0 || SrcAlign >= 16)))) {
1771 if (Subtarget->hasInt256())
1773 if (Subtarget->hasFp256())
1776 if (Subtarget->hasSSE2())
1778 if (Subtarget->hasSSE1())
1780 } else if (!MemcpyStrSrc && Size >= 8 &&
1781 !Subtarget->is64Bit() &&
1782 Subtarget->hasSSE2()) {
1783 // Do not use f64 to lower memcpy if source is string constant. It's
1784 // better to use i32 to avoid the loads.
1788 if (Subtarget->is64Bit() && Size >= 8)
1793 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1795 return X86ScalarSSEf32;
1796 else if (VT == MVT::f64)
1797 return X86ScalarSSEf64;
1802 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1807 *Fast = Subtarget->isUnalignedMemAccessFast();
1811 /// Return the entry encoding for a jump table in the
1812 /// current function. The returned value is a member of the
1813 /// MachineJumpTableInfo::JTEntryKind enum.
1814 unsigned X86TargetLowering::getJumpTableEncoding() const {
1815 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1817 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1818 Subtarget->isPICStyleGOT())
1819 return MachineJumpTableInfo::EK_Custom32;
1821 // Otherwise, use the normal jump table encoding heuristics.
1822 return TargetLowering::getJumpTableEncoding();
1825 bool X86TargetLowering::useSoftFloat() const {
1826 return Subtarget->useSoftFloat();
1830 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1831 const MachineBasicBlock *MBB,
1832 unsigned uid,MCContext &Ctx) const{
1833 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1834 Subtarget->isPICStyleGOT());
1835 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1837 return MCSymbolRefExpr::create(MBB->getSymbol(),
1838 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1841 /// Returns relocation base for the given PIC jumptable.
1842 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1843 SelectionDAG &DAG) const {
1844 if (!Subtarget->is64Bit())
1845 // This doesn't have SDLoc associated with it, but is not really the
1846 // same as a Register.
1847 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1851 /// This returns the relocation base for the given PIC jumptable,
1852 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1853 const MCExpr *X86TargetLowering::
1854 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1855 MCContext &Ctx) const {
1856 // X86-64 uses RIP relative addressing based on the jump table label.
1857 if (Subtarget->isPICStyleRIPRel())
1858 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1860 // Otherwise, the reference is relative to the PIC base.
1861 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
1864 std::pair<const TargetRegisterClass *, uint8_t>
1865 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1867 const TargetRegisterClass *RRC = nullptr;
1869 switch (VT.SimpleTy) {
1871 return TargetLowering::findRepresentativeClass(TRI, VT);
1872 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1873 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1876 RRC = &X86::VR64RegClass;
1878 case MVT::f32: case MVT::f64:
1879 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1880 case MVT::v4f32: case MVT::v2f64:
1881 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1883 RRC = &X86::VR128RegClass;
1886 return std::make_pair(RRC, Cost);
1889 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1890 unsigned &Offset) const {
1891 if (!Subtarget->isTargetLinux())
1894 if (Subtarget->is64Bit()) {
1895 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1897 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1909 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1910 unsigned DestAS) const {
1911 assert(SrcAS != DestAS && "Expected different address spaces!");
1913 return SrcAS < 256 && DestAS < 256;
1916 //===----------------------------------------------------------------------===//
1917 // Return Value Calling Convention Implementation
1918 //===----------------------------------------------------------------------===//
1920 #include "X86GenCallingConv.inc"
1923 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1924 MachineFunction &MF, bool isVarArg,
1925 const SmallVectorImpl<ISD::OutputArg> &Outs,
1926 LLVMContext &Context) const {
1927 SmallVector<CCValAssign, 16> RVLocs;
1928 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1929 return CCInfo.CheckReturn(Outs, RetCC_X86);
1932 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1933 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1938 X86TargetLowering::LowerReturn(SDValue Chain,
1939 CallingConv::ID CallConv, bool isVarArg,
1940 const SmallVectorImpl<ISD::OutputArg> &Outs,
1941 const SmallVectorImpl<SDValue> &OutVals,
1942 SDLoc dl, SelectionDAG &DAG) const {
1943 MachineFunction &MF = DAG.getMachineFunction();
1944 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1946 SmallVector<CCValAssign, 16> RVLocs;
1947 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1948 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1951 SmallVector<SDValue, 6> RetOps;
1952 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1953 // Operand #1 = Bytes To Pop
1954 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
1957 // Copy the result values into the output registers.
1958 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1959 CCValAssign &VA = RVLocs[i];
1960 assert(VA.isRegLoc() && "Can only return in registers!");
1961 SDValue ValToCopy = OutVals[i];
1962 EVT ValVT = ValToCopy.getValueType();
1964 // Promote values to the appropriate types.
1965 if (VA.getLocInfo() == CCValAssign::SExt)
1966 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1967 else if (VA.getLocInfo() == CCValAssign::ZExt)
1968 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1969 else if (VA.getLocInfo() == CCValAssign::AExt) {
1970 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
1971 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1973 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1975 else if (VA.getLocInfo() == CCValAssign::BCvt)
1976 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
1978 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1979 "Unexpected FP-extend for return value.");
1981 // If this is x86-64, and we disabled SSE, we can't return FP values,
1982 // or SSE or MMX vectors.
1983 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1984 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1985 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1986 report_fatal_error("SSE register return with SSE disabled");
1988 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1989 // llvm-gcc has never done it right and no one has noticed, so this
1990 // should be OK for now.
1991 if (ValVT == MVT::f64 &&
1992 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1993 report_fatal_error("SSE2 register return with SSE2 disabled");
1995 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1996 // the RET instruction and handled by the FP Stackifier.
1997 if (VA.getLocReg() == X86::FP0 ||
1998 VA.getLocReg() == X86::FP1) {
1999 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2000 // change the value to the FP stack register class.
2001 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2002 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2003 RetOps.push_back(ValToCopy);
2004 // Don't emit a copytoreg.
2008 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2009 // which is returned in RAX / RDX.
2010 if (Subtarget->is64Bit()) {
2011 if (ValVT == MVT::x86mmx) {
2012 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2013 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2014 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2016 // If we don't have SSE2 available, convert to v4f32 so the generated
2017 // register is legal.
2018 if (!Subtarget->hasSSE2())
2019 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2024 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2025 Flag = Chain.getValue(1);
2026 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2029 // All x86 ABIs require that for returning structs by value we copy
2030 // the sret argument into %rax/%eax (depending on ABI) for the return.
2031 // We saved the argument into a virtual register in the entry block,
2032 // so now we copy the value out and into %rax/%eax.
2034 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2035 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2036 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2037 // either case FuncInfo->setSRetReturnReg() will have been called.
2038 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2039 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2042 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2043 X86::RAX : X86::EAX;
2044 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2045 Flag = Chain.getValue(1);
2047 // RAX/EAX now acts like a return value.
2048 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2051 RetOps[0] = Chain; // Update chain.
2053 // Add the flag if we have it.
2055 RetOps.push_back(Flag);
2057 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2060 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2061 if (N->getNumValues() != 1)
2063 if (!N->hasNUsesOfValue(1, 0))
2066 SDValue TCChain = Chain;
2067 SDNode *Copy = *N->use_begin();
2068 if (Copy->getOpcode() == ISD::CopyToReg) {
2069 // If the copy has a glue operand, we conservatively assume it isn't safe to
2070 // perform a tail call.
2071 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2073 TCChain = Copy->getOperand(0);
2074 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2077 bool HasRet = false;
2078 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2080 if (UI->getOpcode() != X86ISD::RET_FLAG)
2082 // If we are returning more than one value, we can definitely
2083 // not make a tail call see PR19530
2084 if (UI->getNumOperands() > 4)
2086 if (UI->getNumOperands() == 4 &&
2087 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2100 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2101 ISD::NodeType ExtendKind) const {
2103 // TODO: Is this also valid on 32-bit?
2104 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2105 ReturnMVT = MVT::i8;
2107 ReturnMVT = MVT::i32;
2109 EVT MinVT = getRegisterType(Context, ReturnMVT);
2110 return VT.bitsLT(MinVT) ? MinVT : VT;
2113 /// Lower the result values of a call into the
2114 /// appropriate copies out of appropriate physical registers.
2117 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2118 CallingConv::ID CallConv, bool isVarArg,
2119 const SmallVectorImpl<ISD::InputArg> &Ins,
2120 SDLoc dl, SelectionDAG &DAG,
2121 SmallVectorImpl<SDValue> &InVals) const {
2123 // Assign locations to each value returned by this call.
2124 SmallVector<CCValAssign, 16> RVLocs;
2125 bool Is64Bit = Subtarget->is64Bit();
2126 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2128 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2130 // Copy all of the result registers out of their specified physreg.
2131 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2132 CCValAssign &VA = RVLocs[i];
2133 EVT CopyVT = VA.getLocVT();
2135 // If this is x86-64, and we disabled SSE, we can't return FP values
2136 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2137 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2138 report_fatal_error("SSE register return with SSE disabled");
2141 // If we prefer to use the value in xmm registers, copy it out as f80 and
2142 // use a truncate to move it from fp stack reg to xmm reg.
2143 bool RoundAfterCopy = false;
2144 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2145 isScalarFPTypeInSSEReg(VA.getValVT())) {
2147 RoundAfterCopy = (CopyVT != VA.getLocVT());
2150 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2151 CopyVT, InFlag).getValue(1);
2152 SDValue Val = Chain.getValue(0);
2155 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2156 // This truncation won't change the value.
2157 DAG.getIntPtrConstant(1, dl));
2159 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2160 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2162 InFlag = Chain.getValue(2);
2163 InVals.push_back(Val);
2169 //===----------------------------------------------------------------------===//
2170 // C & StdCall & Fast Calling Convention implementation
2171 //===----------------------------------------------------------------------===//
2172 // StdCall calling convention seems to be standard for many Windows' API
2173 // routines and around. It differs from C calling convention just a little:
2174 // callee should clean up the stack, not caller. Symbols should be also
2175 // decorated in some fancy way :) It doesn't support any vector arguments.
2176 // For info on fast calling convention see Fast Calling Convention (tail call)
2177 // implementation LowerX86_32FastCCCallTo.
2179 /// CallIsStructReturn - Determines whether a call uses struct return
2181 enum StructReturnType {
2186 static StructReturnType
2187 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2189 return NotStructReturn;
2191 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2192 if (!Flags.isSRet())
2193 return NotStructReturn;
2194 if (Flags.isInReg())
2195 return RegStructReturn;
2196 return StackStructReturn;
2199 /// Determines whether a function uses struct return semantics.
2200 static StructReturnType
2201 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2203 return NotStructReturn;
2205 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2206 if (!Flags.isSRet())
2207 return NotStructReturn;
2208 if (Flags.isInReg())
2209 return RegStructReturn;
2210 return StackStructReturn;
2213 /// Make a copy of an aggregate at address specified by "Src" to address
2214 /// "Dst" with size and alignment information specified by the specific
2215 /// parameter attribute. The copy will be passed as a byval function parameter.
2217 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2218 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2220 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2222 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2223 /*isVolatile*/false, /*AlwaysInline=*/true,
2224 /*isTailCall*/false,
2225 MachinePointerInfo(), MachinePointerInfo());
2228 /// Return true if the calling convention is one that
2229 /// supports tail call optimization.
2230 static bool IsTailCallConvention(CallingConv::ID CC) {
2231 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2232 CC == CallingConv::HiPE);
2235 /// \brief Return true if the calling convention is a C calling convention.
2236 static bool IsCCallConvention(CallingConv::ID CC) {
2237 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2238 CC == CallingConv::X86_64_SysV);
2241 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2242 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2246 CallingConv::ID CalleeCC = CS.getCallingConv();
2247 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2253 /// Return true if the function is being made into
2254 /// a tailcall target by changing its ABI.
2255 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2256 bool GuaranteedTailCallOpt) {
2257 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2261 X86TargetLowering::LowerMemArgument(SDValue Chain,
2262 CallingConv::ID CallConv,
2263 const SmallVectorImpl<ISD::InputArg> &Ins,
2264 SDLoc dl, SelectionDAG &DAG,
2265 const CCValAssign &VA,
2266 MachineFrameInfo *MFI,
2268 // Create the nodes corresponding to a load from this parameter slot.
2269 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2270 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2271 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2272 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2275 // If value is passed by pointer we have address passed instead of the value
2277 bool ExtendedInMem = VA.isExtInLoc() &&
2278 VA.getValVT().getScalarType() == MVT::i1;
2280 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2281 ValVT = VA.getLocVT();
2283 ValVT = VA.getValVT();
2285 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2286 // changed with more analysis.
2287 // In case of tail call optimization mark all arguments mutable. Since they
2288 // could be overwritten by lowering of arguments in case of a tail call.
2289 if (Flags.isByVal()) {
2290 unsigned Bytes = Flags.getByValSize();
2291 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2292 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2293 return DAG.getFrameIndex(FI, getPointerTy());
2295 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2296 VA.getLocMemOffset(), isImmutable);
2297 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2298 SDValue Val = DAG.getLoad(ValVT, dl, Chain, FIN,
2299 MachinePointerInfo::getFixedStack(FI),
2300 false, false, false, 0);
2301 return ExtendedInMem ?
2302 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2306 // FIXME: Get this from tablegen.
2307 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2308 const X86Subtarget *Subtarget) {
2309 assert(Subtarget->is64Bit());
2311 if (Subtarget->isCallingConvWin64(CallConv)) {
2312 static const MCPhysReg GPR64ArgRegsWin64[] = {
2313 X86::RCX, X86::RDX, X86::R8, X86::R9
2315 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2318 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2319 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2321 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2324 // FIXME: Get this from tablegen.
2325 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2326 CallingConv::ID CallConv,
2327 const X86Subtarget *Subtarget) {
2328 assert(Subtarget->is64Bit());
2329 if (Subtarget->isCallingConvWin64(CallConv)) {
2330 // The XMM registers which might contain var arg parameters are shadowed
2331 // in their paired GPR. So we only need to save the GPR to their home
2333 // TODO: __vectorcall will change this.
2337 const Function *Fn = MF.getFunction();
2338 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2339 bool isSoftFloat = Subtarget->useSoftFloat();
2340 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2341 "SSE register cannot be used when SSE is disabled!");
2342 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2343 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2347 static const MCPhysReg XMMArgRegs64Bit[] = {
2348 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2349 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2351 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2355 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2356 CallingConv::ID CallConv,
2358 const SmallVectorImpl<ISD::InputArg> &Ins,
2361 SmallVectorImpl<SDValue> &InVals)
2363 MachineFunction &MF = DAG.getMachineFunction();
2364 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2365 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2367 const Function* Fn = MF.getFunction();
2368 if (Fn->hasExternalLinkage() &&
2369 Subtarget->isTargetCygMing() &&
2370 Fn->getName() == "main")
2371 FuncInfo->setForceFramePointer(true);
2373 MachineFrameInfo *MFI = MF.getFrameInfo();
2374 bool Is64Bit = Subtarget->is64Bit();
2375 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2377 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2378 "Var args not supported with calling convention fastcc, ghc or hipe");
2380 // Assign locations to all of the incoming arguments.
2381 SmallVector<CCValAssign, 16> ArgLocs;
2382 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2384 // Allocate shadow area for Win64
2386 CCInfo.AllocateStack(32, 8);
2388 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2390 unsigned LastVal = ~0U;
2392 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2393 CCValAssign &VA = ArgLocs[i];
2394 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2396 assert(VA.getValNo() != LastVal &&
2397 "Don't support value assigned to multiple locs yet");
2399 LastVal = VA.getValNo();
2401 if (VA.isRegLoc()) {
2402 EVT RegVT = VA.getLocVT();
2403 const TargetRegisterClass *RC;
2404 if (RegVT == MVT::i32)
2405 RC = &X86::GR32RegClass;
2406 else if (Is64Bit && RegVT == MVT::i64)
2407 RC = &X86::GR64RegClass;
2408 else if (RegVT == MVT::f32)
2409 RC = &X86::FR32RegClass;
2410 else if (RegVT == MVT::f64)
2411 RC = &X86::FR64RegClass;
2412 else if (RegVT.is512BitVector())
2413 RC = &X86::VR512RegClass;
2414 else if (RegVT.is256BitVector())
2415 RC = &X86::VR256RegClass;
2416 else if (RegVT.is128BitVector())
2417 RC = &X86::VR128RegClass;
2418 else if (RegVT == MVT::x86mmx)
2419 RC = &X86::VR64RegClass;
2420 else if (RegVT == MVT::i1)
2421 RC = &X86::VK1RegClass;
2422 else if (RegVT == MVT::v8i1)
2423 RC = &X86::VK8RegClass;
2424 else if (RegVT == MVT::v16i1)
2425 RC = &X86::VK16RegClass;
2426 else if (RegVT == MVT::v32i1)
2427 RC = &X86::VK32RegClass;
2428 else if (RegVT == MVT::v64i1)
2429 RC = &X86::VK64RegClass;
2431 llvm_unreachable("Unknown argument type!");
2433 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2434 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2436 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2437 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2439 if (VA.getLocInfo() == CCValAssign::SExt)
2440 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2441 DAG.getValueType(VA.getValVT()));
2442 else if (VA.getLocInfo() == CCValAssign::ZExt)
2443 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2444 DAG.getValueType(VA.getValVT()));
2445 else if (VA.getLocInfo() == CCValAssign::BCvt)
2446 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2448 if (VA.isExtInLoc()) {
2449 // Handle MMX values passed in XMM regs.
2450 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2451 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2453 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2456 assert(VA.isMemLoc());
2457 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2460 // If value is passed via pointer - do a load.
2461 if (VA.getLocInfo() == CCValAssign::Indirect)
2462 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2463 MachinePointerInfo(), false, false, false, 0);
2465 InVals.push_back(ArgValue);
2468 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2469 // All x86 ABIs require that for returning structs by value we copy the
2470 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2471 // the argument into a virtual register so that we can access it from the
2473 if (Ins[i].Flags.isSRet()) {
2474 unsigned Reg = FuncInfo->getSRetReturnReg();
2476 MVT PtrTy = getPointerTy();
2477 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2478 FuncInfo->setSRetReturnReg(Reg);
2480 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2481 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2486 unsigned StackSize = CCInfo.getNextStackOffset();
2487 // Align stack specially for tail calls.
2488 if (FuncIsMadeTailCallSafe(CallConv,
2489 MF.getTarget().Options.GuaranteedTailCallOpt))
2490 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2492 // If the function takes variable number of arguments, make a frame index for
2493 // the start of the first vararg value... for expansion of llvm.va_start. We
2494 // can skip this if there are no va_start calls.
2495 if (MFI->hasVAStart() &&
2496 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2497 CallConv != CallingConv::X86_ThisCall))) {
2498 FuncInfo->setVarArgsFrameIndex(
2499 MFI->CreateFixedObject(1, StackSize, true));
2502 MachineModuleInfo &MMI = MF.getMMI();
2503 const Function *WinEHParent = nullptr;
2504 if (IsWin64 && MMI.hasWinEHFuncInfo(Fn))
2505 WinEHParent = MMI.getWinEHParent(Fn);
2506 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2507 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2509 // Figure out if XMM registers are in use.
2510 assert(!(Subtarget->useSoftFloat() &&
2511 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2512 "SSE register cannot be used when SSE is disabled!");
2514 // 64-bit calling conventions support varargs and register parameters, so we
2515 // have to do extra work to spill them in the prologue.
2516 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2517 // Find the first unallocated argument registers.
2518 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2519 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2520 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2521 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2522 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2523 "SSE register cannot be used when SSE is disabled!");
2525 // Gather all the live in physical registers.
2526 SmallVector<SDValue, 6> LiveGPRs;
2527 SmallVector<SDValue, 8> LiveXMMRegs;
2529 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2530 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2532 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2534 if (!ArgXMMs.empty()) {
2535 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2536 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2537 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2538 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2539 LiveXMMRegs.push_back(
2540 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2545 // Get to the caller-allocated home save location. Add 8 to account
2546 // for the return address.
2547 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2548 FuncInfo->setRegSaveFrameIndex(
2549 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2550 // Fixup to set vararg frame on shadow area (4 x i64).
2552 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2554 // For X86-64, if there are vararg parameters that are passed via
2555 // registers, then we must store them to their spots on the stack so
2556 // they may be loaded by deferencing the result of va_next.
2557 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2558 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2559 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2560 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2563 // Store the integer parameter registers.
2564 SmallVector<SDValue, 8> MemOps;
2565 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2567 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2568 for (SDValue Val : LiveGPRs) {
2569 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2570 DAG.getIntPtrConstant(Offset, dl));
2572 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2573 MachinePointerInfo::getFixedStack(
2574 FuncInfo->getRegSaveFrameIndex(), Offset),
2576 MemOps.push_back(Store);
2580 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2581 // Now store the XMM (fp + vector) parameter registers.
2582 SmallVector<SDValue, 12> SaveXMMOps;
2583 SaveXMMOps.push_back(Chain);
2584 SaveXMMOps.push_back(ALVal);
2585 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2586 FuncInfo->getRegSaveFrameIndex(), dl));
2587 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2588 FuncInfo->getVarArgsFPOffset(), dl));
2589 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2591 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2592 MVT::Other, SaveXMMOps));
2595 if (!MemOps.empty())
2596 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2597 } else if (IsWinEHOutlined) {
2598 // Get to the caller-allocated home save location. Add 8 to account
2599 // for the return address.
2600 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2601 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2602 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2604 MMI.getWinEHFuncInfo(Fn)
2605 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2606 FuncInfo->getRegSaveFrameIndex();
2608 // Store the second integer parameter (rdx) into rsp+16 relative to the
2609 // stack pointer at the entry of the function.
2611 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), getPointerTy());
2612 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2613 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2614 Chain = DAG.getStore(
2615 Val.getValue(1), dl, Val, RSFIN,
2616 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2617 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2620 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2621 // Find the largest legal vector type.
2622 MVT VecVT = MVT::Other;
2623 // FIXME: Only some x86_32 calling conventions support AVX512.
2624 if (Subtarget->hasAVX512() &&
2625 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2626 CallConv == CallingConv::Intel_OCL_BI)))
2627 VecVT = MVT::v16f32;
2628 else if (Subtarget->hasAVX())
2630 else if (Subtarget->hasSSE2())
2633 // We forward some GPRs and some vector types.
2634 SmallVector<MVT, 2> RegParmTypes;
2635 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2636 RegParmTypes.push_back(IntVT);
2637 if (VecVT != MVT::Other)
2638 RegParmTypes.push_back(VecVT);
2640 // Compute the set of forwarded registers. The rest are scratch.
2641 SmallVectorImpl<ForwardedRegister> &Forwards =
2642 FuncInfo->getForwardedMustTailRegParms();
2643 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2645 // Conservatively forward AL on x86_64, since it might be used for varargs.
2646 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2647 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2648 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2651 // Copy all forwards from physical to virtual registers.
2652 for (ForwardedRegister &F : Forwards) {
2653 // FIXME: Can we use a less constrained schedule?
2654 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2655 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2656 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2660 // Some CCs need callee pop.
2661 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2662 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2663 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2665 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2666 // If this is an sret function, the return should pop the hidden pointer.
2667 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2668 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2669 argsAreStructReturn(Ins) == StackStructReturn)
2670 FuncInfo->setBytesToPopOnReturn(4);
2674 // RegSaveFrameIndex is X86-64 only.
2675 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2676 if (CallConv == CallingConv::X86_FastCall ||
2677 CallConv == CallingConv::X86_ThisCall)
2678 // fastcc functions can't have varargs.
2679 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2682 FuncInfo->setArgumentStackSize(StackSize);
2684 if (IsWinEHParent) {
2685 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2686 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2687 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2688 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2689 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2690 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2691 /*isVolatile=*/true,
2692 /*isNonTemporal=*/false, /*Alignment=*/0);
2699 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2700 SDValue StackPtr, SDValue Arg,
2701 SDLoc dl, SelectionDAG &DAG,
2702 const CCValAssign &VA,
2703 ISD::ArgFlagsTy Flags) const {
2704 unsigned LocMemOffset = VA.getLocMemOffset();
2705 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2706 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2707 if (Flags.isByVal())
2708 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2710 return DAG.getStore(Chain, dl, Arg, PtrOff,
2711 MachinePointerInfo::getStack(LocMemOffset),
2715 /// Emit a load of return address if tail call
2716 /// optimization is performed and it is required.
2718 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2719 SDValue &OutRetAddr, SDValue Chain,
2720 bool IsTailCall, bool Is64Bit,
2721 int FPDiff, SDLoc dl) const {
2722 // Adjust the Return address stack slot.
2723 EVT VT = getPointerTy();
2724 OutRetAddr = getReturnAddressFrameIndex(DAG);
2726 // Load the "old" Return address.
2727 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2728 false, false, false, 0);
2729 return SDValue(OutRetAddr.getNode(), 1);
2732 /// Emit a store of the return address if tail call
2733 /// optimization is performed and it is required (FPDiff!=0).
2734 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2735 SDValue Chain, SDValue RetAddrFrIdx,
2736 EVT PtrVT, unsigned SlotSize,
2737 int FPDiff, SDLoc dl) {
2738 // Store the return address to the appropriate stack slot.
2739 if (!FPDiff) return Chain;
2740 // Calculate the new stack slot for the return address.
2741 int NewReturnAddrFI =
2742 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2744 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2745 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2746 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2752 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2753 SmallVectorImpl<SDValue> &InVals) const {
2754 SelectionDAG &DAG = CLI.DAG;
2756 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2757 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2758 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2759 SDValue Chain = CLI.Chain;
2760 SDValue Callee = CLI.Callee;
2761 CallingConv::ID CallConv = CLI.CallConv;
2762 bool &isTailCall = CLI.IsTailCall;
2763 bool isVarArg = CLI.IsVarArg;
2765 MachineFunction &MF = DAG.getMachineFunction();
2766 bool Is64Bit = Subtarget->is64Bit();
2767 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2768 StructReturnType SR = callIsStructReturn(Outs);
2769 bool IsSibcall = false;
2770 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2772 if (MF.getTarget().Options.DisableTailCalls)
2775 if (Subtarget->isPICStyleGOT() &&
2776 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2777 // If we are using a GOT, disable tail calls to external symbols with
2778 // default visibility. Tail calling such a symbol requires using a GOT
2779 // relocation, which forces early binding of the symbol. This breaks code
2780 // that require lazy function symbol resolution. Using musttail or
2781 // GuaranteedTailCallOpt will override this.
2782 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2783 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2784 G->getGlobal()->hasDefaultVisibility()))
2788 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2790 // Force this to be a tail call. The verifier rules are enough to ensure
2791 // that we can lower this successfully without moving the return address
2794 } else if (isTailCall) {
2795 // Check if it's really possible to do a tail call.
2796 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2797 isVarArg, SR != NotStructReturn,
2798 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2799 Outs, OutVals, Ins, DAG);
2801 // Sibcalls are automatically detected tailcalls which do not require
2803 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2810 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2811 "Var args not supported with calling convention fastcc, ghc or hipe");
2813 // Analyze operands of the call, assigning locations to each operand.
2814 SmallVector<CCValAssign, 16> ArgLocs;
2815 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2817 // Allocate shadow area for Win64
2819 CCInfo.AllocateStack(32, 8);
2821 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2823 // Get a count of how many bytes are to be pushed on the stack.
2824 unsigned NumBytes = CCInfo.getNextStackOffset();
2826 // This is a sibcall. The memory operands are available in caller's
2827 // own caller's stack.
2829 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2830 IsTailCallConvention(CallConv))
2831 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2834 if (isTailCall && !IsSibcall && !IsMustTail) {
2835 // Lower arguments at fp - stackoffset + fpdiff.
2836 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2838 FPDiff = NumBytesCallerPushed - NumBytes;
2840 // Set the delta of movement of the returnaddr stackslot.
2841 // But only set if delta is greater than previous delta.
2842 if (FPDiff < X86Info->getTCReturnAddrDelta())
2843 X86Info->setTCReturnAddrDelta(FPDiff);
2846 unsigned NumBytesToPush = NumBytes;
2847 unsigned NumBytesToPop = NumBytes;
2849 // If we have an inalloca argument, all stack space has already been allocated
2850 // for us and be right at the top of the stack. We don't support multiple
2851 // arguments passed in memory when using inalloca.
2852 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2854 if (!ArgLocs.back().isMemLoc())
2855 report_fatal_error("cannot use inalloca attribute on a register "
2857 if (ArgLocs.back().getLocMemOffset() != 0)
2858 report_fatal_error("any parameter with the inalloca attribute must be "
2859 "the only memory argument");
2863 Chain = DAG.getCALLSEQ_START(
2864 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
2866 SDValue RetAddrFrIdx;
2867 // Load return address for tail calls.
2868 if (isTailCall && FPDiff)
2869 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2870 Is64Bit, FPDiff, dl);
2872 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2873 SmallVector<SDValue, 8> MemOpChains;
2876 // Walk the register/memloc assignments, inserting copies/loads. In the case
2877 // of tail call optimization arguments are handle later.
2878 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2879 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2880 // Skip inalloca arguments, they have already been written.
2881 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2882 if (Flags.isInAlloca())
2885 CCValAssign &VA = ArgLocs[i];
2886 EVT RegVT = VA.getLocVT();
2887 SDValue Arg = OutVals[i];
2888 bool isByVal = Flags.isByVal();
2890 // Promote the value if needed.
2891 switch (VA.getLocInfo()) {
2892 default: llvm_unreachable("Unknown loc info!");
2893 case CCValAssign::Full: break;
2894 case CCValAssign::SExt:
2895 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2897 case CCValAssign::ZExt:
2898 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2900 case CCValAssign::AExt:
2901 if (Arg.getValueType().isVector() &&
2902 Arg.getValueType().getScalarType() == MVT::i1)
2903 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2904 else if (RegVT.is128BitVector()) {
2905 // Special case: passing MMX values in XMM registers.
2906 Arg = DAG.getBitcast(MVT::i64, Arg);
2907 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2908 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2910 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2912 case CCValAssign::BCvt:
2913 Arg = DAG.getBitcast(RegVT, Arg);
2915 case CCValAssign::Indirect: {
2916 // Store the argument.
2917 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2918 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2919 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2920 MachinePointerInfo::getFixedStack(FI),
2927 if (VA.isRegLoc()) {
2928 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2929 if (isVarArg && IsWin64) {
2930 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2931 // shadow reg if callee is a varargs function.
2932 unsigned ShadowReg = 0;
2933 switch (VA.getLocReg()) {
2934 case X86::XMM0: ShadowReg = X86::RCX; break;
2935 case X86::XMM1: ShadowReg = X86::RDX; break;
2936 case X86::XMM2: ShadowReg = X86::R8; break;
2937 case X86::XMM3: ShadowReg = X86::R9; break;
2940 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2942 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2943 assert(VA.isMemLoc());
2944 if (!StackPtr.getNode())
2945 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2947 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2948 dl, DAG, VA, Flags));
2952 if (!MemOpChains.empty())
2953 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2955 if (Subtarget->isPICStyleGOT()) {
2956 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2959 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2960 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2962 // If we are tail calling and generating PIC/GOT style code load the
2963 // address of the callee into ECX. The value in ecx is used as target of
2964 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2965 // for tail calls on PIC/GOT architectures. Normally we would just put the
2966 // address of GOT into ebx and then call target@PLT. But for tail calls
2967 // ebx would be restored (since ebx is callee saved) before jumping to the
2970 // Note: The actual moving to ECX is done further down.
2971 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2972 if (G && !G->getGlobal()->hasLocalLinkage() &&
2973 G->getGlobal()->hasDefaultVisibility())
2974 Callee = LowerGlobalAddress(Callee, DAG);
2975 else if (isa<ExternalSymbolSDNode>(Callee))
2976 Callee = LowerExternalSymbol(Callee, DAG);
2980 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2981 // From AMD64 ABI document:
2982 // For calls that may call functions that use varargs or stdargs
2983 // (prototype-less calls or calls to functions containing ellipsis (...) in
2984 // the declaration) %al is used as hidden argument to specify the number
2985 // of SSE registers used. The contents of %al do not need to match exactly
2986 // the number of registers, but must be an ubound on the number of SSE
2987 // registers used and is in the range 0 - 8 inclusive.
2989 // Count the number of XMM registers allocated.
2990 static const MCPhysReg XMMArgRegs[] = {
2991 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2992 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2994 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2995 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2996 && "SSE registers cannot be used when SSE is disabled");
2998 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2999 DAG.getConstant(NumXMMRegs, dl,
3003 if (isVarArg && IsMustTail) {
3004 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3005 for (const auto &F : Forwards) {
3006 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3007 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3011 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3012 // don't need this because the eligibility check rejects calls that require
3013 // shuffling arguments passed in memory.
3014 if (!IsSibcall && isTailCall) {
3015 // Force all the incoming stack arguments to be loaded from the stack
3016 // before any new outgoing arguments are stored to the stack, because the
3017 // outgoing stack slots may alias the incoming argument stack slots, and
3018 // the alias isn't otherwise explicit. This is slightly more conservative
3019 // than necessary, because it means that each store effectively depends
3020 // on every argument instead of just those arguments it would clobber.
3021 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3023 SmallVector<SDValue, 8> MemOpChains2;
3026 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3027 CCValAssign &VA = ArgLocs[i];
3030 assert(VA.isMemLoc());
3031 SDValue Arg = OutVals[i];
3032 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3033 // Skip inalloca arguments. They don't require any work.
3034 if (Flags.isInAlloca())
3036 // Create frame index.
3037 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3038 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3039 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3040 FIN = DAG.getFrameIndex(FI, getPointerTy());
3042 if (Flags.isByVal()) {
3043 // Copy relative to framepointer.
3044 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3045 if (!StackPtr.getNode())
3046 StackPtr = DAG.getCopyFromReg(Chain, dl,
3047 RegInfo->getStackRegister(),
3049 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3051 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3055 // Store relative to framepointer.
3056 MemOpChains2.push_back(
3057 DAG.getStore(ArgChain, dl, Arg, FIN,
3058 MachinePointerInfo::getFixedStack(FI),
3063 if (!MemOpChains2.empty())
3064 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3066 // Store the return address to the appropriate stack slot.
3067 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3068 getPointerTy(), RegInfo->getSlotSize(),
3072 // Build a sequence of copy-to-reg nodes chained together with token chain
3073 // and flag operands which copy the outgoing args into registers.
3075 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3076 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3077 RegsToPass[i].second, InFlag);
3078 InFlag = Chain.getValue(1);
3081 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3082 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3083 // In the 64-bit large code model, we have to make all calls
3084 // through a register, since the call instruction's 32-bit
3085 // pc-relative offset may not be large enough to hold the whole
3087 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3088 // If the callee is a GlobalAddress node (quite common, every direct call
3089 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3091 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3093 // We should use extra load for direct calls to dllimported functions in
3095 const GlobalValue *GV = G->getGlobal();
3096 if (!GV->hasDLLImportStorageClass()) {
3097 unsigned char OpFlags = 0;
3098 bool ExtraLoad = false;
3099 unsigned WrapperKind = ISD::DELETED_NODE;
3101 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3102 // external symbols most go through the PLT in PIC mode. If the symbol
3103 // has hidden or protected visibility, or if it is static or local, then
3104 // we don't need to use the PLT - we can directly call it.
3105 if (Subtarget->isTargetELF() &&
3106 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3107 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3108 OpFlags = X86II::MO_PLT;
3109 } else if (Subtarget->isPICStyleStubAny() &&
3110 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3111 (!Subtarget->getTargetTriple().isMacOSX() ||
3112 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3113 // PC-relative references to external symbols should go through $stub,
3114 // unless we're building with the leopard linker or later, which
3115 // automatically synthesizes these stubs.
3116 OpFlags = X86II::MO_DARWIN_STUB;
3117 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3118 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3119 // If the function is marked as non-lazy, generate an indirect call
3120 // which loads from the GOT directly. This avoids runtime overhead
3121 // at the cost of eager binding (and one extra byte of encoding).
3122 OpFlags = X86II::MO_GOTPCREL;
3123 WrapperKind = X86ISD::WrapperRIP;
3127 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3128 G->getOffset(), OpFlags);
3130 // Add a wrapper if needed.
3131 if (WrapperKind != ISD::DELETED_NODE)
3132 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3133 // Add extra indirection if needed.
3135 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3136 MachinePointerInfo::getGOT(),
3137 false, false, false, 0);
3139 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3140 unsigned char OpFlags = 0;
3142 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3143 // external symbols should go through the PLT.
3144 if (Subtarget->isTargetELF() &&
3145 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3146 OpFlags = X86II::MO_PLT;
3147 } else if (Subtarget->isPICStyleStubAny() &&
3148 (!Subtarget->getTargetTriple().isMacOSX() ||
3149 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3150 // PC-relative references to external symbols should go through $stub,
3151 // unless we're building with the leopard linker or later, which
3152 // automatically synthesizes these stubs.
3153 OpFlags = X86II::MO_DARWIN_STUB;
3156 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3158 } else if (Subtarget->isTarget64BitILP32() &&
3159 Callee->getValueType(0) == MVT::i32) {
3160 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3161 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3164 // Returns a chain & a flag for retval copy to use.
3165 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3166 SmallVector<SDValue, 8> Ops;
3168 if (!IsSibcall && isTailCall) {
3169 Chain = DAG.getCALLSEQ_END(Chain,
3170 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3171 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3172 InFlag = Chain.getValue(1);
3175 Ops.push_back(Chain);
3176 Ops.push_back(Callee);
3179 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3181 // Add argument registers to the end of the list so that they are known live
3183 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3184 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3185 RegsToPass[i].second.getValueType()));
3187 // Add a register mask operand representing the call-preserved registers.
3188 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3189 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
3190 assert(Mask && "Missing call preserved mask for calling convention");
3191 Ops.push_back(DAG.getRegisterMask(Mask));
3193 if (InFlag.getNode())
3194 Ops.push_back(InFlag);
3198 //// If this is the first return lowered for this function, add the regs
3199 //// to the liveout set for the function.
3200 // This isn't right, although it's probably harmless on x86; liveouts
3201 // should be computed from returns not tail calls. Consider a void
3202 // function making a tail call to a function returning int.
3203 MF.getFrameInfo()->setHasTailCall();
3204 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3207 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3208 InFlag = Chain.getValue(1);
3210 // Create the CALLSEQ_END node.
3211 unsigned NumBytesForCalleeToPop;
3212 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3213 DAG.getTarget().Options.GuaranteedTailCallOpt))
3214 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3215 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3216 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3217 SR == StackStructReturn)
3218 // If this is a call to a struct-return function, the callee
3219 // pops the hidden struct pointer, so we have to push it back.
3220 // This is common for Darwin/X86, Linux & Mingw32 targets.
3221 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3222 NumBytesForCalleeToPop = 4;
3224 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3226 // Returns a flag for retval copy to use.
3228 Chain = DAG.getCALLSEQ_END(Chain,
3229 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3230 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3233 InFlag = Chain.getValue(1);
3236 // Handle result values, copying them out of physregs into vregs that we
3238 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3239 Ins, dl, DAG, InVals);
3242 //===----------------------------------------------------------------------===//
3243 // Fast Calling Convention (tail call) implementation
3244 //===----------------------------------------------------------------------===//
3246 // Like std call, callee cleans arguments, convention except that ECX is
3247 // reserved for storing the tail called function address. Only 2 registers are
3248 // free for argument passing (inreg). Tail call optimization is performed
3250 // * tailcallopt is enabled
3251 // * caller/callee are fastcc
3252 // On X86_64 architecture with GOT-style position independent code only local
3253 // (within module) calls are supported at the moment.
3254 // To keep the stack aligned according to platform abi the function
3255 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3256 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3257 // If a tail called function callee has more arguments than the caller the
3258 // caller needs to make sure that there is room to move the RETADDR to. This is
3259 // achieved by reserving an area the size of the argument delta right after the
3260 // original RETADDR, but before the saved framepointer or the spilled registers
3261 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3273 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3274 /// for a 16 byte align requirement.
3276 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3277 SelectionDAG& DAG) const {
3278 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3279 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3280 unsigned StackAlignment = TFI.getStackAlignment();
3281 uint64_t AlignMask = StackAlignment - 1;
3282 int64_t Offset = StackSize;
3283 unsigned SlotSize = RegInfo->getSlotSize();
3284 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3285 // Number smaller than 12 so just add the difference.
3286 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3288 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3289 Offset = ((~AlignMask) & Offset) + StackAlignment +
3290 (StackAlignment-SlotSize);
3295 /// MatchingStackOffset - Return true if the given stack call argument is
3296 /// already available in the same position (relatively) of the caller's
3297 /// incoming argument stack.
3299 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3300 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3301 const X86InstrInfo *TII) {
3302 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3304 if (Arg.getOpcode() == ISD::CopyFromReg) {
3305 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3306 if (!TargetRegisterInfo::isVirtualRegister(VR))
3308 MachineInstr *Def = MRI->getVRegDef(VR);
3311 if (!Flags.isByVal()) {
3312 if (!TII->isLoadFromStackSlot(Def, FI))
3315 unsigned Opcode = Def->getOpcode();
3316 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3317 Opcode == X86::LEA64_32r) &&
3318 Def->getOperand(1).isFI()) {
3319 FI = Def->getOperand(1).getIndex();
3320 Bytes = Flags.getByValSize();
3324 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3325 if (Flags.isByVal())
3326 // ByVal argument is passed in as a pointer but it's now being
3327 // dereferenced. e.g.
3328 // define @foo(%struct.X* %A) {
3329 // tail call @bar(%struct.X* byval %A)
3332 SDValue Ptr = Ld->getBasePtr();
3333 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3336 FI = FINode->getIndex();
3337 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3338 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3339 FI = FINode->getIndex();
3340 Bytes = Flags.getByValSize();
3344 assert(FI != INT_MAX);
3345 if (!MFI->isFixedObjectIndex(FI))
3347 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3350 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3351 /// for tail call optimization. Targets which want to do tail call
3352 /// optimization should implement this function.
3354 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3355 CallingConv::ID CalleeCC,
3357 bool isCalleeStructRet,
3358 bool isCallerStructRet,
3360 const SmallVectorImpl<ISD::OutputArg> &Outs,
3361 const SmallVectorImpl<SDValue> &OutVals,
3362 const SmallVectorImpl<ISD::InputArg> &Ins,
3363 SelectionDAG &DAG) const {
3364 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3367 // If -tailcallopt is specified, make fastcc functions tail-callable.
3368 const MachineFunction &MF = DAG.getMachineFunction();
3369 const Function *CallerF = MF.getFunction();
3371 // If the function return type is x86_fp80 and the callee return type is not,
3372 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3373 // perform a tailcall optimization here.
3374 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3377 CallingConv::ID CallerCC = CallerF->getCallingConv();
3378 bool CCMatch = CallerCC == CalleeCC;
3379 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3380 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3382 // Win64 functions have extra shadow space for argument homing. Don't do the
3383 // sibcall if the caller and callee have mismatched expectations for this
3385 if (IsCalleeWin64 != IsCallerWin64)
3388 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3389 if (IsTailCallConvention(CalleeCC) && CCMatch)
3394 // Look for obvious safe cases to perform tail call optimization that do not
3395 // require ABI changes. This is what gcc calls sibcall.
3397 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3398 // emit a special epilogue.
3399 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3400 if (RegInfo->needsStackRealignment(MF))
3403 // Also avoid sibcall optimization if either caller or callee uses struct
3404 // return semantics.
3405 if (isCalleeStructRet || isCallerStructRet)
3408 // An stdcall/thiscall caller is expected to clean up its arguments; the
3409 // callee isn't going to do that.
3410 // FIXME: this is more restrictive than needed. We could produce a tailcall
3411 // when the stack adjustment matches. For example, with a thiscall that takes
3412 // only one argument.
3413 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3414 CallerCC == CallingConv::X86_ThisCall))
3417 // Do not sibcall optimize vararg calls unless all arguments are passed via
3419 if (isVarArg && !Outs.empty()) {
3421 // Optimizing for varargs on Win64 is unlikely to be safe without
3422 // additional testing.
3423 if (IsCalleeWin64 || IsCallerWin64)
3426 SmallVector<CCValAssign, 16> ArgLocs;
3427 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3430 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3431 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3432 if (!ArgLocs[i].isRegLoc())
3436 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3437 // stack. Therefore, if it's not used by the call it is not safe to optimize
3438 // this into a sibcall.
3439 bool Unused = false;
3440 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3447 SmallVector<CCValAssign, 16> RVLocs;
3448 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3450 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3451 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3452 CCValAssign &VA = RVLocs[i];
3453 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3458 // If the calling conventions do not match, then we'd better make sure the
3459 // results are returned in the same way as what the caller expects.
3461 SmallVector<CCValAssign, 16> RVLocs1;
3462 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3464 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3466 SmallVector<CCValAssign, 16> RVLocs2;
3467 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3469 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3471 if (RVLocs1.size() != RVLocs2.size())
3473 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3474 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3476 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3478 if (RVLocs1[i].isRegLoc()) {
3479 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3482 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3488 // If the callee takes no arguments then go on to check the results of the
3490 if (!Outs.empty()) {
3491 // Check if stack adjustment is needed. For now, do not do this if any
3492 // argument is passed on the stack.
3493 SmallVector<CCValAssign, 16> ArgLocs;
3494 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3497 // Allocate shadow area for Win64
3499 CCInfo.AllocateStack(32, 8);
3501 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3502 if (CCInfo.getNextStackOffset()) {
3503 MachineFunction &MF = DAG.getMachineFunction();
3504 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3507 // Check if the arguments are already laid out in the right way as
3508 // the caller's fixed stack objects.
3509 MachineFrameInfo *MFI = MF.getFrameInfo();
3510 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3511 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3512 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3513 CCValAssign &VA = ArgLocs[i];
3514 SDValue Arg = OutVals[i];
3515 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3516 if (VA.getLocInfo() == CCValAssign::Indirect)
3518 if (!VA.isRegLoc()) {
3519 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3526 // If the tailcall address may be in a register, then make sure it's
3527 // possible to register allocate for it. In 32-bit, the call address can
3528 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3529 // callee-saved registers are restored. These happen to be the same
3530 // registers used to pass 'inreg' arguments so watch out for those.
3531 if (!Subtarget->is64Bit() &&
3532 ((!isa<GlobalAddressSDNode>(Callee) &&
3533 !isa<ExternalSymbolSDNode>(Callee)) ||
3534 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3535 unsigned NumInRegs = 0;
3536 // In PIC we need an extra register to formulate the address computation
3538 unsigned MaxInRegs =
3539 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3541 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3542 CCValAssign &VA = ArgLocs[i];
3545 unsigned Reg = VA.getLocReg();
3548 case X86::EAX: case X86::EDX: case X86::ECX:
3549 if (++NumInRegs == MaxInRegs)
3561 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3562 const TargetLibraryInfo *libInfo) const {
3563 return X86::createFastISel(funcInfo, libInfo);
3566 //===----------------------------------------------------------------------===//
3567 // Other Lowering Hooks
3568 //===----------------------------------------------------------------------===//
3570 static bool MayFoldLoad(SDValue Op) {
3571 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3574 static bool MayFoldIntoStore(SDValue Op) {
3575 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3578 static bool isTargetShuffle(unsigned Opcode) {
3580 default: return false;
3581 case X86ISD::BLENDI:
3582 case X86ISD::PSHUFB:
3583 case X86ISD::PSHUFD:
3584 case X86ISD::PSHUFHW:
3585 case X86ISD::PSHUFLW:
3587 case X86ISD::PALIGNR:
3588 case X86ISD::MOVLHPS:
3589 case X86ISD::MOVLHPD:
3590 case X86ISD::MOVHLPS:
3591 case X86ISD::MOVLPS:
3592 case X86ISD::MOVLPD:
3593 case X86ISD::MOVSHDUP:
3594 case X86ISD::MOVSLDUP:
3595 case X86ISD::MOVDDUP:
3598 case X86ISD::UNPCKL:
3599 case X86ISD::UNPCKH:
3600 case X86ISD::VPERMILPI:
3601 case X86ISD::VPERM2X128:
3602 case X86ISD::VPERMI:
3607 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3608 SDValue V1, unsigned TargetMask,
3609 SelectionDAG &DAG) {
3611 default: llvm_unreachable("Unknown x86 shuffle node");
3612 case X86ISD::PSHUFD:
3613 case X86ISD::PSHUFHW:
3614 case X86ISD::PSHUFLW:
3615 case X86ISD::VPERMILPI:
3616 case X86ISD::VPERMI:
3617 return DAG.getNode(Opc, dl, VT, V1,
3618 DAG.getConstant(TargetMask, dl, MVT::i8));
3622 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3623 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3625 default: llvm_unreachable("Unknown x86 shuffle node");
3626 case X86ISD::MOVLHPS:
3627 case X86ISD::MOVLHPD:
3628 case X86ISD::MOVHLPS:
3629 case X86ISD::MOVLPS:
3630 case X86ISD::MOVLPD:
3633 case X86ISD::UNPCKL:
3634 case X86ISD::UNPCKH:
3635 return DAG.getNode(Opc, dl, VT, V1, V2);
3639 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3640 MachineFunction &MF = DAG.getMachineFunction();
3641 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3642 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3643 int ReturnAddrIndex = FuncInfo->getRAIndex();
3645 if (ReturnAddrIndex == 0) {
3646 // Set up a frame object for the return address.
3647 unsigned SlotSize = RegInfo->getSlotSize();
3648 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3651 FuncInfo->setRAIndex(ReturnAddrIndex);
3654 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3657 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3658 bool hasSymbolicDisplacement) {
3659 // Offset should fit into 32 bit immediate field.
3660 if (!isInt<32>(Offset))
3663 // If we don't have a symbolic displacement - we don't have any extra
3665 if (!hasSymbolicDisplacement)
3668 // FIXME: Some tweaks might be needed for medium code model.
3669 if (M != CodeModel::Small && M != CodeModel::Kernel)
3672 // For small code model we assume that latest object is 16MB before end of 31
3673 // bits boundary. We may also accept pretty large negative constants knowing
3674 // that all objects are in the positive half of address space.
3675 if (M == CodeModel::Small && Offset < 16*1024*1024)
3678 // For kernel code model we know that all object resist in the negative half
3679 // of 32bits address space. We may not accept negative offsets, since they may
3680 // be just off and we may accept pretty large positive ones.
3681 if (M == CodeModel::Kernel && Offset >= 0)
3687 /// isCalleePop - Determines whether the callee is required to pop its
3688 /// own arguments. Callee pop is necessary to support tail calls.
3689 bool X86::isCalleePop(CallingConv::ID CallingConv,
3690 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3691 switch (CallingConv) {
3694 case CallingConv::X86_StdCall:
3695 case CallingConv::X86_FastCall:
3696 case CallingConv::X86_ThisCall:
3698 case CallingConv::Fast:
3699 case CallingConv::GHC:
3700 case CallingConv::HiPE:
3707 /// \brief Return true if the condition is an unsigned comparison operation.
3708 static bool isX86CCUnsigned(unsigned X86CC) {
3710 default: llvm_unreachable("Invalid integer condition!");
3711 case X86::COND_E: return true;
3712 case X86::COND_G: return false;
3713 case X86::COND_GE: return false;
3714 case X86::COND_L: return false;
3715 case X86::COND_LE: return false;
3716 case X86::COND_NE: return true;
3717 case X86::COND_B: return true;
3718 case X86::COND_A: return true;
3719 case X86::COND_BE: return true;
3720 case X86::COND_AE: return true;
3722 llvm_unreachable("covered switch fell through?!");
3725 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3726 /// specific condition code, returning the condition code and the LHS/RHS of the
3727 /// comparison to make.
3728 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3729 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3731 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3732 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3733 // X > -1 -> X == 0, jump !sign.
3734 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3735 return X86::COND_NS;
3737 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3738 // X < 0 -> X == 0, jump on sign.
3741 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3743 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3744 return X86::COND_LE;
3748 switch (SetCCOpcode) {
3749 default: llvm_unreachable("Invalid integer condition!");
3750 case ISD::SETEQ: return X86::COND_E;
3751 case ISD::SETGT: return X86::COND_G;
3752 case ISD::SETGE: return X86::COND_GE;
3753 case ISD::SETLT: return X86::COND_L;
3754 case ISD::SETLE: return X86::COND_LE;
3755 case ISD::SETNE: return X86::COND_NE;
3756 case ISD::SETULT: return X86::COND_B;
3757 case ISD::SETUGT: return X86::COND_A;
3758 case ISD::SETULE: return X86::COND_BE;
3759 case ISD::SETUGE: return X86::COND_AE;
3763 // First determine if it is required or is profitable to flip the operands.
3765 // If LHS is a foldable load, but RHS is not, flip the condition.
3766 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3767 !ISD::isNON_EXTLoad(RHS.getNode())) {
3768 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3769 std::swap(LHS, RHS);
3772 switch (SetCCOpcode) {
3778 std::swap(LHS, RHS);
3782 // On a floating point condition, the flags are set as follows:
3784 // 0 | 0 | 0 | X > Y
3785 // 0 | 0 | 1 | X < Y
3786 // 1 | 0 | 0 | X == Y
3787 // 1 | 1 | 1 | unordered
3788 switch (SetCCOpcode) {
3789 default: llvm_unreachable("Condcode should be pre-legalized away");
3791 case ISD::SETEQ: return X86::COND_E;
3792 case ISD::SETOLT: // flipped
3794 case ISD::SETGT: return X86::COND_A;
3795 case ISD::SETOLE: // flipped
3797 case ISD::SETGE: return X86::COND_AE;
3798 case ISD::SETUGT: // flipped
3800 case ISD::SETLT: return X86::COND_B;
3801 case ISD::SETUGE: // flipped
3803 case ISD::SETLE: return X86::COND_BE;
3805 case ISD::SETNE: return X86::COND_NE;
3806 case ISD::SETUO: return X86::COND_P;
3807 case ISD::SETO: return X86::COND_NP;
3809 case ISD::SETUNE: return X86::COND_INVALID;
3813 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3814 /// code. Current x86 isa includes the following FP cmov instructions:
3815 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3816 static bool hasFPCMov(unsigned X86CC) {
3832 /// isFPImmLegal - Returns true if the target can instruction select the
3833 /// specified FP immediate natively. If false, the legalizer will
3834 /// materialize the FP immediate as a load from a constant pool.
3835 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3836 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3837 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3843 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3844 ISD::LoadExtType ExtTy,
3846 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3847 // relocation target a movq or addq instruction: don't let the load shrink.
3848 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3849 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3850 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3851 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3855 /// \brief Returns true if it is beneficial to convert a load of a constant
3856 /// to just the constant itself.
3857 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3859 assert(Ty->isIntegerTy());
3861 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3862 if (BitSize == 0 || BitSize > 64)
3867 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3868 unsigned Index) const {
3869 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3872 return (Index == 0 || Index == ResVT.getVectorNumElements());
3875 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3876 // Speculate cttz only if we can directly use TZCNT.
3877 return Subtarget->hasBMI();
3880 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3881 // Speculate ctlz only if we can directly use LZCNT.
3882 return Subtarget->hasLZCNT();
3885 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3886 /// the specified range (L, H].
3887 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3888 return (Val < 0) || (Val >= Low && Val < Hi);
3891 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3892 /// specified value.
3893 static bool isUndefOrEqual(int Val, int CmpVal) {
3894 return (Val < 0 || Val == CmpVal);
3897 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3898 /// from position Pos and ending in Pos+Size, falls within the specified
3899 /// sequential range (Low, Low+Size]. or is undef.
3900 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3901 unsigned Pos, unsigned Size, int Low) {
3902 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3903 if (!isUndefOrEqual(Mask[i], Low))
3908 /// isVEXTRACTIndex - Return true if the specified
3909 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3910 /// suitable for instruction that extract 128 or 256 bit vectors
3911 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3912 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3913 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3916 // The index should be aligned on a vecWidth-bit boundary.
3918 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3920 MVT VT = N->getSimpleValueType(0);
3921 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3922 bool Result = (Index * ElSize) % vecWidth == 0;
3927 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3928 /// operand specifies a subvector insert that is suitable for input to
3929 /// insertion of 128 or 256-bit subvectors
3930 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3931 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3932 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3934 // The index should be aligned on a vecWidth-bit boundary.
3936 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3938 MVT VT = N->getSimpleValueType(0);
3939 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3940 bool Result = (Index * ElSize) % vecWidth == 0;
3945 bool X86::isVINSERT128Index(SDNode *N) {
3946 return isVINSERTIndex(N, 128);
3949 bool X86::isVINSERT256Index(SDNode *N) {
3950 return isVINSERTIndex(N, 256);
3953 bool X86::isVEXTRACT128Index(SDNode *N) {
3954 return isVEXTRACTIndex(N, 128);
3957 bool X86::isVEXTRACT256Index(SDNode *N) {
3958 return isVEXTRACTIndex(N, 256);
3961 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3962 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3963 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3964 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3967 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3969 MVT VecVT = N->getOperand(0).getSimpleValueType();
3970 MVT ElVT = VecVT.getVectorElementType();
3972 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3973 return Index / NumElemsPerChunk;
3976 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3977 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3978 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3979 llvm_unreachable("Illegal insert subvector for VINSERT");
3982 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3984 MVT VecVT = N->getSimpleValueType(0);
3985 MVT ElVT = VecVT.getVectorElementType();
3987 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3988 return Index / NumElemsPerChunk;
3991 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
3992 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3993 /// and VINSERTI128 instructions.
3994 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
3995 return getExtractVEXTRACTImmediate(N, 128);
3998 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
3999 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4000 /// and VINSERTI64x4 instructions.
4001 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4002 return getExtractVEXTRACTImmediate(N, 256);
4005 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4006 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4007 /// and VINSERTI128 instructions.
4008 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4009 return getInsertVINSERTImmediate(N, 128);
4012 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4013 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4014 /// and VINSERTI64x4 instructions.
4015 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4016 return getInsertVINSERTImmediate(N, 256);
4019 /// isZero - Returns true if Elt is a constant integer zero
4020 static bool isZero(SDValue V) {
4021 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4022 return C && C->isNullValue();
4025 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4027 bool X86::isZeroNode(SDValue Elt) {
4030 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4031 return CFP->getValueAPF().isPosZero();
4035 /// getZeroVector - Returns a vector of specified type with all zero elements.
4037 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4038 SelectionDAG &DAG, SDLoc dl) {
4039 assert(VT.isVector() && "Expected a vector type");
4041 // Always build SSE zero vectors as <4 x i32> bitcasted
4042 // to their dest type. This ensures they get CSE'd.
4044 if (VT.is128BitVector()) { // SSE
4045 if (Subtarget->hasSSE2()) { // SSE2
4046 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4047 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4049 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4050 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4052 } else if (VT.is256BitVector()) { // AVX
4053 if (Subtarget->hasInt256()) { // AVX2
4054 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4055 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4056 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4058 // 256-bit logic and arithmetic instructions in AVX are all
4059 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4060 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4061 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4062 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4064 } else if (VT.is512BitVector()) { // AVX-512
4065 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4066 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4067 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4068 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4069 } else if (VT.getScalarType() == MVT::i1) {
4071 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4072 && "Unexpected vector type");
4073 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4074 && "Unexpected vector type");
4075 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4076 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4077 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4079 llvm_unreachable("Unexpected vector type");
4081 return DAG.getBitcast(VT, Vec);
4084 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4085 SelectionDAG &DAG, SDLoc dl,
4086 unsigned vectorWidth) {
4087 assert((vectorWidth == 128 || vectorWidth == 256) &&
4088 "Unsupported vector width");
4089 EVT VT = Vec.getValueType();
4090 EVT ElVT = VT.getVectorElementType();
4091 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4092 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4093 VT.getVectorNumElements()/Factor);
4095 // Extract from UNDEF is UNDEF.
4096 if (Vec.getOpcode() == ISD::UNDEF)
4097 return DAG.getUNDEF(ResultVT);
4099 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4100 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4102 // This is the index of the first element of the vectorWidth-bit chunk
4104 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4107 // If the input is a buildvector just emit a smaller one.
4108 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4109 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4110 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4113 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4114 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4117 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4118 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4119 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4120 /// instructions or a simple subregister reference. Idx is an index in the
4121 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4122 /// lowering EXTRACT_VECTOR_ELT operations easier.
4123 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4124 SelectionDAG &DAG, SDLoc dl) {
4125 assert((Vec.getValueType().is256BitVector() ||
4126 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4127 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4130 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4131 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4132 SelectionDAG &DAG, SDLoc dl) {
4133 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4134 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4137 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4138 unsigned IdxVal, SelectionDAG &DAG,
4139 SDLoc dl, unsigned vectorWidth) {
4140 assert((vectorWidth == 128 || vectorWidth == 256) &&
4141 "Unsupported vector width");
4142 // Inserting UNDEF is Result
4143 if (Vec.getOpcode() == ISD::UNDEF)
4145 EVT VT = Vec.getValueType();
4146 EVT ElVT = VT.getVectorElementType();
4147 EVT ResultVT = Result.getValueType();
4149 // Insert the relevant vectorWidth bits.
4150 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4152 // This is the index of the first element of the vectorWidth-bit chunk
4154 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4157 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4158 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4161 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4162 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4163 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4164 /// simple superregister reference. Idx is an index in the 128 bits
4165 /// we want. It need not be aligned to a 128-bit boundary. That makes
4166 /// lowering INSERT_VECTOR_ELT operations easier.
4167 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4168 SelectionDAG &DAG, SDLoc dl) {
4169 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4171 // For insertion into the zero index (low half) of a 256-bit vector, it is
4172 // more efficient to generate a blend with immediate instead of an insert*128.
4173 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4174 // extend the subvector to the size of the result vector. Make sure that
4175 // we are not recursing on that node by checking for undef here.
4176 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4177 Result.getOpcode() != ISD::UNDEF) {
4178 EVT ResultVT = Result.getValueType();
4179 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4180 SDValue Undef = DAG.getUNDEF(ResultVT);
4181 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4184 // The blend instruction, and therefore its mask, depend on the data type.
4185 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4186 if (ScalarType.isFloatingPoint()) {
4187 // Choose either vblendps (float) or vblendpd (double).
4188 unsigned ScalarSize = ScalarType.getSizeInBits();
4189 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4190 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4191 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4192 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4195 const X86Subtarget &Subtarget =
4196 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4198 // AVX2 is needed for 256-bit integer blend support.
4199 // Integers must be cast to 32-bit because there is only vpblendd;
4200 // vpblendw can't be used for this because it has a handicapped mask.
4202 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4203 // is still more efficient than using the wrong domain vinsertf128 that
4204 // will be created by InsertSubVector().
4205 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4207 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4208 Vec256 = DAG.getBitcast(CastVT, Vec256);
4209 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4210 return DAG.getBitcast(ResultVT, Vec256);
4213 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4216 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4217 SelectionDAG &DAG, SDLoc dl) {
4218 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4219 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4222 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4223 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4224 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4225 /// large BUILD_VECTORS.
4226 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4227 unsigned NumElems, SelectionDAG &DAG,
4229 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4230 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4233 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4234 unsigned NumElems, SelectionDAG &DAG,
4236 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4237 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4240 /// getOnesVector - Returns a vector of specified type with all bits set.
4241 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4242 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4243 /// Then bitcast to their original type, ensuring they get CSE'd.
4244 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4246 assert(VT.isVector() && "Expected a vector type");
4248 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4250 if (VT.is256BitVector()) {
4251 if (HasInt256) { // AVX2
4252 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4253 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4255 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4256 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4258 } else if (VT.is128BitVector()) {
4259 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4261 llvm_unreachable("Unexpected vector type");
4263 return DAG.getBitcast(VT, Vec);
4266 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4267 /// operation of specified width.
4268 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4270 unsigned NumElems = VT.getVectorNumElements();
4271 SmallVector<int, 8> Mask;
4272 Mask.push_back(NumElems);
4273 for (unsigned i = 1; i != NumElems; ++i)
4275 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4278 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4279 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4281 unsigned NumElems = VT.getVectorNumElements();
4282 SmallVector<int, 8> Mask;
4283 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4285 Mask.push_back(i + NumElems);
4287 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4290 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4291 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4293 unsigned NumElems = VT.getVectorNumElements();
4294 SmallVector<int, 8> Mask;
4295 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4296 Mask.push_back(i + Half);
4297 Mask.push_back(i + NumElems + Half);
4299 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4302 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4303 /// vector of zero or undef vector. This produces a shuffle where the low
4304 /// element of V2 is swizzled into the zero/undef vector, landing at element
4305 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4306 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4308 const X86Subtarget *Subtarget,
4309 SelectionDAG &DAG) {
4310 MVT VT = V2.getSimpleValueType();
4312 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4313 unsigned NumElems = VT.getVectorNumElements();
4314 SmallVector<int, 16> MaskVec;
4315 for (unsigned i = 0; i != NumElems; ++i)
4316 // If this is the insertion idx, put the low elt of V2 here.
4317 MaskVec.push_back(i == Idx ? NumElems : i);
4318 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4321 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4322 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4323 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4324 /// shuffles which use a single input multiple times, and in those cases it will
4325 /// adjust the mask to only have indices within that single input.
4326 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4327 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4328 unsigned NumElems = VT.getVectorNumElements();
4332 bool IsFakeUnary = false;
4333 switch(N->getOpcode()) {
4334 case X86ISD::BLENDI:
4335 ImmN = N->getOperand(N->getNumOperands()-1);
4336 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4339 ImmN = N->getOperand(N->getNumOperands()-1);
4340 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4341 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4343 case X86ISD::UNPCKH:
4344 DecodeUNPCKHMask(VT, Mask);
4345 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4347 case X86ISD::UNPCKL:
4348 DecodeUNPCKLMask(VT, Mask);
4349 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4351 case X86ISD::MOVHLPS:
4352 DecodeMOVHLPSMask(NumElems, Mask);
4353 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4355 case X86ISD::MOVLHPS:
4356 DecodeMOVLHPSMask(NumElems, Mask);
4357 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4359 case X86ISD::PALIGNR:
4360 ImmN = N->getOperand(N->getNumOperands()-1);
4361 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4363 case X86ISD::PSHUFD:
4364 case X86ISD::VPERMILPI:
4365 ImmN = N->getOperand(N->getNumOperands()-1);
4366 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4369 case X86ISD::PSHUFHW:
4370 ImmN = N->getOperand(N->getNumOperands()-1);
4371 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4374 case X86ISD::PSHUFLW:
4375 ImmN = N->getOperand(N->getNumOperands()-1);
4376 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4379 case X86ISD::PSHUFB: {
4381 SDValue MaskNode = N->getOperand(1);
4382 while (MaskNode->getOpcode() == ISD::BITCAST)
4383 MaskNode = MaskNode->getOperand(0);
4385 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4386 // If we have a build-vector, then things are easy.
4387 EVT VT = MaskNode.getValueType();
4388 assert(VT.isVector() &&
4389 "Can't produce a non-vector with a build_vector!");
4390 if (!VT.isInteger())
4393 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4395 SmallVector<uint64_t, 32> RawMask;
4396 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4397 SDValue Op = MaskNode->getOperand(i);
4398 if (Op->getOpcode() == ISD::UNDEF) {
4399 RawMask.push_back((uint64_t)SM_SentinelUndef);
4402 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4405 APInt MaskElement = CN->getAPIntValue();
4407 // We now have to decode the element which could be any integer size and
4408 // extract each byte of it.
4409 for (int j = 0; j < NumBytesPerElement; ++j) {
4410 // Note that this is x86 and so always little endian: the low byte is
4411 // the first byte of the mask.
4412 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4413 MaskElement = MaskElement.lshr(8);
4416 DecodePSHUFBMask(RawMask, Mask);
4420 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4424 SDValue Ptr = MaskLoad->getBasePtr();
4425 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4426 Ptr->getOpcode() == X86ISD::WrapperRIP)
4427 Ptr = Ptr->getOperand(0);
4429 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4430 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4433 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4434 DecodePSHUFBMask(C, Mask);
4442 case X86ISD::VPERMI:
4443 ImmN = N->getOperand(N->getNumOperands()-1);
4444 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4449 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4451 case X86ISD::VPERM2X128:
4452 ImmN = N->getOperand(N->getNumOperands()-1);
4453 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4454 if (Mask.empty()) return false;
4456 case X86ISD::MOVSLDUP:
4457 DecodeMOVSLDUPMask(VT, Mask);
4460 case X86ISD::MOVSHDUP:
4461 DecodeMOVSHDUPMask(VT, Mask);
4464 case X86ISD::MOVDDUP:
4465 DecodeMOVDDUPMask(VT, Mask);
4468 case X86ISD::MOVLHPD:
4469 case X86ISD::MOVLPD:
4470 case X86ISD::MOVLPS:
4471 // Not yet implemented
4473 default: llvm_unreachable("unknown target shuffle node");
4476 // If we have a fake unary shuffle, the shuffle mask is spread across two
4477 // inputs that are actually the same node. Re-map the mask to always point
4478 // into the first input.
4481 if (M >= (int)Mask.size())
4487 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4488 /// element of the result of the vector shuffle.
4489 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4492 return SDValue(); // Limit search depth.
4494 SDValue V = SDValue(N, 0);
4495 EVT VT = V.getValueType();
4496 unsigned Opcode = V.getOpcode();
4498 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4499 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4500 int Elt = SV->getMaskElt(Index);
4503 return DAG.getUNDEF(VT.getVectorElementType());
4505 unsigned NumElems = VT.getVectorNumElements();
4506 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4507 : SV->getOperand(1);
4508 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4511 // Recurse into target specific vector shuffles to find scalars.
4512 if (isTargetShuffle(Opcode)) {
4513 MVT ShufVT = V.getSimpleValueType();
4514 unsigned NumElems = ShufVT.getVectorNumElements();
4515 SmallVector<int, 16> ShuffleMask;
4518 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4521 int Elt = ShuffleMask[Index];
4523 return DAG.getUNDEF(ShufVT.getVectorElementType());
4525 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4527 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4531 // Actual nodes that may contain scalar elements
4532 if (Opcode == ISD::BITCAST) {
4533 V = V.getOperand(0);
4534 EVT SrcVT = V.getValueType();
4535 unsigned NumElems = VT.getVectorNumElements();
4537 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4541 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4542 return (Index == 0) ? V.getOperand(0)
4543 : DAG.getUNDEF(VT.getVectorElementType());
4545 if (V.getOpcode() == ISD::BUILD_VECTOR)
4546 return V.getOperand(Index);
4551 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4553 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4554 unsigned NumNonZero, unsigned NumZero,
4556 const X86Subtarget* Subtarget,
4557 const TargetLowering &TLI) {
4565 // SSE4.1 - use PINSRB to insert each byte directly.
4566 if (Subtarget->hasSSE41()) {
4567 for (unsigned i = 0; i < 16; ++i) {
4568 bool isNonZero = (NonZeros & (1 << i)) != 0;
4572 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4574 V = DAG.getUNDEF(MVT::v16i8);
4577 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4578 MVT::v16i8, V, Op.getOperand(i),
4579 DAG.getIntPtrConstant(i, dl));
4586 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4587 for (unsigned i = 0; i < 16; ++i) {
4588 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4589 if (ThisIsNonZero && First) {
4591 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4593 V = DAG.getUNDEF(MVT::v8i16);
4598 SDValue ThisElt, LastElt;
4599 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4600 if (LastIsNonZero) {
4601 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4602 MVT::i16, Op.getOperand(i-1));
4604 if (ThisIsNonZero) {
4605 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4606 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4607 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4609 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4613 if (ThisElt.getNode())
4614 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4615 DAG.getIntPtrConstant(i/2, dl));
4619 return DAG.getBitcast(MVT::v16i8, V);
4622 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4624 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4625 unsigned NumNonZero, unsigned NumZero,
4627 const X86Subtarget* Subtarget,
4628 const TargetLowering &TLI) {
4635 for (unsigned i = 0; i < 8; ++i) {
4636 bool isNonZero = (NonZeros & (1 << i)) != 0;
4640 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4642 V = DAG.getUNDEF(MVT::v8i16);
4645 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4646 MVT::v8i16, V, Op.getOperand(i),
4647 DAG.getIntPtrConstant(i, dl));
4654 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4655 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4656 const X86Subtarget *Subtarget,
4657 const TargetLowering &TLI) {
4658 // Find all zeroable elements.
4659 std::bitset<4> Zeroable;
4660 for (int i=0; i < 4; ++i) {
4661 SDValue Elt = Op->getOperand(i);
4662 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4664 assert(Zeroable.size() - Zeroable.count() > 1 &&
4665 "We expect at least two non-zero elements!");
4667 // We only know how to deal with build_vector nodes where elements are either
4668 // zeroable or extract_vector_elt with constant index.
4669 SDValue FirstNonZero;
4670 unsigned FirstNonZeroIdx;
4671 for (unsigned i=0; i < 4; ++i) {
4674 SDValue Elt = Op->getOperand(i);
4675 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4676 !isa<ConstantSDNode>(Elt.getOperand(1)))
4678 // Make sure that this node is extracting from a 128-bit vector.
4679 MVT VT = Elt.getOperand(0).getSimpleValueType();
4680 if (!VT.is128BitVector())
4682 if (!FirstNonZero.getNode()) {
4684 FirstNonZeroIdx = i;
4688 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4689 SDValue V1 = FirstNonZero.getOperand(0);
4690 MVT VT = V1.getSimpleValueType();
4692 // See if this build_vector can be lowered as a blend with zero.
4694 unsigned EltMaskIdx, EltIdx;
4696 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4697 if (Zeroable[EltIdx]) {
4698 // The zero vector will be on the right hand side.
4699 Mask[EltIdx] = EltIdx+4;
4703 Elt = Op->getOperand(EltIdx);
4704 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4705 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4706 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4708 Mask[EltIdx] = EltIdx;
4712 // Let the shuffle legalizer deal with blend operations.
4713 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4714 if (V1.getSimpleValueType() != VT)
4715 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4716 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4719 // See if we can lower this build_vector to a INSERTPS.
4720 if (!Subtarget->hasSSE41())
4723 SDValue V2 = Elt.getOperand(0);
4724 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4727 bool CanFold = true;
4728 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4732 SDValue Current = Op->getOperand(i);
4733 SDValue SrcVector = Current->getOperand(0);
4736 CanFold = SrcVector == V1 &&
4737 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4743 assert(V1.getNode() && "Expected at least two non-zero elements!");
4744 if (V1.getSimpleValueType() != MVT::v4f32)
4745 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4746 if (V2.getSimpleValueType() != MVT::v4f32)
4747 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4749 // Ok, we can emit an INSERTPS instruction.
4750 unsigned ZMask = Zeroable.to_ulong();
4752 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4753 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4755 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4756 DAG.getIntPtrConstant(InsertPSMask, DL));
4757 return DAG.getBitcast(VT, Result);
4760 /// Return a vector logical shift node.
4761 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4762 unsigned NumBits, SelectionDAG &DAG,
4763 const TargetLowering &TLI, SDLoc dl) {
4764 assert(VT.is128BitVector() && "Unknown type for VShift");
4765 MVT ShVT = MVT::v2i64;
4766 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4767 SrcOp = DAG.getBitcast(ShVT, SrcOp);
4768 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4769 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4770 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4771 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4775 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4777 // Check if the scalar load can be widened into a vector load. And if
4778 // the address is "base + cst" see if the cst can be "absorbed" into
4779 // the shuffle mask.
4780 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4781 SDValue Ptr = LD->getBasePtr();
4782 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4784 EVT PVT = LD->getValueType(0);
4785 if (PVT != MVT::i32 && PVT != MVT::f32)
4790 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4791 FI = FINode->getIndex();
4793 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4794 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4795 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4796 Offset = Ptr.getConstantOperandVal(1);
4797 Ptr = Ptr.getOperand(0);
4802 // FIXME: 256-bit vector instructions don't require a strict alignment,
4803 // improve this code to support it better.
4804 unsigned RequiredAlign = VT.getSizeInBits()/8;
4805 SDValue Chain = LD->getChain();
4806 // Make sure the stack object alignment is at least 16 or 32.
4807 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4808 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4809 if (MFI->isFixedObjectIndex(FI)) {
4810 // Can't change the alignment. FIXME: It's possible to compute
4811 // the exact stack offset and reference FI + adjust offset instead.
4812 // If someone *really* cares about this. That's the way to implement it.
4815 MFI->setObjectAlignment(FI, RequiredAlign);
4819 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4820 // Ptr + (Offset & ~15).
4823 if ((Offset % RequiredAlign) & 3)
4825 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4828 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4829 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4832 int EltNo = (Offset - StartOffset) >> 2;
4833 unsigned NumElems = VT.getVectorNumElements();
4835 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4836 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4837 LD->getPointerInfo().getWithOffset(StartOffset),
4838 false, false, false, 0);
4840 SmallVector<int, 8> Mask(NumElems, EltNo);
4842 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4848 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4849 /// elements can be replaced by a single large load which has the same value as
4850 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4852 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4854 /// FIXME: we'd also like to handle the case where the last elements are zero
4855 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4856 /// There's even a handy isZeroNode for that purpose.
4857 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4858 SDLoc &DL, SelectionDAG &DAG,
4859 bool isAfterLegalize) {
4860 unsigned NumElems = Elts.size();
4862 LoadSDNode *LDBase = nullptr;
4863 unsigned LastLoadedElt = -1U;
4865 // For each element in the initializer, see if we've found a load or an undef.
4866 // If we don't find an initial load element, or later load elements are
4867 // non-consecutive, bail out.
4868 for (unsigned i = 0; i < NumElems; ++i) {
4869 SDValue Elt = Elts[i];
4870 // Look through a bitcast.
4871 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4872 Elt = Elt.getOperand(0);
4873 if (!Elt.getNode() ||
4874 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4877 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4879 LDBase = cast<LoadSDNode>(Elt.getNode());
4883 if (Elt.getOpcode() == ISD::UNDEF)
4886 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4887 EVT LdVT = Elt.getValueType();
4888 // Each loaded element must be the correct fractional portion of the
4889 // requested vector load.
4890 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4892 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4897 // If we have found an entire vector of loads and undefs, then return a large
4898 // load of the entire vector width starting at the base pointer. If we found
4899 // consecutive loads for the low half, generate a vzext_load node.
4900 if (LastLoadedElt == NumElems - 1) {
4901 assert(LDBase && "Did not find base load for merging consecutive loads");
4902 EVT EltVT = LDBase->getValueType(0);
4903 // Ensure that the input vector size for the merged loads matches the
4904 // cumulative size of the input elements.
4905 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4908 if (isAfterLegalize &&
4909 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4912 SDValue NewLd = SDValue();
4914 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4915 LDBase->getPointerInfo(), LDBase->isVolatile(),
4916 LDBase->isNonTemporal(), LDBase->isInvariant(),
4917 LDBase->getAlignment());
4919 if (LDBase->hasAnyUseOfValue(1)) {
4920 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4922 SDValue(NewLd.getNode(), 1));
4923 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4924 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4925 SDValue(NewLd.getNode(), 1));
4931 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4932 //of a v4i32 / v4f32. It's probably worth generalizing.
4933 EVT EltVT = VT.getVectorElementType();
4934 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4935 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4936 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4937 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4939 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4940 LDBase->getPointerInfo(),
4941 LDBase->getAlignment(),
4942 false/*isVolatile*/, true/*ReadMem*/,
4945 // Make sure the newly-created LOAD is in the same position as LDBase in
4946 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4947 // update uses of LDBase's output chain to use the TokenFactor.
4948 if (LDBase->hasAnyUseOfValue(1)) {
4949 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4950 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4951 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4952 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4953 SDValue(ResNode.getNode(), 1));
4956 return DAG.getBitcast(VT, ResNode);
4961 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4962 /// to generate a splat value for the following cases:
4963 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4964 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4965 /// a scalar load, or a constant.
4966 /// The VBROADCAST node is returned when a pattern is found,
4967 /// or SDValue() otherwise.
4968 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4969 SelectionDAG &DAG) {
4970 // VBROADCAST requires AVX.
4971 // TODO: Splats could be generated for non-AVX CPUs using SSE
4972 // instructions, but there's less potential gain for only 128-bit vectors.
4973 if (!Subtarget->hasAVX())
4976 MVT VT = Op.getSimpleValueType();
4979 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4980 "Unsupported vector type for broadcast.");
4985 switch (Op.getOpcode()) {
4987 // Unknown pattern found.
4990 case ISD::BUILD_VECTOR: {
4991 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4992 BitVector UndefElements;
4993 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4995 // We need a splat of a single value to use broadcast, and it doesn't
4996 // make any sense if the value is only in one element of the vector.
4997 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5001 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5002 Ld.getOpcode() == ISD::ConstantFP);
5004 // Make sure that all of the users of a non-constant load are from the
5005 // BUILD_VECTOR node.
5006 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5011 case ISD::VECTOR_SHUFFLE: {
5012 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5014 // Shuffles must have a splat mask where the first element is
5016 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5019 SDValue Sc = Op.getOperand(0);
5020 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5021 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5023 if (!Subtarget->hasInt256())
5026 // Use the register form of the broadcast instruction available on AVX2.
5027 if (VT.getSizeInBits() >= 256)
5028 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5029 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5032 Ld = Sc.getOperand(0);
5033 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5034 Ld.getOpcode() == ISD::ConstantFP);
5036 // The scalar_to_vector node and the suspected
5037 // load node must have exactly one user.
5038 // Constants may have multiple users.
5040 // AVX-512 has register version of the broadcast
5041 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5042 Ld.getValueType().getSizeInBits() >= 32;
5043 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5050 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5051 bool IsGE256 = (VT.getSizeInBits() >= 256);
5053 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5054 // instruction to save 8 or more bytes of constant pool data.
5055 // TODO: If multiple splats are generated to load the same constant,
5056 // it may be detrimental to overall size. There needs to be a way to detect
5057 // that condition to know if this is truly a size win.
5058 const Function *F = DAG.getMachineFunction().getFunction();
5059 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
5061 // Handle broadcasting a single constant scalar from the constant pool
5063 // On Sandybridge (no AVX2), it is still better to load a constant vector
5064 // from the constant pool and not to broadcast it from a scalar.
5065 // But override that restriction when optimizing for size.
5066 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5067 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5068 EVT CVT = Ld.getValueType();
5069 assert(!CVT.isVector() && "Must not broadcast a vector type");
5071 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5072 // For size optimization, also splat v2f64 and v2i64, and for size opt
5073 // with AVX2, also splat i8 and i16.
5074 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5075 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5076 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5077 const Constant *C = nullptr;
5078 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5079 C = CI->getConstantIntValue();
5080 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5081 C = CF->getConstantFPValue();
5083 assert(C && "Invalid constant type");
5085 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5086 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5087 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5088 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5089 MachinePointerInfo::getConstantPool(),
5090 false, false, false, Alignment);
5092 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5096 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5098 // Handle AVX2 in-register broadcasts.
5099 if (!IsLoad && Subtarget->hasInt256() &&
5100 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5101 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5103 // The scalar source must be a normal load.
5107 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5108 (Subtarget->hasVLX() && ScalarSize == 64))
5109 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5111 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5112 // double since there is no vbroadcastsd xmm
5113 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5114 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5115 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5118 // Unsupported broadcast.
5122 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5123 /// underlying vector and index.
5125 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5127 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5129 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5130 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5133 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5135 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5137 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5138 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5141 // In this case the vector is the extract_subvector expression and the index
5142 // is 2, as specified by the shuffle.
5143 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5144 SDValue ShuffleVec = SVOp->getOperand(0);
5145 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5146 assert(ShuffleVecVT.getVectorElementType() ==
5147 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5149 int ShuffleIdx = SVOp->getMaskElt(Idx);
5150 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5151 ExtractedFromVec = ShuffleVec;
5157 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5158 MVT VT = Op.getSimpleValueType();
5160 // Skip if insert_vec_elt is not supported.
5161 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5162 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5166 unsigned NumElems = Op.getNumOperands();
5170 SmallVector<unsigned, 4> InsertIndices;
5171 SmallVector<int, 8> Mask(NumElems, -1);
5173 for (unsigned i = 0; i != NumElems; ++i) {
5174 unsigned Opc = Op.getOperand(i).getOpcode();
5176 if (Opc == ISD::UNDEF)
5179 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5180 // Quit if more than 1 elements need inserting.
5181 if (InsertIndices.size() > 1)
5184 InsertIndices.push_back(i);
5188 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5189 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5190 // Quit if non-constant index.
5191 if (!isa<ConstantSDNode>(ExtIdx))
5193 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5195 // Quit if extracted from vector of different type.
5196 if (ExtractedFromVec.getValueType() != VT)
5199 if (!VecIn1.getNode())
5200 VecIn1 = ExtractedFromVec;
5201 else if (VecIn1 != ExtractedFromVec) {
5202 if (!VecIn2.getNode())
5203 VecIn2 = ExtractedFromVec;
5204 else if (VecIn2 != ExtractedFromVec)
5205 // Quit if more than 2 vectors to shuffle
5209 if (ExtractedFromVec == VecIn1)
5211 else if (ExtractedFromVec == VecIn2)
5212 Mask[i] = Idx + NumElems;
5215 if (!VecIn1.getNode())
5218 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5219 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5220 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5221 unsigned Idx = InsertIndices[i];
5222 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5223 DAG.getIntPtrConstant(Idx, DL));
5229 static SDValue ConvertI1VectorToInterger(SDValue Op, SelectionDAG &DAG) {
5230 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5231 Op.getScalarValueSizeInBits() == 1 &&
5232 "Can not convert non-constant vector");
5233 uint64_t Immediate = 0;
5234 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5235 SDValue In = Op.getOperand(idx);
5236 if (In.getOpcode() != ISD::UNDEF)
5237 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5241 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5242 return DAG.getConstant(Immediate, dl, VT);
5244 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5246 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5248 MVT VT = Op.getSimpleValueType();
5249 assert((VT.getVectorElementType() == MVT::i1) &&
5250 "Unexpected type in LowerBUILD_VECTORvXi1!");
5253 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5254 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5255 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5256 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5259 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5260 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5261 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5262 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5265 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5266 SDValue Imm = ConvertI1VectorToInterger(Op, DAG);
5267 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5268 return DAG.getBitcast(VT, Imm);
5269 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5270 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5271 DAG.getIntPtrConstant(0, dl));
5274 // Vector has one or more non-const elements
5275 uint64_t Immediate = 0;
5276 SmallVector<unsigned, 16> NonConstIdx;
5277 bool IsSplat = true;
5278 bool HasConstElts = false;
5280 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5281 SDValue In = Op.getOperand(idx);
5282 if (In.getOpcode() == ISD::UNDEF)
5284 if (!isa<ConstantSDNode>(In))
5285 NonConstIdx.push_back(idx);
5287 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5288 HasConstElts = true;
5292 else if (In != Op.getOperand(SplatIdx))
5296 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5298 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5299 DAG.getConstant(1, dl, VT),
5300 DAG.getConstant(0, dl, VT));
5302 // insert elements one by one
5306 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5307 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5309 else if (HasConstElts)
5310 Imm = DAG.getConstant(0, dl, VT);
5312 Imm = DAG.getUNDEF(VT);
5313 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5314 DstVec = DAG.getBitcast(VT, Imm);
5316 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5317 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5318 DAG.getIntPtrConstant(0, dl));
5321 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5322 unsigned InsertIdx = NonConstIdx[i];
5323 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5324 Op.getOperand(InsertIdx),
5325 DAG.getIntPtrConstant(InsertIdx, dl));
5330 /// \brief Return true if \p N implements a horizontal binop and return the
5331 /// operands for the horizontal binop into V0 and V1.
5333 /// This is a helper function of LowerToHorizontalOp().
5334 /// This function checks that the build_vector \p N in input implements a
5335 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5336 /// operation to match.
5337 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5338 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5339 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5342 /// This function only analyzes elements of \p N whose indices are
5343 /// in range [BaseIdx, LastIdx).
5344 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5346 unsigned BaseIdx, unsigned LastIdx,
5347 SDValue &V0, SDValue &V1) {
5348 EVT VT = N->getValueType(0);
5350 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5351 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5352 "Invalid Vector in input!");
5354 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5355 bool CanFold = true;
5356 unsigned ExpectedVExtractIdx = BaseIdx;
5357 unsigned NumElts = LastIdx - BaseIdx;
5358 V0 = DAG.getUNDEF(VT);
5359 V1 = DAG.getUNDEF(VT);
5361 // Check if N implements a horizontal binop.
5362 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5363 SDValue Op = N->getOperand(i + BaseIdx);
5366 if (Op->getOpcode() == ISD::UNDEF) {
5367 // Update the expected vector extract index.
5368 if (i * 2 == NumElts)
5369 ExpectedVExtractIdx = BaseIdx;
5370 ExpectedVExtractIdx += 2;
5374 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5379 SDValue Op0 = Op.getOperand(0);
5380 SDValue Op1 = Op.getOperand(1);
5382 // Try to match the following pattern:
5383 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5384 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5385 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5386 Op0.getOperand(0) == Op1.getOperand(0) &&
5387 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5388 isa<ConstantSDNode>(Op1.getOperand(1)));
5392 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5393 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5395 if (i * 2 < NumElts) {
5396 if (V0.getOpcode() == ISD::UNDEF) {
5397 V0 = Op0.getOperand(0);
5398 if (V0.getValueType() != VT)
5402 if (V1.getOpcode() == ISD::UNDEF) {
5403 V1 = Op0.getOperand(0);
5404 if (V1.getValueType() != VT)
5407 if (i * 2 == NumElts)
5408 ExpectedVExtractIdx = BaseIdx;
5411 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5412 if (I0 == ExpectedVExtractIdx)
5413 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5414 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5415 // Try to match the following dag sequence:
5416 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5417 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5421 ExpectedVExtractIdx += 2;
5427 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5428 /// a concat_vector.
5430 /// This is a helper function of LowerToHorizontalOp().
5431 /// This function expects two 256-bit vectors called V0 and V1.
5432 /// At first, each vector is split into two separate 128-bit vectors.
5433 /// Then, the resulting 128-bit vectors are used to implement two
5434 /// horizontal binary operations.
5436 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5438 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5439 /// the two new horizontal binop.
5440 /// When Mode is set, the first horizontal binop dag node would take as input
5441 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5442 /// horizontal binop dag node would take as input the lower 128-bit of V1
5443 /// and the upper 128-bit of V1.
5445 /// HADD V0_LO, V0_HI
5446 /// HADD V1_LO, V1_HI
5448 /// Otherwise, the first horizontal binop dag node takes as input the lower
5449 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5450 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5452 /// HADD V0_LO, V1_LO
5453 /// HADD V0_HI, V1_HI
5455 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5456 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5457 /// the upper 128-bits of the result.
5458 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5459 SDLoc DL, SelectionDAG &DAG,
5460 unsigned X86Opcode, bool Mode,
5461 bool isUndefLO, bool isUndefHI) {
5462 EVT VT = V0.getValueType();
5463 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5464 "Invalid nodes in input!");
5466 unsigned NumElts = VT.getVectorNumElements();
5467 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5468 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5469 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5470 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5471 EVT NewVT = V0_LO.getValueType();
5473 SDValue LO = DAG.getUNDEF(NewVT);
5474 SDValue HI = DAG.getUNDEF(NewVT);
5477 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5478 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5479 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5480 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5481 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5483 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5484 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5485 V1_LO->getOpcode() != ISD::UNDEF))
5486 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5488 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5489 V1_HI->getOpcode() != ISD::UNDEF))
5490 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5493 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5496 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5498 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5499 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5500 EVT VT = BV->getValueType(0);
5501 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5502 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5506 unsigned NumElts = VT.getVectorNumElements();
5507 SDValue InVec0 = DAG.getUNDEF(VT);
5508 SDValue InVec1 = DAG.getUNDEF(VT);
5510 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5511 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5513 // Odd-numbered elements in the input build vector are obtained from
5514 // adding two integer/float elements.
5515 // Even-numbered elements in the input build vector are obtained from
5516 // subtracting two integer/float elements.
5517 unsigned ExpectedOpcode = ISD::FSUB;
5518 unsigned NextExpectedOpcode = ISD::FADD;
5519 bool AddFound = false;
5520 bool SubFound = false;
5522 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5523 SDValue Op = BV->getOperand(i);
5525 // Skip 'undef' values.
5526 unsigned Opcode = Op.getOpcode();
5527 if (Opcode == ISD::UNDEF) {
5528 std::swap(ExpectedOpcode, NextExpectedOpcode);
5532 // Early exit if we found an unexpected opcode.
5533 if (Opcode != ExpectedOpcode)
5536 SDValue Op0 = Op.getOperand(0);
5537 SDValue Op1 = Op.getOperand(1);
5539 // Try to match the following pattern:
5540 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5541 // Early exit if we cannot match that sequence.
5542 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5543 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5544 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5545 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5546 Op0.getOperand(1) != Op1.getOperand(1))
5549 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5553 // We found a valid add/sub node. Update the information accordingly.
5559 // Update InVec0 and InVec1.
5560 if (InVec0.getOpcode() == ISD::UNDEF) {
5561 InVec0 = Op0.getOperand(0);
5562 if (InVec0.getValueType() != VT)
5565 if (InVec1.getOpcode() == ISD::UNDEF) {
5566 InVec1 = Op1.getOperand(0);
5567 if (InVec1.getValueType() != VT)
5571 // Make sure that operands in input to each add/sub node always
5572 // come from a same pair of vectors.
5573 if (InVec0 != Op0.getOperand(0)) {
5574 if (ExpectedOpcode == ISD::FSUB)
5577 // FADD is commutable. Try to commute the operands
5578 // and then test again.
5579 std::swap(Op0, Op1);
5580 if (InVec0 != Op0.getOperand(0))
5584 if (InVec1 != Op1.getOperand(0))
5587 // Update the pair of expected opcodes.
5588 std::swap(ExpectedOpcode, NextExpectedOpcode);
5591 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5592 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5593 InVec1.getOpcode() != ISD::UNDEF)
5594 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5599 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5600 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5601 const X86Subtarget *Subtarget,
5602 SelectionDAG &DAG) {
5603 EVT VT = BV->getValueType(0);
5604 unsigned NumElts = VT.getVectorNumElements();
5605 unsigned NumUndefsLO = 0;
5606 unsigned NumUndefsHI = 0;
5607 unsigned Half = NumElts/2;
5609 // Count the number of UNDEF operands in the build_vector in input.
5610 for (unsigned i = 0, e = Half; i != e; ++i)
5611 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5614 for (unsigned i = Half, e = NumElts; i != e; ++i)
5615 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5618 // Early exit if this is either a build_vector of all UNDEFs or all the
5619 // operands but one are UNDEF.
5620 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5624 SDValue InVec0, InVec1;
5625 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5626 // Try to match an SSE3 float HADD/HSUB.
5627 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5628 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5630 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5631 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5632 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5633 // Try to match an SSSE3 integer HADD/HSUB.
5634 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5635 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5637 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5638 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5641 if (!Subtarget->hasAVX())
5644 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5645 // Try to match an AVX horizontal add/sub of packed single/double
5646 // precision floating point values from 256-bit vectors.
5647 SDValue InVec2, InVec3;
5648 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5649 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5650 ((InVec0.getOpcode() == ISD::UNDEF ||
5651 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5652 ((InVec1.getOpcode() == ISD::UNDEF ||
5653 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5654 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5656 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5657 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5658 ((InVec0.getOpcode() == ISD::UNDEF ||
5659 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5660 ((InVec1.getOpcode() == ISD::UNDEF ||
5661 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5662 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5663 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5664 // Try to match an AVX2 horizontal add/sub of signed integers.
5665 SDValue InVec2, InVec3;
5667 bool CanFold = true;
5669 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5670 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5671 ((InVec0.getOpcode() == ISD::UNDEF ||
5672 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5673 ((InVec1.getOpcode() == ISD::UNDEF ||
5674 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5675 X86Opcode = X86ISD::HADD;
5676 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5677 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5678 ((InVec0.getOpcode() == ISD::UNDEF ||
5679 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5680 ((InVec1.getOpcode() == ISD::UNDEF ||
5681 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5682 X86Opcode = X86ISD::HSUB;
5687 // Fold this build_vector into a single horizontal add/sub.
5688 // Do this only if the target has AVX2.
5689 if (Subtarget->hasAVX2())
5690 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5692 // Do not try to expand this build_vector into a pair of horizontal
5693 // add/sub if we can emit a pair of scalar add/sub.
5694 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5697 // Convert this build_vector into a pair of horizontal binop followed by
5699 bool isUndefLO = NumUndefsLO == Half;
5700 bool isUndefHI = NumUndefsHI == Half;
5701 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5702 isUndefLO, isUndefHI);
5706 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5707 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5709 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5710 X86Opcode = X86ISD::HADD;
5711 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5712 X86Opcode = X86ISD::HSUB;
5713 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5714 X86Opcode = X86ISD::FHADD;
5715 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5716 X86Opcode = X86ISD::FHSUB;
5720 // Don't try to expand this build_vector into a pair of horizontal add/sub
5721 // if we can simply emit a pair of scalar add/sub.
5722 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5725 // Convert this build_vector into two horizontal add/sub followed by
5727 bool isUndefLO = NumUndefsLO == Half;
5728 bool isUndefHI = NumUndefsHI == Half;
5729 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5730 isUndefLO, isUndefHI);
5737 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5740 MVT VT = Op.getSimpleValueType();
5741 MVT ExtVT = VT.getVectorElementType();
5742 unsigned NumElems = Op.getNumOperands();
5744 // Generate vectors for predicate vectors.
5745 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5746 return LowerBUILD_VECTORvXi1(Op, DAG);
5748 // Vectors containing all zeros can be matched by pxor and xorps later
5749 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5750 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5751 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5752 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5755 return getZeroVector(VT, Subtarget, DAG, dl);
5758 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5759 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5760 // vpcmpeqd on 256-bit vectors.
5761 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5762 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5765 if (!VT.is512BitVector())
5766 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5769 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5770 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5772 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5773 return HorizontalOp;
5774 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5777 unsigned EVTBits = ExtVT.getSizeInBits();
5779 unsigned NumZero = 0;
5780 unsigned NumNonZero = 0;
5781 unsigned NonZeros = 0;
5782 bool IsAllConstants = true;
5783 SmallSet<SDValue, 8> Values;
5784 for (unsigned i = 0; i < NumElems; ++i) {
5785 SDValue Elt = Op.getOperand(i);
5786 if (Elt.getOpcode() == ISD::UNDEF)
5789 if (Elt.getOpcode() != ISD::Constant &&
5790 Elt.getOpcode() != ISD::ConstantFP)
5791 IsAllConstants = false;
5792 if (X86::isZeroNode(Elt))
5795 NonZeros |= (1 << i);
5800 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5801 if (NumNonZero == 0)
5802 return DAG.getUNDEF(VT);
5804 // Special case for single non-zero, non-undef, element.
5805 if (NumNonZero == 1) {
5806 unsigned Idx = countTrailingZeros(NonZeros);
5807 SDValue Item = Op.getOperand(Idx);
5809 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5810 // the value are obviously zero, truncate the value to i32 and do the
5811 // insertion that way. Only do this if the value is non-constant or if the
5812 // value is a constant being inserted into element 0. It is cheaper to do
5813 // a constant pool load than it is to do a movd + shuffle.
5814 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5815 (!IsAllConstants || Idx == 0)) {
5816 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5818 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5819 EVT VecVT = MVT::v4i32;
5821 // Truncate the value (which may itself be a constant) to i32, and
5822 // convert it to a vector with movd (S2V+shuffle to zero extend).
5823 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5824 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5825 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
5826 Item, Idx * 2, true, Subtarget, DAG));
5830 // If we have a constant or non-constant insertion into the low element of
5831 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5832 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5833 // depending on what the source datatype is.
5836 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5838 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5839 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5840 if (VT.is512BitVector()) {
5841 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5842 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5843 Item, DAG.getIntPtrConstant(0, dl));
5845 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5846 "Expected an SSE value type!");
5847 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5848 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5849 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5852 // We can't directly insert an i8 or i16 into a vector, so zero extend
5854 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5855 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5856 if (VT.is256BitVector()) {
5857 if (Subtarget->hasAVX()) {
5858 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5859 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5861 // Without AVX, we need to extend to a 128-bit vector and then
5862 // insert into the 256-bit vector.
5863 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5864 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5865 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5868 assert(VT.is128BitVector() && "Expected an SSE value type!");
5869 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5870 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5872 return DAG.getBitcast(VT, Item);
5876 // Is it a vector logical left shift?
5877 if (NumElems == 2 && Idx == 1 &&
5878 X86::isZeroNode(Op.getOperand(0)) &&
5879 !X86::isZeroNode(Op.getOperand(1))) {
5880 unsigned NumBits = VT.getSizeInBits();
5881 return getVShift(true, VT,
5882 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5883 VT, Op.getOperand(1)),
5884 NumBits/2, DAG, *this, dl);
5887 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5890 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5891 // is a non-constant being inserted into an element other than the low one,
5892 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5893 // movd/movss) to move this into the low element, then shuffle it into
5895 if (EVTBits == 32) {
5896 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5897 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5901 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5902 if (Values.size() == 1) {
5903 if (EVTBits == 32) {
5904 // Instead of a shuffle like this:
5905 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5906 // Check if it's possible to issue this instead.
5907 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5908 unsigned Idx = countTrailingZeros(NonZeros);
5909 SDValue Item = Op.getOperand(Idx);
5910 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5911 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5916 // A vector full of immediates; various special cases are already
5917 // handled, so this is best done with a single constant-pool load.
5921 // For AVX-length vectors, see if we can use a vector load to get all of the
5922 // elements, otherwise build the individual 128-bit pieces and use
5923 // shuffles to put them in place.
5924 if (VT.is256BitVector() || VT.is512BitVector()) {
5925 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5927 // Check for a build vector of consecutive loads.
5928 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5931 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5933 // Build both the lower and upper subvector.
5934 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5935 makeArrayRef(&V[0], NumElems/2));
5936 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5937 makeArrayRef(&V[NumElems / 2], NumElems/2));
5939 // Recreate the wider vector with the lower and upper part.
5940 if (VT.is256BitVector())
5941 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5942 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5945 // Let legalizer expand 2-wide build_vectors.
5946 if (EVTBits == 64) {
5947 if (NumNonZero == 1) {
5948 // One half is zero or undef.
5949 unsigned Idx = countTrailingZeros(NonZeros);
5950 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5951 Op.getOperand(Idx));
5952 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5957 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5958 if (EVTBits == 8 && NumElems == 16)
5959 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5963 if (EVTBits == 16 && NumElems == 8)
5964 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5968 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5969 if (EVTBits == 32 && NumElems == 4)
5970 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
5973 // If element VT is == 32 bits, turn it into a number of shuffles.
5974 SmallVector<SDValue, 8> V(NumElems);
5975 if (NumElems == 4 && NumZero > 0) {
5976 for (unsigned i = 0; i < 4; ++i) {
5977 bool isZero = !(NonZeros & (1 << i));
5979 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5981 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5984 for (unsigned i = 0; i < 2; ++i) {
5985 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5988 V[i] = V[i*2]; // Must be a zero vector.
5991 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5994 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5997 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6002 bool Reverse1 = (NonZeros & 0x3) == 2;
6003 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6007 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6008 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6010 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6013 if (Values.size() > 1 && VT.is128BitVector()) {
6014 // Check for a build vector of consecutive loads.
6015 for (unsigned i = 0; i < NumElems; ++i)
6016 V[i] = Op.getOperand(i);
6018 // Check for elements which are consecutive loads.
6019 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6022 // Check for a build vector from mostly shuffle plus few inserting.
6023 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6026 // For SSE 4.1, use insertps to put the high elements into the low element.
6027 if (Subtarget->hasSSE41()) {
6029 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6030 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6032 Result = DAG.getUNDEF(VT);
6034 for (unsigned i = 1; i < NumElems; ++i) {
6035 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6036 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6037 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6042 // Otherwise, expand into a number of unpckl*, start by extending each of
6043 // our (non-undef) elements to the full vector width with the element in the
6044 // bottom slot of the vector (which generates no code for SSE).
6045 for (unsigned i = 0; i < NumElems; ++i) {
6046 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6047 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6049 V[i] = DAG.getUNDEF(VT);
6052 // Next, we iteratively mix elements, e.g. for v4f32:
6053 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6054 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6055 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6056 unsigned EltStride = NumElems >> 1;
6057 while (EltStride != 0) {
6058 for (unsigned i = 0; i < EltStride; ++i) {
6059 // If V[i+EltStride] is undef and this is the first round of mixing,
6060 // then it is safe to just drop this shuffle: V[i] is already in the
6061 // right place, the one element (since it's the first round) being
6062 // inserted as undef can be dropped. This isn't safe for successive
6063 // rounds because they will permute elements within both vectors.
6064 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6065 EltStride == NumElems/2)
6068 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6077 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6078 // to create 256-bit vectors from two other 128-bit ones.
6079 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6081 MVT ResVT = Op.getSimpleValueType();
6083 assert((ResVT.is256BitVector() ||
6084 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6086 SDValue V1 = Op.getOperand(0);
6087 SDValue V2 = Op.getOperand(1);
6088 unsigned NumElems = ResVT.getVectorNumElements();
6089 if (ResVT.is256BitVector())
6090 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6092 if (Op.getNumOperands() == 4) {
6093 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6094 ResVT.getVectorNumElements()/2);
6095 SDValue V3 = Op.getOperand(2);
6096 SDValue V4 = Op.getOperand(3);
6097 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6098 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6100 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6103 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6104 const X86Subtarget *Subtarget,
6105 SelectionDAG & DAG) {
6107 MVT ResVT = Op.getSimpleValueType();
6108 unsigned NumOfOperands = Op.getNumOperands();
6110 assert(isPowerOf2_32(NumOfOperands) &&
6111 "Unexpected number of operands in CONCAT_VECTORS");
6113 if (NumOfOperands > 2) {
6114 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6115 ResVT.getVectorNumElements()/2);
6116 SmallVector<SDValue, 2> Ops;
6117 for (unsigned i = 0; i < NumOfOperands/2; i++)
6118 Ops.push_back(Op.getOperand(i));
6119 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6121 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6122 Ops.push_back(Op.getOperand(i));
6123 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6124 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6127 SDValue V1 = Op.getOperand(0);
6128 SDValue V2 = Op.getOperand(1);
6129 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6130 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6132 if (IsZeroV1 && IsZeroV2)
6133 return getZeroVector(ResVT, Subtarget, DAG, dl);
6135 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6136 SDValue Undef = DAG.getUNDEF(ResVT);
6137 unsigned NumElems = ResVT.getVectorNumElements();
6138 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6140 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6141 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6145 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6146 // Zero the upper bits of V1
6147 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6148 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6151 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6154 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6155 const X86Subtarget *Subtarget,
6156 SelectionDAG &DAG) {
6157 MVT VT = Op.getSimpleValueType();
6158 if (VT.getVectorElementType() == MVT::i1)
6159 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6161 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6162 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6163 Op.getNumOperands() == 4)));
6165 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6166 // from two other 128-bit ones.
6168 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6169 return LowerAVXCONCAT_VECTORS(Op, DAG);
6173 //===----------------------------------------------------------------------===//
6174 // Vector shuffle lowering
6176 // This is an experimental code path for lowering vector shuffles on x86. It is
6177 // designed to handle arbitrary vector shuffles and blends, gracefully
6178 // degrading performance as necessary. It works hard to recognize idiomatic
6179 // shuffles and lower them to optimal instruction patterns without leaving
6180 // a framework that allows reasonably efficient handling of all vector shuffle
6182 //===----------------------------------------------------------------------===//
6184 /// \brief Tiny helper function to identify a no-op mask.
6186 /// This is a somewhat boring predicate function. It checks whether the mask
6187 /// array input, which is assumed to be a single-input shuffle mask of the kind
6188 /// used by the X86 shuffle instructions (not a fully general
6189 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6190 /// in-place shuffle are 'no-op's.
6191 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6192 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6193 if (Mask[i] != -1 && Mask[i] != i)
6198 /// \brief Helper function to classify a mask as a single-input mask.
6200 /// This isn't a generic single-input test because in the vector shuffle
6201 /// lowering we canonicalize single inputs to be the first input operand. This
6202 /// means we can more quickly test for a single input by only checking whether
6203 /// an input from the second operand exists. We also assume that the size of
6204 /// mask corresponds to the size of the input vectors which isn't true in the
6205 /// fully general case.
6206 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6208 if (M >= (int)Mask.size())
6213 /// \brief Test whether there are elements crossing 128-bit lanes in this
6216 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6217 /// and we routinely test for these.
6218 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6219 int LaneSize = 128 / VT.getScalarSizeInBits();
6220 int Size = Mask.size();
6221 for (int i = 0; i < Size; ++i)
6222 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6227 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6229 /// This checks a shuffle mask to see if it is performing the same
6230 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6231 /// that it is also not lane-crossing. It may however involve a blend from the
6232 /// same lane of a second vector.
6234 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6235 /// non-trivial to compute in the face of undef lanes. The representation is
6236 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6237 /// entries from both V1 and V2 inputs to the wider mask.
6239 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6240 SmallVectorImpl<int> &RepeatedMask) {
6241 int LaneSize = 128 / VT.getScalarSizeInBits();
6242 RepeatedMask.resize(LaneSize, -1);
6243 int Size = Mask.size();
6244 for (int i = 0; i < Size; ++i) {
6247 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6248 // This entry crosses lanes, so there is no way to model this shuffle.
6251 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6252 if (RepeatedMask[i % LaneSize] == -1)
6253 // This is the first non-undef entry in this slot of a 128-bit lane.
6254 RepeatedMask[i % LaneSize] =
6255 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6256 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6257 // Found a mismatch with the repeated mask.
6263 /// \brief Test whether a shuffle mask is equivalent within each 256-bit lane.
6265 /// This checks a shuffle mask to see if it is performing the same
6266 /// 256-bit lane-relative shuffle in each 256-bit lane. This trivially implies
6267 /// that it is also not lane-crossing. It may however involve a blend from the
6268 /// same lane of a second vector.
6270 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6271 /// non-trivial to compute in the face of undef lanes. The representation is
6272 /// *not* suitable for use with existing 256-bit shuffles as it will contain
6273 /// entries from both V1 and V2 inputs to the wider mask.
6275 is256BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6276 SmallVectorImpl<int> &RepeatedMask) {
6277 int LaneSize = 256 / VT.getScalarSizeInBits();
6278 RepeatedMask.resize(LaneSize, -1);
6279 int Size = Mask.size();
6280 for (int i = 0; i < Size; ++i) {
6283 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6284 // This entry crosses lanes, so there is no way to model this shuffle.
6287 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6288 if (RepeatedMask[i % LaneSize] == -1)
6289 // This is the first non-undef entry in this slot of a 256-bit lane.
6290 RepeatedMask[i % LaneSize] =
6291 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6292 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6293 // Found a mismatch with the repeated mask.
6299 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6302 /// This is a fast way to test a shuffle mask against a fixed pattern:
6304 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6306 /// It returns true if the mask is exactly as wide as the argument list, and
6307 /// each element of the mask is either -1 (signifying undef) or the value given
6308 /// in the argument.
6309 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6310 ArrayRef<int> ExpectedMask) {
6311 if (Mask.size() != ExpectedMask.size())
6314 int Size = Mask.size();
6316 // If the values are build vectors, we can look through them to find
6317 // equivalent inputs that make the shuffles equivalent.
6318 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6319 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6321 for (int i = 0; i < Size; ++i)
6322 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6323 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6324 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6325 if (!MaskBV || !ExpectedBV ||
6326 MaskBV->getOperand(Mask[i] % Size) !=
6327 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6334 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6336 /// This helper function produces an 8-bit shuffle immediate corresponding to
6337 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6338 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6341 /// NB: We rely heavily on "undef" masks preserving the input lane.
6342 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6343 SelectionDAG &DAG) {
6344 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6345 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6346 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6347 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6348 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6351 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6352 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6353 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6354 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6355 return DAG.getConstant(Imm, DL, MVT::i8);
6358 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6360 /// This is used as a fallback approach when first class blend instructions are
6361 /// unavailable. Currently it is only suitable for integer vectors, but could
6362 /// be generalized for floating point vectors if desirable.
6363 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6364 SDValue V2, ArrayRef<int> Mask,
6365 SelectionDAG &DAG) {
6366 assert(VT.isInteger() && "Only supports integer vector types!");
6367 MVT EltVT = VT.getScalarType();
6368 int NumEltBits = EltVT.getSizeInBits();
6369 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6370 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6372 SmallVector<SDValue, 16> MaskOps;
6373 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6374 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6375 return SDValue(); // Shuffled input!
6376 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6379 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6380 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6381 // We have to cast V2 around.
6382 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6383 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6384 DAG.getBitcast(MaskVT, V1Mask),
6385 DAG.getBitcast(MaskVT, V2)));
6386 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6389 /// \brief Try to emit a blend instruction for a shuffle.
6391 /// This doesn't do any checks for the availability of instructions for blending
6392 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6393 /// be matched in the backend with the type given. What it does check for is
6394 /// that the shuffle mask is in fact a blend.
6395 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6396 SDValue V2, ArrayRef<int> Mask,
6397 const X86Subtarget *Subtarget,
6398 SelectionDAG &DAG) {
6399 unsigned BlendMask = 0;
6400 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6401 if (Mask[i] >= Size) {
6402 if (Mask[i] != i + Size)
6403 return SDValue(); // Shuffled V2 input!
6404 BlendMask |= 1u << i;
6407 if (Mask[i] >= 0 && Mask[i] != i)
6408 return SDValue(); // Shuffled V1 input!
6410 switch (VT.SimpleTy) {
6415 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6416 DAG.getConstant(BlendMask, DL, MVT::i8));
6420 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6424 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6425 // that instruction.
6426 if (Subtarget->hasAVX2()) {
6427 // Scale the blend by the number of 32-bit dwords per element.
6428 int Scale = VT.getScalarSizeInBits() / 32;
6430 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6431 if (Mask[i] >= Size)
6432 for (int j = 0; j < Scale; ++j)
6433 BlendMask |= 1u << (i * Scale + j);
6435 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6436 V1 = DAG.getBitcast(BlendVT, V1);
6437 V2 = DAG.getBitcast(BlendVT, V2);
6438 return DAG.getBitcast(
6439 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6440 DAG.getConstant(BlendMask, DL, MVT::i8)));
6444 // For integer shuffles we need to expand the mask and cast the inputs to
6445 // v8i16s prior to blending.
6446 int Scale = 8 / VT.getVectorNumElements();
6448 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6449 if (Mask[i] >= Size)
6450 for (int j = 0; j < Scale; ++j)
6451 BlendMask |= 1u << (i * Scale + j);
6453 V1 = DAG.getBitcast(MVT::v8i16, V1);
6454 V2 = DAG.getBitcast(MVT::v8i16, V2);
6455 return DAG.getBitcast(VT,
6456 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6457 DAG.getConstant(BlendMask, DL, MVT::i8)));
6461 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6462 SmallVector<int, 8> RepeatedMask;
6463 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6464 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6465 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6467 for (int i = 0; i < 8; ++i)
6468 if (RepeatedMask[i] >= 16)
6469 BlendMask |= 1u << i;
6470 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6471 DAG.getConstant(BlendMask, DL, MVT::i8));
6477 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6478 "256-bit byte-blends require AVX2 support!");
6480 // Scale the blend by the number of bytes per element.
6481 int Scale = VT.getScalarSizeInBits() / 8;
6483 // This form of blend is always done on bytes. Compute the byte vector
6485 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6487 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6488 // mix of LLVM's code generator and the x86 backend. We tell the code
6489 // generator that boolean values in the elements of an x86 vector register
6490 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6491 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6492 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6493 // of the element (the remaining are ignored) and 0 in that high bit would
6494 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6495 // the LLVM model for boolean values in vector elements gets the relevant
6496 // bit set, it is set backwards and over constrained relative to x86's
6498 SmallVector<SDValue, 32> VSELECTMask;
6499 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6500 for (int j = 0; j < Scale; ++j)
6501 VSELECTMask.push_back(
6502 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6503 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6506 V1 = DAG.getBitcast(BlendVT, V1);
6507 V2 = DAG.getBitcast(BlendVT, V2);
6508 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6509 DAG.getNode(ISD::BUILD_VECTOR, DL,
6510 BlendVT, VSELECTMask),
6515 llvm_unreachable("Not a supported integer vector type!");
6519 /// \brief Try to lower as a blend of elements from two inputs followed by
6520 /// a single-input permutation.
6522 /// This matches the pattern where we can blend elements from two inputs and
6523 /// then reduce the shuffle to a single-input permutation.
6524 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6527 SelectionDAG &DAG) {
6528 // We build up the blend mask while checking whether a blend is a viable way
6529 // to reduce the shuffle.
6530 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6531 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6533 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6537 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6539 if (BlendMask[Mask[i] % Size] == -1)
6540 BlendMask[Mask[i] % Size] = Mask[i];
6541 else if (BlendMask[Mask[i] % Size] != Mask[i])
6542 return SDValue(); // Can't blend in the needed input!
6544 PermuteMask[i] = Mask[i] % Size;
6547 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6548 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6551 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6552 /// blends and permutes.
6554 /// This matches the extremely common pattern for handling combined
6555 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6556 /// operations. It will try to pick the best arrangement of shuffles and
6558 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6562 SelectionDAG &DAG) {
6563 // Shuffle the input elements into the desired positions in V1 and V2 and
6564 // blend them together.
6565 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6566 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6567 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6568 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6569 if (Mask[i] >= 0 && Mask[i] < Size) {
6570 V1Mask[i] = Mask[i];
6572 } else if (Mask[i] >= Size) {
6573 V2Mask[i] = Mask[i] - Size;
6574 BlendMask[i] = i + Size;
6577 // Try to lower with the simpler initial blend strategy unless one of the
6578 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6579 // shuffle may be able to fold with a load or other benefit. However, when
6580 // we'll have to do 2x as many shuffles in order to achieve this, blending
6581 // first is a better strategy.
6582 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6583 if (SDValue BlendPerm =
6584 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6587 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6588 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6589 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6592 /// \brief Try to lower a vector shuffle as a byte rotation.
6594 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6595 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6596 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6597 /// try to generically lower a vector shuffle through such an pattern. It
6598 /// does not check for the profitability of lowering either as PALIGNR or
6599 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6600 /// This matches shuffle vectors that look like:
6602 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6604 /// Essentially it concatenates V1 and V2, shifts right by some number of
6605 /// elements, and takes the low elements as the result. Note that while this is
6606 /// specified as a *right shift* because x86 is little-endian, it is a *left
6607 /// rotate* of the vector lanes.
6608 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6611 const X86Subtarget *Subtarget,
6612 SelectionDAG &DAG) {
6613 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6615 int NumElts = Mask.size();
6616 int NumLanes = VT.getSizeInBits() / 128;
6617 int NumLaneElts = NumElts / NumLanes;
6619 // We need to detect various ways of spelling a rotation:
6620 // [11, 12, 13, 14, 15, 0, 1, 2]
6621 // [-1, 12, 13, 14, -1, -1, 1, -1]
6622 // [-1, -1, -1, -1, -1, -1, 1, 2]
6623 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6624 // [-1, 4, 5, 6, -1, -1, 9, -1]
6625 // [-1, 4, 5, 6, -1, -1, -1, -1]
6628 for (int l = 0; l < NumElts; l += NumLaneElts) {
6629 for (int i = 0; i < NumLaneElts; ++i) {
6630 if (Mask[l + i] == -1)
6632 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6634 // Get the mod-Size index and lane correct it.
6635 int LaneIdx = (Mask[l + i] % NumElts) - l;
6636 // Make sure it was in this lane.
6637 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6640 // Determine where a rotated vector would have started.
6641 int StartIdx = i - LaneIdx;
6643 // The identity rotation isn't interesting, stop.
6646 // If we found the tail of a vector the rotation must be the missing
6647 // front. If we found the head of a vector, it must be how much of the
6649 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6652 Rotation = CandidateRotation;
6653 else if (Rotation != CandidateRotation)
6654 // The rotations don't match, so we can't match this mask.
6657 // Compute which value this mask is pointing at.
6658 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6660 // Compute which of the two target values this index should be assigned
6661 // to. This reflects whether the high elements are remaining or the low
6662 // elements are remaining.
6663 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6665 // Either set up this value if we've not encountered it before, or check
6666 // that it remains consistent.
6669 else if (TargetV != MaskV)
6670 // This may be a rotation, but it pulls from the inputs in some
6671 // unsupported interleaving.
6676 // Check that we successfully analyzed the mask, and normalize the results.
6677 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6678 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6684 // The actual rotate instruction rotates bytes, so we need to scale the
6685 // rotation based on how many bytes are in the vector lane.
6686 int Scale = 16 / NumLaneElts;
6688 // SSSE3 targets can use the palignr instruction.
6689 if (Subtarget->hasSSSE3()) {
6690 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6691 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6692 Lo = DAG.getBitcast(AlignVT, Lo);
6693 Hi = DAG.getBitcast(AlignVT, Hi);
6695 return DAG.getBitcast(
6696 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6697 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6700 assert(VT.getSizeInBits() == 128 &&
6701 "Rotate-based lowering only supports 128-bit lowering!");
6702 assert(Mask.size() <= 16 &&
6703 "Can shuffle at most 16 bytes in a 128-bit vector!");
6705 // Default SSE2 implementation
6706 int LoByteShift = 16 - Rotation * Scale;
6707 int HiByteShift = Rotation * Scale;
6709 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6710 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6711 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6713 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6714 DAG.getConstant(LoByteShift, DL, MVT::i8));
6715 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6716 DAG.getConstant(HiByteShift, DL, MVT::i8));
6717 return DAG.getBitcast(VT,
6718 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6721 /// \brief Compute whether each element of a shuffle is zeroable.
6723 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6724 /// Either it is an undef element in the shuffle mask, the element of the input
6725 /// referenced is undef, or the element of the input referenced is known to be
6726 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6727 /// as many lanes with this technique as possible to simplify the remaining
6729 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6730 SDValue V1, SDValue V2) {
6731 SmallBitVector Zeroable(Mask.size(), false);
6733 while (V1.getOpcode() == ISD::BITCAST)
6734 V1 = V1->getOperand(0);
6735 while (V2.getOpcode() == ISD::BITCAST)
6736 V2 = V2->getOperand(0);
6738 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6739 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6741 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6743 // Handle the easy cases.
6744 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6749 // If this is an index into a build_vector node (which has the same number
6750 // of elements), dig out the input value and use it.
6751 SDValue V = M < Size ? V1 : V2;
6752 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6755 SDValue Input = V.getOperand(M % Size);
6756 // The UNDEF opcode check really should be dead code here, but not quite
6757 // worth asserting on (it isn't invalid, just unexpected).
6758 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6765 /// \brief Try to emit a bitmask instruction for a shuffle.
6767 /// This handles cases where we can model a blend exactly as a bitmask due to
6768 /// one of the inputs being zeroable.
6769 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6770 SDValue V2, ArrayRef<int> Mask,
6771 SelectionDAG &DAG) {
6772 MVT EltVT = VT.getScalarType();
6773 int NumEltBits = EltVT.getSizeInBits();
6774 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6775 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6776 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6778 if (EltVT.isFloatingPoint()) {
6779 Zero = DAG.getBitcast(EltVT, Zero);
6780 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6782 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6783 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6785 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6788 if (Mask[i] % Size != i)
6789 return SDValue(); // Not a blend.
6791 V = Mask[i] < Size ? V1 : V2;
6792 else if (V != (Mask[i] < Size ? V1 : V2))
6793 return SDValue(); // Can only let one input through the mask.
6795 VMaskOps[i] = AllOnes;
6798 return SDValue(); // No non-zeroable elements!
6800 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6801 V = DAG.getNode(VT.isFloatingPoint()
6802 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6807 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6809 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6810 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6811 /// matches elements from one of the input vectors shuffled to the left or
6812 /// right with zeroable elements 'shifted in'. It handles both the strictly
6813 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6816 /// PSHL : (little-endian) left bit shift.
6817 /// [ zz, 0, zz, 2 ]
6818 /// [ -1, 4, zz, -1 ]
6819 /// PSRL : (little-endian) right bit shift.
6821 /// [ -1, -1, 7, zz]
6822 /// PSLLDQ : (little-endian) left byte shift
6823 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6824 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6825 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6826 /// PSRLDQ : (little-endian) right byte shift
6827 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6828 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6829 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6830 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6831 SDValue V2, ArrayRef<int> Mask,
6832 SelectionDAG &DAG) {
6833 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6835 int Size = Mask.size();
6836 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6838 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6839 for (int i = 0; i < Size; i += Scale)
6840 for (int j = 0; j < Shift; ++j)
6841 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6847 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6848 for (int i = 0; i != Size; i += Scale) {
6849 unsigned Pos = Left ? i + Shift : i;
6850 unsigned Low = Left ? i : i + Shift;
6851 unsigned Len = Scale - Shift;
6852 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6853 Low + (V == V1 ? 0 : Size)))
6857 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6858 bool ByteShift = ShiftEltBits > 64;
6859 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6860 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6861 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6863 // Normalize the scale for byte shifts to still produce an i64 element
6865 Scale = ByteShift ? Scale / 2 : Scale;
6867 // We need to round trip through the appropriate type for the shift.
6868 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6869 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6870 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6871 "Illegal integer vector type");
6872 V = DAG.getBitcast(ShiftVT, V);
6874 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6875 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6876 return DAG.getBitcast(VT, V);
6879 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6880 // keep doubling the size of the integer elements up to that. We can
6881 // then shift the elements of the integer vector by whole multiples of
6882 // their width within the elements of the larger integer vector. Test each
6883 // multiple to see if we can find a match with the moved element indices
6884 // and that the shifted in elements are all zeroable.
6885 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6886 for (int Shift = 1; Shift != Scale; ++Shift)
6887 for (bool Left : {true, false})
6888 if (CheckZeros(Shift, Scale, Left))
6889 for (SDValue V : {V1, V2})
6890 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6897 /// \brief Lower a vector shuffle as a zero or any extension.
6899 /// Given a specific number of elements, element bit width, and extension
6900 /// stride, produce either a zero or any extension based on the available
6901 /// features of the subtarget.
6902 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6903 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6904 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6905 assert(Scale > 1 && "Need a scale to extend.");
6906 int NumElements = VT.getVectorNumElements();
6907 int EltBits = VT.getScalarSizeInBits();
6908 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6909 "Only 8, 16, and 32 bit elements can be extended.");
6910 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6912 // Found a valid zext mask! Try various lowering strategies based on the
6913 // input type and available ISA extensions.
6914 if (Subtarget->hasSSE41()) {
6915 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6916 NumElements / Scale);
6917 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6920 // For any extends we can cheat for larger element sizes and use shuffle
6921 // instructions that can fold with a load and/or copy.
6922 if (AnyExt && EltBits == 32) {
6923 int PSHUFDMask[4] = {0, -1, 1, -1};
6924 return DAG.getBitcast(
6925 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6926 DAG.getBitcast(MVT::v4i32, InputV),
6927 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
6929 if (AnyExt && EltBits == 16 && Scale > 2) {
6930 int PSHUFDMask[4] = {0, -1, 0, -1};
6931 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6932 DAG.getBitcast(MVT::v4i32, InputV),
6933 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
6934 int PSHUFHWMask[4] = {1, -1, -1, -1};
6935 return DAG.getBitcast(
6936 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6937 DAG.getBitcast(MVT::v8i16, InputV),
6938 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
6941 // If this would require more than 2 unpack instructions to expand, use
6942 // pshufb when available. We can only use more than 2 unpack instructions
6943 // when zero extending i8 elements which also makes it easier to use pshufb.
6944 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6945 assert(NumElements == 16 && "Unexpected byte vector width!");
6946 SDValue PSHUFBMask[16];
6947 for (int i = 0; i < 16; ++i)
6949 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
6950 InputV = DAG.getBitcast(MVT::v16i8, InputV);
6951 return DAG.getBitcast(VT,
6952 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6953 DAG.getNode(ISD::BUILD_VECTOR, DL,
6954 MVT::v16i8, PSHUFBMask)));
6957 // Otherwise emit a sequence of unpacks.
6959 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6960 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6961 : getZeroVector(InputVT, Subtarget, DAG, DL);
6962 InputV = DAG.getBitcast(InputVT, InputV);
6963 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6967 } while (Scale > 1);
6968 return DAG.getBitcast(VT, InputV);
6971 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6973 /// This routine will try to do everything in its power to cleverly lower
6974 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6975 /// check for the profitability of this lowering, it tries to aggressively
6976 /// match this pattern. It will use all of the micro-architectural details it
6977 /// can to emit an efficient lowering. It handles both blends with all-zero
6978 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6979 /// masking out later).
6981 /// The reason we have dedicated lowering for zext-style shuffles is that they
6982 /// are both incredibly common and often quite performance sensitive.
6983 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6984 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6985 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6986 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6988 int Bits = VT.getSizeInBits();
6989 int NumElements = VT.getVectorNumElements();
6990 assert(VT.getScalarSizeInBits() <= 32 &&
6991 "Exceeds 32-bit integer zero extension limit");
6992 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6994 // Define a helper function to check a particular ext-scale and lower to it if
6996 auto Lower = [&](int Scale) -> SDValue {
6999 for (int i = 0; i < NumElements; ++i) {
7001 continue; // Valid anywhere but doesn't tell us anything.
7002 if (i % Scale != 0) {
7003 // Each of the extended elements need to be zeroable.
7007 // We no longer are in the anyext case.
7012 // Each of the base elements needs to be consecutive indices into the
7013 // same input vector.
7014 SDValue V = Mask[i] < NumElements ? V1 : V2;
7017 else if (InputV != V)
7018 return SDValue(); // Flip-flopping inputs.
7020 if (Mask[i] % NumElements != i / Scale)
7021 return SDValue(); // Non-consecutive strided elements.
7024 // If we fail to find an input, we have a zero-shuffle which should always
7025 // have already been handled.
7026 // FIXME: Maybe handle this here in case during blending we end up with one?
7030 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7031 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
7034 // The widest scale possible for extending is to a 64-bit integer.
7035 assert(Bits % 64 == 0 &&
7036 "The number of bits in a vector must be divisible by 64 on x86!");
7037 int NumExtElements = Bits / 64;
7039 // Each iteration, try extending the elements half as much, but into twice as
7041 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7042 assert(NumElements % NumExtElements == 0 &&
7043 "The input vector size must be divisible by the extended size.");
7044 if (SDValue V = Lower(NumElements / NumExtElements))
7048 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7052 // Returns one of the source operands if the shuffle can be reduced to a
7053 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7054 auto CanZExtLowHalf = [&]() {
7055 for (int i = NumElements / 2; i != NumElements; ++i)
7058 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7060 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7065 if (SDValue V = CanZExtLowHalf()) {
7066 V = DAG.getBitcast(MVT::v2i64, V);
7067 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7068 return DAG.getBitcast(VT, V);
7071 // No viable ext lowering found.
7075 /// \brief Try to get a scalar value for a specific element of a vector.
7077 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7078 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7079 SelectionDAG &DAG) {
7080 MVT VT = V.getSimpleValueType();
7081 MVT EltVT = VT.getVectorElementType();
7082 while (V.getOpcode() == ISD::BITCAST)
7083 V = V.getOperand(0);
7084 // If the bitcasts shift the element size, we can't extract an equivalent
7086 MVT NewVT = V.getSimpleValueType();
7087 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7090 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7091 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7092 // Ensure the scalar operand is the same size as the destination.
7093 // FIXME: Add support for scalar truncation where possible.
7094 SDValue S = V.getOperand(Idx);
7095 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7096 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7102 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7104 /// This is particularly important because the set of instructions varies
7105 /// significantly based on whether the operand is a load or not.
7106 static bool isShuffleFoldableLoad(SDValue V) {
7107 while (V.getOpcode() == ISD::BITCAST)
7108 V = V.getOperand(0);
7110 return ISD::isNON_EXTLoad(V.getNode());
7113 /// \brief Try to lower insertion of a single element into a zero vector.
7115 /// This is a common pattern that we have especially efficient patterns to lower
7116 /// across all subtarget feature sets.
7117 static SDValue lowerVectorShuffleAsElementInsertion(
7118 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7119 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7120 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7122 MVT EltVT = VT.getVectorElementType();
7124 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7125 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7127 bool IsV1Zeroable = true;
7128 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7129 if (i != V2Index && !Zeroable[i]) {
7130 IsV1Zeroable = false;
7134 // Check for a single input from a SCALAR_TO_VECTOR node.
7135 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7136 // all the smarts here sunk into that routine. However, the current
7137 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7138 // vector shuffle lowering is dead.
7139 if (SDValue V2S = getScalarValueForVectorElement(
7140 V2, Mask[V2Index] - Mask.size(), DAG)) {
7141 // We need to zext the scalar if it is smaller than an i32.
7142 V2S = DAG.getBitcast(EltVT, V2S);
7143 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7144 // Using zext to expand a narrow element won't work for non-zero
7149 // Zero-extend directly to i32.
7151 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7153 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7154 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7155 EltVT == MVT::i16) {
7156 // Either not inserting from the low element of the input or the input
7157 // element size is too small to use VZEXT_MOVL to clear the high bits.
7161 if (!IsV1Zeroable) {
7162 // If V1 can't be treated as a zero vector we have fewer options to lower
7163 // this. We can't support integer vectors or non-zero targets cheaply, and
7164 // the V1 elements can't be permuted in any way.
7165 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7166 if (!VT.isFloatingPoint() || V2Index != 0)
7168 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7169 V1Mask[V2Index] = -1;
7170 if (!isNoopShuffleMask(V1Mask))
7172 // This is essentially a special case blend operation, but if we have
7173 // general purpose blend operations, they are always faster. Bail and let
7174 // the rest of the lowering handle these as blends.
7175 if (Subtarget->hasSSE41())
7178 // Otherwise, use MOVSD or MOVSS.
7179 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7180 "Only two types of floating point element types to handle!");
7181 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7185 // This lowering only works for the low element with floating point vectors.
7186 if (VT.isFloatingPoint() && V2Index != 0)
7189 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7191 V2 = DAG.getBitcast(VT, V2);
7194 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7195 // the desired position. Otherwise it is more efficient to do a vector
7196 // shift left. We know that we can do a vector shift left because all
7197 // the inputs are zero.
7198 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7199 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7200 V2Shuffle[V2Index] = 0;
7201 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7203 V2 = DAG.getBitcast(MVT::v2i64, V2);
7205 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7207 V2Index * EltVT.getSizeInBits()/8, DL,
7208 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7209 V2 = DAG.getBitcast(VT, V2);
7215 /// \brief Try to lower broadcast of a single element.
7217 /// For convenience, this code also bundles all of the subtarget feature set
7218 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7219 /// a convenient way to factor it out.
7220 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7222 const X86Subtarget *Subtarget,
7223 SelectionDAG &DAG) {
7224 if (!Subtarget->hasAVX())
7226 if (VT.isInteger() && !Subtarget->hasAVX2())
7229 // Check that the mask is a broadcast.
7230 int BroadcastIdx = -1;
7232 if (M >= 0 && BroadcastIdx == -1)
7234 else if (M >= 0 && M != BroadcastIdx)
7237 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7238 "a sorted mask where the broadcast "
7241 // Go up the chain of (vector) values to find a scalar load that we can
7242 // combine with the broadcast.
7244 switch (V.getOpcode()) {
7245 case ISD::CONCAT_VECTORS: {
7246 int OperandSize = Mask.size() / V.getNumOperands();
7247 V = V.getOperand(BroadcastIdx / OperandSize);
7248 BroadcastIdx %= OperandSize;
7252 case ISD::INSERT_SUBVECTOR: {
7253 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7254 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7258 int BeginIdx = (int)ConstantIdx->getZExtValue();
7260 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7261 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7262 BroadcastIdx -= BeginIdx;
7273 // Check if this is a broadcast of a scalar. We special case lowering
7274 // for scalars so that we can more effectively fold with loads.
7275 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7276 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7277 V = V.getOperand(BroadcastIdx);
7279 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7280 // Only AVX2 has register broadcasts.
7281 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7283 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7284 // We can't broadcast from a vector register without AVX2, and we can only
7285 // broadcast from the zero-element of a vector register.
7289 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7292 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7293 // INSERTPS when the V1 elements are already in the correct locations
7294 // because otherwise we can just always use two SHUFPS instructions which
7295 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7296 // perform INSERTPS if a single V1 element is out of place and all V2
7297 // elements are zeroable.
7298 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7300 SelectionDAG &DAG) {
7301 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7302 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7303 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7304 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7306 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7309 int V1DstIndex = -1;
7310 int V2DstIndex = -1;
7311 bool V1UsedInPlace = false;
7313 for (int i = 0; i < 4; ++i) {
7314 // Synthesize a zero mask from the zeroable elements (includes undefs).
7320 // Flag if we use any V1 inputs in place.
7322 V1UsedInPlace = true;
7326 // We can only insert a single non-zeroable element.
7327 if (V1DstIndex != -1 || V2DstIndex != -1)
7331 // V1 input out of place for insertion.
7334 // V2 input for insertion.
7339 // Don't bother if we have no (non-zeroable) element for insertion.
7340 if (V1DstIndex == -1 && V2DstIndex == -1)
7343 // Determine element insertion src/dst indices. The src index is from the
7344 // start of the inserted vector, not the start of the concatenated vector.
7345 unsigned V2SrcIndex = 0;
7346 if (V1DstIndex != -1) {
7347 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7348 // and don't use the original V2 at all.
7349 V2SrcIndex = Mask[V1DstIndex];
7350 V2DstIndex = V1DstIndex;
7353 V2SrcIndex = Mask[V2DstIndex] - 4;
7356 // If no V1 inputs are used in place, then the result is created only from
7357 // the zero mask and the V2 insertion - so remove V1 dependency.
7359 V1 = DAG.getUNDEF(MVT::v4f32);
7361 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7362 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7364 // Insert the V2 element into the desired position.
7366 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7367 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7370 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7371 /// UNPCK instruction.
7373 /// This specifically targets cases where we end up with alternating between
7374 /// the two inputs, and so can permute them into something that feeds a single
7375 /// UNPCK instruction. Note that this routine only targets integer vectors
7376 /// because for floating point vectors we have a generalized SHUFPS lowering
7377 /// strategy that handles everything that doesn't *exactly* match an unpack,
7378 /// making this clever lowering unnecessary.
7379 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7380 SDValue V2, ArrayRef<int> Mask,
7381 SelectionDAG &DAG) {
7382 assert(!VT.isFloatingPoint() &&
7383 "This routine only supports integer vectors.");
7384 assert(!isSingleInputShuffleMask(Mask) &&
7385 "This routine should only be used when blending two inputs.");
7386 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7388 int Size = Mask.size();
7390 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7391 return M >= 0 && M % Size < Size / 2;
7393 int NumHiInputs = std::count_if(
7394 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7396 bool UnpackLo = NumLoInputs >= NumHiInputs;
7398 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7399 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7400 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7402 for (int i = 0; i < Size; ++i) {
7406 // Each element of the unpack contains Scale elements from this mask.
7407 int UnpackIdx = i / Scale;
7409 // We only handle the case where V1 feeds the first slots of the unpack.
7410 // We rely on canonicalization to ensure this is the case.
7411 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7414 // Setup the mask for this input. The indexing is tricky as we have to
7415 // handle the unpack stride.
7416 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7417 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7421 // If we will have to shuffle both inputs to use the unpack, check whether
7422 // we can just unpack first and shuffle the result. If so, skip this unpack.
7423 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7424 !isNoopShuffleMask(V2Mask))
7427 // Shuffle the inputs into place.
7428 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7429 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7431 // Cast the inputs to the type we will use to unpack them.
7432 V1 = DAG.getBitcast(UnpackVT, V1);
7433 V2 = DAG.getBitcast(UnpackVT, V2);
7435 // Unpack the inputs and cast the result back to the desired type.
7436 return DAG.getBitcast(
7437 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7441 // We try each unpack from the largest to the smallest to try and find one
7442 // that fits this mask.
7443 int OrigNumElements = VT.getVectorNumElements();
7444 int OrigScalarSize = VT.getScalarSizeInBits();
7445 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7446 int Scale = ScalarSize / OrigScalarSize;
7447 int NumElements = OrigNumElements / Scale;
7448 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7449 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7453 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7455 if (NumLoInputs == 0 || NumHiInputs == 0) {
7456 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7457 "We have to have *some* inputs!");
7458 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7460 // FIXME: We could consider the total complexity of the permute of each
7461 // possible unpacking. Or at the least we should consider how many
7462 // half-crossings are created.
7463 // FIXME: We could consider commuting the unpacks.
7465 SmallVector<int, 32> PermMask;
7466 PermMask.assign(Size, -1);
7467 for (int i = 0; i < Size; ++i) {
7471 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7474 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7476 return DAG.getVectorShuffle(
7477 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7479 DAG.getUNDEF(VT), PermMask);
7485 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7487 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7488 /// support for floating point shuffles but not integer shuffles. These
7489 /// instructions will incur a domain crossing penalty on some chips though so
7490 /// it is better to avoid lowering through this for integer vectors where
7492 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7493 const X86Subtarget *Subtarget,
7494 SelectionDAG &DAG) {
7496 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7497 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7498 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7499 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7500 ArrayRef<int> Mask = SVOp->getMask();
7501 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7503 if (isSingleInputShuffleMask(Mask)) {
7504 // Use low duplicate instructions for masks that match their pattern.
7505 if (Subtarget->hasSSE3())
7506 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7507 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7509 // Straight shuffle of a single input vector. Simulate this by using the
7510 // single input as both of the "inputs" to this instruction..
7511 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7513 if (Subtarget->hasAVX()) {
7514 // If we have AVX, we can use VPERMILPS which will allow folding a load
7515 // into the shuffle.
7516 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7517 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7520 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7521 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7523 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7524 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7526 // If we have a single input, insert that into V1 if we can do so cheaply.
7527 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7528 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7529 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7531 // Try inverting the insertion since for v2 masks it is easy to do and we
7532 // can't reliably sort the mask one way or the other.
7533 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7534 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7535 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7536 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7540 // Try to use one of the special instruction patterns to handle two common
7541 // blend patterns if a zero-blend above didn't work.
7542 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7543 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7544 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7545 // We can either use a special instruction to load over the low double or
7546 // to move just the low double.
7548 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7550 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7552 if (Subtarget->hasSSE41())
7553 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7557 // Use dedicated unpack instructions for masks that match their pattern.
7558 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7559 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7560 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7561 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7563 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7564 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7565 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7568 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7570 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7571 /// the integer unit to minimize domain crossing penalties. However, for blends
7572 /// it falls back to the floating point shuffle operation with appropriate bit
7574 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7575 const X86Subtarget *Subtarget,
7576 SelectionDAG &DAG) {
7578 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7579 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7580 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7581 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7582 ArrayRef<int> Mask = SVOp->getMask();
7583 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7585 if (isSingleInputShuffleMask(Mask)) {
7586 // Check for being able to broadcast a single element.
7587 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7588 Mask, Subtarget, DAG))
7591 // Straight shuffle of a single input vector. For everything from SSE2
7592 // onward this has a single fast instruction with no scary immediates.
7593 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7594 V1 = DAG.getBitcast(MVT::v4i32, V1);
7595 int WidenedMask[4] = {
7596 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7597 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7598 return DAG.getBitcast(
7600 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7601 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7603 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7604 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7605 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7606 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7608 // If we have a blend of two PACKUS operations an the blend aligns with the
7609 // low and half halves, we can just merge the PACKUS operations. This is
7610 // particularly important as it lets us merge shuffles that this routine itself
7612 auto GetPackNode = [](SDValue V) {
7613 while (V.getOpcode() == ISD::BITCAST)
7614 V = V.getOperand(0);
7616 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7618 if (SDValue V1Pack = GetPackNode(V1))
7619 if (SDValue V2Pack = GetPackNode(V2))
7620 return DAG.getBitcast(MVT::v2i64,
7621 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7622 Mask[0] == 0 ? V1Pack.getOperand(0)
7623 : V1Pack.getOperand(1),
7624 Mask[1] == 2 ? V2Pack.getOperand(0)
7625 : V2Pack.getOperand(1)));
7627 // Try to use shift instructions.
7629 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7632 // When loading a scalar and then shuffling it into a vector we can often do
7633 // the insertion cheaply.
7634 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7635 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7637 // Try inverting the insertion since for v2 masks it is easy to do and we
7638 // can't reliably sort the mask one way or the other.
7639 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7640 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7641 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7644 // We have different paths for blend lowering, but they all must use the
7645 // *exact* same predicate.
7646 bool IsBlendSupported = Subtarget->hasSSE41();
7647 if (IsBlendSupported)
7648 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7652 // Use dedicated unpack instructions for masks that match their pattern.
7653 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7654 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7655 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7656 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7658 // Try to use byte rotation instructions.
7659 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7660 if (Subtarget->hasSSSE3())
7661 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7662 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7665 // If we have direct support for blends, we should lower by decomposing into
7666 // a permute. That will be faster than the domain cross.
7667 if (IsBlendSupported)
7668 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7671 // We implement this with SHUFPD which is pretty lame because it will likely
7672 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7673 // However, all the alternatives are still more cycles and newer chips don't
7674 // have this problem. It would be really nice if x86 had better shuffles here.
7675 V1 = DAG.getBitcast(MVT::v2f64, V1);
7676 V2 = DAG.getBitcast(MVT::v2f64, V2);
7677 return DAG.getBitcast(MVT::v2i64,
7678 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7681 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7683 /// This is used to disable more specialized lowerings when the shufps lowering
7684 /// will happen to be efficient.
7685 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7686 // This routine only handles 128-bit shufps.
7687 assert(Mask.size() == 4 && "Unsupported mask size!");
7689 // To lower with a single SHUFPS we need to have the low half and high half
7690 // each requiring a single input.
7691 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7693 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7699 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7701 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7702 /// It makes no assumptions about whether this is the *best* lowering, it simply
7704 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7705 ArrayRef<int> Mask, SDValue V1,
7706 SDValue V2, SelectionDAG &DAG) {
7707 SDValue LowV = V1, HighV = V2;
7708 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7711 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7713 if (NumV2Elements == 1) {
7715 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7718 // Compute the index adjacent to V2Index and in the same half by toggling
7720 int V2AdjIndex = V2Index ^ 1;
7722 if (Mask[V2AdjIndex] == -1) {
7723 // Handles all the cases where we have a single V2 element and an undef.
7724 // This will only ever happen in the high lanes because we commute the
7725 // vector otherwise.
7727 std::swap(LowV, HighV);
7728 NewMask[V2Index] -= 4;
7730 // Handle the case where the V2 element ends up adjacent to a V1 element.
7731 // To make this work, blend them together as the first step.
7732 int V1Index = V2AdjIndex;
7733 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7734 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7735 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7737 // Now proceed to reconstruct the final blend as we have the necessary
7738 // high or low half formed.
7745 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7746 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7748 } else if (NumV2Elements == 2) {
7749 if (Mask[0] < 4 && Mask[1] < 4) {
7750 // Handle the easy case where we have V1 in the low lanes and V2 in the
7754 } else if (Mask[2] < 4 && Mask[3] < 4) {
7755 // We also handle the reversed case because this utility may get called
7756 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7757 // arrange things in the right direction.
7763 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7764 // trying to place elements directly, just blend them and set up the final
7765 // shuffle to place them.
7767 // The first two blend mask elements are for V1, the second two are for
7769 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7770 Mask[2] < 4 ? Mask[2] : Mask[3],
7771 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7772 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7773 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7774 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7776 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7779 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7780 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7781 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7782 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7785 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7786 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
7789 /// \brief Lower 4-lane 32-bit floating point shuffles.
7791 /// Uses instructions exclusively from the floating point unit to minimize
7792 /// domain crossing penalties, as these are sufficient to implement all v4f32
7794 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7795 const X86Subtarget *Subtarget,
7796 SelectionDAG &DAG) {
7798 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7799 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7800 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7801 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7802 ArrayRef<int> Mask = SVOp->getMask();
7803 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7806 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7808 if (NumV2Elements == 0) {
7809 // Check for being able to broadcast a single element.
7810 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7811 Mask, Subtarget, DAG))
7814 // Use even/odd duplicate instructions for masks that match their pattern.
7815 if (Subtarget->hasSSE3()) {
7816 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7817 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7818 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7819 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7822 if (Subtarget->hasAVX()) {
7823 // If we have AVX, we can use VPERMILPS which will allow folding a load
7824 // into the shuffle.
7825 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7826 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7829 // Otherwise, use a straight shuffle of a single input vector. We pass the
7830 // input vector to both operands to simulate this with a SHUFPS.
7831 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7832 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7835 // There are special ways we can lower some single-element blends. However, we
7836 // have custom ways we can lower more complex single-element blends below that
7837 // we defer to if both this and BLENDPS fail to match, so restrict this to
7838 // when the V2 input is targeting element 0 of the mask -- that is the fast
7840 if (NumV2Elements == 1 && Mask[0] >= 4)
7841 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7842 Mask, Subtarget, DAG))
7845 if (Subtarget->hasSSE41()) {
7846 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7850 // Use INSERTPS if we can complete the shuffle efficiently.
7851 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7854 if (!isSingleSHUFPSMask(Mask))
7855 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7856 DL, MVT::v4f32, V1, V2, Mask, DAG))
7860 // Use dedicated unpack instructions for masks that match their pattern.
7861 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7862 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7863 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7864 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7865 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7866 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7867 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7868 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7870 // Otherwise fall back to a SHUFPS lowering strategy.
7871 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7874 /// \brief Lower 4-lane i32 vector shuffles.
7876 /// We try to handle these with integer-domain shuffles where we can, but for
7877 /// blends we use the floating point domain blend instructions.
7878 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7879 const X86Subtarget *Subtarget,
7880 SelectionDAG &DAG) {
7882 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7883 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7884 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7885 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7886 ArrayRef<int> Mask = SVOp->getMask();
7887 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7889 // Whenever we can lower this as a zext, that instruction is strictly faster
7890 // than any alternative. It also allows us to fold memory operands into the
7891 // shuffle in many cases.
7892 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7893 Mask, Subtarget, DAG))
7897 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7899 if (NumV2Elements == 0) {
7900 // Check for being able to broadcast a single element.
7901 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7902 Mask, Subtarget, DAG))
7905 // Straight shuffle of a single input vector. For everything from SSE2
7906 // onward this has a single fast instruction with no scary immediates.
7907 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7908 // but we aren't actually going to use the UNPCK instruction because doing
7909 // so prevents folding a load into this instruction or making a copy.
7910 const int UnpackLoMask[] = {0, 0, 1, 1};
7911 const int UnpackHiMask[] = {2, 2, 3, 3};
7912 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7913 Mask = UnpackLoMask;
7914 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7915 Mask = UnpackHiMask;
7917 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7918 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7921 // Try to use shift instructions.
7923 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7926 // There are special ways we can lower some single-element blends.
7927 if (NumV2Elements == 1)
7928 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7929 Mask, Subtarget, DAG))
7932 // We have different paths for blend lowering, but they all must use the
7933 // *exact* same predicate.
7934 bool IsBlendSupported = Subtarget->hasSSE41();
7935 if (IsBlendSupported)
7936 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7940 if (SDValue Masked =
7941 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7944 // Use dedicated unpack instructions for masks that match their pattern.
7945 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7946 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7947 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7948 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7949 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7950 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7951 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7952 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7954 // Try to use byte rotation instructions.
7955 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7956 if (Subtarget->hasSSSE3())
7957 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7958 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7961 // If we have direct support for blends, we should lower by decomposing into
7962 // a permute. That will be faster than the domain cross.
7963 if (IsBlendSupported)
7964 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7967 // Try to lower by permuting the inputs into an unpack instruction.
7968 if (SDValue Unpack =
7969 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7972 // We implement this with SHUFPS because it can blend from two vectors.
7973 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7974 // up the inputs, bypassing domain shift penalties that we would encur if we
7975 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7977 return DAG.getBitcast(
7979 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
7980 DAG.getBitcast(MVT::v4f32, V2), Mask));
7983 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7984 /// shuffle lowering, and the most complex part.
7986 /// The lowering strategy is to try to form pairs of input lanes which are
7987 /// targeted at the same half of the final vector, and then use a dword shuffle
7988 /// to place them onto the right half, and finally unpack the paired lanes into
7989 /// their final position.
7991 /// The exact breakdown of how to form these dword pairs and align them on the
7992 /// correct sides is really tricky. See the comments within the function for
7993 /// more of the details.
7995 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
7996 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
7997 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
7998 /// vector, form the analogous 128-bit 8-element Mask.
7999 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8000 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8001 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8002 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8003 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8005 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8006 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8007 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8009 SmallVector<int, 4> LoInputs;
8010 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8011 [](int M) { return M >= 0; });
8012 std::sort(LoInputs.begin(), LoInputs.end());
8013 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8014 SmallVector<int, 4> HiInputs;
8015 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8016 [](int M) { return M >= 0; });
8017 std::sort(HiInputs.begin(), HiInputs.end());
8018 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8020 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8021 int NumHToL = LoInputs.size() - NumLToL;
8023 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8024 int NumHToH = HiInputs.size() - NumLToH;
8025 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8026 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8027 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8028 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8030 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8031 // such inputs we can swap two of the dwords across the half mark and end up
8032 // with <=2 inputs to each half in each half. Once there, we can fall through
8033 // to the generic code below. For example:
8035 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8036 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8038 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8039 // and an existing 2-into-2 on the other half. In this case we may have to
8040 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8041 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8042 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8043 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8044 // half than the one we target for fixing) will be fixed when we re-enter this
8045 // path. We will also combine away any sequence of PSHUFD instructions that
8046 // result into a single instruction. Here is an example of the tricky case:
8048 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8049 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8051 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8053 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8054 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8056 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8057 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8059 // The result is fine to be handled by the generic logic.
8060 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8061 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8062 int AOffset, int BOffset) {
8063 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8064 "Must call this with A having 3 or 1 inputs from the A half.");
8065 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8066 "Must call this with B having 1 or 3 inputs from the B half.");
8067 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8068 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8070 // Compute the index of dword with only one word among the three inputs in
8071 // a half by taking the sum of the half with three inputs and subtracting
8072 // the sum of the actual three inputs. The difference is the remaining
8075 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8076 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8077 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8078 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8079 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8080 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8081 int TripleNonInputIdx =
8082 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8083 TripleDWord = TripleNonInputIdx / 2;
8085 // We use xor with one to compute the adjacent DWord to whichever one the
8087 OneInputDWord = (OneInput / 2) ^ 1;
8089 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8090 // and BToA inputs. If there is also such a problem with the BToB and AToB
8091 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8092 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8093 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8094 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8095 // Compute how many inputs will be flipped by swapping these DWords. We
8097 // to balance this to ensure we don't form a 3-1 shuffle in the other
8099 int NumFlippedAToBInputs =
8100 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8101 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8102 int NumFlippedBToBInputs =
8103 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8104 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8105 if ((NumFlippedAToBInputs == 1 &&
8106 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8107 (NumFlippedBToBInputs == 1 &&
8108 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8109 // We choose whether to fix the A half or B half based on whether that
8110 // half has zero flipped inputs. At zero, we may not be able to fix it
8111 // with that half. We also bias towards fixing the B half because that
8112 // will more commonly be the high half, and we have to bias one way.
8113 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8114 ArrayRef<int> Inputs) {
8115 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8116 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8117 PinnedIdx ^ 1) != Inputs.end();
8118 // Determine whether the free index is in the flipped dword or the
8119 // unflipped dword based on where the pinned index is. We use this bit
8120 // in an xor to conditionally select the adjacent dword.
8121 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8122 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8123 FixFreeIdx) != Inputs.end();
8124 if (IsFixIdxInput == IsFixFreeIdxInput)
8126 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8127 FixFreeIdx) != Inputs.end();
8128 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8129 "We need to be changing the number of flipped inputs!");
8130 int PSHUFHalfMask[] = {0, 1, 2, 3};
8131 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8132 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8134 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8137 if (M != -1 && M == FixIdx)
8139 else if (M != -1 && M == FixFreeIdx)
8142 if (NumFlippedBToBInputs != 0) {
8144 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8145 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8147 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8149 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8150 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8155 int PSHUFDMask[] = {0, 1, 2, 3};
8156 PSHUFDMask[ADWord] = BDWord;
8157 PSHUFDMask[BDWord] = ADWord;
8160 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8161 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8163 // Adjust the mask to match the new locations of A and B.
8165 if (M != -1 && M/2 == ADWord)
8166 M = 2 * BDWord + M % 2;
8167 else if (M != -1 && M/2 == BDWord)
8168 M = 2 * ADWord + M % 2;
8170 // Recurse back into this routine to re-compute state now that this isn't
8171 // a 3 and 1 problem.
8172 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8175 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8176 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8177 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8178 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8180 // At this point there are at most two inputs to the low and high halves from
8181 // each half. That means the inputs can always be grouped into dwords and
8182 // those dwords can then be moved to the correct half with a dword shuffle.
8183 // We use at most one low and one high word shuffle to collect these paired
8184 // inputs into dwords, and finally a dword shuffle to place them.
8185 int PSHUFLMask[4] = {-1, -1, -1, -1};
8186 int PSHUFHMask[4] = {-1, -1, -1, -1};
8187 int PSHUFDMask[4] = {-1, -1, -1, -1};
8189 // First fix the masks for all the inputs that are staying in their
8190 // original halves. This will then dictate the targets of the cross-half
8192 auto fixInPlaceInputs =
8193 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8194 MutableArrayRef<int> SourceHalfMask,
8195 MutableArrayRef<int> HalfMask, int HalfOffset) {
8196 if (InPlaceInputs.empty())
8198 if (InPlaceInputs.size() == 1) {
8199 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8200 InPlaceInputs[0] - HalfOffset;
8201 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8204 if (IncomingInputs.empty()) {
8205 // Just fix all of the in place inputs.
8206 for (int Input : InPlaceInputs) {
8207 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8208 PSHUFDMask[Input / 2] = Input / 2;
8213 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8214 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8215 InPlaceInputs[0] - HalfOffset;
8216 // Put the second input next to the first so that they are packed into
8217 // a dword. We find the adjacent index by toggling the low bit.
8218 int AdjIndex = InPlaceInputs[0] ^ 1;
8219 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8220 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8221 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8223 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8224 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8226 // Now gather the cross-half inputs and place them into a free dword of
8227 // their target half.
8228 // FIXME: This operation could almost certainly be simplified dramatically to
8229 // look more like the 3-1 fixing operation.
8230 auto moveInputsToRightHalf = [&PSHUFDMask](
8231 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8232 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8233 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8235 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8236 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8238 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8240 int LowWord = Word & ~1;
8241 int HighWord = Word | 1;
8242 return isWordClobbered(SourceHalfMask, LowWord) ||
8243 isWordClobbered(SourceHalfMask, HighWord);
8246 if (IncomingInputs.empty())
8249 if (ExistingInputs.empty()) {
8250 // Map any dwords with inputs from them into the right half.
8251 for (int Input : IncomingInputs) {
8252 // If the source half mask maps over the inputs, turn those into
8253 // swaps and use the swapped lane.
8254 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8255 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8256 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8257 Input - SourceOffset;
8258 // We have to swap the uses in our half mask in one sweep.
8259 for (int &M : HalfMask)
8260 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8262 else if (M == Input)
8263 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8265 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8266 Input - SourceOffset &&
8267 "Previous placement doesn't match!");
8269 // Note that this correctly re-maps both when we do a swap and when
8270 // we observe the other side of the swap above. We rely on that to
8271 // avoid swapping the members of the input list directly.
8272 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8275 // Map the input's dword into the correct half.
8276 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8277 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8279 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8281 "Previous placement doesn't match!");
8284 // And just directly shift any other-half mask elements to be same-half
8285 // as we will have mirrored the dword containing the element into the
8286 // same position within that half.
8287 for (int &M : HalfMask)
8288 if (M >= SourceOffset && M < SourceOffset + 4) {
8289 M = M - SourceOffset + DestOffset;
8290 assert(M >= 0 && "This should never wrap below zero!");
8295 // Ensure we have the input in a viable dword of its current half. This
8296 // is particularly tricky because the original position may be clobbered
8297 // by inputs being moved and *staying* in that half.
8298 if (IncomingInputs.size() == 1) {
8299 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8300 int InputFixed = std::find(std::begin(SourceHalfMask),
8301 std::end(SourceHalfMask), -1) -
8302 std::begin(SourceHalfMask) + SourceOffset;
8303 SourceHalfMask[InputFixed - SourceOffset] =
8304 IncomingInputs[0] - SourceOffset;
8305 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8307 IncomingInputs[0] = InputFixed;
8309 } else if (IncomingInputs.size() == 2) {
8310 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8311 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8312 // We have two non-adjacent or clobbered inputs we need to extract from
8313 // the source half. To do this, we need to map them into some adjacent
8314 // dword slot in the source mask.
8315 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8316 IncomingInputs[1] - SourceOffset};
8318 // If there is a free slot in the source half mask adjacent to one of
8319 // the inputs, place the other input in it. We use (Index XOR 1) to
8320 // compute an adjacent index.
8321 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8322 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8323 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8324 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8325 InputsFixed[1] = InputsFixed[0] ^ 1;
8326 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8327 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8328 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8329 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8330 InputsFixed[0] = InputsFixed[1] ^ 1;
8331 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8332 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8333 // The two inputs are in the same DWord but it is clobbered and the
8334 // adjacent DWord isn't used at all. Move both inputs to the free
8336 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8337 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8338 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8339 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8341 // The only way we hit this point is if there is no clobbering
8342 // (because there are no off-half inputs to this half) and there is no
8343 // free slot adjacent to one of the inputs. In this case, we have to
8344 // swap an input with a non-input.
8345 for (int i = 0; i < 4; ++i)
8346 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8347 "We can't handle any clobbers here!");
8348 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8349 "Cannot have adjacent inputs here!");
8351 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8352 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8354 // We also have to update the final source mask in this case because
8355 // it may need to undo the above swap.
8356 for (int &M : FinalSourceHalfMask)
8357 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8358 M = InputsFixed[1] + SourceOffset;
8359 else if (M == InputsFixed[1] + SourceOffset)
8360 M = (InputsFixed[0] ^ 1) + SourceOffset;
8362 InputsFixed[1] = InputsFixed[0] ^ 1;
8365 // Point everything at the fixed inputs.
8366 for (int &M : HalfMask)
8367 if (M == IncomingInputs[0])
8368 M = InputsFixed[0] + SourceOffset;
8369 else if (M == IncomingInputs[1])
8370 M = InputsFixed[1] + SourceOffset;
8372 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8373 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8376 llvm_unreachable("Unhandled input size!");
8379 // Now hoist the DWord down to the right half.
8380 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8381 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8382 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8383 for (int &M : HalfMask)
8384 for (int Input : IncomingInputs)
8386 M = FreeDWord * 2 + Input % 2;
8388 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8389 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8390 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8391 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8393 // Now enact all the shuffles we've computed to move the inputs into their
8395 if (!isNoopShuffleMask(PSHUFLMask))
8396 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8397 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8398 if (!isNoopShuffleMask(PSHUFHMask))
8399 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8400 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8401 if (!isNoopShuffleMask(PSHUFDMask))
8404 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8405 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8407 // At this point, each half should contain all its inputs, and we can then
8408 // just shuffle them into their final position.
8409 assert(std::count_if(LoMask.begin(), LoMask.end(),
8410 [](int M) { return M >= 4; }) == 0 &&
8411 "Failed to lift all the high half inputs to the low mask!");
8412 assert(std::count_if(HiMask.begin(), HiMask.end(),
8413 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8414 "Failed to lift all the low half inputs to the high mask!");
8416 // Do a half shuffle for the low mask.
8417 if (!isNoopShuffleMask(LoMask))
8418 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8419 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8421 // Do a half shuffle with the high mask after shifting its values down.
8422 for (int &M : HiMask)
8425 if (!isNoopShuffleMask(HiMask))
8426 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8427 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8432 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8433 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8434 SDValue V2, ArrayRef<int> Mask,
8435 SelectionDAG &DAG, bool &V1InUse,
8437 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8443 int Size = Mask.size();
8444 int Scale = 16 / Size;
8445 for (int i = 0; i < 16; ++i) {
8446 if (Mask[i / Scale] == -1) {
8447 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8449 const int ZeroMask = 0x80;
8450 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8452 int V2Idx = Mask[i / Scale] < Size
8454 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8455 if (Zeroable[i / Scale])
8456 V1Idx = V2Idx = ZeroMask;
8457 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8458 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8459 V1InUse |= (ZeroMask != V1Idx);
8460 V2InUse |= (ZeroMask != V2Idx);
8465 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8466 DAG.getBitcast(MVT::v16i8, V1),
8467 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8469 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8470 DAG.getBitcast(MVT::v16i8, V2),
8471 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8473 // If we need shuffled inputs from both, blend the two.
8475 if (V1InUse && V2InUse)
8476 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8478 V = V1InUse ? V1 : V2;
8480 // Cast the result back to the correct type.
8481 return DAG.getBitcast(VT, V);
8484 /// \brief Generic lowering of 8-lane i16 shuffles.
8486 /// This handles both single-input shuffles and combined shuffle/blends with
8487 /// two inputs. The single input shuffles are immediately delegated to
8488 /// a dedicated lowering routine.
8490 /// The blends are lowered in one of three fundamental ways. If there are few
8491 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8492 /// of the input is significantly cheaper when lowered as an interleaving of
8493 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8494 /// halves of the inputs separately (making them have relatively few inputs)
8495 /// and then concatenate them.
8496 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8497 const X86Subtarget *Subtarget,
8498 SelectionDAG &DAG) {
8500 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8501 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8502 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8503 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8504 ArrayRef<int> OrigMask = SVOp->getMask();
8505 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8506 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8507 MutableArrayRef<int> Mask(MaskStorage);
8509 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8511 // Whenever we can lower this as a zext, that instruction is strictly faster
8512 // than any alternative.
8513 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8514 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8517 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8519 auto isV2 = [](int M) { return M >= 8; };
8521 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8523 if (NumV2Inputs == 0) {
8524 // Check for being able to broadcast a single element.
8525 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8526 Mask, Subtarget, DAG))
8529 // Try to use shift instructions.
8531 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8534 // Use dedicated unpack instructions for masks that match their pattern.
8535 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8536 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8537 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8538 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8540 // Try to use byte rotation instructions.
8541 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8542 Mask, Subtarget, DAG))
8545 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8549 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8550 "All single-input shuffles should be canonicalized to be V1-input "
8553 // Try to use shift instructions.
8555 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8558 // There are special ways we can lower some single-element blends.
8559 if (NumV2Inputs == 1)
8560 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8561 Mask, Subtarget, DAG))
8564 // We have different paths for blend lowering, but they all must use the
8565 // *exact* same predicate.
8566 bool IsBlendSupported = Subtarget->hasSSE41();
8567 if (IsBlendSupported)
8568 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8572 if (SDValue Masked =
8573 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8576 // Use dedicated unpack instructions for masks that match their pattern.
8577 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8578 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8579 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8580 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8582 // Try to use byte rotation instructions.
8583 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8584 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8587 if (SDValue BitBlend =
8588 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8591 if (SDValue Unpack =
8592 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8595 // If we can't directly blend but can use PSHUFB, that will be better as it
8596 // can both shuffle and set up the inefficient blend.
8597 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8598 bool V1InUse, V2InUse;
8599 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8603 // We can always bit-blend if we have to so the fallback strategy is to
8604 // decompose into single-input permutes and blends.
8605 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8609 /// \brief Check whether a compaction lowering can be done by dropping even
8610 /// elements and compute how many times even elements must be dropped.
8612 /// This handles shuffles which take every Nth element where N is a power of
8613 /// two. Example shuffle masks:
8615 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8616 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8617 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8618 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8619 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8620 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8622 /// Any of these lanes can of course be undef.
8624 /// This routine only supports N <= 3.
8625 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8628 /// \returns N above, or the number of times even elements must be dropped if
8629 /// there is such a number. Otherwise returns zero.
8630 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8631 // Figure out whether we're looping over two inputs or just one.
8632 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8634 // The modulus for the shuffle vector entries is based on whether this is
8635 // a single input or not.
8636 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8637 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8638 "We should only be called with masks with a power-of-2 size!");
8640 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8642 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8643 // and 2^3 simultaneously. This is because we may have ambiguity with
8644 // partially undef inputs.
8645 bool ViableForN[3] = {true, true, true};
8647 for (int i = 0, e = Mask.size(); i < e; ++i) {
8648 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8653 bool IsAnyViable = false;
8654 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8655 if (ViableForN[j]) {
8658 // The shuffle mask must be equal to (i * 2^N) % M.
8659 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8662 ViableForN[j] = false;
8664 // Early exit if we exhaust the possible powers of two.
8669 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8673 // Return 0 as there is no viable power of two.
8677 /// \brief Generic lowering of v16i8 shuffles.
8679 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8680 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8681 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8682 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8684 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8685 const X86Subtarget *Subtarget,
8686 SelectionDAG &DAG) {
8688 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8689 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8690 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8691 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8692 ArrayRef<int> Mask = SVOp->getMask();
8693 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8695 // Try to use shift instructions.
8697 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8700 // Try to use byte rotation instructions.
8701 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8702 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8705 // Try to use a zext lowering.
8706 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8707 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8711 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8713 // For single-input shuffles, there are some nicer lowering tricks we can use.
8714 if (NumV2Elements == 0) {
8715 // Check for being able to broadcast a single element.
8716 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8717 Mask, Subtarget, DAG))
8720 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8721 // Notably, this handles splat and partial-splat shuffles more efficiently.
8722 // However, it only makes sense if the pre-duplication shuffle simplifies
8723 // things significantly. Currently, this means we need to be able to
8724 // express the pre-duplication shuffle as an i16 shuffle.
8726 // FIXME: We should check for other patterns which can be widened into an
8727 // i16 shuffle as well.
8728 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8729 for (int i = 0; i < 16; i += 2)
8730 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8735 auto tryToWidenViaDuplication = [&]() -> SDValue {
8736 if (!canWidenViaDuplication(Mask))
8738 SmallVector<int, 4> LoInputs;
8739 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8740 [](int M) { return M >= 0 && M < 8; });
8741 std::sort(LoInputs.begin(), LoInputs.end());
8742 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8744 SmallVector<int, 4> HiInputs;
8745 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8746 [](int M) { return M >= 8; });
8747 std::sort(HiInputs.begin(), HiInputs.end());
8748 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8751 bool TargetLo = LoInputs.size() >= HiInputs.size();
8752 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8753 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8755 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8756 SmallDenseMap<int, int, 8> LaneMap;
8757 for (int I : InPlaceInputs) {
8758 PreDupI16Shuffle[I/2] = I/2;
8761 int j = TargetLo ? 0 : 4, je = j + 4;
8762 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8763 // Check if j is already a shuffle of this input. This happens when
8764 // there are two adjacent bytes after we move the low one.
8765 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8766 // If we haven't yet mapped the input, search for a slot into which
8768 while (j < je && PreDupI16Shuffle[j] != -1)
8772 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8775 // Map this input with the i16 shuffle.
8776 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8779 // Update the lane map based on the mapping we ended up with.
8780 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8782 V1 = DAG.getBitcast(
8784 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
8785 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8787 // Unpack the bytes to form the i16s that will be shuffled into place.
8788 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8789 MVT::v16i8, V1, V1);
8791 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8792 for (int i = 0; i < 16; ++i)
8793 if (Mask[i] != -1) {
8794 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8795 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8796 if (PostDupI16Shuffle[i / 2] == -1)
8797 PostDupI16Shuffle[i / 2] = MappedMask;
8799 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8800 "Conflicting entrties in the original shuffle!");
8802 return DAG.getBitcast(
8804 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
8805 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8807 if (SDValue V = tryToWidenViaDuplication())
8811 // Use dedicated unpack instructions for masks that match their pattern.
8812 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8813 0, 16, 1, 17, 2, 18, 3, 19,
8815 4, 20, 5, 21, 6, 22, 7, 23}))
8816 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8817 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8818 8, 24, 9, 25, 10, 26, 11, 27,
8820 12, 28, 13, 29, 14, 30, 15, 31}))
8821 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8823 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8824 // with PSHUFB. It is important to do this before we attempt to generate any
8825 // blends but after all of the single-input lowerings. If the single input
8826 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8827 // want to preserve that and we can DAG combine any longer sequences into
8828 // a PSHUFB in the end. But once we start blending from multiple inputs,
8829 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8830 // and there are *very* few patterns that would actually be faster than the
8831 // PSHUFB approach because of its ability to zero lanes.
8833 // FIXME: The only exceptions to the above are blends which are exact
8834 // interleavings with direct instructions supporting them. We currently don't
8835 // handle those well here.
8836 if (Subtarget->hasSSSE3()) {
8837 bool V1InUse = false;
8838 bool V2InUse = false;
8840 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8841 DAG, V1InUse, V2InUse);
8843 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8844 // do so. This avoids using them to handle blends-with-zero which is
8845 // important as a single pshufb is significantly faster for that.
8846 if (V1InUse && V2InUse) {
8847 if (Subtarget->hasSSE41())
8848 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8849 Mask, Subtarget, DAG))
8852 // We can use an unpack to do the blending rather than an or in some
8853 // cases. Even though the or may be (very minorly) more efficient, we
8854 // preference this lowering because there are common cases where part of
8855 // the complexity of the shuffles goes away when we do the final blend as
8857 // FIXME: It might be worth trying to detect if the unpack-feeding
8858 // shuffles will both be pshufb, in which case we shouldn't bother with
8860 if (SDValue Unpack =
8861 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8868 // There are special ways we can lower some single-element blends.
8869 if (NumV2Elements == 1)
8870 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8871 Mask, Subtarget, DAG))
8874 if (SDValue BitBlend =
8875 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8878 // Check whether a compaction lowering can be done. This handles shuffles
8879 // which take every Nth element for some even N. See the helper function for
8882 // We special case these as they can be particularly efficiently handled with
8883 // the PACKUSB instruction on x86 and they show up in common patterns of
8884 // rearranging bytes to truncate wide elements.
8885 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8886 // NumEvenDrops is the power of two stride of the elements. Another way of
8887 // thinking about it is that we need to drop the even elements this many
8888 // times to get the original input.
8889 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8891 // First we need to zero all the dropped bytes.
8892 assert(NumEvenDrops <= 3 &&
8893 "No support for dropping even elements more than 3 times.");
8894 // We use the mask type to pick which bytes are preserved based on how many
8895 // elements are dropped.
8896 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8897 SDValue ByteClearMask = DAG.getBitcast(
8898 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
8899 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8901 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8903 // Now pack things back together.
8904 V1 = DAG.getBitcast(MVT::v8i16, V1);
8905 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
8906 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8907 for (int i = 1; i < NumEvenDrops; ++i) {
8908 Result = DAG.getBitcast(MVT::v8i16, Result);
8909 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8915 // Handle multi-input cases by blending single-input shuffles.
8916 if (NumV2Elements > 0)
8917 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8920 // The fallback path for single-input shuffles widens this into two v8i16
8921 // vectors with unpacks, shuffles those, and then pulls them back together
8925 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8926 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8927 for (int i = 0; i < 16; ++i)
8929 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8931 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8933 SDValue VLoHalf, VHiHalf;
8934 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8935 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8937 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8938 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8939 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8940 [](int M) { return M >= 0 && M % 2 == 1; })) {
8941 // Use a mask to drop the high bytes.
8942 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
8943 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8944 DAG.getConstant(0x00FF, DL, MVT::v8i16));
8946 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8947 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8949 // Squash the masks to point directly into VLoHalf.
8950 for (int &M : LoBlendMask)
8953 for (int &M : HiBlendMask)
8957 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8958 // VHiHalf so that we can blend them as i16s.
8959 VLoHalf = DAG.getBitcast(
8960 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8961 VHiHalf = DAG.getBitcast(
8962 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8965 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8966 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8968 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8971 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8973 /// This routine breaks down the specific type of 128-bit shuffle and
8974 /// dispatches to the lowering routines accordingly.
8975 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8976 MVT VT, const X86Subtarget *Subtarget,
8977 SelectionDAG &DAG) {
8978 switch (VT.SimpleTy) {
8980 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8982 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8984 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8986 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8988 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8990 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8993 llvm_unreachable("Unimplemented!");
8997 /// \brief Helper function to test whether a shuffle mask could be
8998 /// simplified by widening the elements being shuffled.
9000 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9001 /// leaves it in an unspecified state.
9003 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9004 /// shuffle masks. The latter have the special property of a '-2' representing
9005 /// a zero-ed lane of a vector.
9006 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9007 SmallVectorImpl<int> &WidenedMask) {
9008 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9009 // If both elements are undef, its trivial.
9010 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9011 WidenedMask.push_back(SM_SentinelUndef);
9015 // Check for an undef mask and a mask value properly aligned to fit with
9016 // a pair of values. If we find such a case, use the non-undef mask's value.
9017 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9018 WidenedMask.push_back(Mask[i + 1] / 2);
9021 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9022 WidenedMask.push_back(Mask[i] / 2);
9026 // When zeroing, we need to spread the zeroing across both lanes to widen.
9027 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9028 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9029 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9030 WidenedMask.push_back(SM_SentinelZero);
9036 // Finally check if the two mask values are adjacent and aligned with
9038 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9039 WidenedMask.push_back(Mask[i] / 2);
9043 // Otherwise we can't safely widen the elements used in this shuffle.
9046 assert(WidenedMask.size() == Mask.size() / 2 &&
9047 "Incorrect size of mask after widening the elements!");
9052 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9054 /// This routine just extracts two subvectors, shuffles them independently, and
9055 /// then concatenates them back together. This should work effectively with all
9056 /// AVX vector shuffle types.
9057 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9058 SDValue V2, ArrayRef<int> Mask,
9059 SelectionDAG &DAG) {
9060 assert(VT.getSizeInBits() >= 256 &&
9061 "Only for 256-bit or wider vector shuffles!");
9062 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9063 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9065 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9066 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9068 int NumElements = VT.getVectorNumElements();
9069 int SplitNumElements = NumElements / 2;
9070 MVT ScalarVT = VT.getScalarType();
9071 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9073 // Rather than splitting build-vectors, just build two narrower build
9074 // vectors. This helps shuffling with splats and zeros.
9075 auto SplitVector = [&](SDValue V) {
9076 while (V.getOpcode() == ISD::BITCAST)
9077 V = V->getOperand(0);
9079 MVT OrigVT = V.getSimpleValueType();
9080 int OrigNumElements = OrigVT.getVectorNumElements();
9081 int OrigSplitNumElements = OrigNumElements / 2;
9082 MVT OrigScalarVT = OrigVT.getScalarType();
9083 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9087 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9089 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9090 DAG.getIntPtrConstant(0, DL));
9091 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9092 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9095 SmallVector<SDValue, 16> LoOps, HiOps;
9096 for (int i = 0; i < OrigSplitNumElements; ++i) {
9097 LoOps.push_back(BV->getOperand(i));
9098 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9100 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9101 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9103 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9104 DAG.getBitcast(SplitVT, HiV));
9107 SDValue LoV1, HiV1, LoV2, HiV2;
9108 std::tie(LoV1, HiV1) = SplitVector(V1);
9109 std::tie(LoV2, HiV2) = SplitVector(V2);
9111 // Now create two 4-way blends of these half-width vectors.
9112 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9113 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9114 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9115 for (int i = 0; i < SplitNumElements; ++i) {
9116 int M = HalfMask[i];
9117 if (M >= NumElements) {
9118 if (M >= NumElements + SplitNumElements)
9122 V2BlendMask.push_back(M - NumElements);
9123 V1BlendMask.push_back(-1);
9124 BlendMask.push_back(SplitNumElements + i);
9125 } else if (M >= 0) {
9126 if (M >= SplitNumElements)
9130 V2BlendMask.push_back(-1);
9131 V1BlendMask.push_back(M);
9132 BlendMask.push_back(i);
9134 V2BlendMask.push_back(-1);
9135 V1BlendMask.push_back(-1);
9136 BlendMask.push_back(-1);
9140 // Because the lowering happens after all combining takes place, we need to
9141 // manually combine these blend masks as much as possible so that we create
9142 // a minimal number of high-level vector shuffle nodes.
9144 // First try just blending the halves of V1 or V2.
9145 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9146 return DAG.getUNDEF(SplitVT);
9147 if (!UseLoV2 && !UseHiV2)
9148 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9149 if (!UseLoV1 && !UseHiV1)
9150 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9152 SDValue V1Blend, V2Blend;
9153 if (UseLoV1 && UseHiV1) {
9155 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9157 // We only use half of V1 so map the usage down into the final blend mask.
9158 V1Blend = UseLoV1 ? LoV1 : HiV1;
9159 for (int i = 0; i < SplitNumElements; ++i)
9160 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9161 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9163 if (UseLoV2 && UseHiV2) {
9165 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9167 // We only use half of V2 so map the usage down into the final blend mask.
9168 V2Blend = UseLoV2 ? LoV2 : HiV2;
9169 for (int i = 0; i < SplitNumElements; ++i)
9170 if (BlendMask[i] >= SplitNumElements)
9171 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9173 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9175 SDValue Lo = HalfBlend(LoMask);
9176 SDValue Hi = HalfBlend(HiMask);
9177 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9180 /// \brief Either split a vector in halves or decompose the shuffles and the
9183 /// This is provided as a good fallback for many lowerings of non-single-input
9184 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9185 /// between splitting the shuffle into 128-bit components and stitching those
9186 /// back together vs. extracting the single-input shuffles and blending those
9188 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9189 SDValue V2, ArrayRef<int> Mask,
9190 SelectionDAG &DAG) {
9191 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9192 "lower single-input shuffles as it "
9193 "could then recurse on itself.");
9194 int Size = Mask.size();
9196 // If this can be modeled as a broadcast of two elements followed by a blend,
9197 // prefer that lowering. This is especially important because broadcasts can
9198 // often fold with memory operands.
9199 auto DoBothBroadcast = [&] {
9200 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9203 if (V2BroadcastIdx == -1)
9204 V2BroadcastIdx = M - Size;
9205 else if (M - Size != V2BroadcastIdx)
9207 } else if (M >= 0) {
9208 if (V1BroadcastIdx == -1)
9210 else if (M != V1BroadcastIdx)
9215 if (DoBothBroadcast())
9216 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9219 // If the inputs all stem from a single 128-bit lane of each input, then we
9220 // split them rather than blending because the split will decompose to
9221 // unusually few instructions.
9222 int LaneCount = VT.getSizeInBits() / 128;
9223 int LaneSize = Size / LaneCount;
9224 SmallBitVector LaneInputs[2];
9225 LaneInputs[0].resize(LaneCount, false);
9226 LaneInputs[1].resize(LaneCount, false);
9227 for (int i = 0; i < Size; ++i)
9229 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9230 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9231 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9233 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9234 // that the decomposed single-input shuffles don't end up here.
9235 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9238 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9239 /// a permutation and blend of those lanes.
9241 /// This essentially blends the out-of-lane inputs to each lane into the lane
9242 /// from a permuted copy of the vector. This lowering strategy results in four
9243 /// instructions in the worst case for a single-input cross lane shuffle which
9244 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9245 /// of. Special cases for each particular shuffle pattern should be handled
9246 /// prior to trying this lowering.
9247 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9248 SDValue V1, SDValue V2,
9250 SelectionDAG &DAG) {
9251 // FIXME: This should probably be generalized for 512-bit vectors as well.
9252 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9253 int LaneSize = Mask.size() / 2;
9255 // If there are only inputs from one 128-bit lane, splitting will in fact be
9256 // less expensive. The flags track whether the given lane contains an element
9257 // that crosses to another lane.
9258 bool LaneCrossing[2] = {false, false};
9259 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9260 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9261 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9262 if (!LaneCrossing[0] || !LaneCrossing[1])
9263 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9265 if (isSingleInputShuffleMask(Mask)) {
9266 SmallVector<int, 32> FlippedBlendMask;
9267 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9268 FlippedBlendMask.push_back(
9269 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9271 : Mask[i] % LaneSize +
9272 (i / LaneSize) * LaneSize + Size));
9274 // Flip the vector, and blend the results which should now be in-lane. The
9275 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9276 // 5 for the high source. The value 3 selects the high half of source 2 and
9277 // the value 2 selects the low half of source 2. We only use source 2 to
9278 // allow folding it into a memory operand.
9279 unsigned PERMMask = 3 | 2 << 4;
9280 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9281 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9282 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9285 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9286 // will be handled by the above logic and a blend of the results, much like
9287 // other patterns in AVX.
9288 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9291 /// \brief Handle lowering 2-lane 128-bit shuffles.
9292 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9293 SDValue V2, ArrayRef<int> Mask,
9294 const X86Subtarget *Subtarget,
9295 SelectionDAG &DAG) {
9296 // TODO: If minimizing size and one of the inputs is a zero vector and the
9297 // the zero vector has only one use, we could use a VPERM2X128 to save the
9298 // instruction bytes needed to explicitly generate the zero vector.
9300 // Blends are faster and handle all the non-lane-crossing cases.
9301 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9305 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9306 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9308 // If either input operand is a zero vector, use VPERM2X128 because its mask
9309 // allows us to replace the zero input with an implicit zero.
9310 if (!IsV1Zero && !IsV2Zero) {
9311 // Check for patterns which can be matched with a single insert of a 128-bit
9313 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9314 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9315 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9316 VT.getVectorNumElements() / 2);
9317 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9318 DAG.getIntPtrConstant(0, DL));
9319 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9320 OnlyUsesV1 ? V1 : V2,
9321 DAG.getIntPtrConstant(0, DL));
9322 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9326 // Otherwise form a 128-bit permutation. After accounting for undefs,
9327 // convert the 64-bit shuffle mask selection values into 128-bit
9328 // selection bits by dividing the indexes by 2 and shifting into positions
9329 // defined by a vperm2*128 instruction's immediate control byte.
9331 // The immediate permute control byte looks like this:
9332 // [1:0] - select 128 bits from sources for low half of destination
9334 // [3] - zero low half of destination
9335 // [5:4] - select 128 bits from sources for high half of destination
9337 // [7] - zero high half of destination
9339 int MaskLO = Mask[0];
9340 if (MaskLO == SM_SentinelUndef)
9341 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9343 int MaskHI = Mask[2];
9344 if (MaskHI == SM_SentinelUndef)
9345 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9347 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9349 // If either input is a zero vector, replace it with an undef input.
9350 // Shuffle mask values < 4 are selecting elements of V1.
9351 // Shuffle mask values >= 4 are selecting elements of V2.
9352 // Adjust each half of the permute mask by clearing the half that was
9353 // selecting the zero vector and setting the zero mask bit.
9355 V1 = DAG.getUNDEF(VT);
9357 PermMask = (PermMask & 0xf0) | 0x08;
9359 PermMask = (PermMask & 0x0f) | 0x80;
9362 V2 = DAG.getUNDEF(VT);
9364 PermMask = (PermMask & 0xf0) | 0x08;
9366 PermMask = (PermMask & 0x0f) | 0x80;
9369 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9370 DAG.getConstant(PermMask, DL, MVT::i8));
9373 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9374 /// shuffling each lane.
9376 /// This will only succeed when the result of fixing the 128-bit lanes results
9377 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9378 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9379 /// the lane crosses early and then use simpler shuffles within each lane.
9381 /// FIXME: It might be worthwhile at some point to support this without
9382 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9383 /// in x86 only floating point has interesting non-repeating shuffles, and even
9384 /// those are still *marginally* more expensive.
9385 static SDValue lowerVectorShuffleByMerging128BitLanes(
9386 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9387 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9388 assert(!isSingleInputShuffleMask(Mask) &&
9389 "This is only useful with multiple inputs.");
9391 int Size = Mask.size();
9392 int LaneSize = 128 / VT.getScalarSizeInBits();
9393 int NumLanes = Size / LaneSize;
9394 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9396 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9397 // check whether the in-128-bit lane shuffles share a repeating pattern.
9398 SmallVector<int, 4> Lanes;
9399 Lanes.resize(NumLanes, -1);
9400 SmallVector<int, 4> InLaneMask;
9401 InLaneMask.resize(LaneSize, -1);
9402 for (int i = 0; i < Size; ++i) {
9406 int j = i / LaneSize;
9409 // First entry we've seen for this lane.
9410 Lanes[j] = Mask[i] / LaneSize;
9411 } else if (Lanes[j] != Mask[i] / LaneSize) {
9412 // This doesn't match the lane selected previously!
9416 // Check that within each lane we have a consistent shuffle mask.
9417 int k = i % LaneSize;
9418 if (InLaneMask[k] < 0) {
9419 InLaneMask[k] = Mask[i] % LaneSize;
9420 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9421 // This doesn't fit a repeating in-lane mask.
9426 // First shuffle the lanes into place.
9427 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9428 VT.getSizeInBits() / 64);
9429 SmallVector<int, 8> LaneMask;
9430 LaneMask.resize(NumLanes * 2, -1);
9431 for (int i = 0; i < NumLanes; ++i)
9432 if (Lanes[i] >= 0) {
9433 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9434 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9437 V1 = DAG.getBitcast(LaneVT, V1);
9438 V2 = DAG.getBitcast(LaneVT, V2);
9439 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9441 // Cast it back to the type we actually want.
9442 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9444 // Now do a simple shuffle that isn't lane crossing.
9445 SmallVector<int, 8> NewMask;
9446 NewMask.resize(Size, -1);
9447 for (int i = 0; i < Size; ++i)
9449 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9450 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9451 "Must not introduce lane crosses at this point!");
9453 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9456 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9459 /// This returns true if the elements from a particular input are already in the
9460 /// slot required by the given mask and require no permutation.
9461 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9462 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9463 int Size = Mask.size();
9464 for (int i = 0; i < Size; ++i)
9465 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9471 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9473 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9474 /// isn't available.
9475 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9476 const X86Subtarget *Subtarget,
9477 SelectionDAG &DAG) {
9479 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9480 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9481 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9482 ArrayRef<int> Mask = SVOp->getMask();
9483 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9485 SmallVector<int, 4> WidenedMask;
9486 if (canWidenShuffleElements(Mask, WidenedMask))
9487 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9490 if (isSingleInputShuffleMask(Mask)) {
9491 // Check for being able to broadcast a single element.
9492 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9493 Mask, Subtarget, DAG))
9496 // Use low duplicate instructions for masks that match their pattern.
9497 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9498 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9500 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9501 // Non-half-crossing single input shuffles can be lowerid with an
9502 // interleaved permutation.
9503 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9504 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9505 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9506 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9509 // With AVX2 we have direct support for this permutation.
9510 if (Subtarget->hasAVX2())
9511 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9512 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9514 // Otherwise, fall back.
9515 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9519 // X86 has dedicated unpack instructions that can handle specific blend
9520 // operations: UNPCKH and UNPCKL.
9521 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9522 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9523 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9524 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9525 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9526 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9527 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9528 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9530 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9534 // Check if the blend happens to exactly fit that of SHUFPD.
9535 if ((Mask[0] == -1 || Mask[0] < 2) &&
9536 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9537 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9538 (Mask[3] == -1 || Mask[3] >= 6)) {
9539 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9540 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9541 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9542 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9544 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9545 (Mask[1] == -1 || Mask[1] < 2) &&
9546 (Mask[2] == -1 || Mask[2] >= 6) &&
9547 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9548 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9549 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9550 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9551 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9554 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9555 // shuffle. However, if we have AVX2 and either inputs are already in place,
9556 // we will be able to shuffle even across lanes the other input in a single
9557 // instruction so skip this pattern.
9558 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9559 isShuffleMaskInputInPlace(1, Mask))))
9560 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9561 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9564 // If we have AVX2 then we always want to lower with a blend because an v4 we
9565 // can fully permute the elements.
9566 if (Subtarget->hasAVX2())
9567 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9570 // Otherwise fall back on generic lowering.
9571 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9574 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9576 /// This routine is only called when we have AVX2 and thus a reasonable
9577 /// instruction set for v4i64 shuffling..
9578 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9579 const X86Subtarget *Subtarget,
9580 SelectionDAG &DAG) {
9582 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9583 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9584 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9585 ArrayRef<int> Mask = SVOp->getMask();
9586 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9587 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9589 SmallVector<int, 4> WidenedMask;
9590 if (canWidenShuffleElements(Mask, WidenedMask))
9591 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9594 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9598 // Check for being able to broadcast a single element.
9599 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9600 Mask, Subtarget, DAG))
9603 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9604 // use lower latency instructions that will operate on both 128-bit lanes.
9605 SmallVector<int, 2> RepeatedMask;
9606 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9607 if (isSingleInputShuffleMask(Mask)) {
9608 int PSHUFDMask[] = {-1, -1, -1, -1};
9609 for (int i = 0; i < 2; ++i)
9610 if (RepeatedMask[i] >= 0) {
9611 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9612 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9614 return DAG.getBitcast(
9616 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9617 DAG.getBitcast(MVT::v8i32, V1),
9618 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9622 // AVX2 provides a direct instruction for permuting a single input across
9624 if (isSingleInputShuffleMask(Mask))
9625 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9626 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9628 // Try to use shift instructions.
9630 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9633 // Use dedicated unpack instructions for masks that match their pattern.
9634 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9635 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9636 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9637 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9638 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9639 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9640 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9641 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9643 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9644 // shuffle. However, if we have AVX2 and either inputs are already in place,
9645 // we will be able to shuffle even across lanes the other input in a single
9646 // instruction so skip this pattern.
9647 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9648 isShuffleMaskInputInPlace(1, Mask))))
9649 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9650 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9653 // Otherwise fall back on generic blend lowering.
9654 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9658 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9660 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9661 /// isn't available.
9662 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9663 const X86Subtarget *Subtarget,
9664 SelectionDAG &DAG) {
9666 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9667 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9668 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9669 ArrayRef<int> Mask = SVOp->getMask();
9670 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9672 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9676 // Check for being able to broadcast a single element.
9677 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9678 Mask, Subtarget, DAG))
9681 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9682 // options to efficiently lower the shuffle.
9683 SmallVector<int, 4> RepeatedMask;
9684 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9685 assert(RepeatedMask.size() == 4 &&
9686 "Repeated masks must be half the mask width!");
9688 // Use even/odd duplicate instructions for masks that match their pattern.
9689 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9690 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9691 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9692 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9694 if (isSingleInputShuffleMask(Mask))
9695 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9696 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9698 // Use dedicated unpack instructions for masks that match their pattern.
9699 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9700 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9701 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9702 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9703 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9704 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9705 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9706 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9708 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9709 // have already handled any direct blends. We also need to squash the
9710 // repeated mask into a simulated v4f32 mask.
9711 for (int i = 0; i < 4; ++i)
9712 if (RepeatedMask[i] >= 8)
9713 RepeatedMask[i] -= 4;
9714 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9717 // If we have a single input shuffle with different shuffle patterns in the
9718 // two 128-bit lanes use the variable mask to VPERMILPS.
9719 if (isSingleInputShuffleMask(Mask)) {
9720 SDValue VPermMask[8];
9721 for (int i = 0; i < 8; ++i)
9722 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9723 : DAG.getConstant(Mask[i], DL, MVT::i32);
9724 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9726 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9727 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9729 if (Subtarget->hasAVX2())
9731 X86ISD::VPERMV, DL, MVT::v8f32,
9732 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
9733 MVT::v8i32, VPermMask)),
9736 // Otherwise, fall back.
9737 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9741 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9743 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9744 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9747 // If we have AVX2 then we always want to lower with a blend because at v8 we
9748 // can fully permute the elements.
9749 if (Subtarget->hasAVX2())
9750 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9753 // Otherwise fall back on generic lowering.
9754 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9757 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9759 /// This routine is only called when we have AVX2 and thus a reasonable
9760 /// instruction set for v8i32 shuffling..
9761 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9762 const X86Subtarget *Subtarget,
9763 SelectionDAG &DAG) {
9765 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9766 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9767 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9768 ArrayRef<int> Mask = SVOp->getMask();
9769 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9770 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9772 // Whenever we can lower this as a zext, that instruction is strictly faster
9773 // than any alternative. It also allows us to fold memory operands into the
9774 // shuffle in many cases.
9775 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9776 Mask, Subtarget, DAG))
9779 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9783 // Check for being able to broadcast a single element.
9784 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9785 Mask, Subtarget, DAG))
9788 // If the shuffle mask is repeated in each 128-bit lane we can use more
9789 // efficient instructions that mirror the shuffles across the two 128-bit
9791 SmallVector<int, 4> RepeatedMask;
9792 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9793 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9794 if (isSingleInputShuffleMask(Mask))
9795 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9796 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9798 // Use dedicated unpack instructions for masks that match their pattern.
9799 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9800 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9801 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9802 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9803 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9804 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9805 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9806 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9809 // Try to use shift instructions.
9811 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9814 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9815 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9818 // If the shuffle patterns aren't repeated but it is a single input, directly
9819 // generate a cross-lane VPERMD instruction.
9820 if (isSingleInputShuffleMask(Mask)) {
9821 SDValue VPermMask[8];
9822 for (int i = 0; i < 8; ++i)
9823 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9824 : DAG.getConstant(Mask[i], DL, MVT::i32);
9826 X86ISD::VPERMV, DL, MVT::v8i32,
9827 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9830 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9832 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9833 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9836 // Otherwise fall back on generic blend lowering.
9837 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9841 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9843 /// This routine is only called when we have AVX2 and thus a reasonable
9844 /// instruction set for v16i16 shuffling..
9845 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9846 const X86Subtarget *Subtarget,
9847 SelectionDAG &DAG) {
9849 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9850 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9851 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9852 ArrayRef<int> Mask = SVOp->getMask();
9853 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9854 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9856 // Whenever we can lower this as a zext, that instruction is strictly faster
9857 // than any alternative. It also allows us to fold memory operands into the
9858 // shuffle in many cases.
9859 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9860 Mask, Subtarget, DAG))
9863 // Check for being able to broadcast a single element.
9864 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9865 Mask, Subtarget, DAG))
9868 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9872 // Use dedicated unpack instructions for masks that match their pattern.
9873 if (isShuffleEquivalent(V1, V2, Mask,
9874 {// First 128-bit lane:
9875 0, 16, 1, 17, 2, 18, 3, 19,
9876 // Second 128-bit lane:
9877 8, 24, 9, 25, 10, 26, 11, 27}))
9878 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9879 if (isShuffleEquivalent(V1, V2, Mask,
9880 {// First 128-bit lane:
9881 4, 20, 5, 21, 6, 22, 7, 23,
9882 // Second 128-bit lane:
9883 12, 28, 13, 29, 14, 30, 15, 31}))
9884 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9886 // Try to use shift instructions.
9888 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9891 // Try to use byte rotation instructions.
9892 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9893 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9896 if (isSingleInputShuffleMask(Mask)) {
9897 // There are no generalized cross-lane shuffle operations available on i16
9899 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9900 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9903 SmallVector<int, 8> RepeatedMask;
9904 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
9905 // As this is a single-input shuffle, the repeated mask should be
9906 // a strictly valid v8i16 mask that we can pass through to the v8i16
9907 // lowering to handle even the v16 case.
9908 return lowerV8I16GeneralSingleInputVectorShuffle(
9909 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
9912 SDValue PSHUFBMask[32];
9913 for (int i = 0; i < 16; ++i) {
9914 if (Mask[i] == -1) {
9915 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9919 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9920 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9921 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
9922 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
9924 return DAG.getBitcast(MVT::v16i16,
9925 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
9926 DAG.getBitcast(MVT::v32i8, V1),
9927 DAG.getNode(ISD::BUILD_VECTOR, DL,
9928 MVT::v32i8, PSHUFBMask)));
9931 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9933 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9934 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9937 // Otherwise fall back on generic lowering.
9938 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9941 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9943 /// This routine is only called when we have AVX2 and thus a reasonable
9944 /// instruction set for v32i8 shuffling..
9945 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9946 const X86Subtarget *Subtarget,
9947 SelectionDAG &DAG) {
9949 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9950 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9951 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9952 ArrayRef<int> Mask = SVOp->getMask();
9953 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9954 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9956 // Whenever we can lower this as a zext, that instruction is strictly faster
9957 // than any alternative. It also allows us to fold memory operands into the
9958 // shuffle in many cases.
9959 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9960 Mask, Subtarget, DAG))
9963 // Check for being able to broadcast a single element.
9964 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9965 Mask, Subtarget, DAG))
9968 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9972 // Use dedicated unpack instructions for masks that match their pattern.
9973 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9975 if (isShuffleEquivalent(
9977 {// First 128-bit lane:
9978 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9979 // Second 128-bit lane:
9980 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9981 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9982 if (isShuffleEquivalent(
9984 {// First 128-bit lane:
9985 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9986 // Second 128-bit lane:
9987 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9988 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9990 // Try to use shift instructions.
9992 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9995 // Try to use byte rotation instructions.
9996 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9997 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10000 if (isSingleInputShuffleMask(Mask)) {
10001 // There are no generalized cross-lane shuffle operations available on i8
10003 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10004 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10007 SDValue PSHUFBMask[32];
10008 for (int i = 0; i < 32; ++i)
10011 ? DAG.getUNDEF(MVT::i8)
10012 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10015 return DAG.getNode(
10016 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10017 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10020 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10022 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10023 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10026 // Otherwise fall back on generic lowering.
10027 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10030 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10032 /// This routine either breaks down the specific type of a 256-bit x86 vector
10033 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10034 /// together based on the available instructions.
10035 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10036 MVT VT, const X86Subtarget *Subtarget,
10037 SelectionDAG &DAG) {
10039 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10040 ArrayRef<int> Mask = SVOp->getMask();
10042 // If we have a single input to the zero element, insert that into V1 if we
10043 // can do so cheaply.
10044 int NumElts = VT.getVectorNumElements();
10045 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10046 return M >= NumElts;
10049 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10050 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10051 DL, VT, V1, V2, Mask, Subtarget, DAG))
10054 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10055 // check for those subtargets here and avoid much of the subtarget querying in
10056 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10057 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10058 // floating point types there eventually, just immediately cast everything to
10059 // a float and operate entirely in that domain.
10060 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10061 int ElementBits = VT.getScalarSizeInBits();
10062 if (ElementBits < 32)
10063 // No floating point type available, decompose into 128-bit vectors.
10064 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10066 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10067 VT.getVectorNumElements());
10068 V1 = DAG.getBitcast(FpVT, V1);
10069 V2 = DAG.getBitcast(FpVT, V2);
10070 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10073 switch (VT.SimpleTy) {
10075 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10077 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10079 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10081 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10083 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10085 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10088 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10092 static SDValue lowerVectorShuffleWithVALIGN(SDLoc DL, MVT VT,
10093 ArrayRef<int> Mask, SDValue V1,
10094 SDValue V2, SelectionDAG &DAG) {
10096 assert(VT.getScalarSizeInBits() >= 32 && "Unexpected data type for VALIGN");
10097 // VALIGN pattern 2, 3, 4, 5, .. (sequential, shifted right)
10099 for (int i = 0; i < (signed)VT.getVectorNumElements(); ++i) {
10104 if (AlignVal == -1)
10105 AlignVal = Mask[i] - i;
10106 else if (Mask[i] - i != AlignVal)
10109 return DAG.getNode(X86ISD::VALIGN, DL, VT, V1, V2,
10110 DAG.getConstant(AlignVal, DL, MVT::i8));
10113 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10114 ArrayRef<int> Mask, SDValue V1,
10115 SDValue V2, SelectionDAG &DAG) {
10117 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10119 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10120 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10122 SmallVector<SDValue, 32> VPermMask;
10123 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i)
10124 VPermMask.push_back(Mask[i] < 0 ? DAG.getUNDEF(MaskEltVT) :
10125 DAG.getConstant(Mask[i], DL,MaskEltVT));
10126 SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecVT,
10128 if (isSingleInputShuffleMask(Mask))
10129 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10131 return DAG.getNode(X86ISD::VPERMV3, DL, VT, MaskNode, V1, V2);
10135 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10136 static SDValue lowerV8X64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10137 const X86Subtarget *Subtarget,
10138 SelectionDAG &DAG) {
10140 MVT VT = Op.getSimpleValueType();
10141 assert((V1.getSimpleValueType() == MVT::v8f64 ||
10142 V1.getSimpleValueType() == MVT::v8i64) && "Bad operand type!");
10143 assert((V2.getSimpleValueType() == MVT::v8f64 ||
10144 V2.getSimpleValueType() == MVT::v8i64) && "Bad operand type!");
10145 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10146 ArrayRef<int> Mask = SVOp->getMask();
10147 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10149 // X86 has dedicated unpack instructions that can handle specific blend
10150 // operations: UNPCKH and UNPCKL.
10151 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10152 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
10153 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10154 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
10156 if (SDValue Op = lowerVectorShuffleWithVALIGN(DL, VT, Mask, V1, V2, DAG))
10159 // VSHUFPD instruction - mask 0/1, 8/9, 2/3, 10/11, 4/5, 12/13, 6/7, 14/15
10160 bool ShufpdMask = true;
10161 unsigned Immediate = 0;
10162 for (int i = 0; i < 8; ++i) {
10165 int Val = (i & 6) + 8 * (i & 1);
10166 if (Mask[i] < Val || Mask[i] > Val+1) {
10167 ShufpdMask = false;
10170 Immediate |= (Mask[i]%2) << i;
10173 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10174 DAG.getConstant(Immediate, DL, MVT::i8));
10176 // PERMILPD instruction - mask 0/1, 0/1, 2/3, 2/3, 4/5, 4/5, 6/7, 6/7
10177 if (isSingleInputShuffleMask(Mask)) {
10178 bool PermilMask = true;
10179 unsigned Immediate = 0;
10180 for (int i = 0; i < 8; ++i) {
10184 if (Mask[i] < Val || Mask[i] > Val+1) {
10185 PermilMask = false;
10188 Immediate |= (Mask[i]%2) << i;
10191 return DAG.getNode(X86ISD::VPERMILPI, DL, VT, V1,
10192 DAG.getConstant(Immediate, DL, MVT::i8));
10194 SmallVector<int, 4> RepeatedMask;
10195 if (is256BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask)) {
10196 unsigned Immediate = 0;
10197 for (int i = 0; i < 4; ++i)
10198 if (RepeatedMask[i] > 0)
10199 Immediate |= (RepeatedMask[i] & 3) << (i*2);
10200 return DAG.getNode(X86ISD::VPERMI, DL, VT, V1,
10201 DAG.getConstant(Immediate, DL, MVT::i8));
10204 return lowerVectorShuffleWithPERMV(DL, VT, Mask, V1, V2, DAG);
10207 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10208 static SDValue lowerV16X32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10209 const X86Subtarget *Subtarget,
10210 SelectionDAG &DAG) {
10211 MVT VT = Op.getSimpleValueType();
10213 assert((V1.getSimpleValueType() == MVT::v16i32 ||
10214 V1.getSimpleValueType() == MVT::v16f32) && "Bad operand type!");
10215 assert((V2.getSimpleValueType() == MVT::v16i32 ||
10216 V2.getSimpleValueType() == MVT::v16f32) && "Bad operand type!");
10217 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10218 ArrayRef<int> Mask = SVOp->getMask();
10219 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10221 // Use dedicated unpack instructions for masks that match their pattern.
10222 if (isShuffleEquivalent(V1, V2, Mask,
10223 {// First 128-bit lane.
10224 0, 16, 1, 17, 4, 20, 5, 21,
10225 // Second 128-bit lane.
10226 8, 24, 9, 25, 12, 28, 13, 29}))
10227 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
10228 if (isShuffleEquivalent(V1, V2, Mask,
10229 {// First 128-bit lane.
10230 2, 18, 3, 19, 6, 22, 7, 23,
10231 // Second 128-bit lane.
10232 10, 26, 11, 27, 14, 30, 15, 31}))
10233 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
10235 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10,
10237 return DAG.getNode(X86ISD::MOVSLDUP, DL, VT, V1);
10238 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11,
10240 return DAG.getNode(X86ISD::MOVSHDUP, DL, VT, V1);
10242 SmallVector<int, 4> RepeatedMask;
10243 if (is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask)) {
10244 if (isSingleInputShuffleMask(Mask)) {
10245 unsigned Opc = VT.isInteger() ? X86ISD::PSHUFD : X86ISD::VPERMILPI;
10246 return DAG.getNode(Opc, DL, VT, V1,
10247 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10250 for (int i = 0; i < 4; ++i) {
10251 if (RepeatedMask[i] >= 16)
10252 RepeatedMask[i] -= 12;
10254 return lowerVectorShuffleWithSHUFPS(DL, VT, RepeatedMask, V1, V2, DAG);
10257 if (SDValue Op = lowerVectorShuffleWithVALIGN(DL, VT, Mask, V1, V2, DAG))
10260 return lowerVectorShuffleWithPERMV(DL, VT, Mask, V1, V2, DAG);
10263 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10264 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10265 const X86Subtarget *Subtarget,
10266 SelectionDAG &DAG) {
10268 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10269 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10270 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10271 ArrayRef<int> Mask = SVOp->getMask();
10272 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10273 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10275 // FIXME: Implement direct support for this type!
10276 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10279 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10280 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10281 const X86Subtarget *Subtarget,
10282 SelectionDAG &DAG) {
10284 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10285 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10286 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10287 ArrayRef<int> Mask = SVOp->getMask();
10288 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10289 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10291 // FIXME: Implement direct support for this type!
10292 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10295 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10297 /// This routine either breaks down the specific type of a 512-bit x86 vector
10298 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10299 /// together based on the available instructions.
10300 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10301 MVT VT, const X86Subtarget *Subtarget,
10302 SelectionDAG &DAG) {
10304 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10305 ArrayRef<int> Mask = SVOp->getMask();
10306 assert(Subtarget->hasAVX512() &&
10307 "Cannot lower 512-bit vectors w/ basic ISA!");
10309 // Check for being able to broadcast a single element.
10310 if (SDValue Broadcast =
10311 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10314 // Dispatch to each element type for lowering. If we don't have supprot for
10315 // specific element type shuffles at 512 bits, immediately split them and
10316 // lower them. Each lowering routine of a given type is allowed to assume that
10317 // the requisite ISA extensions for that element type are available.
10318 switch (VT.SimpleTy) {
10321 return lowerV8X64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10324 return lowerV16X32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10326 if (Subtarget->hasBWI())
10327 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10330 if (Subtarget->hasBWI())
10331 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10335 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10338 // Otherwise fall back on splitting.
10339 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10342 /// \brief Top-level lowering for x86 vector shuffles.
10344 /// This handles decomposition, canonicalization, and lowering of all x86
10345 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10346 /// above in helper routines. The canonicalization attempts to widen shuffles
10347 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10348 /// s.t. only one of the two inputs needs to be tested, etc.
10349 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10350 SelectionDAG &DAG) {
10351 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10352 ArrayRef<int> Mask = SVOp->getMask();
10353 SDValue V1 = Op.getOperand(0);
10354 SDValue V2 = Op.getOperand(1);
10355 MVT VT = Op.getSimpleValueType();
10356 int NumElements = VT.getVectorNumElements();
10359 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10361 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10362 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10363 if (V1IsUndef && V2IsUndef)
10364 return DAG.getUNDEF(VT);
10366 // When we create a shuffle node we put the UNDEF node to second operand,
10367 // but in some cases the first operand may be transformed to UNDEF.
10368 // In this case we should just commute the node.
10370 return DAG.getCommutedVectorShuffle(*SVOp);
10372 // Check for non-undef masks pointing at an undef vector and make the masks
10373 // undef as well. This makes it easier to match the shuffle based solely on
10377 if (M >= NumElements) {
10378 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10379 for (int &M : NewMask)
10380 if (M >= NumElements)
10382 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10385 // We actually see shuffles that are entirely re-arrangements of a set of
10386 // zero inputs. This mostly happens while decomposing complex shuffles into
10387 // simple ones. Directly lower these as a buildvector of zeros.
10388 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10389 if (Zeroable.all())
10390 return getZeroVector(VT, Subtarget, DAG, dl);
10392 // Try to collapse shuffles into using a vector type with fewer elements but
10393 // wider element types. We cap this to not form integers or floating point
10394 // elements wider than 64 bits, but it might be interesting to form i128
10395 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10396 SmallVector<int, 16> WidenedMask;
10397 if (VT.getScalarSizeInBits() < 64 &&
10398 canWidenShuffleElements(Mask, WidenedMask)) {
10399 MVT NewEltVT = VT.isFloatingPoint()
10400 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10401 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10402 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10403 // Make sure that the new vector type is legal. For example, v2f64 isn't
10405 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10406 V1 = DAG.getBitcast(NewVT, V1);
10407 V2 = DAG.getBitcast(NewVT, V2);
10408 return DAG.getBitcast(
10409 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10413 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10414 for (int M : SVOp->getMask())
10416 ++NumUndefElements;
10417 else if (M < NumElements)
10422 // Commute the shuffle as needed such that more elements come from V1 than
10423 // V2. This allows us to match the shuffle pattern strictly on how many
10424 // elements come from V1 without handling the symmetric cases.
10425 if (NumV2Elements > NumV1Elements)
10426 return DAG.getCommutedVectorShuffle(*SVOp);
10428 // When the number of V1 and V2 elements are the same, try to minimize the
10429 // number of uses of V2 in the low half of the vector. When that is tied,
10430 // ensure that the sum of indices for V1 is equal to or lower than the sum
10431 // indices for V2. When those are equal, try to ensure that the number of odd
10432 // indices for V1 is lower than the number of odd indices for V2.
10433 if (NumV1Elements == NumV2Elements) {
10434 int LowV1Elements = 0, LowV2Elements = 0;
10435 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10436 if (M >= NumElements)
10440 if (LowV2Elements > LowV1Elements) {
10441 return DAG.getCommutedVectorShuffle(*SVOp);
10442 } else if (LowV2Elements == LowV1Elements) {
10443 int SumV1Indices = 0, SumV2Indices = 0;
10444 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10445 if (SVOp->getMask()[i] >= NumElements)
10447 else if (SVOp->getMask()[i] >= 0)
10449 if (SumV2Indices < SumV1Indices) {
10450 return DAG.getCommutedVectorShuffle(*SVOp);
10451 } else if (SumV2Indices == SumV1Indices) {
10452 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10453 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10454 if (SVOp->getMask()[i] >= NumElements)
10455 NumV2OddIndices += i % 2;
10456 else if (SVOp->getMask()[i] >= 0)
10457 NumV1OddIndices += i % 2;
10458 if (NumV2OddIndices < NumV1OddIndices)
10459 return DAG.getCommutedVectorShuffle(*SVOp);
10464 // For each vector width, delegate to a specialized lowering routine.
10465 if (VT.getSizeInBits() == 128)
10466 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10468 if (VT.getSizeInBits() == 256)
10469 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10471 // Force AVX-512 vectors to be scalarized for now.
10472 // FIXME: Implement AVX-512 support!
10473 if (VT.getSizeInBits() == 512)
10474 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10476 llvm_unreachable("Unimplemented!");
10479 // This function assumes its argument is a BUILD_VECTOR of constants or
10480 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10482 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10483 unsigned &MaskValue) {
10485 unsigned NumElems = BuildVector->getNumOperands();
10486 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10487 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10488 unsigned NumElemsInLane = NumElems / NumLanes;
10490 // Blend for v16i16 should be symetric for the both lanes.
10491 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10492 SDValue EltCond = BuildVector->getOperand(i);
10493 SDValue SndLaneEltCond =
10494 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10496 int Lane1Cond = -1, Lane2Cond = -1;
10497 if (isa<ConstantSDNode>(EltCond))
10498 Lane1Cond = !isZero(EltCond);
10499 if (isa<ConstantSDNode>(SndLaneEltCond))
10500 Lane2Cond = !isZero(SndLaneEltCond);
10502 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10503 // Lane1Cond != 0, means we want the first argument.
10504 // Lane1Cond == 0, means we want the second argument.
10505 // The encoding of this argument is 0 for the first argument, 1
10506 // for the second. Therefore, invert the condition.
10507 MaskValue |= !Lane1Cond << i;
10508 else if (Lane1Cond < 0)
10509 MaskValue |= !Lane2Cond << i;
10516 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10517 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10518 const X86Subtarget *Subtarget,
10519 SelectionDAG &DAG) {
10520 SDValue Cond = Op.getOperand(0);
10521 SDValue LHS = Op.getOperand(1);
10522 SDValue RHS = Op.getOperand(2);
10524 MVT VT = Op.getSimpleValueType();
10526 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10528 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10530 // Only non-legal VSELECTs reach this lowering, convert those into generic
10531 // shuffles and re-use the shuffle lowering path for blends.
10532 SmallVector<int, 32> Mask;
10533 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10534 SDValue CondElt = CondBV->getOperand(i);
10536 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10538 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10541 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10542 // A vselect where all conditions and data are constants can be optimized into
10543 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10544 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10545 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10546 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10549 // Try to lower this to a blend-style vector shuffle. This can handle all
10550 // constant condition cases.
10551 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10554 // Variable blends are only legal from SSE4.1 onward.
10555 if (!Subtarget->hasSSE41())
10558 // Only some types will be legal on some subtargets. If we can emit a legal
10559 // VSELECT-matching blend, return Op, and but if we need to expand, return
10561 switch (Op.getSimpleValueType().SimpleTy) {
10563 // Most of the vector types have blends past SSE4.1.
10567 // The byte blends for AVX vectors were introduced only in AVX2.
10568 if (Subtarget->hasAVX2())
10575 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10576 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10579 // FIXME: We should custom lower this by fixing the condition and using i8
10585 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10586 MVT VT = Op.getSimpleValueType();
10589 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10592 if (VT.getSizeInBits() == 8) {
10593 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10594 Op.getOperand(0), Op.getOperand(1));
10595 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10596 DAG.getValueType(VT));
10597 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10600 if (VT.getSizeInBits() == 16) {
10601 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10602 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10604 return DAG.getNode(
10605 ISD::TRUNCATE, dl, MVT::i16,
10606 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10607 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10608 Op.getOperand(1)));
10609 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10610 Op.getOperand(0), Op.getOperand(1));
10611 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10612 DAG.getValueType(VT));
10613 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10616 if (VT == MVT::f32) {
10617 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10618 // the result back to FR32 register. It's only worth matching if the
10619 // result has a single use which is a store or a bitcast to i32. And in
10620 // the case of a store, it's not worth it if the index is a constant 0,
10621 // because a MOVSSmr can be used instead, which is smaller and faster.
10622 if (!Op.hasOneUse())
10624 SDNode *User = *Op.getNode()->use_begin();
10625 if ((User->getOpcode() != ISD::STORE ||
10626 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10627 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10628 (User->getOpcode() != ISD::BITCAST ||
10629 User->getValueType(0) != MVT::i32))
10631 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10632 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10634 return DAG.getBitcast(MVT::f32, Extract);
10637 if (VT == MVT::i32 || VT == MVT::i64) {
10638 // ExtractPS/pextrq works with constant index.
10639 if (isa<ConstantSDNode>(Op.getOperand(1)))
10645 /// Extract one bit from mask vector, like v16i1 or v8i1.
10646 /// AVX-512 feature.
10648 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10649 SDValue Vec = Op.getOperand(0);
10651 MVT VecVT = Vec.getSimpleValueType();
10652 SDValue Idx = Op.getOperand(1);
10653 MVT EltVT = Op.getSimpleValueType();
10655 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10656 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10657 "Unexpected vector type in ExtractBitFromMaskVector");
10659 // variable index can't be handled in mask registers,
10660 // extend vector to VR512
10661 if (!isa<ConstantSDNode>(Idx)) {
10662 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10663 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10664 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10665 ExtVT.getVectorElementType(), Ext, Idx);
10666 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10669 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10670 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10671 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10672 rc = getRegClassFor(MVT::v16i1);
10673 unsigned MaxSift = rc->getSize()*8 - 1;
10674 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10675 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10676 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10677 DAG.getConstant(MaxSift, dl, MVT::i8));
10678 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10679 DAG.getIntPtrConstant(0, dl));
10683 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10684 SelectionDAG &DAG) const {
10686 SDValue Vec = Op.getOperand(0);
10687 MVT VecVT = Vec.getSimpleValueType();
10688 SDValue Idx = Op.getOperand(1);
10690 if (Op.getSimpleValueType() == MVT::i1)
10691 return ExtractBitFromMaskVector(Op, DAG);
10693 if (!isa<ConstantSDNode>(Idx)) {
10694 if (VecVT.is512BitVector() ||
10695 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10696 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10699 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10700 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10701 MaskEltVT.getSizeInBits());
10703 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10704 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10705 getZeroVector(MaskVT, Subtarget, DAG, dl),
10706 Idx, DAG.getConstant(0, dl, getPointerTy()));
10707 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10708 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10709 Perm, DAG.getConstant(0, dl, getPointerTy()));
10714 // If this is a 256-bit vector result, first extract the 128-bit vector and
10715 // then extract the element from the 128-bit vector.
10716 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10718 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10719 // Get the 128-bit vector.
10720 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10721 MVT EltVT = VecVT.getVectorElementType();
10723 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10725 //if (IdxVal >= NumElems/2)
10726 // IdxVal -= NumElems/2;
10727 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10728 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10729 DAG.getConstant(IdxVal, dl, MVT::i32));
10732 assert(VecVT.is128BitVector() && "Unexpected vector length");
10734 if (Subtarget->hasSSE41()) {
10735 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10740 MVT VT = Op.getSimpleValueType();
10741 // TODO: handle v16i8.
10742 if (VT.getSizeInBits() == 16) {
10743 SDValue Vec = Op.getOperand(0);
10744 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10746 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10747 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10748 DAG.getBitcast(MVT::v4i32, Vec),
10749 Op.getOperand(1)));
10750 // Transform it so it match pextrw which produces a 32-bit result.
10751 MVT EltVT = MVT::i32;
10752 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10753 Op.getOperand(0), Op.getOperand(1));
10754 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10755 DAG.getValueType(VT));
10756 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10759 if (VT.getSizeInBits() == 32) {
10760 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10764 // SHUFPS the element to the lowest double word, then movss.
10765 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10766 MVT VVT = Op.getOperand(0).getSimpleValueType();
10767 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10768 DAG.getUNDEF(VVT), Mask);
10769 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10770 DAG.getIntPtrConstant(0, dl));
10773 if (VT.getSizeInBits() == 64) {
10774 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10775 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10776 // to match extract_elt for f64.
10777 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10781 // UNPCKHPD the element to the lowest double word, then movsd.
10782 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10783 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10784 int Mask[2] = { 1, -1 };
10785 MVT VVT = Op.getOperand(0).getSimpleValueType();
10786 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10787 DAG.getUNDEF(VVT), Mask);
10788 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10789 DAG.getIntPtrConstant(0, dl));
10795 /// Insert one bit to mask vector, like v16i1 or v8i1.
10796 /// AVX-512 feature.
10798 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10800 SDValue Vec = Op.getOperand(0);
10801 SDValue Elt = Op.getOperand(1);
10802 SDValue Idx = Op.getOperand(2);
10803 MVT VecVT = Vec.getSimpleValueType();
10805 if (!isa<ConstantSDNode>(Idx)) {
10806 // Non constant index. Extend source and destination,
10807 // insert element and then truncate the result.
10808 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10809 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10810 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10811 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10812 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10813 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10816 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10817 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10819 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10820 DAG.getConstant(IdxVal, dl, MVT::i8));
10821 if (Vec.getOpcode() == ISD::UNDEF)
10823 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10826 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10827 SelectionDAG &DAG) const {
10828 MVT VT = Op.getSimpleValueType();
10829 MVT EltVT = VT.getVectorElementType();
10831 if (EltVT == MVT::i1)
10832 return InsertBitToMaskVector(Op, DAG);
10835 SDValue N0 = Op.getOperand(0);
10836 SDValue N1 = Op.getOperand(1);
10837 SDValue N2 = Op.getOperand(2);
10838 if (!isa<ConstantSDNode>(N2))
10840 auto *N2C = cast<ConstantSDNode>(N2);
10841 unsigned IdxVal = N2C->getZExtValue();
10843 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10844 // into that, and then insert the subvector back into the result.
10845 if (VT.is256BitVector() || VT.is512BitVector()) {
10846 // With a 256-bit vector, we can insert into the zero element efficiently
10847 // using a blend if we have AVX or AVX2 and the right data type.
10848 if (VT.is256BitVector() && IdxVal == 0) {
10849 // TODO: It is worthwhile to cast integer to floating point and back
10850 // and incur a domain crossing penalty if that's what we'll end up
10851 // doing anyway after extracting to a 128-bit vector.
10852 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
10853 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
10854 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
10855 N2 = DAG.getIntPtrConstant(1, dl);
10856 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
10860 // Get the desired 128-bit vector chunk.
10861 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10863 // Insert the element into the desired chunk.
10864 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10865 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10867 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10868 DAG.getConstant(IdxIn128, dl, MVT::i32));
10870 // Insert the changed part back into the bigger vector
10871 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10873 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10875 if (Subtarget->hasSSE41()) {
10876 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10878 if (VT == MVT::v8i16) {
10879 Opc = X86ISD::PINSRW;
10881 assert(VT == MVT::v16i8);
10882 Opc = X86ISD::PINSRB;
10885 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10887 if (N1.getValueType() != MVT::i32)
10888 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10889 if (N2.getValueType() != MVT::i32)
10890 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10891 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10894 if (EltVT == MVT::f32) {
10895 // Bits [7:6] of the constant are the source select. This will always be
10896 // zero here. The DAG Combiner may combine an extract_elt index into
10897 // these bits. For example (insert (extract, 3), 2) could be matched by
10898 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
10899 // Bits [5:4] of the constant are the destination select. This is the
10900 // value of the incoming immediate.
10901 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10902 // combine either bitwise AND or insert of float 0.0 to set these bits.
10904 const Function *F = DAG.getMachineFunction().getFunction();
10905 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
10906 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
10907 // If this is an insertion of 32-bits into the low 32-bits of
10908 // a vector, we prefer to generate a blend with immediate rather
10909 // than an insertps. Blends are simpler operations in hardware and so
10910 // will always have equal or better performance than insertps.
10911 // But if optimizing for size and there's a load folding opportunity,
10912 // generate insertps because blendps does not have a 32-bit memory
10914 N2 = DAG.getIntPtrConstant(1, dl);
10915 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10916 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
10918 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
10919 // Create this as a scalar to vector..
10920 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10921 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10924 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10925 // PINSR* works with constant index.
10930 if (EltVT == MVT::i8)
10933 if (EltVT.getSizeInBits() == 16) {
10934 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10935 // as its second argument.
10936 if (N1.getValueType() != MVT::i32)
10937 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10938 if (N2.getValueType() != MVT::i32)
10939 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10940 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10945 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10947 MVT OpVT = Op.getSimpleValueType();
10949 // If this is a 256-bit vector result, first insert into a 128-bit
10950 // vector and then insert into the 256-bit vector.
10951 if (!OpVT.is128BitVector()) {
10952 // Insert into a 128-bit vector.
10953 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10954 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10955 OpVT.getVectorNumElements() / SizeFactor);
10957 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10959 // Insert the 128-bit vector.
10960 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10963 if (OpVT == MVT::v1i64 &&
10964 Op.getOperand(0).getValueType() == MVT::i64)
10965 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10967 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10968 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10969 return DAG.getBitcast(
10970 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
10973 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10974 // a simple subregister reference or explicit instructions to grab
10975 // upper bits of a vector.
10976 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10977 SelectionDAG &DAG) {
10979 SDValue In = Op.getOperand(0);
10980 SDValue Idx = Op.getOperand(1);
10981 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10982 MVT ResVT = Op.getSimpleValueType();
10983 MVT InVT = In.getSimpleValueType();
10985 if (Subtarget->hasFp256()) {
10986 if (ResVT.is128BitVector() &&
10987 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10988 isa<ConstantSDNode>(Idx)) {
10989 return Extract128BitVector(In, IdxVal, DAG, dl);
10991 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10992 isa<ConstantSDNode>(Idx)) {
10993 return Extract256BitVector(In, IdxVal, DAG, dl);
10999 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11000 // simple superregister reference or explicit instructions to insert
11001 // the upper bits of a vector.
11002 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11003 SelectionDAG &DAG) {
11004 if (!Subtarget->hasAVX())
11008 SDValue Vec = Op.getOperand(0);
11009 SDValue SubVec = Op.getOperand(1);
11010 SDValue Idx = Op.getOperand(2);
11012 if (!isa<ConstantSDNode>(Idx))
11015 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11016 MVT OpVT = Op.getSimpleValueType();
11017 MVT SubVecVT = SubVec.getSimpleValueType();
11019 // Fold two 16-byte subvector loads into one 32-byte load:
11020 // (insert_subvector (insert_subvector undef, (load addr), 0),
11021 // (load addr + 16), Elts/2)
11023 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11024 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11025 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
11026 !Subtarget->isUnalignedMem32Slow()) {
11027 SDValue SubVec2 = Vec.getOperand(1);
11028 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
11029 if (Idx2->getZExtValue() == 0) {
11030 SDValue Ops[] = { SubVec2, SubVec };
11031 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
11038 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11039 SubVecVT.is128BitVector())
11040 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11042 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11043 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11045 if (OpVT.getVectorElementType() == MVT::i1) {
11046 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11048 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11049 SDValue Undef = DAG.getUNDEF(OpVT);
11050 unsigned NumElems = OpVT.getVectorNumElements();
11051 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11053 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11054 // Zero upper bits of the Vec
11055 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11056 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11058 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11060 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11061 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11064 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11066 // Zero upper bits of the Vec2
11067 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11068 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11069 // Zero lower bits of the Vec
11070 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11071 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11072 // Merge them together
11073 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11079 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11080 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11081 // one of the above mentioned nodes. It has to be wrapped because otherwise
11082 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11083 // be used to form addressing mode. These wrapped nodes will be selected
11086 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11087 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11089 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11090 // global base reg.
11091 unsigned char OpFlag = 0;
11092 unsigned WrapperKind = X86ISD::Wrapper;
11093 CodeModel::Model M = DAG.getTarget().getCodeModel();
11095 if (Subtarget->isPICStyleRIPRel() &&
11096 (M == CodeModel::Small || M == CodeModel::Kernel))
11097 WrapperKind = X86ISD::WrapperRIP;
11098 else if (Subtarget->isPICStyleGOT())
11099 OpFlag = X86II::MO_GOTOFF;
11100 else if (Subtarget->isPICStyleStubPIC())
11101 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11103 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
11104 CP->getAlignment(),
11105 CP->getOffset(), OpFlag);
11107 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11108 // With PIC, the address is actually $g + Offset.
11110 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11111 DAG.getNode(X86ISD::GlobalBaseReg,
11112 SDLoc(), getPointerTy()),
11119 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11120 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11122 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11123 // global base reg.
11124 unsigned char OpFlag = 0;
11125 unsigned WrapperKind = X86ISD::Wrapper;
11126 CodeModel::Model M = DAG.getTarget().getCodeModel();
11128 if (Subtarget->isPICStyleRIPRel() &&
11129 (M == CodeModel::Small || M == CodeModel::Kernel))
11130 WrapperKind = X86ISD::WrapperRIP;
11131 else if (Subtarget->isPICStyleGOT())
11132 OpFlag = X86II::MO_GOTOFF;
11133 else if (Subtarget->isPICStyleStubPIC())
11134 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11136 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
11139 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11141 // With PIC, the address is actually $g + Offset.
11143 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11144 DAG.getNode(X86ISD::GlobalBaseReg,
11145 SDLoc(), getPointerTy()),
11152 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11153 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11155 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11156 // global base reg.
11157 unsigned char OpFlag = 0;
11158 unsigned WrapperKind = X86ISD::Wrapper;
11159 CodeModel::Model M = DAG.getTarget().getCodeModel();
11161 if (Subtarget->isPICStyleRIPRel() &&
11162 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11163 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11164 OpFlag = X86II::MO_GOTPCREL;
11165 WrapperKind = X86ISD::WrapperRIP;
11166 } else if (Subtarget->isPICStyleGOT()) {
11167 OpFlag = X86II::MO_GOT;
11168 } else if (Subtarget->isPICStyleStubPIC()) {
11169 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11170 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11171 OpFlag = X86II::MO_DARWIN_NONLAZY;
11174 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
11177 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11179 // With PIC, the address is actually $g + Offset.
11180 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11181 !Subtarget->is64Bit()) {
11182 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11183 DAG.getNode(X86ISD::GlobalBaseReg,
11184 SDLoc(), getPointerTy()),
11188 // For symbols that require a load from a stub to get the address, emit the
11190 if (isGlobalStubReference(OpFlag))
11191 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
11192 MachinePointerInfo::getGOT(), false, false, false, 0);
11198 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11199 // Create the TargetBlockAddressAddress node.
11200 unsigned char OpFlags =
11201 Subtarget->ClassifyBlockAddressReference();
11202 CodeModel::Model M = DAG.getTarget().getCodeModel();
11203 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11204 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11206 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
11209 if (Subtarget->isPICStyleRIPRel() &&
11210 (M == CodeModel::Small || M == CodeModel::Kernel))
11211 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11213 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11215 // With PIC, the address is actually $g + Offset.
11216 if (isGlobalRelativeToPICBase(OpFlags)) {
11217 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11218 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11226 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11227 int64_t Offset, SelectionDAG &DAG) const {
11228 // Create the TargetGlobalAddress node, folding in the constant
11229 // offset if it is legal.
11230 unsigned char OpFlags =
11231 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11232 CodeModel::Model M = DAG.getTarget().getCodeModel();
11234 if (OpFlags == X86II::MO_NO_FLAG &&
11235 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11236 // A direct static reference to a global.
11237 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
11240 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
11243 if (Subtarget->isPICStyleRIPRel() &&
11244 (M == CodeModel::Small || M == CodeModel::Kernel))
11245 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11247 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11249 // With PIC, the address is actually $g + Offset.
11250 if (isGlobalRelativeToPICBase(OpFlags)) {
11251 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11252 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11256 // For globals that require a load from a stub to get the address, emit the
11258 if (isGlobalStubReference(OpFlags))
11259 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
11260 MachinePointerInfo::getGOT(), false, false, false, 0);
11262 // If there was a non-zero offset that we didn't fold, create an explicit
11263 // addition for it.
11265 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
11266 DAG.getConstant(Offset, dl, getPointerTy()));
11272 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11273 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11274 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11275 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11279 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11280 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11281 unsigned char OperandFlags, bool LocalDynamic = false) {
11282 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11283 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11285 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11286 GA->getValueType(0),
11290 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11294 SDValue Ops[] = { Chain, TGA, *InFlag };
11295 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11297 SDValue Ops[] = { Chain, TGA };
11298 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11301 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11302 MFI->setAdjustsStack(true);
11303 MFI->setHasCalls(true);
11305 SDValue Flag = Chain.getValue(1);
11306 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11309 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11311 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11314 SDLoc dl(GA); // ? function entry point might be better
11315 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11316 DAG.getNode(X86ISD::GlobalBaseReg,
11317 SDLoc(), PtrVT), InFlag);
11318 InFlag = Chain.getValue(1);
11320 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11323 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11325 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11327 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11328 X86::RAX, X86II::MO_TLSGD);
11331 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11337 // Get the start address of the TLS block for this module.
11338 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11339 .getInfo<X86MachineFunctionInfo>();
11340 MFI->incNumLocalDynamicTLSAccesses();
11344 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11345 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11348 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11349 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11350 InFlag = Chain.getValue(1);
11351 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11352 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11355 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11359 unsigned char OperandFlags = X86II::MO_DTPOFF;
11360 unsigned WrapperKind = X86ISD::Wrapper;
11361 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11362 GA->getValueType(0),
11363 GA->getOffset(), OperandFlags);
11364 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11366 // Add x@dtpoff with the base.
11367 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11370 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11371 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11372 const EVT PtrVT, TLSModel::Model model,
11373 bool is64Bit, bool isPIC) {
11376 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11377 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11378 is64Bit ? 257 : 256));
11380 SDValue ThreadPointer =
11381 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11382 MachinePointerInfo(Ptr), false, false, false, 0);
11384 unsigned char OperandFlags = 0;
11385 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11387 unsigned WrapperKind = X86ISD::Wrapper;
11388 if (model == TLSModel::LocalExec) {
11389 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11390 } else if (model == TLSModel::InitialExec) {
11392 OperandFlags = X86II::MO_GOTTPOFF;
11393 WrapperKind = X86ISD::WrapperRIP;
11395 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11398 llvm_unreachable("Unexpected model");
11401 // emit "addl x@ntpoff,%eax" (local exec)
11402 // or "addl x@indntpoff,%eax" (initial exec)
11403 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11405 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11406 GA->getOffset(), OperandFlags);
11407 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11409 if (model == TLSModel::InitialExec) {
11410 if (isPIC && !is64Bit) {
11411 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11412 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11416 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11417 MachinePointerInfo::getGOT(), false, false, false, 0);
11420 // The address of the thread local variable is the add of the thread
11421 // pointer with the offset of the variable.
11422 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11426 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11428 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11429 const GlobalValue *GV = GA->getGlobal();
11431 if (Subtarget->isTargetELF()) {
11432 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11434 case TLSModel::GeneralDynamic:
11435 if (Subtarget->is64Bit())
11436 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
11437 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
11438 case TLSModel::LocalDynamic:
11439 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
11440 Subtarget->is64Bit());
11441 case TLSModel::InitialExec:
11442 case TLSModel::LocalExec:
11443 return LowerToTLSExecModel(
11444 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
11445 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
11447 llvm_unreachable("Unknown TLS model.");
11450 if (Subtarget->isTargetDarwin()) {
11451 // Darwin only has one model of TLS. Lower to that.
11452 unsigned char OpFlag = 0;
11453 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11454 X86ISD::WrapperRIP : X86ISD::Wrapper;
11456 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11457 // global base reg.
11458 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11459 !Subtarget->is64Bit();
11461 OpFlag = X86II::MO_TLVP_PIC_BASE;
11463 OpFlag = X86II::MO_TLVP;
11465 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11466 GA->getValueType(0),
11467 GA->getOffset(), OpFlag);
11468 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11470 // With PIC32, the address is actually $g + Offset.
11472 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11473 DAG.getNode(X86ISD::GlobalBaseReg,
11474 SDLoc(), getPointerTy()),
11477 // Lowering the machine isd will make sure everything is in the right
11479 SDValue Chain = DAG.getEntryNode();
11480 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11481 SDValue Args[] = { Chain, Offset };
11482 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11484 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11485 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11486 MFI->setAdjustsStack(true);
11488 // And our return value (tls address) is in the standard call return value
11490 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11491 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11492 Chain.getValue(1));
11495 if (Subtarget->isTargetKnownWindowsMSVC() ||
11496 Subtarget->isTargetWindowsGNU()) {
11497 // Just use the implicit TLS architecture
11498 // Need to generate someting similar to:
11499 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11501 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11502 // mov rcx, qword [rdx+rcx*8]
11503 // mov eax, .tls$:tlsvar
11504 // [rax+rcx] contains the address
11505 // Windows 64bit: gs:0x58
11506 // Windows 32bit: fs:__tls_array
11509 SDValue Chain = DAG.getEntryNode();
11511 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11512 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11513 // use its literal value of 0x2C.
11514 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11515 ? Type::getInt8PtrTy(*DAG.getContext(),
11517 : Type::getInt32PtrTy(*DAG.getContext(),
11521 Subtarget->is64Bit()
11522 ? DAG.getIntPtrConstant(0x58, dl)
11523 : (Subtarget->isTargetWindowsGNU()
11524 ? DAG.getIntPtrConstant(0x2C, dl)
11525 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11527 SDValue ThreadPointer =
11528 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11529 MachinePointerInfo(Ptr), false, false, false, 0);
11532 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11533 res = ThreadPointer;
11535 // Load the _tls_index variable
11536 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11537 if (Subtarget->is64Bit())
11538 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain, IDX,
11539 MachinePointerInfo(), MVT::i32, false, false,
11542 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11543 false, false, false, 0);
11545 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()), dl,
11547 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11549 res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11552 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11553 false, false, false, 0);
11555 // Get the offset of start of .tls section
11556 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11557 GA->getValueType(0),
11558 GA->getOffset(), X86II::MO_SECREL);
11559 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11561 // The address of the thread local variable is the add of the thread
11562 // pointer with the offset of the variable.
11563 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11566 llvm_unreachable("TLS not implemented for this target.");
11569 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11570 /// and take a 2 x i32 value to shift plus a shift amount.
11571 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11572 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11573 MVT VT = Op.getSimpleValueType();
11574 unsigned VTBits = VT.getSizeInBits();
11576 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11577 SDValue ShOpLo = Op.getOperand(0);
11578 SDValue ShOpHi = Op.getOperand(1);
11579 SDValue ShAmt = Op.getOperand(2);
11580 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11581 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11583 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11584 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11585 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11586 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11587 : DAG.getConstant(0, dl, VT);
11589 SDValue Tmp2, Tmp3;
11590 if (Op.getOpcode() == ISD::SHL_PARTS) {
11591 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11592 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11594 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11595 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11598 // If the shift amount is larger or equal than the width of a part we can't
11599 // rely on the results of shld/shrd. Insert a test and select the appropriate
11600 // values for large shift amounts.
11601 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11602 DAG.getConstant(VTBits, dl, MVT::i8));
11603 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11604 AndNode, DAG.getConstant(0, dl, MVT::i8));
11607 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11608 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11609 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11611 if (Op.getOpcode() == ISD::SHL_PARTS) {
11612 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11613 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11615 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11616 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11619 SDValue Ops[2] = { Lo, Hi };
11620 return DAG.getMergeValues(Ops, dl);
11623 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11624 SelectionDAG &DAG) const {
11625 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11628 if (SrcVT.isVector()) {
11629 if (SrcVT.getVectorElementType() == MVT::i1) {
11630 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11631 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11632 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11633 Op.getOperand(0)));
11638 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11639 "Unknown SINT_TO_FP to lower!");
11641 // These are really Legal; return the operand so the caller accepts it as
11643 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11645 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11646 Subtarget->is64Bit()) {
11650 unsigned Size = SrcVT.getSizeInBits()/8;
11651 MachineFunction &MF = DAG.getMachineFunction();
11652 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11653 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11654 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11656 MachinePointerInfo::getFixedStack(SSFI),
11658 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11661 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11663 SelectionDAG &DAG) const {
11667 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11669 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11671 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11673 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11675 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11676 MachineMemOperand *MMO;
11678 int SSFI = FI->getIndex();
11680 DAG.getMachineFunction()
11681 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11682 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11684 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11685 StackSlot = StackSlot.getOperand(1);
11687 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11688 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11690 Tys, Ops, SrcVT, MMO);
11693 Chain = Result.getValue(1);
11694 SDValue InFlag = Result.getValue(2);
11696 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11697 // shouldn't be necessary except that RFP cannot be live across
11698 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11699 MachineFunction &MF = DAG.getMachineFunction();
11700 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11701 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11702 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11703 Tys = DAG.getVTList(MVT::Other);
11705 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11707 MachineMemOperand *MMO =
11708 DAG.getMachineFunction()
11709 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11710 MachineMemOperand::MOStore, SSFISize, SSFISize);
11712 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11713 Ops, Op.getValueType(), MMO);
11714 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11715 MachinePointerInfo::getFixedStack(SSFI),
11716 false, false, false, 0);
11722 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11723 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11724 SelectionDAG &DAG) const {
11725 // This algorithm is not obvious. Here it is what we're trying to output:
11728 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11729 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11731 haddpd %xmm0, %xmm0
11733 pshufd $0x4e, %xmm0, %xmm1
11739 LLVMContext *Context = DAG.getContext();
11741 // Build some magic constants.
11742 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11743 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11744 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11746 SmallVector<Constant*,2> CV1;
11748 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11749 APInt(64, 0x4330000000000000ULL))));
11751 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11752 APInt(64, 0x4530000000000000ULL))));
11753 Constant *C1 = ConstantVector::get(CV1);
11754 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11756 // Load the 64-bit value into an XMM register.
11757 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11759 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11760 MachinePointerInfo::getConstantPool(),
11761 false, false, false, 16);
11763 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
11765 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11766 MachinePointerInfo::getConstantPool(),
11767 false, false, false, 16);
11768 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
11769 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11772 if (Subtarget->hasSSE3()) {
11773 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11774 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11776 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
11777 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11779 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11780 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
11783 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11784 DAG.getIntPtrConstant(0, dl));
11787 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11788 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11789 SelectionDAG &DAG) const {
11791 // FP constant to bias correct the final result.
11792 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
11795 // Load the 32-bit value into an XMM register.
11796 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11799 // Zero out the upper parts of the register.
11800 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11802 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11803 DAG.getBitcast(MVT::v2f64, Load),
11804 DAG.getIntPtrConstant(0, dl));
11806 // Or the load with the bias.
11807 SDValue Or = DAG.getNode(
11808 ISD::OR, dl, MVT::v2i64,
11809 DAG.getBitcast(MVT::v2i64,
11810 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
11811 DAG.getBitcast(MVT::v2i64,
11812 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
11814 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11815 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
11817 // Subtract the bias.
11818 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11820 // Handle final rounding.
11821 EVT DestVT = Op.getValueType();
11823 if (DestVT.bitsLT(MVT::f64))
11824 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11825 DAG.getIntPtrConstant(0, dl));
11826 if (DestVT.bitsGT(MVT::f64))
11827 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11829 // Handle final rounding.
11833 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11834 const X86Subtarget &Subtarget) {
11835 // The algorithm is the following:
11836 // #ifdef __SSE4_1__
11837 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11838 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11839 // (uint4) 0x53000000, 0xaa);
11841 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11842 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11844 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11845 // return (float4) lo + fhi;
11848 SDValue V = Op->getOperand(0);
11849 EVT VecIntVT = V.getValueType();
11850 bool Is128 = VecIntVT == MVT::v4i32;
11851 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11852 // If we convert to something else than the supported type, e.g., to v4f64,
11854 if (VecFloatVT != Op->getValueType(0))
11857 unsigned NumElts = VecIntVT.getVectorNumElements();
11858 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11859 "Unsupported custom type");
11860 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11862 // In the #idef/#else code, we have in common:
11863 // - The vector of constants:
11869 // Create the splat vector for 0x4b000000.
11870 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
11871 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11872 CstLow, CstLow, CstLow, CstLow};
11873 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11874 makeArrayRef(&CstLowArray[0], NumElts));
11875 // Create the splat vector for 0x53000000.
11876 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
11877 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11878 CstHigh, CstHigh, CstHigh, CstHigh};
11879 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11880 makeArrayRef(&CstHighArray[0], NumElts));
11882 // Create the right shift.
11883 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
11884 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11885 CstShift, CstShift, CstShift, CstShift};
11886 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11887 makeArrayRef(&CstShiftArray[0], NumElts));
11888 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11891 if (Subtarget.hasSSE41()) {
11892 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11893 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11894 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
11895 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
11896 // Low will be bitcasted right away, so do not bother bitcasting back to its
11898 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11899 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11900 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11901 // (uint4) 0x53000000, 0xaa);
11902 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
11903 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
11904 // High will be bitcasted right away, so do not bother bitcasting back to
11905 // its original type.
11906 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11907 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11909 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
11910 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11911 CstMask, CstMask, CstMask);
11912 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11913 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11914 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11916 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11917 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11920 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11921 SDValue CstFAdd = DAG.getConstantFP(
11922 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
11923 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11924 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11925 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11926 makeArrayRef(&CstFAddArray[0], NumElts));
11928 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11929 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
11931 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11932 // return (float4) lo + fhi;
11933 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
11934 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11937 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11938 SelectionDAG &DAG) const {
11939 SDValue N0 = Op.getOperand(0);
11940 MVT SVT = N0.getSimpleValueType();
11943 switch (SVT.SimpleTy) {
11945 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11950 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11951 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11952 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11956 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11959 if (Subtarget->hasAVX512())
11960 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
11961 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
11963 llvm_unreachable(nullptr);
11966 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11967 SelectionDAG &DAG) const {
11968 SDValue N0 = Op.getOperand(0);
11971 if (Op.getValueType().isVector())
11972 return lowerUINT_TO_FP_vec(Op, DAG);
11974 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11975 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11976 // the optimization here.
11977 if (DAG.SignBitIsZero(N0))
11978 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11980 MVT SrcVT = N0.getSimpleValueType();
11981 MVT DstVT = Op.getSimpleValueType();
11982 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11983 return LowerUINT_TO_FP_i64(Op, DAG);
11984 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11985 return LowerUINT_TO_FP_i32(Op, DAG);
11986 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11989 // Make a 64-bit buffer, and use it to build an FILD.
11990 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11991 if (SrcVT == MVT::i32) {
11992 SDValue WordOff = DAG.getConstant(4, dl, getPointerTy());
11993 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11994 getPointerTy(), StackSlot, WordOff);
11995 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11996 StackSlot, MachinePointerInfo(),
11998 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
11999 OffsetSlot, MachinePointerInfo(),
12001 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12005 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12006 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12007 StackSlot, MachinePointerInfo(),
12009 // For i64 source, we need to add the appropriate power of 2 if the input
12010 // was negative. This is the same as the optimization in
12011 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12012 // we must be careful to do the computation in x87 extended precision, not
12013 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12014 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12015 MachineMemOperand *MMO =
12016 DAG.getMachineFunction()
12017 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12018 MachineMemOperand::MOLoad, 8, 8);
12020 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12021 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12022 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12025 APInt FF(32, 0x5F800000ULL);
12027 // Check whether the sign bit is set.
12028 SDValue SignSet = DAG.getSetCC(dl,
12029 getSetCCResultType(*DAG.getContext(), MVT::i64),
12031 DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12033 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12034 SDValue FudgePtr = DAG.getConstantPool(
12035 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
12038 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12039 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12040 SDValue Four = DAG.getIntPtrConstant(4, dl);
12041 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12043 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
12045 // Load the value out, extending it from f32 to f80.
12046 // FIXME: Avoid the extend by constructing the right constant pool?
12047 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
12048 FudgePtr, MachinePointerInfo::getConstantPool(),
12049 MVT::f32, false, false, false, 4);
12050 // Extend everything to 80 bits to force it to be done on x87.
12051 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12052 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12053 DAG.getIntPtrConstant(0, dl));
12056 std::pair<SDValue,SDValue>
12057 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12058 bool IsSigned, bool IsReplace) const {
12061 EVT DstTy = Op.getValueType();
12063 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12064 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12068 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12069 DstTy.getSimpleVT() >= MVT::i16 &&
12070 "Unknown FP_TO_INT to lower!");
12072 // These are really Legal.
12073 if (DstTy == MVT::i32 &&
12074 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12075 return std::make_pair(SDValue(), SDValue());
12076 if (Subtarget->is64Bit() &&
12077 DstTy == MVT::i64 &&
12078 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12079 return std::make_pair(SDValue(), SDValue());
12081 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12082 // stack slot, or into the FTOL runtime function.
12083 MachineFunction &MF = DAG.getMachineFunction();
12084 unsigned MemSize = DstTy.getSizeInBits()/8;
12085 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12086 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12089 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12090 Opc = X86ISD::WIN_FTOL;
12092 switch (DstTy.getSimpleVT().SimpleTy) {
12093 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12094 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12095 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12096 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12099 SDValue Chain = DAG.getEntryNode();
12100 SDValue Value = Op.getOperand(0);
12101 EVT TheVT = Op.getOperand(0).getValueType();
12102 // FIXME This causes a redundant load/store if the SSE-class value is already
12103 // in memory, such as if it is on the callstack.
12104 if (isScalarFPTypeInSSEReg(TheVT)) {
12105 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12106 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12107 MachinePointerInfo::getFixedStack(SSFI),
12109 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12111 Chain, StackSlot, DAG.getValueType(TheVT)
12114 MachineMemOperand *MMO =
12115 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12116 MachineMemOperand::MOLoad, MemSize, MemSize);
12117 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12118 Chain = Value.getValue(1);
12119 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12120 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12123 MachineMemOperand *MMO =
12124 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12125 MachineMemOperand::MOStore, MemSize, MemSize);
12127 if (Opc != X86ISD::WIN_FTOL) {
12128 // Build the FP_TO_INT*_IN_MEM
12129 SDValue Ops[] = { Chain, Value, StackSlot };
12130 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12132 return std::make_pair(FIST, StackSlot);
12134 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12135 DAG.getVTList(MVT::Other, MVT::Glue),
12137 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12138 MVT::i32, ftol.getValue(1));
12139 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12140 MVT::i32, eax.getValue(2));
12141 SDValue Ops[] = { eax, edx };
12142 SDValue pair = IsReplace
12143 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12144 : DAG.getMergeValues(Ops, DL);
12145 return std::make_pair(pair, SDValue());
12149 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12150 const X86Subtarget *Subtarget) {
12151 MVT VT = Op->getSimpleValueType(0);
12152 SDValue In = Op->getOperand(0);
12153 MVT InVT = In.getSimpleValueType();
12156 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12157 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12159 // Optimize vectors in AVX mode:
12162 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12163 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12164 // Concat upper and lower parts.
12167 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12168 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12169 // Concat upper and lower parts.
12172 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12173 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12174 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12177 if (Subtarget->hasInt256())
12178 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12180 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12181 SDValue Undef = DAG.getUNDEF(InVT);
12182 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12183 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12184 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12186 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12187 VT.getVectorNumElements()/2);
12189 OpLo = DAG.getBitcast(HVT, OpLo);
12190 OpHi = DAG.getBitcast(HVT, OpHi);
12192 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12195 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12196 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12197 MVT VT = Op->getSimpleValueType(0);
12198 SDValue In = Op->getOperand(0);
12199 MVT InVT = In.getSimpleValueType();
12201 unsigned int NumElts = VT.getVectorNumElements();
12202 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12205 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12206 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12208 assert(InVT.getVectorElementType() == MVT::i1);
12209 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12211 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12213 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12215 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12216 if (VT.is512BitVector())
12218 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12221 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12222 SelectionDAG &DAG) {
12223 if (Subtarget->hasFp256()) {
12224 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12232 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12233 SelectionDAG &DAG) {
12235 MVT VT = Op.getSimpleValueType();
12236 SDValue In = Op.getOperand(0);
12237 MVT SVT = In.getSimpleValueType();
12239 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12240 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12242 if (Subtarget->hasFp256()) {
12243 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12248 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12249 VT.getVectorNumElements() != SVT.getVectorNumElements());
12253 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12255 MVT VT = Op.getSimpleValueType();
12256 SDValue In = Op.getOperand(0);
12257 MVT InVT = In.getSimpleValueType();
12259 if (VT == MVT::i1) {
12260 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12261 "Invalid scalar TRUNCATE operation");
12262 if (InVT.getSizeInBits() >= 32)
12264 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12265 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12267 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12268 "Invalid TRUNCATE operation");
12270 // move vector to mask - truncate solution for SKX
12271 if (VT.getVectorElementType() == MVT::i1) {
12272 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12273 Subtarget->hasBWI())
12274 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12275 if ((InVT.is256BitVector() || InVT.is128BitVector())
12276 && InVT.getScalarSizeInBits() <= 16 &&
12277 Subtarget->hasBWI() && Subtarget->hasVLX())
12278 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12279 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12280 Subtarget->hasDQI())
12281 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12282 if ((InVT.is256BitVector() || InVT.is128BitVector())
12283 && InVT.getScalarSizeInBits() >= 32 &&
12284 Subtarget->hasDQI() && Subtarget->hasVLX())
12285 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12287 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
12288 if (VT.getVectorElementType().getSizeInBits() >=8)
12289 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12291 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12292 unsigned NumElts = InVT.getVectorNumElements();
12293 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12294 if (InVT.getSizeInBits() < 512) {
12295 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12296 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12301 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12302 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12303 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12306 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12307 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12308 if (Subtarget->hasInt256()) {
12309 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12310 In = DAG.getBitcast(MVT::v8i32, In);
12311 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12313 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12314 DAG.getIntPtrConstant(0, DL));
12317 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12318 DAG.getIntPtrConstant(0, DL));
12319 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12320 DAG.getIntPtrConstant(2, DL));
12321 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12322 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12323 static const int ShufMask[] = {0, 2, 4, 6};
12324 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12327 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12328 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12329 if (Subtarget->hasInt256()) {
12330 In = DAG.getBitcast(MVT::v32i8, In);
12332 SmallVector<SDValue,32> pshufbMask;
12333 for (unsigned i = 0; i < 2; ++i) {
12334 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12335 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12336 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12337 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12338 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12339 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12340 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12341 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12342 for (unsigned j = 0; j < 8; ++j)
12343 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12345 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12346 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12347 In = DAG.getBitcast(MVT::v4i64, In);
12349 static const int ShufMask[] = {0, 2, -1, -1};
12350 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12352 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12353 DAG.getIntPtrConstant(0, DL));
12354 return DAG.getBitcast(VT, In);
12357 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12358 DAG.getIntPtrConstant(0, DL));
12360 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12361 DAG.getIntPtrConstant(4, DL));
12363 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12364 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12366 // The PSHUFB mask:
12367 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12368 -1, -1, -1, -1, -1, -1, -1, -1};
12370 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12371 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12372 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12374 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12375 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12377 // The MOVLHPS Mask:
12378 static const int ShufMask2[] = {0, 1, 4, 5};
12379 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12380 return DAG.getBitcast(MVT::v8i16, res);
12383 // Handle truncation of V256 to V128 using shuffles.
12384 if (!VT.is128BitVector() || !InVT.is256BitVector())
12387 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12389 unsigned NumElems = VT.getVectorNumElements();
12390 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12392 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12393 // Prepare truncation shuffle mask
12394 for (unsigned i = 0; i != NumElems; ++i)
12395 MaskVec[i] = i * 2;
12396 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12397 DAG.getUNDEF(NVT), &MaskVec[0]);
12398 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12399 DAG.getIntPtrConstant(0, DL));
12402 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12403 SelectionDAG &DAG) const {
12404 assert(!Op.getSimpleValueType().isVector());
12406 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12407 /*IsSigned=*/ true, /*IsReplace=*/ false);
12408 SDValue FIST = Vals.first, StackSlot = Vals.second;
12409 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12410 if (!FIST.getNode()) return Op;
12412 if (StackSlot.getNode())
12413 // Load the result.
12414 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12415 FIST, StackSlot, MachinePointerInfo(),
12416 false, false, false, 0);
12418 // The node is the result.
12422 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12423 SelectionDAG &DAG) const {
12424 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12425 /*IsSigned=*/ false, /*IsReplace=*/ false);
12426 SDValue FIST = Vals.first, StackSlot = Vals.second;
12427 assert(FIST.getNode() && "Unexpected failure");
12429 if (StackSlot.getNode())
12430 // Load the result.
12431 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12432 FIST, StackSlot, MachinePointerInfo(),
12433 false, false, false, 0);
12435 // The node is the result.
12439 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12441 MVT VT = Op.getSimpleValueType();
12442 SDValue In = Op.getOperand(0);
12443 MVT SVT = In.getSimpleValueType();
12445 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12447 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12448 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12449 In, DAG.getUNDEF(SVT)));
12452 /// The only differences between FABS and FNEG are the mask and the logic op.
12453 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12454 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12455 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12456 "Wrong opcode for lowering FABS or FNEG.");
12458 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12460 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12461 // into an FNABS. We'll lower the FABS after that if it is still in use.
12463 for (SDNode *User : Op->uses())
12464 if (User->getOpcode() == ISD::FNEG)
12467 SDValue Op0 = Op.getOperand(0);
12468 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12471 MVT VT = Op.getSimpleValueType();
12472 // Assume scalar op for initialization; update for vector if needed.
12473 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
12474 // generate a 16-byte vector constant and logic op even for the scalar case.
12475 // Using a 16-byte mask allows folding the load of the mask with
12476 // the logic op, so it can save (~4 bytes) on code size.
12478 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
12479 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12480 // decide if we should generate a 16-byte constant mask when we only need 4 or
12481 // 8 bytes for the scalar case.
12482 if (VT.isVector()) {
12483 EltVT = VT.getVectorElementType();
12484 NumElts = VT.getVectorNumElements();
12487 unsigned EltBits = EltVT.getSizeInBits();
12488 LLVMContext *Context = DAG.getContext();
12489 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12491 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12492 Constant *C = ConstantInt::get(*Context, MaskElt);
12493 C = ConstantVector::getSplat(NumElts, C);
12494 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12495 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
12496 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12497 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12498 MachinePointerInfo::getConstantPool(),
12499 false, false, false, Alignment);
12501 if (VT.isVector()) {
12502 // For a vector, cast operands to a vector type, perform the logic op,
12503 // and cast the result back to the original value type.
12504 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12505 SDValue MaskCasted = DAG.getBitcast(VecVT, Mask);
12506 SDValue Operand = IsFNABS ? DAG.getBitcast(VecVT, Op0.getOperand(0))
12507 : DAG.getBitcast(VecVT, Op0);
12508 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12509 return DAG.getBitcast(VT,
12510 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12513 // If not vector, then scalar.
12514 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12515 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12516 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12519 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12520 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12521 LLVMContext *Context = DAG.getContext();
12522 SDValue Op0 = Op.getOperand(0);
12523 SDValue Op1 = Op.getOperand(1);
12525 MVT VT = Op.getSimpleValueType();
12526 MVT SrcVT = Op1.getSimpleValueType();
12528 // If second operand is smaller, extend it first.
12529 if (SrcVT.bitsLT(VT)) {
12530 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12533 // And if it is bigger, shrink it first.
12534 if (SrcVT.bitsGT(VT)) {
12535 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12539 // At this point the operands and the result should have the same
12540 // type, and that won't be f80 since that is not custom lowered.
12542 const fltSemantics &Sem =
12543 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12544 const unsigned SizeInBits = VT.getSizeInBits();
12546 SmallVector<Constant *, 4> CV(
12547 VT == MVT::f64 ? 2 : 4,
12548 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12550 // First, clear all bits but the sign bit from the second operand (sign).
12551 CV[0] = ConstantFP::get(*Context,
12552 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12553 Constant *C = ConstantVector::get(CV);
12554 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12555 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12556 MachinePointerInfo::getConstantPool(),
12557 false, false, false, 16);
12558 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12560 // Next, clear the sign bit from the first operand (magnitude).
12561 // If it's a constant, we can clear it here.
12562 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12563 APFloat APF = Op0CN->getValueAPF();
12564 // If the magnitude is a positive zero, the sign bit alone is enough.
12565 if (APF.isPosZero())
12568 CV[0] = ConstantFP::get(*Context, APF);
12570 CV[0] = ConstantFP::get(
12572 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12574 C = ConstantVector::get(CV);
12575 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12576 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12577 MachinePointerInfo::getConstantPool(),
12578 false, false, false, 16);
12579 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12580 if (!isa<ConstantFPSDNode>(Op0))
12581 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12583 // OR the magnitude value with the sign bit.
12584 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12587 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12588 SDValue N0 = Op.getOperand(0);
12590 MVT VT = Op.getSimpleValueType();
12592 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12593 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12594 DAG.getConstant(1, dl, VT));
12595 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12598 // Check whether an OR'd tree is PTEST-able.
12599 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12600 SelectionDAG &DAG) {
12601 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12603 if (!Subtarget->hasSSE41())
12606 if (!Op->hasOneUse())
12609 SDNode *N = Op.getNode();
12612 SmallVector<SDValue, 8> Opnds;
12613 DenseMap<SDValue, unsigned> VecInMap;
12614 SmallVector<SDValue, 8> VecIns;
12615 EVT VT = MVT::Other;
12617 // Recognize a special case where a vector is casted into wide integer to
12619 Opnds.push_back(N->getOperand(0));
12620 Opnds.push_back(N->getOperand(1));
12622 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12623 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12624 // BFS traverse all OR'd operands.
12625 if (I->getOpcode() == ISD::OR) {
12626 Opnds.push_back(I->getOperand(0));
12627 Opnds.push_back(I->getOperand(1));
12628 // Re-evaluate the number of nodes to be traversed.
12629 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12633 // Quit if a non-EXTRACT_VECTOR_ELT
12634 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12637 // Quit if without a constant index.
12638 SDValue Idx = I->getOperand(1);
12639 if (!isa<ConstantSDNode>(Idx))
12642 SDValue ExtractedFromVec = I->getOperand(0);
12643 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12644 if (M == VecInMap.end()) {
12645 VT = ExtractedFromVec.getValueType();
12646 // Quit if not 128/256-bit vector.
12647 if (!VT.is128BitVector() && !VT.is256BitVector())
12649 // Quit if not the same type.
12650 if (VecInMap.begin() != VecInMap.end() &&
12651 VT != VecInMap.begin()->first.getValueType())
12653 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12654 VecIns.push_back(ExtractedFromVec);
12656 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12659 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12660 "Not extracted from 128-/256-bit vector.");
12662 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12664 for (DenseMap<SDValue, unsigned>::const_iterator
12665 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12666 // Quit if not all elements are used.
12667 if (I->second != FullMask)
12671 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12673 // Cast all vectors into TestVT for PTEST.
12674 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12675 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
12677 // If more than one full vectors are evaluated, OR them first before PTEST.
12678 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12679 // Each iteration will OR 2 nodes and append the result until there is only
12680 // 1 node left, i.e. the final OR'd value of all vectors.
12681 SDValue LHS = VecIns[Slot];
12682 SDValue RHS = VecIns[Slot + 1];
12683 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12686 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12687 VecIns.back(), VecIns.back());
12690 /// \brief return true if \c Op has a use that doesn't just read flags.
12691 static bool hasNonFlagsUse(SDValue Op) {
12692 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12694 SDNode *User = *UI;
12695 unsigned UOpNo = UI.getOperandNo();
12696 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12697 // Look pass truncate.
12698 UOpNo = User->use_begin().getOperandNo();
12699 User = *User->use_begin();
12702 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12703 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12709 /// Emit nodes that will be selected as "test Op0,Op0", or something
12711 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12712 SelectionDAG &DAG) const {
12713 if (Op.getValueType() == MVT::i1) {
12714 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12715 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12716 DAG.getConstant(0, dl, MVT::i8));
12718 // CF and OF aren't always set the way we want. Determine which
12719 // of these we need.
12720 bool NeedCF = false;
12721 bool NeedOF = false;
12724 case X86::COND_A: case X86::COND_AE:
12725 case X86::COND_B: case X86::COND_BE:
12728 case X86::COND_G: case X86::COND_GE:
12729 case X86::COND_L: case X86::COND_LE:
12730 case X86::COND_O: case X86::COND_NO: {
12731 // Check if we really need to set the
12732 // Overflow flag. If NoSignedWrap is present
12733 // that is not actually needed.
12734 switch (Op->getOpcode()) {
12739 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
12740 if (BinNode->Flags.hasNoSignedWrap())
12750 // See if we can use the EFLAGS value from the operand instead of
12751 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12752 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12753 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12754 // Emit a CMP with 0, which is the TEST pattern.
12755 //if (Op.getValueType() == MVT::i1)
12756 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12757 // DAG.getConstant(0, MVT::i1));
12758 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12759 DAG.getConstant(0, dl, Op.getValueType()));
12761 unsigned Opcode = 0;
12762 unsigned NumOperands = 0;
12764 // Truncate operations may prevent the merge of the SETCC instruction
12765 // and the arithmetic instruction before it. Attempt to truncate the operands
12766 // of the arithmetic instruction and use a reduced bit-width instruction.
12767 bool NeedTruncation = false;
12768 SDValue ArithOp = Op;
12769 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12770 SDValue Arith = Op->getOperand(0);
12771 // Both the trunc and the arithmetic op need to have one user each.
12772 if (Arith->hasOneUse())
12773 switch (Arith.getOpcode()) {
12780 NeedTruncation = true;
12786 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12787 // which may be the result of a CAST. We use the variable 'Op', which is the
12788 // non-casted variable when we check for possible users.
12789 switch (ArithOp.getOpcode()) {
12791 // Due to an isel shortcoming, be conservative if this add is likely to be
12792 // selected as part of a load-modify-store instruction. When the root node
12793 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12794 // uses of other nodes in the match, such as the ADD in this case. This
12795 // leads to the ADD being left around and reselected, with the result being
12796 // two adds in the output. Alas, even if none our users are stores, that
12797 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12798 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12799 // climbing the DAG back to the root, and it doesn't seem to be worth the
12801 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12802 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12803 if (UI->getOpcode() != ISD::CopyToReg &&
12804 UI->getOpcode() != ISD::SETCC &&
12805 UI->getOpcode() != ISD::STORE)
12808 if (ConstantSDNode *C =
12809 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12810 // An add of one will be selected as an INC.
12811 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12812 Opcode = X86ISD::INC;
12817 // An add of negative one (subtract of one) will be selected as a DEC.
12818 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12819 Opcode = X86ISD::DEC;
12825 // Otherwise use a regular EFLAGS-setting add.
12826 Opcode = X86ISD::ADD;
12831 // If we have a constant logical shift that's only used in a comparison
12832 // against zero turn it into an equivalent AND. This allows turning it into
12833 // a TEST instruction later.
12834 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12835 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12836 EVT VT = Op.getValueType();
12837 unsigned BitWidth = VT.getSizeInBits();
12838 unsigned ShAmt = Op->getConstantOperandVal(1);
12839 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12841 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12842 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12843 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12844 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12846 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12847 DAG.getConstant(Mask, dl, VT));
12848 DAG.ReplaceAllUsesWith(Op, New);
12854 // If the primary and result isn't used, don't bother using X86ISD::AND,
12855 // because a TEST instruction will be better.
12856 if (!hasNonFlagsUse(Op))
12862 // Due to the ISEL shortcoming noted above, be conservative if this op is
12863 // likely to be selected as part of a load-modify-store instruction.
12864 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12865 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12866 if (UI->getOpcode() == ISD::STORE)
12869 // Otherwise use a regular EFLAGS-setting instruction.
12870 switch (ArithOp.getOpcode()) {
12871 default: llvm_unreachable("unexpected operator!");
12872 case ISD::SUB: Opcode = X86ISD::SUB; break;
12873 case ISD::XOR: Opcode = X86ISD::XOR; break;
12874 case ISD::AND: Opcode = X86ISD::AND; break;
12876 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12877 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12878 if (EFLAGS.getNode())
12881 Opcode = X86ISD::OR;
12895 return SDValue(Op.getNode(), 1);
12901 // If we found that truncation is beneficial, perform the truncation and
12903 if (NeedTruncation) {
12904 EVT VT = Op.getValueType();
12905 SDValue WideVal = Op->getOperand(0);
12906 EVT WideVT = WideVal.getValueType();
12907 unsigned ConvertedOp = 0;
12908 // Use a target machine opcode to prevent further DAGCombine
12909 // optimizations that may separate the arithmetic operations
12910 // from the setcc node.
12911 switch (WideVal.getOpcode()) {
12913 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12914 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12915 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12916 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12917 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12921 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12922 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12923 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12924 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12925 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12931 // Emit a CMP with 0, which is the TEST pattern.
12932 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12933 DAG.getConstant(0, dl, Op.getValueType()));
12935 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12936 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12938 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12939 DAG.ReplaceAllUsesWith(Op, New);
12940 return SDValue(New.getNode(), 1);
12943 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12945 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12946 SDLoc dl, SelectionDAG &DAG) const {
12947 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12948 if (C->getAPIntValue() == 0)
12949 return EmitTest(Op0, X86CC, dl, DAG);
12951 if (Op0.getValueType() == MVT::i1)
12952 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12955 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12956 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12957 // Do the comparison at i32 if it's smaller, besides the Atom case.
12958 // This avoids subregister aliasing issues. Keep the smaller reference
12959 // if we're optimizing for size, however, as that'll allow better folding
12960 // of memory operations.
12961 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12962 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12963 Attribute::MinSize) &&
12964 !Subtarget->isAtom()) {
12965 unsigned ExtendOp =
12966 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12967 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12968 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12970 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12971 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12972 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12974 return SDValue(Sub.getNode(), 1);
12976 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12979 /// Convert a comparison if required by the subtarget.
12980 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12981 SelectionDAG &DAG) const {
12982 // If the subtarget does not support the FUCOMI instruction, floating-point
12983 // comparisons have to be converted.
12984 if (Subtarget->hasCMov() ||
12985 Cmp.getOpcode() != X86ISD::CMP ||
12986 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12987 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12990 // The instruction selector will select an FUCOM instruction instead of
12991 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12992 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12993 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12995 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12996 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12997 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12998 DAG.getConstant(8, dl, MVT::i8));
12999 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13000 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13003 /// The minimum architected relative accuracy is 2^-12. We need one
13004 /// Newton-Raphson step to have a good float result (24 bits of precision).
13005 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13006 DAGCombinerInfo &DCI,
13007 unsigned &RefinementSteps,
13008 bool &UseOneConstNR) const {
13009 // FIXME: We should use instruction latency models to calculate the cost of
13010 // each potential sequence, but this is very hard to do reliably because
13011 // at least Intel's Core* chips have variable timing based on the number of
13012 // significant digits in the divisor and/or sqrt operand.
13013 if (!Subtarget->useSqrtEst())
13016 EVT VT = Op.getValueType();
13018 // SSE1 has rsqrtss and rsqrtps.
13019 // TODO: Add support for AVX512 (v16f32).
13020 // It is likely not profitable to do this for f64 because a double-precision
13021 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13022 // instructions: convert to single, rsqrtss, convert back to double, refine
13023 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13024 // along with FMA, this could be a throughput win.
13025 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
13026 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
13027 RefinementSteps = 1;
13028 UseOneConstNR = false;
13029 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13034 /// The minimum architected relative accuracy is 2^-12. We need one
13035 /// Newton-Raphson step to have a good float result (24 bits of precision).
13036 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13037 DAGCombinerInfo &DCI,
13038 unsigned &RefinementSteps) const {
13039 // FIXME: We should use instruction latency models to calculate the cost of
13040 // each potential sequence, but this is very hard to do reliably because
13041 // at least Intel's Core* chips have variable timing based on the number of
13042 // significant digits in the divisor.
13043 if (!Subtarget->useReciprocalEst())
13046 EVT VT = Op.getValueType();
13048 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13049 // TODO: Add support for AVX512 (v16f32).
13050 // It is likely not profitable to do this for f64 because a double-precision
13051 // reciprocal estimate with refinement on x86 prior to FMA requires
13052 // 15 instructions: convert to single, rcpss, convert back to double, refine
13053 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13054 // along with FMA, this could be a throughput win.
13055 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
13056 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
13057 RefinementSteps = ReciprocalEstimateRefinementSteps;
13058 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13063 /// If we have at least two divisions that use the same divisor, convert to
13064 /// multplication by a reciprocal. This may need to be adjusted for a given
13065 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13066 /// This is because we still need one division to calculate the reciprocal and
13067 /// then we need two multiplies by that reciprocal as replacements for the
13068 /// original divisions.
13069 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
13070 return NumUsers > 1;
13073 static bool isAllOnes(SDValue V) {
13074 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13075 return C && C->isAllOnesValue();
13078 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13079 /// if it's possible.
13080 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13081 SDLoc dl, SelectionDAG &DAG) const {
13082 SDValue Op0 = And.getOperand(0);
13083 SDValue Op1 = And.getOperand(1);
13084 if (Op0.getOpcode() == ISD::TRUNCATE)
13085 Op0 = Op0.getOperand(0);
13086 if (Op1.getOpcode() == ISD::TRUNCATE)
13087 Op1 = Op1.getOperand(0);
13090 if (Op1.getOpcode() == ISD::SHL)
13091 std::swap(Op0, Op1);
13092 if (Op0.getOpcode() == ISD::SHL) {
13093 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13094 if (And00C->getZExtValue() == 1) {
13095 // If we looked past a truncate, check that it's only truncating away
13097 unsigned BitWidth = Op0.getValueSizeInBits();
13098 unsigned AndBitWidth = And.getValueSizeInBits();
13099 if (BitWidth > AndBitWidth) {
13101 DAG.computeKnownBits(Op0, Zeros, Ones);
13102 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13106 RHS = Op0.getOperand(1);
13108 } else if (Op1.getOpcode() == ISD::Constant) {
13109 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13110 uint64_t AndRHSVal = AndRHS->getZExtValue();
13111 SDValue AndLHS = Op0;
13113 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13114 LHS = AndLHS.getOperand(0);
13115 RHS = AndLHS.getOperand(1);
13118 // Use BT if the immediate can't be encoded in a TEST instruction.
13119 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13121 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13125 if (LHS.getNode()) {
13126 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13127 // instruction. Since the shift amount is in-range-or-undefined, we know
13128 // that doing a bittest on the i32 value is ok. We extend to i32 because
13129 // the encoding for the i16 version is larger than the i32 version.
13130 // Also promote i16 to i32 for performance / code size reason.
13131 if (LHS.getValueType() == MVT::i8 ||
13132 LHS.getValueType() == MVT::i16)
13133 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13135 // If the operand types disagree, extend the shift amount to match. Since
13136 // BT ignores high bits (like shifts) we can use anyextend.
13137 if (LHS.getValueType() != RHS.getValueType())
13138 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13140 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13141 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13142 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13143 DAG.getConstant(Cond, dl, MVT::i8), BT);
13149 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13151 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13156 // SSE Condition code mapping:
13165 switch (SetCCOpcode) {
13166 default: llvm_unreachable("Unexpected SETCC condition");
13168 case ISD::SETEQ: SSECC = 0; break;
13170 case ISD::SETGT: Swap = true; // Fallthrough
13172 case ISD::SETOLT: SSECC = 1; break;
13174 case ISD::SETGE: Swap = true; // Fallthrough
13176 case ISD::SETOLE: SSECC = 2; break;
13177 case ISD::SETUO: SSECC = 3; break;
13179 case ISD::SETNE: SSECC = 4; break;
13180 case ISD::SETULE: Swap = true; // Fallthrough
13181 case ISD::SETUGE: SSECC = 5; break;
13182 case ISD::SETULT: Swap = true; // Fallthrough
13183 case ISD::SETUGT: SSECC = 6; break;
13184 case ISD::SETO: SSECC = 7; break;
13186 case ISD::SETONE: SSECC = 8; break;
13189 std::swap(Op0, Op1);
13194 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13195 // ones, and then concatenate the result back.
13196 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13197 MVT VT = Op.getSimpleValueType();
13199 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13200 "Unsupported value type for operation");
13202 unsigned NumElems = VT.getVectorNumElements();
13204 SDValue CC = Op.getOperand(2);
13206 // Extract the LHS vectors
13207 SDValue LHS = Op.getOperand(0);
13208 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13209 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13211 // Extract the RHS vectors
13212 SDValue RHS = Op.getOperand(1);
13213 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13214 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13216 // Issue the operation on the smaller types and concatenate the result back
13217 MVT EltVT = VT.getVectorElementType();
13218 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13219 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13220 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13221 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13224 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13225 SDValue Op0 = Op.getOperand(0);
13226 SDValue Op1 = Op.getOperand(1);
13227 SDValue CC = Op.getOperand(2);
13228 MVT VT = Op.getSimpleValueType();
13231 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13232 "Unexpected type for boolean compare operation");
13233 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13234 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13235 DAG.getConstant(-1, dl, VT));
13236 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13237 DAG.getConstant(-1, dl, VT));
13238 switch (SetCCOpcode) {
13239 default: llvm_unreachable("Unexpected SETCC condition");
13241 // (x != y) -> ~(x ^ y)
13242 return DAG.getNode(ISD::XOR, dl, VT,
13243 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13244 DAG.getConstant(-1, dl, VT));
13246 // (x == y) -> (x ^ y)
13247 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13250 // (x > y) -> (x & ~y)
13251 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13254 // (x < y) -> (~x & y)
13255 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13258 // (x <= y) -> (~x | y)
13259 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13262 // (x >=y) -> (x | ~y)
13263 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13267 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13268 const X86Subtarget *Subtarget) {
13269 SDValue Op0 = Op.getOperand(0);
13270 SDValue Op1 = Op.getOperand(1);
13271 SDValue CC = Op.getOperand(2);
13272 MVT VT = Op.getSimpleValueType();
13275 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13276 Op.getValueType().getScalarType() == MVT::i1 &&
13277 "Cannot set masked compare for this operation");
13279 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13281 bool Unsigned = false;
13284 switch (SetCCOpcode) {
13285 default: llvm_unreachable("Unexpected SETCC condition");
13286 case ISD::SETNE: SSECC = 4; break;
13287 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13288 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13289 case ISD::SETLT: Swap = true; //fall-through
13290 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13291 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13292 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13293 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13294 case ISD::SETULE: Unsigned = true; //fall-through
13295 case ISD::SETLE: SSECC = 2; break;
13299 std::swap(Op0, Op1);
13301 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13302 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13303 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13304 DAG.getConstant(SSECC, dl, MVT::i8));
13307 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13308 /// operand \p Op1. If non-trivial (for example because it's not constant)
13309 /// return an empty value.
13310 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13312 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13316 MVT VT = Op1.getSimpleValueType();
13317 MVT EVT = VT.getVectorElementType();
13318 unsigned n = VT.getVectorNumElements();
13319 SmallVector<SDValue, 8> ULTOp1;
13321 for (unsigned i = 0; i < n; ++i) {
13322 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13323 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13326 // Avoid underflow.
13327 APInt Val = Elt->getAPIntValue();
13331 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13334 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13337 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13338 SelectionDAG &DAG) {
13339 SDValue Op0 = Op.getOperand(0);
13340 SDValue Op1 = Op.getOperand(1);
13341 SDValue CC = Op.getOperand(2);
13342 MVT VT = Op.getSimpleValueType();
13343 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13344 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13349 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13350 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13353 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13354 unsigned Opc = X86ISD::CMPP;
13355 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13356 assert(VT.getVectorNumElements() <= 16);
13357 Opc = X86ISD::CMPM;
13359 // In the two special cases we can't handle, emit two comparisons.
13362 unsigned CombineOpc;
13363 if (SetCCOpcode == ISD::SETUEQ) {
13364 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13366 assert(SetCCOpcode == ISD::SETONE);
13367 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13370 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13371 DAG.getConstant(CC0, dl, MVT::i8));
13372 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13373 DAG.getConstant(CC1, dl, MVT::i8));
13374 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13376 // Handle all other FP comparisons here.
13377 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13378 DAG.getConstant(SSECC, dl, MVT::i8));
13381 // Break 256-bit integer vector compare into smaller ones.
13382 if (VT.is256BitVector() && !Subtarget->hasInt256())
13383 return Lower256IntVSETCC(Op, DAG);
13385 EVT OpVT = Op1.getValueType();
13386 if (OpVT.getVectorElementType() == MVT::i1)
13387 return LowerBoolVSETCC_AVX512(Op, DAG);
13389 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13390 if (Subtarget->hasAVX512()) {
13391 if (Op1.getValueType().is512BitVector() ||
13392 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13393 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13394 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13396 // In AVX-512 architecture setcc returns mask with i1 elements,
13397 // But there is no compare instruction for i8 and i16 elements in KNL.
13398 // We are not talking about 512-bit operands in this case, these
13399 // types are illegal.
13401 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13402 OpVT.getVectorElementType().getSizeInBits() >= 8))
13403 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13404 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13407 // We are handling one of the integer comparisons here. Since SSE only has
13408 // GT and EQ comparisons for integer, swapping operands and multiple
13409 // operations may be required for some comparisons.
13411 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13412 bool Subus = false;
13414 switch (SetCCOpcode) {
13415 default: llvm_unreachable("Unexpected SETCC condition");
13416 case ISD::SETNE: Invert = true;
13417 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13418 case ISD::SETLT: Swap = true;
13419 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13420 case ISD::SETGE: Swap = true;
13421 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13422 Invert = true; break;
13423 case ISD::SETULT: Swap = true;
13424 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13425 FlipSigns = true; break;
13426 case ISD::SETUGE: Swap = true;
13427 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13428 FlipSigns = true; Invert = true; break;
13431 // Special case: Use min/max operations for SETULE/SETUGE
13432 MVT VET = VT.getVectorElementType();
13434 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13435 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13438 switch (SetCCOpcode) {
13440 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
13441 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
13444 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13447 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13448 if (!MinMax && hasSubus) {
13449 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13451 // t = psubus Op0, Op1
13452 // pcmpeq t, <0..0>
13453 switch (SetCCOpcode) {
13455 case ISD::SETULT: {
13456 // If the comparison is against a constant we can turn this into a
13457 // setule. With psubus, setule does not require a swap. This is
13458 // beneficial because the constant in the register is no longer
13459 // destructed as the destination so it can be hoisted out of a loop.
13460 // Only do this pre-AVX since vpcmp* is no longer destructive.
13461 if (Subtarget->hasAVX())
13463 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13464 if (ULEOp1.getNode()) {
13466 Subus = true; Invert = false; Swap = false;
13470 // Psubus is better than flip-sign because it requires no inversion.
13471 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13472 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13476 Opc = X86ISD::SUBUS;
13482 std::swap(Op0, Op1);
13484 // Check that the operation in question is available (most are plain SSE2,
13485 // but PCMPGTQ and PCMPEQQ have different requirements).
13486 if (VT == MVT::v2i64) {
13487 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13488 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13490 // First cast everything to the right type.
13491 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13492 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13494 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13495 // bits of the inputs before performing those operations. The lower
13496 // compare is always unsigned.
13499 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13501 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13502 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13503 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13504 Sign, Zero, Sign, Zero);
13506 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13507 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13509 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13510 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13511 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13513 // Create masks for only the low parts/high parts of the 64 bit integers.
13514 static const int MaskHi[] = { 1, 1, 3, 3 };
13515 static const int MaskLo[] = { 0, 0, 2, 2 };
13516 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13517 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13518 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13520 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13521 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13524 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13526 return DAG.getBitcast(VT, Result);
13529 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13530 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13531 // pcmpeqd + pshufd + pand.
13532 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13534 // First cast everything to the right type.
13535 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13536 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13539 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13541 // Make sure the lower and upper halves are both all-ones.
13542 static const int Mask[] = { 1, 0, 3, 2 };
13543 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13544 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13547 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13549 return DAG.getBitcast(VT, Result);
13553 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13554 // bits of the inputs before performing those operations.
13556 EVT EltVT = VT.getVectorElementType();
13557 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13559 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13560 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13563 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13565 // If the logical-not of the result is required, perform that now.
13567 Result = DAG.getNOT(dl, Result, VT);
13570 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13573 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13574 getZeroVector(VT, Subtarget, DAG, dl));
13579 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13581 MVT VT = Op.getSimpleValueType();
13583 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13585 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13586 && "SetCC type must be 8-bit or 1-bit integer");
13587 SDValue Op0 = Op.getOperand(0);
13588 SDValue Op1 = Op.getOperand(1);
13590 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13592 // Optimize to BT if possible.
13593 // Lower (X & (1 << N)) == 0 to BT(X, N).
13594 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13595 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13596 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13597 Op1.getOpcode() == ISD::Constant &&
13598 cast<ConstantSDNode>(Op1)->isNullValue() &&
13599 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13600 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13601 if (NewSetCC.getNode()) {
13603 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13608 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13610 if (Op1.getOpcode() == ISD::Constant &&
13611 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13612 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13613 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13615 // If the input is a setcc, then reuse the input setcc or use a new one with
13616 // the inverted condition.
13617 if (Op0.getOpcode() == X86ISD::SETCC) {
13618 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13619 bool Invert = (CC == ISD::SETNE) ^
13620 cast<ConstantSDNode>(Op1)->isNullValue();
13624 CCode = X86::GetOppositeBranchCondition(CCode);
13625 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13626 DAG.getConstant(CCode, dl, MVT::i8),
13627 Op0.getOperand(1));
13629 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13633 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13634 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13635 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13637 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13638 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13641 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13642 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13643 if (X86CC == X86::COND_INVALID)
13646 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13647 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13648 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13649 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13651 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13655 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13656 static bool isX86LogicalCmp(SDValue Op) {
13657 unsigned Opc = Op.getNode()->getOpcode();
13658 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13659 Opc == X86ISD::SAHF)
13661 if (Op.getResNo() == 1 &&
13662 (Opc == X86ISD::ADD ||
13663 Opc == X86ISD::SUB ||
13664 Opc == X86ISD::ADC ||
13665 Opc == X86ISD::SBB ||
13666 Opc == X86ISD::SMUL ||
13667 Opc == X86ISD::UMUL ||
13668 Opc == X86ISD::INC ||
13669 Opc == X86ISD::DEC ||
13670 Opc == X86ISD::OR ||
13671 Opc == X86ISD::XOR ||
13672 Opc == X86ISD::AND))
13675 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13681 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13682 if (V.getOpcode() != ISD::TRUNCATE)
13685 SDValue VOp0 = V.getOperand(0);
13686 unsigned InBits = VOp0.getValueSizeInBits();
13687 unsigned Bits = V.getValueSizeInBits();
13688 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13691 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13692 bool addTest = true;
13693 SDValue Cond = Op.getOperand(0);
13694 SDValue Op1 = Op.getOperand(1);
13695 SDValue Op2 = Op.getOperand(2);
13697 EVT VT = Op1.getValueType();
13700 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13701 // are available or VBLENDV if AVX is available.
13702 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13703 if (Cond.getOpcode() == ISD::SETCC &&
13704 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13705 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13706 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13707 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13708 int SSECC = translateX86FSETCC(
13709 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13712 if (Subtarget->hasAVX512()) {
13713 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13714 DAG.getConstant(SSECC, DL, MVT::i8));
13715 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13718 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13719 DAG.getConstant(SSECC, DL, MVT::i8));
13721 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13722 // of 3 logic instructions for size savings and potentially speed.
13723 // Unfortunately, there is no scalar form of VBLENDV.
13725 // If either operand is a constant, don't try this. We can expect to
13726 // optimize away at least one of the logic instructions later in that
13727 // case, so that sequence would be faster than a variable blend.
13729 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13730 // uses XMM0 as the selection register. That may need just as many
13731 // instructions as the AND/ANDN/OR sequence due to register moves, so
13734 if (Subtarget->hasAVX() &&
13735 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13737 // Convert to vectors, do a VSELECT, and convert back to scalar.
13738 // All of the conversions should be optimized away.
13740 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13741 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13742 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13743 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13745 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13746 VCmp = DAG.getBitcast(VCmpVT, VCmp);
13748 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13750 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13751 VSel, DAG.getIntPtrConstant(0, DL));
13753 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13754 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13755 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13759 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
13761 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
13762 Op1Scalar = ConvertI1VectorToInterger(Op1, DAG);
13763 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
13764 Op1Scalar = Op1.getOperand(0);
13766 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
13767 Op2Scalar = ConvertI1VectorToInterger(Op2, DAG);
13768 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
13769 Op2Scalar = Op2.getOperand(0);
13770 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
13771 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
13772 Op1Scalar.getValueType(),
13773 Cond, Op1Scalar, Op2Scalar);
13774 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
13775 return DAG.getBitcast(VT, newSelect);
13776 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
13777 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
13778 DAG.getIntPtrConstant(0, DL));
13782 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
13783 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
13784 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13785 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
13786 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13787 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
13788 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
13790 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
13793 if (Cond.getOpcode() == ISD::SETCC) {
13794 SDValue NewCond = LowerSETCC(Cond, DAG);
13795 if (NewCond.getNode())
13799 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13800 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13801 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13802 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13803 if (Cond.getOpcode() == X86ISD::SETCC &&
13804 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13805 isZero(Cond.getOperand(1).getOperand(1))) {
13806 SDValue Cmp = Cond.getOperand(1);
13808 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13810 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13811 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13812 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13814 SDValue CmpOp0 = Cmp.getOperand(0);
13815 // Apply further optimizations for special cases
13816 // (select (x != 0), -1, 0) -> neg & sbb
13817 // (select (x == 0), 0, -1) -> neg & sbb
13818 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13819 if (YC->isNullValue() &&
13820 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13821 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13822 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13823 DAG.getConstant(0, DL,
13824 CmpOp0.getValueType()),
13826 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13827 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13828 SDValue(Neg.getNode(), 1));
13832 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13833 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
13834 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13836 SDValue Res = // Res = 0 or -1.
13837 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13838 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
13840 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13841 Res = DAG.getNOT(DL, Res, Res.getValueType());
13843 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13844 if (!N2C || !N2C->isNullValue())
13845 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13850 // Look past (and (setcc_carry (cmp ...)), 1).
13851 if (Cond.getOpcode() == ISD::AND &&
13852 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13853 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13854 if (C && C->getAPIntValue() == 1)
13855 Cond = Cond.getOperand(0);
13858 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13859 // setting operand in place of the X86ISD::SETCC.
13860 unsigned CondOpcode = Cond.getOpcode();
13861 if (CondOpcode == X86ISD::SETCC ||
13862 CondOpcode == X86ISD::SETCC_CARRY) {
13863 CC = Cond.getOperand(0);
13865 SDValue Cmp = Cond.getOperand(1);
13866 unsigned Opc = Cmp.getOpcode();
13867 MVT VT = Op.getSimpleValueType();
13869 bool IllegalFPCMov = false;
13870 if (VT.isFloatingPoint() && !VT.isVector() &&
13871 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13872 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13874 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13875 Opc == X86ISD::BT) { // FIXME
13879 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13880 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13881 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13882 Cond.getOperand(0).getValueType() != MVT::i8)) {
13883 SDValue LHS = Cond.getOperand(0);
13884 SDValue RHS = Cond.getOperand(1);
13885 unsigned X86Opcode;
13888 switch (CondOpcode) {
13889 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13890 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13891 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13892 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13893 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13894 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13895 default: llvm_unreachable("unexpected overflowing operator");
13897 if (CondOpcode == ISD::UMULO)
13898 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13901 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13903 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13905 if (CondOpcode == ISD::UMULO)
13906 Cond = X86Op.getValue(2);
13908 Cond = X86Op.getValue(1);
13910 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
13915 // Look pass the truncate if the high bits are known zero.
13916 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13917 Cond = Cond.getOperand(0);
13919 // We know the result of AND is compared against zero. Try to match
13921 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13922 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13923 if (NewSetCC.getNode()) {
13924 CC = NewSetCC.getOperand(0);
13925 Cond = NewSetCC.getOperand(1);
13932 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
13933 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13936 // a < b ? -1 : 0 -> RES = ~setcc_carry
13937 // a < b ? 0 : -1 -> RES = setcc_carry
13938 // a >= b ? -1 : 0 -> RES = setcc_carry
13939 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13940 if (Cond.getOpcode() == X86ISD::SUB) {
13941 Cond = ConvertCmpIfNecessary(Cond, DAG);
13942 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13944 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13945 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13946 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13947 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13949 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13950 return DAG.getNOT(DL, Res, Res.getValueType());
13955 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13956 // widen the cmov and push the truncate through. This avoids introducing a new
13957 // branch during isel and doesn't add any extensions.
13958 if (Op.getValueType() == MVT::i8 &&
13959 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13960 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13961 if (T1.getValueType() == T2.getValueType() &&
13962 // Blacklist CopyFromReg to avoid partial register stalls.
13963 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13964 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13965 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13966 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13970 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13971 // condition is true.
13972 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13973 SDValue Ops[] = { Op2, Op1, CC, Cond };
13974 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13977 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
13978 const X86Subtarget *Subtarget,
13979 SelectionDAG &DAG) {
13980 MVT VT = Op->getSimpleValueType(0);
13981 SDValue In = Op->getOperand(0);
13982 MVT InVT = In.getSimpleValueType();
13983 MVT VTElt = VT.getVectorElementType();
13984 MVT InVTElt = InVT.getVectorElementType();
13988 if ((InVTElt == MVT::i1) &&
13989 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13990 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13992 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13993 VTElt.getSizeInBits() <= 16)) ||
13995 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13996 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13998 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13999 VTElt.getSizeInBits() >= 32))))
14000 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14002 unsigned int NumElts = VT.getVectorNumElements();
14004 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14007 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14008 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14009 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14010 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14013 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14014 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14016 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14019 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14021 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14022 if (VT.is512BitVector())
14024 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14027 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14028 const X86Subtarget *Subtarget,
14029 SelectionDAG &DAG) {
14030 SDValue In = Op->getOperand(0);
14031 MVT VT = Op->getSimpleValueType(0);
14032 MVT InVT = In.getSimpleValueType();
14033 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14035 MVT InSVT = InVT.getScalarType();
14036 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14038 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14040 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14045 // SSE41 targets can use the pmovsx* instructions directly.
14046 if (Subtarget->hasSSE41())
14047 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14049 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14053 // As SRAI is only available on i16/i32 types, we expand only up to i32
14054 // and handle i64 separately.
14055 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14056 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14057 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14058 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14059 Curr = DAG.getBitcast(CurrVT, Curr);
14062 SDValue SignExt = Curr;
14063 if (CurrVT != InVT) {
14064 unsigned SignExtShift =
14065 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14066 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14067 DAG.getConstant(SignExtShift, dl, MVT::i8));
14073 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14074 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14075 DAG.getConstant(31, dl, MVT::i8));
14076 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14077 return DAG.getBitcast(VT, Ext);
14083 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14084 SelectionDAG &DAG) {
14085 MVT VT = Op->getSimpleValueType(0);
14086 SDValue In = Op->getOperand(0);
14087 MVT InVT = In.getSimpleValueType();
14090 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14091 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14093 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14094 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14095 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14098 if (Subtarget->hasInt256())
14099 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14101 // Optimize vectors in AVX mode
14102 // Sign extend v8i16 to v8i32 and
14105 // Divide input vector into two parts
14106 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14107 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14108 // concat the vectors to original VT
14110 unsigned NumElems = InVT.getVectorNumElements();
14111 SDValue Undef = DAG.getUNDEF(InVT);
14113 SmallVector<int,8> ShufMask1(NumElems, -1);
14114 for (unsigned i = 0; i != NumElems/2; ++i)
14117 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14119 SmallVector<int,8> ShufMask2(NumElems, -1);
14120 for (unsigned i = 0; i != NumElems/2; ++i)
14121 ShufMask2[i] = i + NumElems/2;
14123 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14125 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14126 VT.getVectorNumElements()/2);
14128 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14129 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14131 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14134 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14135 // may emit an illegal shuffle but the expansion is still better than scalar
14136 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14137 // we'll emit a shuffle and a arithmetic shift.
14138 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14139 // TODO: It is possible to support ZExt by zeroing the undef values during
14140 // the shuffle phase or after the shuffle.
14141 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14142 SelectionDAG &DAG) {
14143 MVT RegVT = Op.getSimpleValueType();
14144 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14145 assert(RegVT.isInteger() &&
14146 "We only custom lower integer vector sext loads.");
14148 // Nothing useful we can do without SSE2 shuffles.
14149 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14151 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14153 EVT MemVT = Ld->getMemoryVT();
14154 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14155 unsigned RegSz = RegVT.getSizeInBits();
14157 ISD::LoadExtType Ext = Ld->getExtensionType();
14159 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14160 && "Only anyext and sext are currently implemented.");
14161 assert(MemVT != RegVT && "Cannot extend to the same type");
14162 assert(MemVT.isVector() && "Must load a vector from memory");
14164 unsigned NumElems = RegVT.getVectorNumElements();
14165 unsigned MemSz = MemVT.getSizeInBits();
14166 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14168 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14169 // The only way in which we have a legal 256-bit vector result but not the
14170 // integer 256-bit operations needed to directly lower a sextload is if we
14171 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14172 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14173 // correctly legalized. We do this late to allow the canonical form of
14174 // sextload to persist throughout the rest of the DAG combiner -- it wants
14175 // to fold together any extensions it can, and so will fuse a sign_extend
14176 // of an sextload into a sextload targeting a wider value.
14178 if (MemSz == 128) {
14179 // Just switch this to a normal load.
14180 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14181 "it must be a legal 128-bit vector "
14183 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14184 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14185 Ld->isInvariant(), Ld->getAlignment());
14187 assert(MemSz < 128 &&
14188 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14189 // Do an sext load to a 128-bit vector type. We want to use the same
14190 // number of elements, but elements half as wide. This will end up being
14191 // recursively lowered by this routine, but will succeed as we definitely
14192 // have all the necessary features if we're using AVX1.
14194 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14195 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14197 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14198 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14199 Ld->isNonTemporal(), Ld->isInvariant(),
14200 Ld->getAlignment());
14203 // Replace chain users with the new chain.
14204 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14205 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14207 // Finally, do a normal sign-extend to the desired register.
14208 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14211 // All sizes must be a power of two.
14212 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14213 "Non-power-of-two elements are not custom lowered!");
14215 // Attempt to load the original value using scalar loads.
14216 // Find the largest scalar type that divides the total loaded size.
14217 MVT SclrLoadTy = MVT::i8;
14218 for (MVT Tp : MVT::integer_valuetypes()) {
14219 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14224 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14225 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14227 SclrLoadTy = MVT::f64;
14229 // Calculate the number of scalar loads that we need to perform
14230 // in order to load our vector from memory.
14231 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14233 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14234 "Can only lower sext loads with a single scalar load!");
14236 unsigned loadRegZize = RegSz;
14237 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14240 // Represent our vector as a sequence of elements which are the
14241 // largest scalar that we can load.
14242 EVT LoadUnitVecVT = EVT::getVectorVT(
14243 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14245 // Represent the data using the same element type that is stored in
14246 // memory. In practice, we ''widen'' MemVT.
14248 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14249 loadRegZize / MemVT.getScalarType().getSizeInBits());
14251 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14252 "Invalid vector type");
14254 // We can't shuffle using an illegal type.
14255 assert(TLI.isTypeLegal(WideVecVT) &&
14256 "We only lower types that form legal widened vector types");
14258 SmallVector<SDValue, 8> Chains;
14259 SDValue Ptr = Ld->getBasePtr();
14260 SDValue Increment =
14261 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl, TLI.getPointerTy());
14262 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14264 for (unsigned i = 0; i < NumLoads; ++i) {
14265 // Perform a single load.
14266 SDValue ScalarLoad =
14267 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14268 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14269 Ld->getAlignment());
14270 Chains.push_back(ScalarLoad.getValue(1));
14271 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14272 // another round of DAGCombining.
14274 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14276 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14277 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14279 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14282 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14284 // Bitcast the loaded value to a vector of the original element type, in
14285 // the size of the target vector type.
14286 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14287 unsigned SizeRatio = RegSz / MemSz;
14289 if (Ext == ISD::SEXTLOAD) {
14290 // If we have SSE4.1, we can directly emit a VSEXT node.
14291 if (Subtarget->hasSSE41()) {
14292 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14293 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14297 // Otherwise we'll shuffle the small elements in the high bits of the
14298 // larger type and perform an arithmetic shift. If the shift is not legal
14299 // it's better to scalarize.
14300 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14301 "We can't implement a sext load without an arithmetic right shift!");
14303 // Redistribute the loaded elements into the different locations.
14304 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14305 for (unsigned i = 0; i != NumElems; ++i)
14306 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14308 SDValue Shuff = DAG.getVectorShuffle(
14309 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14311 Shuff = DAG.getBitcast(RegVT, Shuff);
14313 // Build the arithmetic shift.
14314 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14315 MemVT.getVectorElementType().getSizeInBits();
14317 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14318 DAG.getConstant(Amt, dl, RegVT));
14320 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14324 // Redistribute the loaded elements into the different locations.
14325 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14326 for (unsigned i = 0; i != NumElems; ++i)
14327 ShuffleVec[i * SizeRatio] = i;
14329 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14330 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14332 // Bitcast to the requested type.
14333 Shuff = DAG.getBitcast(RegVT, Shuff);
14334 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14338 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14339 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14340 // from the AND / OR.
14341 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14342 Opc = Op.getOpcode();
14343 if (Opc != ISD::OR && Opc != ISD::AND)
14345 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14346 Op.getOperand(0).hasOneUse() &&
14347 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14348 Op.getOperand(1).hasOneUse());
14351 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14352 // 1 and that the SETCC node has a single use.
14353 static bool isXor1OfSetCC(SDValue Op) {
14354 if (Op.getOpcode() != ISD::XOR)
14356 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14357 if (N1C && N1C->getAPIntValue() == 1) {
14358 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14359 Op.getOperand(0).hasOneUse();
14364 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14365 bool addTest = true;
14366 SDValue Chain = Op.getOperand(0);
14367 SDValue Cond = Op.getOperand(1);
14368 SDValue Dest = Op.getOperand(2);
14371 bool Inverted = false;
14373 if (Cond.getOpcode() == ISD::SETCC) {
14374 // Check for setcc([su]{add,sub,mul}o == 0).
14375 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14376 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14377 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14378 Cond.getOperand(0).getResNo() == 1 &&
14379 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14380 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14381 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14382 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14383 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14384 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14386 Cond = Cond.getOperand(0);
14388 SDValue NewCond = LowerSETCC(Cond, DAG);
14389 if (NewCond.getNode())
14394 // FIXME: LowerXALUO doesn't handle these!!
14395 else if (Cond.getOpcode() == X86ISD::ADD ||
14396 Cond.getOpcode() == X86ISD::SUB ||
14397 Cond.getOpcode() == X86ISD::SMUL ||
14398 Cond.getOpcode() == X86ISD::UMUL)
14399 Cond = LowerXALUO(Cond, DAG);
14402 // Look pass (and (setcc_carry (cmp ...)), 1).
14403 if (Cond.getOpcode() == ISD::AND &&
14404 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14405 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14406 if (C && C->getAPIntValue() == 1)
14407 Cond = Cond.getOperand(0);
14410 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14411 // setting operand in place of the X86ISD::SETCC.
14412 unsigned CondOpcode = Cond.getOpcode();
14413 if (CondOpcode == X86ISD::SETCC ||
14414 CondOpcode == X86ISD::SETCC_CARRY) {
14415 CC = Cond.getOperand(0);
14417 SDValue Cmp = Cond.getOperand(1);
14418 unsigned Opc = Cmp.getOpcode();
14419 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14420 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14424 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14428 // These can only come from an arithmetic instruction with overflow,
14429 // e.g. SADDO, UADDO.
14430 Cond = Cond.getNode()->getOperand(1);
14436 CondOpcode = Cond.getOpcode();
14437 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14438 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14439 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14440 Cond.getOperand(0).getValueType() != MVT::i8)) {
14441 SDValue LHS = Cond.getOperand(0);
14442 SDValue RHS = Cond.getOperand(1);
14443 unsigned X86Opcode;
14446 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14447 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14449 switch (CondOpcode) {
14450 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14454 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14457 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14458 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14462 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14465 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14466 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14467 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14468 default: llvm_unreachable("unexpected overflowing operator");
14471 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14472 if (CondOpcode == ISD::UMULO)
14473 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14476 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14478 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14480 if (CondOpcode == ISD::UMULO)
14481 Cond = X86Op.getValue(2);
14483 Cond = X86Op.getValue(1);
14485 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14489 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14490 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14491 if (CondOpc == ISD::OR) {
14492 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14493 // two branches instead of an explicit OR instruction with a
14495 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14496 isX86LogicalCmp(Cmp)) {
14497 CC = Cond.getOperand(0).getOperand(0);
14498 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14499 Chain, Dest, CC, Cmp);
14500 CC = Cond.getOperand(1).getOperand(0);
14504 } else { // ISD::AND
14505 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14506 // two branches instead of an explicit AND instruction with a
14507 // separate test. However, we only do this if this block doesn't
14508 // have a fall-through edge, because this requires an explicit
14509 // jmp when the condition is false.
14510 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14511 isX86LogicalCmp(Cmp) &&
14512 Op.getNode()->hasOneUse()) {
14513 X86::CondCode CCode =
14514 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14515 CCode = X86::GetOppositeBranchCondition(CCode);
14516 CC = DAG.getConstant(CCode, dl, MVT::i8);
14517 SDNode *User = *Op.getNode()->use_begin();
14518 // Look for an unconditional branch following this conditional branch.
14519 // We need this because we need to reverse the successors in order
14520 // to implement FCMP_OEQ.
14521 if (User->getOpcode() == ISD::BR) {
14522 SDValue FalseBB = User->getOperand(1);
14524 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14525 assert(NewBR == User);
14529 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14530 Chain, Dest, CC, Cmp);
14531 X86::CondCode CCode =
14532 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14533 CCode = X86::GetOppositeBranchCondition(CCode);
14534 CC = DAG.getConstant(CCode, dl, MVT::i8);
14540 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14541 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14542 // It should be transformed during dag combiner except when the condition
14543 // is set by a arithmetics with overflow node.
14544 X86::CondCode CCode =
14545 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14546 CCode = X86::GetOppositeBranchCondition(CCode);
14547 CC = DAG.getConstant(CCode, dl, MVT::i8);
14548 Cond = Cond.getOperand(0).getOperand(1);
14550 } else if (Cond.getOpcode() == ISD::SETCC &&
14551 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14552 // For FCMP_OEQ, we can emit
14553 // two branches instead of an explicit AND instruction with a
14554 // separate test. However, we only do this if this block doesn't
14555 // have a fall-through edge, because this requires an explicit
14556 // jmp when the condition is false.
14557 if (Op.getNode()->hasOneUse()) {
14558 SDNode *User = *Op.getNode()->use_begin();
14559 // Look for an unconditional branch following this conditional branch.
14560 // We need this because we need to reverse the successors in order
14561 // to implement FCMP_OEQ.
14562 if (User->getOpcode() == ISD::BR) {
14563 SDValue FalseBB = User->getOperand(1);
14565 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14566 assert(NewBR == User);
14570 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14571 Cond.getOperand(0), Cond.getOperand(1));
14572 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14573 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14574 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14575 Chain, Dest, CC, Cmp);
14576 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14581 } else if (Cond.getOpcode() == ISD::SETCC &&
14582 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14583 // For FCMP_UNE, we can emit
14584 // two branches instead of an explicit AND instruction with a
14585 // separate test. However, we only do this if this block doesn't
14586 // have a fall-through edge, because this requires an explicit
14587 // jmp when the condition is false.
14588 if (Op.getNode()->hasOneUse()) {
14589 SDNode *User = *Op.getNode()->use_begin();
14590 // Look for an unconditional branch following this conditional branch.
14591 // We need this because we need to reverse the successors in order
14592 // to implement FCMP_UNE.
14593 if (User->getOpcode() == ISD::BR) {
14594 SDValue FalseBB = User->getOperand(1);
14596 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14597 assert(NewBR == User);
14600 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14601 Cond.getOperand(0), Cond.getOperand(1));
14602 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14603 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14604 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14605 Chain, Dest, CC, Cmp);
14606 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14616 // Look pass the truncate if the high bits are known zero.
14617 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14618 Cond = Cond.getOperand(0);
14620 // We know the result of AND is compared against zero. Try to match
14622 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14623 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14624 if (NewSetCC.getNode()) {
14625 CC = NewSetCC.getOperand(0);
14626 Cond = NewSetCC.getOperand(1);
14633 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14634 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14635 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14637 Cond = ConvertCmpIfNecessary(Cond, DAG);
14638 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14639 Chain, Dest, CC, Cond);
14642 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14643 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14644 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14645 // that the guard pages used by the OS virtual memory manager are allocated in
14646 // correct sequence.
14648 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14649 SelectionDAG &DAG) const {
14650 MachineFunction &MF = DAG.getMachineFunction();
14651 bool SplitStack = MF.shouldSplitStack();
14652 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14657 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14658 SDNode* Node = Op.getNode();
14660 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14661 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14662 " not tell us which reg is the stack pointer!");
14663 EVT VT = Node->getValueType(0);
14664 SDValue Tmp1 = SDValue(Node, 0);
14665 SDValue Tmp2 = SDValue(Node, 1);
14666 SDValue Tmp3 = Node->getOperand(2);
14667 SDValue Chain = Tmp1.getOperand(0);
14669 // Chain the dynamic stack allocation so that it doesn't modify the stack
14670 // pointer when other instructions are using the stack.
14671 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14674 SDValue Size = Tmp2.getOperand(1);
14675 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14676 Chain = SP.getValue(1);
14677 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14678 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14679 unsigned StackAlign = TFI.getStackAlignment();
14680 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14681 if (Align > StackAlign)
14682 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14683 DAG.getConstant(-(uint64_t)Align, dl, VT));
14684 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14686 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14687 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14690 SDValue Ops[2] = { Tmp1, Tmp2 };
14691 return DAG.getMergeValues(Ops, dl);
14695 SDValue Chain = Op.getOperand(0);
14696 SDValue Size = Op.getOperand(1);
14697 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14698 EVT VT = Op.getNode()->getValueType(0);
14700 bool Is64Bit = Subtarget->is64Bit();
14701 EVT SPTy = getPointerTy();
14704 MachineRegisterInfo &MRI = MF.getRegInfo();
14707 // The 64 bit implementation of segmented stacks needs to clobber both r10
14708 // r11. This makes it impossible to use it along with nested parameters.
14709 const Function *F = MF.getFunction();
14711 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14713 if (I->hasNestAttr())
14714 report_fatal_error("Cannot use segmented stacks with functions that "
14715 "have nested arguments.");
14718 const TargetRegisterClass *AddrRegClass =
14719 getRegClassFor(getPointerTy());
14720 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14721 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14722 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14723 DAG.getRegister(Vreg, SPTy));
14724 SDValue Ops1[2] = { Value, Chain };
14725 return DAG.getMergeValues(Ops1, dl);
14728 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14730 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14731 Flag = Chain.getValue(1);
14732 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14734 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14736 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14737 unsigned SPReg = RegInfo->getStackRegister();
14738 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14739 Chain = SP.getValue(1);
14742 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14743 DAG.getConstant(-(uint64_t)Align, dl, VT));
14744 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14747 SDValue Ops1[2] = { SP, Chain };
14748 return DAG.getMergeValues(Ops1, dl);
14752 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14753 MachineFunction &MF = DAG.getMachineFunction();
14754 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14756 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14759 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14760 // vastart just stores the address of the VarArgsFrameIndex slot into the
14761 // memory location argument.
14762 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14764 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14765 MachinePointerInfo(SV), false, false, 0);
14769 // gp_offset (0 - 6 * 8)
14770 // fp_offset (48 - 48 + 8 * 16)
14771 // overflow_arg_area (point to parameters coming in memory).
14773 SmallVector<SDValue, 8> MemOps;
14774 SDValue FIN = Op.getOperand(1);
14776 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14777 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14779 FIN, MachinePointerInfo(SV), false, false, 0);
14780 MemOps.push_back(Store);
14783 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14784 FIN, DAG.getIntPtrConstant(4, DL));
14785 Store = DAG.getStore(Op.getOperand(0), DL,
14786 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
14788 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14789 MemOps.push_back(Store);
14791 // Store ptr to overflow_arg_area
14792 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14793 FIN, DAG.getIntPtrConstant(4, DL));
14794 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14796 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14797 MachinePointerInfo(SV, 8),
14799 MemOps.push_back(Store);
14801 // Store ptr to reg_save_area.
14802 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14803 FIN, DAG.getIntPtrConstant(8, DL));
14804 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14806 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14807 MachinePointerInfo(SV, 16), false, false, 0);
14808 MemOps.push_back(Store);
14809 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14812 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14813 assert(Subtarget->is64Bit() &&
14814 "LowerVAARG only handles 64-bit va_arg!");
14815 assert((Subtarget->isTargetLinux() ||
14816 Subtarget->isTargetDarwin()) &&
14817 "Unhandled target in LowerVAARG");
14818 assert(Op.getNode()->getNumOperands() == 4);
14819 SDValue Chain = Op.getOperand(0);
14820 SDValue SrcPtr = Op.getOperand(1);
14821 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14822 unsigned Align = Op.getConstantOperandVal(3);
14825 EVT ArgVT = Op.getNode()->getValueType(0);
14826 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14827 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14830 // Decide which area this value should be read from.
14831 // TODO: Implement the AMD64 ABI in its entirety. This simple
14832 // selection mechanism works only for the basic types.
14833 if (ArgVT == MVT::f80) {
14834 llvm_unreachable("va_arg for f80 not yet implemented");
14835 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14836 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14837 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14838 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14840 llvm_unreachable("Unhandled argument type in LowerVAARG");
14843 if (ArgMode == 2) {
14844 // Sanity Check: Make sure using fp_offset makes sense.
14845 assert(!Subtarget->useSoftFloat() &&
14846 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14847 Attribute::NoImplicitFloat)) &&
14848 Subtarget->hasSSE1());
14851 // Insert VAARG_64 node into the DAG
14852 // VAARG_64 returns two values: Variable Argument Address, Chain
14853 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
14854 DAG.getConstant(ArgMode, dl, MVT::i8),
14855 DAG.getConstant(Align, dl, MVT::i32)};
14856 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14857 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14858 VTs, InstOps, MVT::i64,
14859 MachinePointerInfo(SV),
14861 /*Volatile=*/false,
14863 /*WriteMem=*/true);
14864 Chain = VAARG.getValue(1);
14866 // Load the next argument and return it
14867 return DAG.getLoad(ArgVT, dl,
14870 MachinePointerInfo(),
14871 false, false, false, 0);
14874 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14875 SelectionDAG &DAG) {
14876 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14877 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14878 SDValue Chain = Op.getOperand(0);
14879 SDValue DstPtr = Op.getOperand(1);
14880 SDValue SrcPtr = Op.getOperand(2);
14881 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14882 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14885 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14886 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
14888 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14891 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14892 // amount is a constant. Takes immediate version of shift as input.
14893 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14894 SDValue SrcOp, uint64_t ShiftAmt,
14895 SelectionDAG &DAG) {
14896 MVT ElementType = VT.getVectorElementType();
14898 // Fold this packed shift into its first operand if ShiftAmt is 0.
14902 // Check for ShiftAmt >= element width
14903 if (ShiftAmt >= ElementType.getSizeInBits()) {
14904 if (Opc == X86ISD::VSRAI)
14905 ShiftAmt = ElementType.getSizeInBits() - 1;
14907 return DAG.getConstant(0, dl, VT);
14910 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14911 && "Unknown target vector shift-by-constant node");
14913 // Fold this packed vector shift into a build vector if SrcOp is a
14914 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14915 if (VT == SrcOp.getSimpleValueType() &&
14916 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14917 SmallVector<SDValue, 8> Elts;
14918 unsigned NumElts = SrcOp->getNumOperands();
14919 ConstantSDNode *ND;
14922 default: llvm_unreachable(nullptr);
14923 case X86ISD::VSHLI:
14924 for (unsigned i=0; i!=NumElts; ++i) {
14925 SDValue CurrentOp = SrcOp->getOperand(i);
14926 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14927 Elts.push_back(CurrentOp);
14930 ND = cast<ConstantSDNode>(CurrentOp);
14931 const APInt &C = ND->getAPIntValue();
14932 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
14935 case X86ISD::VSRLI:
14936 for (unsigned i=0; i!=NumElts; ++i) {
14937 SDValue CurrentOp = SrcOp->getOperand(i);
14938 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14939 Elts.push_back(CurrentOp);
14942 ND = cast<ConstantSDNode>(CurrentOp);
14943 const APInt &C = ND->getAPIntValue();
14944 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
14947 case X86ISD::VSRAI:
14948 for (unsigned i=0; i!=NumElts; ++i) {
14949 SDValue CurrentOp = SrcOp->getOperand(i);
14950 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14951 Elts.push_back(CurrentOp);
14954 ND = cast<ConstantSDNode>(CurrentOp);
14955 const APInt &C = ND->getAPIntValue();
14956 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
14961 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14964 return DAG.getNode(Opc, dl, VT, SrcOp,
14965 DAG.getConstant(ShiftAmt, dl, MVT::i8));
14968 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14969 // may or may not be a constant. Takes immediate version of shift as input.
14970 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14971 SDValue SrcOp, SDValue ShAmt,
14972 SelectionDAG &DAG) {
14973 MVT SVT = ShAmt.getSimpleValueType();
14974 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14976 // Catch shift-by-constant.
14977 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14978 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14979 CShAmt->getZExtValue(), DAG);
14981 // Change opcode to non-immediate version
14983 default: llvm_unreachable("Unknown target vector shift node");
14984 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14985 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14986 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14989 const X86Subtarget &Subtarget =
14990 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14991 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14992 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14993 // Let the shuffle legalizer expand this shift amount node.
14994 SDValue Op0 = ShAmt.getOperand(0);
14995 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14996 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14998 // Need to build a vector containing shift amount.
14999 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15000 SmallVector<SDValue, 4> ShOps;
15001 ShOps.push_back(ShAmt);
15002 if (SVT == MVT::i32) {
15003 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15004 ShOps.push_back(DAG.getUNDEF(SVT));
15006 ShOps.push_back(DAG.getUNDEF(SVT));
15008 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15009 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15012 // The return type has to be a 128-bit type with the same element
15013 // type as the input type.
15014 MVT EltVT = VT.getVectorElementType();
15015 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15017 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15018 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15021 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15022 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15023 /// necessary casting for \p Mask when lowering masking intrinsics.
15024 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15025 SDValue PreservedSrc,
15026 const X86Subtarget *Subtarget,
15027 SelectionDAG &DAG) {
15028 EVT VT = Op.getValueType();
15029 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15030 MVT::i1, VT.getVectorNumElements());
15031 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15032 Mask.getValueType().getSizeInBits());
15035 assert(MaskVT.isSimple() && "invalid mask type");
15037 if (isAllOnes(Mask))
15040 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15041 // are extracted by EXTRACT_SUBVECTOR.
15042 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15043 DAG.getBitcast(BitcastVT, Mask),
15044 DAG.getIntPtrConstant(0, dl));
15046 switch (Op.getOpcode()) {
15048 case X86ISD::PCMPEQM:
15049 case X86ISD::PCMPGTM:
15051 case X86ISD::CMPMU:
15052 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15054 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15055 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15056 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
15059 /// \brief Creates an SDNode for a predicated scalar operation.
15060 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15061 /// The mask is comming as MVT::i8 and it should be truncated
15062 /// to MVT::i1 while lowering masking intrinsics.
15063 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15064 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
15065 /// a scalar instruction.
15066 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15067 SDValue PreservedSrc,
15068 const X86Subtarget *Subtarget,
15069 SelectionDAG &DAG) {
15070 if (isAllOnes(Mask))
15073 EVT VT = Op.getValueType();
15075 // The mask should be of type MVT::i1
15076 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15078 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15079 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15080 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15083 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15084 SelectionDAG &DAG) {
15086 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15087 EVT VT = Op.getValueType();
15088 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15090 switch(IntrData->Type) {
15091 case INTR_TYPE_1OP:
15092 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15093 case INTR_TYPE_2OP:
15094 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15096 case INTR_TYPE_3OP:
15097 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15098 Op.getOperand(2), Op.getOperand(3));
15099 case INTR_TYPE_1OP_MASK_RM: {
15100 SDValue Src = Op.getOperand(1);
15101 SDValue Src0 = Op.getOperand(2);
15102 SDValue Mask = Op.getOperand(3);
15103 SDValue RoundingMode = Op.getOperand(4);
15104 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15106 Mask, Src0, Subtarget, DAG);
15108 case INTR_TYPE_SCALAR_MASK_RM: {
15109 SDValue Src1 = Op.getOperand(1);
15110 SDValue Src2 = Op.getOperand(2);
15111 SDValue Src0 = Op.getOperand(3);
15112 SDValue Mask = Op.getOperand(4);
15113 // There are 2 kinds of intrinsics in this group:
15114 // (1) With supress-all-exceptions (sae) or rounding mode- 6 operands
15115 // (2) With rounding mode and sae - 7 operands.
15116 if (Op.getNumOperands() == 6) {
15117 SDValue Sae = Op.getOperand(5);
15118 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15119 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15121 Mask, Src0, Subtarget, DAG);
15123 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15124 SDValue RoundingMode = Op.getOperand(5);
15125 SDValue Sae = Op.getOperand(6);
15126 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15127 RoundingMode, Sae),
15128 Mask, Src0, Subtarget, DAG);
15130 case INTR_TYPE_2OP_MASK: {
15131 SDValue Src1 = Op.getOperand(1);
15132 SDValue Src2 = Op.getOperand(2);
15133 SDValue PassThru = Op.getOperand(3);
15134 SDValue Mask = Op.getOperand(4);
15135 // We specify 2 possible opcodes for intrinsics with rounding modes.
15136 // First, we check if the intrinsic may have non-default rounding mode,
15137 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15138 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15139 if (IntrWithRoundingModeOpcode != 0) {
15140 SDValue Rnd = Op.getOperand(5);
15141 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15142 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15143 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15144 dl, Op.getValueType(),
15146 Mask, PassThru, Subtarget, DAG);
15149 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15151 Mask, PassThru, Subtarget, DAG);
15153 case FMA_OP_MASK: {
15154 SDValue Src1 = Op.getOperand(1);
15155 SDValue Src2 = Op.getOperand(2);
15156 SDValue Src3 = Op.getOperand(3);
15157 SDValue Mask = Op.getOperand(4);
15158 // We specify 2 possible opcodes for intrinsics with rounding modes.
15159 // First, we check if the intrinsic may have non-default rounding mode,
15160 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15161 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15162 if (IntrWithRoundingModeOpcode != 0) {
15163 SDValue Rnd = Op.getOperand(5);
15164 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15165 X86::STATIC_ROUNDING::CUR_DIRECTION)
15166 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15167 dl, Op.getValueType(),
15168 Src1, Src2, Src3, Rnd),
15169 Mask, Src1, Subtarget, DAG);
15171 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15172 dl, Op.getValueType(),
15174 Mask, Src1, Subtarget, DAG);
15177 case CMP_MASK_CC: {
15178 // Comparison intrinsics with masks.
15179 // Example of transformation:
15180 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15181 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15183 // (v8i1 (insert_subvector undef,
15184 // (v2i1 (and (PCMPEQM %a, %b),
15185 // (extract_subvector
15186 // (v8i1 (bitcast %mask)), 0))), 0))))
15187 EVT VT = Op.getOperand(1).getValueType();
15188 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15189 VT.getVectorNumElements());
15190 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15191 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15192 Mask.getValueType().getSizeInBits());
15194 if (IntrData->Type == CMP_MASK_CC) {
15195 SDValue CC = Op.getOperand(3);
15196 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15197 // We specify 2 possible opcodes for intrinsics with rounding modes.
15198 // First, we check if the intrinsic may have non-default rounding mode,
15199 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15200 if (IntrData->Opc1 != 0) {
15201 SDValue Rnd = Op.getOperand(5);
15202 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15203 X86::STATIC_ROUNDING::CUR_DIRECTION)
15204 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15205 Op.getOperand(2), CC, Rnd);
15207 //default rounding mode
15209 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15210 Op.getOperand(2), CC);
15213 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15214 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15217 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15218 DAG.getTargetConstant(0, dl,
15221 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15222 DAG.getUNDEF(BitcastVT), CmpMask,
15223 DAG.getIntPtrConstant(0, dl));
15224 return DAG.getBitcast(Op.getValueType(), Res);
15226 case COMI: { // Comparison intrinsics
15227 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15228 SDValue LHS = Op.getOperand(1);
15229 SDValue RHS = Op.getOperand(2);
15230 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15231 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15232 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15233 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15234 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15235 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15238 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15239 Op.getOperand(1), Op.getOperand(2), DAG);
15241 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15242 Op.getSimpleValueType(),
15244 Op.getOperand(2), DAG),
15245 Op.getOperand(4), Op.getOperand(3), Subtarget,
15247 case COMPRESS_EXPAND_IN_REG: {
15248 SDValue Mask = Op.getOperand(3);
15249 SDValue DataToCompress = Op.getOperand(1);
15250 SDValue PassThru = Op.getOperand(2);
15251 if (isAllOnes(Mask)) // return data as is
15252 return Op.getOperand(1);
15253 EVT VT = Op.getValueType();
15254 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15255 VT.getVectorNumElements());
15256 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15257 Mask.getValueType().getSizeInBits());
15259 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15260 DAG.getBitcast(BitcastVT, Mask),
15261 DAG.getIntPtrConstant(0, dl));
15263 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
15267 SDValue Mask = Op.getOperand(3);
15268 EVT VT = Op.getValueType();
15269 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15270 VT.getVectorNumElements());
15271 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15272 Mask.getValueType().getSizeInBits());
15274 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15275 DAG.getBitcast(BitcastVT, Mask),
15276 DAG.getIntPtrConstant(0, dl));
15277 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15286 default: return SDValue(); // Don't custom lower most intrinsics.
15288 case Intrinsic::x86_avx2_permd:
15289 case Intrinsic::x86_avx2_permps:
15290 // Operands intentionally swapped. Mask is last operand to intrinsic,
15291 // but second operand for node/instruction.
15292 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15293 Op.getOperand(2), Op.getOperand(1));
15295 case Intrinsic::x86_avx512_mask_valign_q_512:
15296 case Intrinsic::x86_avx512_mask_valign_d_512:
15297 // Vector source operands are swapped.
15298 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
15299 Op.getValueType(), Op.getOperand(2),
15302 Op.getOperand(5), Op.getOperand(4),
15305 // ptest and testp intrinsics. The intrinsic these come from are designed to
15306 // return an integer value, not just an instruction so lower it to the ptest
15307 // or testp pattern and a setcc for the result.
15308 case Intrinsic::x86_sse41_ptestz:
15309 case Intrinsic::x86_sse41_ptestc:
15310 case Intrinsic::x86_sse41_ptestnzc:
15311 case Intrinsic::x86_avx_ptestz_256:
15312 case Intrinsic::x86_avx_ptestc_256:
15313 case Intrinsic::x86_avx_ptestnzc_256:
15314 case Intrinsic::x86_avx_vtestz_ps:
15315 case Intrinsic::x86_avx_vtestc_ps:
15316 case Intrinsic::x86_avx_vtestnzc_ps:
15317 case Intrinsic::x86_avx_vtestz_pd:
15318 case Intrinsic::x86_avx_vtestc_pd:
15319 case Intrinsic::x86_avx_vtestnzc_pd:
15320 case Intrinsic::x86_avx_vtestz_ps_256:
15321 case Intrinsic::x86_avx_vtestc_ps_256:
15322 case Intrinsic::x86_avx_vtestnzc_ps_256:
15323 case Intrinsic::x86_avx_vtestz_pd_256:
15324 case Intrinsic::x86_avx_vtestc_pd_256:
15325 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15326 bool IsTestPacked = false;
15329 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15330 case Intrinsic::x86_avx_vtestz_ps:
15331 case Intrinsic::x86_avx_vtestz_pd:
15332 case Intrinsic::x86_avx_vtestz_ps_256:
15333 case Intrinsic::x86_avx_vtestz_pd_256:
15334 IsTestPacked = true; // Fallthrough
15335 case Intrinsic::x86_sse41_ptestz:
15336 case Intrinsic::x86_avx_ptestz_256:
15338 X86CC = X86::COND_E;
15340 case Intrinsic::x86_avx_vtestc_ps:
15341 case Intrinsic::x86_avx_vtestc_pd:
15342 case Intrinsic::x86_avx_vtestc_ps_256:
15343 case Intrinsic::x86_avx_vtestc_pd_256:
15344 IsTestPacked = true; // Fallthrough
15345 case Intrinsic::x86_sse41_ptestc:
15346 case Intrinsic::x86_avx_ptestc_256:
15348 X86CC = X86::COND_B;
15350 case Intrinsic::x86_avx_vtestnzc_ps:
15351 case Intrinsic::x86_avx_vtestnzc_pd:
15352 case Intrinsic::x86_avx_vtestnzc_ps_256:
15353 case Intrinsic::x86_avx_vtestnzc_pd_256:
15354 IsTestPacked = true; // Fallthrough
15355 case Intrinsic::x86_sse41_ptestnzc:
15356 case Intrinsic::x86_avx_ptestnzc_256:
15358 X86CC = X86::COND_A;
15362 SDValue LHS = Op.getOperand(1);
15363 SDValue RHS = Op.getOperand(2);
15364 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15365 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15366 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15367 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15368 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15370 case Intrinsic::x86_avx512_kortestz_w:
15371 case Intrinsic::x86_avx512_kortestc_w: {
15372 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15373 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15374 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15375 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15376 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15377 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15378 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15381 case Intrinsic::x86_sse42_pcmpistria128:
15382 case Intrinsic::x86_sse42_pcmpestria128:
15383 case Intrinsic::x86_sse42_pcmpistric128:
15384 case Intrinsic::x86_sse42_pcmpestric128:
15385 case Intrinsic::x86_sse42_pcmpistrio128:
15386 case Intrinsic::x86_sse42_pcmpestrio128:
15387 case Intrinsic::x86_sse42_pcmpistris128:
15388 case Intrinsic::x86_sse42_pcmpestris128:
15389 case Intrinsic::x86_sse42_pcmpistriz128:
15390 case Intrinsic::x86_sse42_pcmpestriz128: {
15394 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15395 case Intrinsic::x86_sse42_pcmpistria128:
15396 Opcode = X86ISD::PCMPISTRI;
15397 X86CC = X86::COND_A;
15399 case Intrinsic::x86_sse42_pcmpestria128:
15400 Opcode = X86ISD::PCMPESTRI;
15401 X86CC = X86::COND_A;
15403 case Intrinsic::x86_sse42_pcmpistric128:
15404 Opcode = X86ISD::PCMPISTRI;
15405 X86CC = X86::COND_B;
15407 case Intrinsic::x86_sse42_pcmpestric128:
15408 Opcode = X86ISD::PCMPESTRI;
15409 X86CC = X86::COND_B;
15411 case Intrinsic::x86_sse42_pcmpistrio128:
15412 Opcode = X86ISD::PCMPISTRI;
15413 X86CC = X86::COND_O;
15415 case Intrinsic::x86_sse42_pcmpestrio128:
15416 Opcode = X86ISD::PCMPESTRI;
15417 X86CC = X86::COND_O;
15419 case Intrinsic::x86_sse42_pcmpistris128:
15420 Opcode = X86ISD::PCMPISTRI;
15421 X86CC = X86::COND_S;
15423 case Intrinsic::x86_sse42_pcmpestris128:
15424 Opcode = X86ISD::PCMPESTRI;
15425 X86CC = X86::COND_S;
15427 case Intrinsic::x86_sse42_pcmpistriz128:
15428 Opcode = X86ISD::PCMPISTRI;
15429 X86CC = X86::COND_E;
15431 case Intrinsic::x86_sse42_pcmpestriz128:
15432 Opcode = X86ISD::PCMPESTRI;
15433 X86CC = X86::COND_E;
15436 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15437 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15438 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15439 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15440 DAG.getConstant(X86CC, dl, MVT::i8),
15441 SDValue(PCMP.getNode(), 1));
15442 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15445 case Intrinsic::x86_sse42_pcmpistri128:
15446 case Intrinsic::x86_sse42_pcmpestri128: {
15448 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15449 Opcode = X86ISD::PCMPISTRI;
15451 Opcode = X86ISD::PCMPESTRI;
15453 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15454 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15455 return DAG.getNode(Opcode, dl, VTs, NewOps);
15458 case Intrinsic::x86_seh_lsda: {
15459 // Compute the symbol for the LSDA. We know it'll get emitted later.
15460 MachineFunction &MF = DAG.getMachineFunction();
15461 SDValue Op1 = Op.getOperand(1);
15462 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15463 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15464 GlobalValue::getRealLinkageName(Fn->getName()));
15465 StringRef Name = LSDASym->getName();
15466 assert(Name.data()[Name.size()] == '\0' && "not null terminated");
15468 // Generate a simple absolute symbol reference. This intrinsic is only
15469 // supported on 32-bit Windows, which isn't PIC.
15471 DAG.getTargetExternalSymbol(Name.data(), VT, X86II::MO_NOPREFIX);
15472 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15477 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15478 SDValue Src, SDValue Mask, SDValue Base,
15479 SDValue Index, SDValue ScaleOp, SDValue Chain,
15480 const X86Subtarget * Subtarget) {
15482 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15483 assert(C && "Invalid scale type");
15484 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15485 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15486 Index.getSimpleValueType().getVectorNumElements());
15488 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15490 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15492 MaskInReg = DAG.getBitcast(MaskVT, Mask);
15493 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15494 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15495 SDValue Segment = DAG.getRegister(0, MVT::i32);
15496 if (Src.getOpcode() == ISD::UNDEF)
15497 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15498 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15499 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15500 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15501 return DAG.getMergeValues(RetOps, dl);
15504 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15505 SDValue Src, SDValue Mask, SDValue Base,
15506 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15508 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15509 assert(C && "Invalid scale type");
15510 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15511 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15512 SDValue Segment = DAG.getRegister(0, MVT::i32);
15513 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15514 Index.getSimpleValueType().getVectorNumElements());
15516 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15518 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15520 MaskInReg = DAG.getBitcast(MaskVT, Mask);
15521 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15522 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15523 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15524 return SDValue(Res, 1);
15527 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15528 SDValue Mask, SDValue Base, SDValue Index,
15529 SDValue ScaleOp, SDValue Chain) {
15531 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15532 assert(C && "Invalid scale type");
15533 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15534 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15535 SDValue Segment = DAG.getRegister(0, MVT::i32);
15537 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15539 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15541 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15543 MaskInReg = DAG.getBitcast(MaskVT, Mask);
15544 //SDVTList VTs = DAG.getVTList(MVT::Other);
15545 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15546 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15547 return SDValue(Res, 0);
15550 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15551 // read performance monitor counters (x86_rdpmc).
15552 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15553 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15554 SmallVectorImpl<SDValue> &Results) {
15555 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15556 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15559 // The ECX register is used to select the index of the performance counter
15561 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
15563 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
15565 // Reads the content of a 64-bit performance counter and returns it in the
15566 // registers EDX:EAX.
15567 if (Subtarget->is64Bit()) {
15568 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15569 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15572 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15573 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15576 Chain = HI.getValue(1);
15578 if (Subtarget->is64Bit()) {
15579 // The EAX register is loaded with the low-order 32 bits. The EDX register
15580 // is loaded with the supported high-order bits of the counter.
15581 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15582 DAG.getConstant(32, DL, MVT::i8));
15583 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15584 Results.push_back(Chain);
15588 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15589 SDValue Ops[] = { LO, HI };
15590 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15591 Results.push_back(Pair);
15592 Results.push_back(Chain);
15595 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
15596 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
15597 // also used to custom lower READCYCLECOUNTER nodes.
15598 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
15599 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15600 SmallVectorImpl<SDValue> &Results) {
15601 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15602 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
15605 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
15606 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
15607 // and the EAX register is loaded with the low-order 32 bits.
15608 if (Subtarget->is64Bit()) {
15609 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15610 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15613 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15614 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15617 SDValue Chain = HI.getValue(1);
15619 if (Opcode == X86ISD::RDTSCP_DAG) {
15620 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15622 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
15623 // the ECX register. Add 'ecx' explicitly to the chain.
15624 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
15626 // Explicitly store the content of ECX at the location passed in input
15627 // to the 'rdtscp' intrinsic.
15628 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
15629 MachinePointerInfo(), false, false, 0);
15632 if (Subtarget->is64Bit()) {
15633 // The EDX register is loaded with the high-order 32 bits of the MSR, and
15634 // the EAX register is loaded with the low-order 32 bits.
15635 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15636 DAG.getConstant(32, DL, MVT::i8));
15637 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15638 Results.push_back(Chain);
15642 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15643 SDValue Ops[] = { LO, HI };
15644 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15645 Results.push_back(Pair);
15646 Results.push_back(Chain);
15649 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
15650 SelectionDAG &DAG) {
15651 SmallVector<SDValue, 2> Results;
15653 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
15655 return DAG.getMergeValues(Results, DL);
15659 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15660 SelectionDAG &DAG) {
15661 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
15663 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
15668 switch(IntrData->Type) {
15670 llvm_unreachable("Unknown Intrinsic Type");
15674 // Emit the node with the right value type.
15675 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
15676 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15678 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
15679 // Otherwise return the value from Rand, which is always 0, casted to i32.
15680 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
15681 DAG.getConstant(1, dl, Op->getValueType(1)),
15682 DAG.getConstant(X86::COND_B, dl, MVT::i32),
15683 SDValue(Result.getNode(), 1) };
15684 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
15685 DAG.getVTList(Op->getValueType(1), MVT::Glue),
15688 // Return { result, isValid, chain }.
15689 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
15690 SDValue(Result.getNode(), 2));
15693 //gather(v1, mask, index, base, scale);
15694 SDValue Chain = Op.getOperand(0);
15695 SDValue Src = Op.getOperand(2);
15696 SDValue Base = Op.getOperand(3);
15697 SDValue Index = Op.getOperand(4);
15698 SDValue Mask = Op.getOperand(5);
15699 SDValue Scale = Op.getOperand(6);
15700 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
15704 //scatter(base, mask, index, v1, scale);
15705 SDValue Chain = Op.getOperand(0);
15706 SDValue Base = Op.getOperand(2);
15707 SDValue Mask = Op.getOperand(3);
15708 SDValue Index = Op.getOperand(4);
15709 SDValue Src = Op.getOperand(5);
15710 SDValue Scale = Op.getOperand(6);
15711 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
15715 SDValue Hint = Op.getOperand(6);
15716 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
15717 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
15718 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
15719 SDValue Chain = Op.getOperand(0);
15720 SDValue Mask = Op.getOperand(2);
15721 SDValue Index = Op.getOperand(3);
15722 SDValue Base = Op.getOperand(4);
15723 SDValue Scale = Op.getOperand(5);
15724 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15726 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15728 SmallVector<SDValue, 2> Results;
15729 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
15731 return DAG.getMergeValues(Results, dl);
15733 // Read Performance Monitoring Counters.
15735 SmallVector<SDValue, 2> Results;
15736 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15737 return DAG.getMergeValues(Results, dl);
15739 // XTEST intrinsics.
15741 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15742 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15743 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15744 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
15746 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15747 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15748 Ret, SDValue(InTrans.getNode(), 1));
15752 SmallVector<SDValue, 2> Results;
15753 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15754 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15755 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15756 DAG.getConstant(-1, dl, MVT::i8));
15757 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15758 Op.getOperand(4), GenCF.getValue(1));
15759 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15760 Op.getOperand(5), MachinePointerInfo(),
15762 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15763 DAG.getConstant(X86::COND_B, dl, MVT::i8),
15765 Results.push_back(SetCC);
15766 Results.push_back(Store);
15767 return DAG.getMergeValues(Results, dl);
15769 case COMPRESS_TO_MEM: {
15771 SDValue Mask = Op.getOperand(4);
15772 SDValue DataToCompress = Op.getOperand(3);
15773 SDValue Addr = Op.getOperand(2);
15774 SDValue Chain = Op.getOperand(0);
15776 EVT VT = DataToCompress.getValueType();
15777 if (isAllOnes(Mask)) // return just a store
15778 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15779 MachinePointerInfo(), false, false,
15780 VT.getScalarSizeInBits()/8);
15782 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15783 VT.getVectorNumElements());
15784 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15785 Mask.getValueType().getSizeInBits());
15786 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15787 DAG.getBitcast(BitcastVT, Mask),
15788 DAG.getIntPtrConstant(0, dl));
15790 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15791 DataToCompress, DAG.getUNDEF(VT));
15792 return DAG.getStore(Chain, dl, Compressed, Addr,
15793 MachinePointerInfo(), false, false,
15794 VT.getScalarSizeInBits()/8);
15796 case EXPAND_FROM_MEM: {
15798 SDValue Mask = Op.getOperand(4);
15799 SDValue PathThru = Op.getOperand(3);
15800 SDValue Addr = Op.getOperand(2);
15801 SDValue Chain = Op.getOperand(0);
15802 EVT VT = Op.getValueType();
15804 if (isAllOnes(Mask)) // return just a load
15805 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15806 false, VT.getScalarSizeInBits()/8);
15807 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15808 VT.getVectorNumElements());
15809 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15810 Mask.getValueType().getSizeInBits());
15811 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15812 DAG.getBitcast(BitcastVT, Mask),
15813 DAG.getIntPtrConstant(0, dl));
15815 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15816 false, false, false,
15817 VT.getScalarSizeInBits()/8);
15819 SDValue Results[] = {
15820 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15822 return DAG.getMergeValues(Results, dl);
15827 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15828 SelectionDAG &DAG) const {
15829 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15830 MFI->setReturnAddressIsTaken(true);
15832 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15835 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15837 EVT PtrVT = getPointerTy();
15840 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15841 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15842 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
15843 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15844 DAG.getNode(ISD::ADD, dl, PtrVT,
15845 FrameAddr, Offset),
15846 MachinePointerInfo(), false, false, false, 0);
15849 // Just load the return address.
15850 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15851 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15852 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15855 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15856 MachineFunction &MF = DAG.getMachineFunction();
15857 MachineFrameInfo *MFI = MF.getFrameInfo();
15858 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15859 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15860 EVT VT = Op.getValueType();
15862 MFI->setFrameAddressIsTaken(true);
15864 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15865 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15866 // is not possible to crawl up the stack without looking at the unwind codes
15868 int FrameAddrIndex = FuncInfo->getFAIndex();
15869 if (!FrameAddrIndex) {
15870 // Set up a frame object for the return address.
15871 unsigned SlotSize = RegInfo->getSlotSize();
15872 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15873 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
15874 FuncInfo->setFAIndex(FrameAddrIndex);
15876 return DAG.getFrameIndex(FrameAddrIndex, VT);
15879 unsigned FrameReg =
15880 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15881 SDLoc dl(Op); // FIXME probably not meaningful
15882 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15883 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15884 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15885 "Invalid Frame Register!");
15886 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15888 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15889 MachinePointerInfo(),
15890 false, false, false, 0);
15894 // FIXME? Maybe this could be a TableGen attribute on some registers and
15895 // this table could be generated automatically from RegInfo.
15896 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15898 unsigned Reg = StringSwitch<unsigned>(RegName)
15899 .Case("esp", X86::ESP)
15900 .Case("rsp", X86::RSP)
15904 report_fatal_error("Invalid register name global variable");
15907 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15908 SelectionDAG &DAG) const {
15909 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15910 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
15913 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15914 SDValue Chain = Op.getOperand(0);
15915 SDValue Offset = Op.getOperand(1);
15916 SDValue Handler = Op.getOperand(2);
15919 EVT PtrVT = getPointerTy();
15920 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15921 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15922 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15923 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15924 "Invalid Frame Register!");
15925 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15926 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15928 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15929 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
15931 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15932 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15934 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15936 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15937 DAG.getRegister(StoreAddrReg, PtrVT));
15940 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15941 SelectionDAG &DAG) const {
15943 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15944 DAG.getVTList(MVT::i32, MVT::Other),
15945 Op.getOperand(0), Op.getOperand(1));
15948 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15949 SelectionDAG &DAG) const {
15951 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15952 Op.getOperand(0), Op.getOperand(1));
15955 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15956 return Op.getOperand(0);
15959 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15960 SelectionDAG &DAG) const {
15961 SDValue Root = Op.getOperand(0);
15962 SDValue Trmp = Op.getOperand(1); // trampoline
15963 SDValue FPtr = Op.getOperand(2); // nested function
15964 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15967 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15968 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15970 if (Subtarget->is64Bit()) {
15971 SDValue OutChains[6];
15973 // Large code-model.
15974 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15975 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15977 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15978 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15980 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15982 // Load the pointer to the nested function into R11.
15983 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15984 SDValue Addr = Trmp;
15985 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15986 Addr, MachinePointerInfo(TrmpAddr),
15989 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15990 DAG.getConstant(2, dl, MVT::i64));
15991 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15992 MachinePointerInfo(TrmpAddr, 2),
15995 // Load the 'nest' parameter value into R10.
15996 // R10 is specified in X86CallingConv.td
15997 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15998 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15999 DAG.getConstant(10, dl, MVT::i64));
16000 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16001 Addr, MachinePointerInfo(TrmpAddr, 10),
16004 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16005 DAG.getConstant(12, dl, MVT::i64));
16006 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16007 MachinePointerInfo(TrmpAddr, 12),
16010 // Jump to the nested function.
16011 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16012 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16013 DAG.getConstant(20, dl, MVT::i64));
16014 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16015 Addr, MachinePointerInfo(TrmpAddr, 20),
16018 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16019 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16020 DAG.getConstant(22, dl, MVT::i64));
16021 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16022 Addr, MachinePointerInfo(TrmpAddr, 22),
16025 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16027 const Function *Func =
16028 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16029 CallingConv::ID CC = Func->getCallingConv();
16034 llvm_unreachable("Unsupported calling convention");
16035 case CallingConv::C:
16036 case CallingConv::X86_StdCall: {
16037 // Pass 'nest' parameter in ECX.
16038 // Must be kept in sync with X86CallingConv.td
16039 NestReg = X86::ECX;
16041 // Check that ECX wasn't needed by an 'inreg' parameter.
16042 FunctionType *FTy = Func->getFunctionType();
16043 const AttributeSet &Attrs = Func->getAttributes();
16045 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16046 unsigned InRegCount = 0;
16049 for (FunctionType::param_iterator I = FTy->param_begin(),
16050 E = FTy->param_end(); I != E; ++I, ++Idx)
16051 if (Attrs.hasAttribute(Idx, Attribute::InReg))
16052 // FIXME: should only count parameters that are lowered to integers.
16053 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
16055 if (InRegCount > 2) {
16056 report_fatal_error("Nest register in use - reduce number of inreg"
16062 case CallingConv::X86_FastCall:
16063 case CallingConv::X86_ThisCall:
16064 case CallingConv::Fast:
16065 // Pass 'nest' parameter in EAX.
16066 // Must be kept in sync with X86CallingConv.td
16067 NestReg = X86::EAX;
16071 SDValue OutChains[4];
16072 SDValue Addr, Disp;
16074 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16075 DAG.getConstant(10, dl, MVT::i32));
16076 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16078 // This is storing the opcode for MOV32ri.
16079 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16080 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16081 OutChains[0] = DAG.getStore(Root, dl,
16082 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16083 Trmp, MachinePointerInfo(TrmpAddr),
16086 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16087 DAG.getConstant(1, dl, MVT::i32));
16088 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16089 MachinePointerInfo(TrmpAddr, 1),
16092 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16093 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16094 DAG.getConstant(5, dl, MVT::i32));
16095 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16096 Addr, MachinePointerInfo(TrmpAddr, 5),
16099 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16100 DAG.getConstant(6, dl, MVT::i32));
16101 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16102 MachinePointerInfo(TrmpAddr, 6),
16105 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16109 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16110 SelectionDAG &DAG) const {
16112 The rounding mode is in bits 11:10 of FPSR, and has the following
16114 00 Round to nearest
16119 FLT_ROUNDS, on the other hand, expects the following:
16126 To perform the conversion, we do:
16127 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16130 MachineFunction &MF = DAG.getMachineFunction();
16131 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16132 unsigned StackAlignment = TFI.getStackAlignment();
16133 MVT VT = Op.getSimpleValueType();
16136 // Save FP Control Word to stack slot
16137 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16138 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
16140 MachineMemOperand *MMO =
16141 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
16142 MachineMemOperand::MOStore, 2, 2);
16144 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16145 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16146 DAG.getVTList(MVT::Other),
16147 Ops, MVT::i16, MMO);
16149 // Load FP Control Word from stack slot
16150 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16151 MachinePointerInfo(), false, false, false, 0);
16153 // Transform as necessary
16155 DAG.getNode(ISD::SRL, DL, MVT::i16,
16156 DAG.getNode(ISD::AND, DL, MVT::i16,
16157 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16158 DAG.getConstant(11, DL, MVT::i8));
16160 DAG.getNode(ISD::SRL, DL, MVT::i16,
16161 DAG.getNode(ISD::AND, DL, MVT::i16,
16162 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16163 DAG.getConstant(9, DL, MVT::i8));
16166 DAG.getNode(ISD::AND, DL, MVT::i16,
16167 DAG.getNode(ISD::ADD, DL, MVT::i16,
16168 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16169 DAG.getConstant(1, DL, MVT::i16)),
16170 DAG.getConstant(3, DL, MVT::i16));
16172 return DAG.getNode((VT.getSizeInBits() < 16 ?
16173 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16176 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16177 MVT VT = Op.getSimpleValueType();
16179 unsigned NumBits = VT.getSizeInBits();
16182 Op = Op.getOperand(0);
16183 if (VT == MVT::i8) {
16184 // Zero extend to i32 since there is not an i8 bsr.
16186 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16189 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16190 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16191 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16193 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16196 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16197 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16200 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16202 // Finally xor with NumBits-1.
16203 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16204 DAG.getConstant(NumBits - 1, dl, OpVT));
16207 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16211 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16212 MVT VT = Op.getSimpleValueType();
16214 unsigned NumBits = VT.getSizeInBits();
16217 Op = Op.getOperand(0);
16218 if (VT == MVT::i8) {
16219 // Zero extend to i32 since there is not an i8 bsr.
16221 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16224 // Issue a bsr (scan bits in reverse).
16225 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16226 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16228 // And xor with NumBits-1.
16229 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16230 DAG.getConstant(NumBits - 1, dl, OpVT));
16233 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16237 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16238 MVT VT = Op.getSimpleValueType();
16239 unsigned NumBits = VT.getSizeInBits();
16241 Op = Op.getOperand(0);
16243 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16244 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16245 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16247 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16250 DAG.getConstant(NumBits, dl, VT),
16251 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16254 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16257 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16258 // ones, and then concatenate the result back.
16259 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16260 MVT VT = Op.getSimpleValueType();
16262 assert(VT.is256BitVector() && VT.isInteger() &&
16263 "Unsupported value type for operation");
16265 unsigned NumElems = VT.getVectorNumElements();
16268 // Extract the LHS vectors
16269 SDValue LHS = Op.getOperand(0);
16270 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16271 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16273 // Extract the RHS vectors
16274 SDValue RHS = Op.getOperand(1);
16275 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16276 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16278 MVT EltVT = VT.getVectorElementType();
16279 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16281 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16282 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16283 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16286 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16287 if (Op.getValueType() == MVT::i1)
16288 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16289 Op.getOperand(0), Op.getOperand(1));
16290 assert(Op.getSimpleValueType().is256BitVector() &&
16291 Op.getSimpleValueType().isInteger() &&
16292 "Only handle AVX 256-bit vector integer operation");
16293 return Lower256IntArith(Op, DAG);
16296 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16297 if (Op.getValueType() == MVT::i1)
16298 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16299 Op.getOperand(0), Op.getOperand(1));
16300 assert(Op.getSimpleValueType().is256BitVector() &&
16301 Op.getSimpleValueType().isInteger() &&
16302 "Only handle AVX 256-bit vector integer operation");
16303 return Lower256IntArith(Op, DAG);
16306 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16307 SelectionDAG &DAG) {
16309 MVT VT = Op.getSimpleValueType();
16312 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
16314 // Decompose 256-bit ops into smaller 128-bit ops.
16315 if (VT.is256BitVector() && !Subtarget->hasInt256())
16316 return Lower256IntArith(Op, DAG);
16318 SDValue A = Op.getOperand(0);
16319 SDValue B = Op.getOperand(1);
16321 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16322 // pairs, multiply and truncate.
16323 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16324 if (Subtarget->hasInt256()) {
16325 if (VT == MVT::v32i8) {
16326 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16327 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16328 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16329 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16330 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16331 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16332 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16333 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16334 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16335 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16338 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16339 return DAG.getNode(
16340 ISD::TRUNCATE, dl, VT,
16341 DAG.getNode(ISD::MUL, dl, ExVT,
16342 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16343 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16346 assert(VT == MVT::v16i8 &&
16347 "Pre-AVX2 support only supports v16i8 multiplication");
16348 MVT ExVT = MVT::v8i16;
16350 // Extract the lo parts and sign extend to i16
16352 if (Subtarget->hasSSE41()) {
16353 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16354 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16356 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16357 -1, 4, -1, 5, -1, 6, -1, 7};
16358 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16359 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16360 ALo = DAG.getBitcast(ExVT, ALo);
16361 BLo = DAG.getBitcast(ExVT, BLo);
16362 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16363 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16366 // Extract the hi parts and sign extend to i16
16368 if (Subtarget->hasSSE41()) {
16369 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16370 -1, -1, -1, -1, -1, -1, -1, -1};
16371 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16372 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16373 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16374 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16376 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16377 -1, 12, -1, 13, -1, 14, -1, 15};
16378 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16379 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16380 AHi = DAG.getBitcast(ExVT, AHi);
16381 BHi = DAG.getBitcast(ExVT, BHi);
16382 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16383 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16386 // Multiply, mask the lower 8bits of the lo/hi results and pack
16387 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16388 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16389 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16390 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16391 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16394 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16395 if (VT == MVT::v4i32) {
16396 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16397 "Should not custom lower when pmuldq is available!");
16399 // Extract the odd parts.
16400 static const int UnpackMask[] = { 1, -1, 3, -1 };
16401 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16402 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16404 // Multiply the even parts.
16405 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16406 // Now multiply odd parts.
16407 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16409 Evens = DAG.getBitcast(VT, Evens);
16410 Odds = DAG.getBitcast(VT, Odds);
16412 // Merge the two vectors back together with a shuffle. This expands into 2
16414 static const int ShufMask[] = { 0, 4, 2, 6 };
16415 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16418 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16419 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16421 // Ahi = psrlqi(a, 32);
16422 // Bhi = psrlqi(b, 32);
16424 // AloBlo = pmuludq(a, b);
16425 // AloBhi = pmuludq(a, Bhi);
16426 // AhiBlo = pmuludq(Ahi, b);
16428 // AloBhi = psllqi(AloBhi, 32);
16429 // AhiBlo = psllqi(AhiBlo, 32);
16430 // return AloBlo + AloBhi + AhiBlo;
16432 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16433 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16435 // Bit cast to 32-bit vectors for MULUDQ
16436 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16437 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16438 A = DAG.getBitcast(MulVT, A);
16439 B = DAG.getBitcast(MulVT, B);
16440 Ahi = DAG.getBitcast(MulVT, Ahi);
16441 Bhi = DAG.getBitcast(MulVT, Bhi);
16443 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16444 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16445 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16447 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16448 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16450 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16451 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16454 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16455 assert(Subtarget->isTargetWin64() && "Unexpected target");
16456 EVT VT = Op.getValueType();
16457 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16458 "Unexpected return type for lowering");
16462 switch (Op->getOpcode()) {
16463 default: llvm_unreachable("Unexpected request for libcall!");
16464 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16465 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16466 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16467 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16468 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16469 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16473 SDValue InChain = DAG.getEntryNode();
16475 TargetLowering::ArgListTy Args;
16476 TargetLowering::ArgListEntry Entry;
16477 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16478 EVT ArgVT = Op->getOperand(i).getValueType();
16479 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16480 "Unexpected argument type for lowering");
16481 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16482 Entry.Node = StackPtr;
16483 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16485 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16486 Entry.Ty = PointerType::get(ArgTy,0);
16487 Entry.isSExt = false;
16488 Entry.isZExt = false;
16489 Args.push_back(Entry);
16492 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16495 TargetLowering::CallLoweringInfo CLI(DAG);
16496 CLI.setDebugLoc(dl).setChain(InChain)
16497 .setCallee(getLibcallCallingConv(LC),
16498 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16499 Callee, std::move(Args), 0)
16500 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16502 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16503 return DAG.getBitcast(VT, CallInfo.first);
16506 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16507 SelectionDAG &DAG) {
16508 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16509 EVT VT = Op0.getValueType();
16512 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16513 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16515 // PMULxD operations multiply each even value (starting at 0) of LHS with
16516 // the related value of RHS and produce a widen result.
16517 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16518 // => <2 x i64> <ae|cg>
16520 // In other word, to have all the results, we need to perform two PMULxD:
16521 // 1. one with the even values.
16522 // 2. one with the odd values.
16523 // To achieve #2, with need to place the odd values at an even position.
16525 // Place the odd value at an even position (basically, shift all values 1
16526 // step to the left):
16527 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16528 // <a|b|c|d> => <b|undef|d|undef>
16529 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16530 // <e|f|g|h> => <f|undef|h|undef>
16531 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16533 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16535 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16536 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16538 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16539 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16540 // => <2 x i64> <ae|cg>
16541 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
16542 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
16543 // => <2 x i64> <bf|dh>
16544 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
16546 // Shuffle it back into the right order.
16547 SDValue Highs, Lows;
16548 if (VT == MVT::v8i32) {
16549 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
16550 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16551 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
16552 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16554 const int HighMask[] = {1, 5, 3, 7};
16555 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16556 const int LowMask[] = {0, 4, 2, 6};
16557 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16560 // If we have a signed multiply but no PMULDQ fix up the high parts of a
16561 // unsigned multiply.
16562 if (IsSigned && !Subtarget->hasSSE41()) {
16564 DAG.getConstant(31, dl,
16565 DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
16566 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
16567 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
16568 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
16569 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
16571 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
16572 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
16575 // The first result of MUL_LOHI is actually the low value, followed by the
16577 SDValue Ops[] = {Lows, Highs};
16578 return DAG.getMergeValues(Ops, dl);
16581 // Return true if the requred (according to Opcode) shift-imm form is natively
16582 // supported by the Subtarget
16583 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
16585 if (VT.getScalarSizeInBits() < 16)
16588 if (VT.is512BitVector() &&
16589 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
16592 bool LShift = VT.is128BitVector() ||
16593 (VT.is256BitVector() && Subtarget->hasInt256());
16595 bool AShift = LShift && (Subtarget->hasVLX() ||
16596 (VT != MVT::v2i64 && VT != MVT::v4i64));
16597 return (Opcode == ISD::SRA) ? AShift : LShift;
16600 // The shift amount is a variable, but it is the same for all vector lanes.
16601 // These instrcutions are defined together with shift-immediate.
16603 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
16605 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
16608 // Return true if the requred (according to Opcode) variable-shift form is
16609 // natively supported by the Subtarget
16610 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
16613 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
16616 // vXi16 supported only on AVX-512, BWI
16617 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
16620 if (VT.is512BitVector() || Subtarget->hasVLX())
16623 bool LShift = VT.is128BitVector() || VT.is256BitVector();
16624 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
16625 return (Opcode == ISD::SRA) ? AShift : LShift;
16628 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
16629 const X86Subtarget *Subtarget) {
16630 MVT VT = Op.getSimpleValueType();
16632 SDValue R = Op.getOperand(0);
16633 SDValue Amt = Op.getOperand(1);
16635 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16636 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16638 // Optimize shl/srl/sra with constant shift amount.
16639 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
16640 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
16641 uint64_t ShiftAmt = ShiftConst->getZExtValue();
16643 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
16644 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16646 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
16647 unsigned NumElts = VT.getVectorNumElements();
16648 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
16650 if (Op.getOpcode() == ISD::SHL) {
16651 // Simple i8 add case
16653 return DAG.getNode(ISD::ADD, dl, VT, R, R);
16655 // Make a large shift.
16656 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
16658 SHL = DAG.getBitcast(VT, SHL);
16659 // Zero out the rightmost bits.
16660 SmallVector<SDValue, 32> V(
16661 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
16662 return DAG.getNode(ISD::AND, dl, VT, SHL,
16663 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16665 if (Op.getOpcode() == ISD::SRL) {
16666 // Make a large shift.
16667 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
16669 SRL = DAG.getBitcast(VT, SRL);
16670 // Zero out the leftmost bits.
16671 SmallVector<SDValue, 32> V(
16672 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
16673 return DAG.getNode(ISD::AND, dl, VT, SRL,
16674 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16676 if (Op.getOpcode() == ISD::SRA) {
16677 if (ShiftAmt == 7) {
16678 // R s>> 7 === R s< 0
16679 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
16680 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
16683 // R s>> a === ((R u>> a) ^ m) - m
16684 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
16685 SmallVector<SDValue, 32> V(NumElts,
16686 DAG.getConstant(128 >> ShiftAmt, dl,
16688 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
16689 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
16690 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
16693 llvm_unreachable("Unknown shift opcode.");
16698 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16699 if (!Subtarget->is64Bit() &&
16700 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
16701 Amt.getOpcode() == ISD::BITCAST &&
16702 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16703 Amt = Amt.getOperand(0);
16704 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16705 VT.getVectorNumElements();
16706 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
16707 uint64_t ShiftAmt = 0;
16708 for (unsigned i = 0; i != Ratio; ++i) {
16709 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
16713 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
16715 // Check remaining shift amounts.
16716 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16717 uint64_t ShAmt = 0;
16718 for (unsigned j = 0; j != Ratio; ++j) {
16719 ConstantSDNode *C =
16720 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
16724 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
16726 if (ShAmt != ShiftAmt)
16729 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16735 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
16736 const X86Subtarget* Subtarget) {
16737 MVT VT = Op.getSimpleValueType();
16739 SDValue R = Op.getOperand(0);
16740 SDValue Amt = Op.getOperand(1);
16742 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16743 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16745 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
16746 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
16748 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
16750 EVT EltVT = VT.getVectorElementType();
16752 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
16753 // Check if this build_vector node is doing a splat.
16754 // If so, then set BaseShAmt equal to the splat value.
16755 BaseShAmt = BV->getSplatValue();
16756 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
16757 BaseShAmt = SDValue();
16759 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
16760 Amt = Amt.getOperand(0);
16762 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
16763 if (SVN && SVN->isSplat()) {
16764 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
16765 SDValue InVec = Amt.getOperand(0);
16766 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
16767 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
16768 "Unexpected shuffle index found!");
16769 BaseShAmt = InVec.getOperand(SplatIdx);
16770 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
16771 if (ConstantSDNode *C =
16772 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
16773 if (C->getZExtValue() == SplatIdx)
16774 BaseShAmt = InVec.getOperand(1);
16779 // Avoid introducing an extract element from a shuffle.
16780 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
16781 DAG.getIntPtrConstant(SplatIdx, dl));
16785 if (BaseShAmt.getNode()) {
16786 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
16787 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
16788 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
16789 else if (EltVT.bitsLT(MVT::i32))
16790 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
16792 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
16796 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16797 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
16798 Amt.getOpcode() == ISD::BITCAST &&
16799 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16800 Amt = Amt.getOperand(0);
16801 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16802 VT.getVectorNumElements();
16803 std::vector<SDValue> Vals(Ratio);
16804 for (unsigned i = 0; i != Ratio; ++i)
16805 Vals[i] = Amt.getOperand(i);
16806 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16807 for (unsigned j = 0; j != Ratio; ++j)
16808 if (Vals[j] != Amt.getOperand(i + j))
16811 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
16816 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16817 SelectionDAG &DAG) {
16818 MVT VT = Op.getSimpleValueType();
16820 SDValue R = Op.getOperand(0);
16821 SDValue Amt = Op.getOperand(1);
16823 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16824 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16826 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
16829 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
16832 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
16835 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
16836 // shifts per-lane and then shuffle the partial results back together.
16837 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
16838 // Splat the shift amounts so the scalar shifts above will catch it.
16839 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
16840 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
16841 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
16842 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
16843 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
16846 // If possible, lower this packed shift into a vector multiply instead of
16847 // expanding it into a sequence of scalar shifts.
16848 // Do this only if the vector shift count is a constant build_vector.
16849 if (Op.getOpcode() == ISD::SHL &&
16850 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16851 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16852 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16853 SmallVector<SDValue, 8> Elts;
16854 EVT SVT = VT.getScalarType();
16855 unsigned SVTBits = SVT.getSizeInBits();
16856 const APInt &One = APInt(SVTBits, 1);
16857 unsigned NumElems = VT.getVectorNumElements();
16859 for (unsigned i=0; i !=NumElems; ++i) {
16860 SDValue Op = Amt->getOperand(i);
16861 if (Op->getOpcode() == ISD::UNDEF) {
16862 Elts.push_back(Op);
16866 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16867 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16868 uint64_t ShAmt = C.getZExtValue();
16869 if (ShAmt >= SVTBits) {
16870 Elts.push_back(DAG.getUNDEF(SVT));
16873 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
16875 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16876 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16879 // Lower SHL with variable shift amount.
16880 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16881 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
16883 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
16884 DAG.getConstant(0x3f800000U, dl, VT));
16885 Op = DAG.getBitcast(MVT::v4f32, Op);
16886 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16887 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16890 // If possible, lower this shift as a sequence of two shifts by
16891 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16893 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16895 // Could be rewritten as:
16896 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16898 // The advantage is that the two shifts from the example would be
16899 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16900 // the vector shift into four scalar shifts plus four pairs of vector
16902 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16903 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16904 unsigned TargetOpcode = X86ISD::MOVSS;
16905 bool CanBeSimplified;
16906 // The splat value for the first packed shift (the 'X' from the example).
16907 SDValue Amt1 = Amt->getOperand(0);
16908 // The splat value for the second packed shift (the 'Y' from the example).
16909 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16910 Amt->getOperand(2);
16912 // See if it is possible to replace this node with a sequence of
16913 // two shifts followed by a MOVSS/MOVSD
16914 if (VT == MVT::v4i32) {
16915 // Check if it is legal to use a MOVSS.
16916 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16917 Amt2 == Amt->getOperand(3);
16918 if (!CanBeSimplified) {
16919 // Otherwise, check if we can still simplify this node using a MOVSD.
16920 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16921 Amt->getOperand(2) == Amt->getOperand(3);
16922 TargetOpcode = X86ISD::MOVSD;
16923 Amt2 = Amt->getOperand(2);
16926 // Do similar checks for the case where the machine value type
16928 CanBeSimplified = Amt1 == Amt->getOperand(1);
16929 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16930 CanBeSimplified = Amt2 == Amt->getOperand(i);
16932 if (!CanBeSimplified) {
16933 TargetOpcode = X86ISD::MOVSD;
16934 CanBeSimplified = true;
16935 Amt2 = Amt->getOperand(4);
16936 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16937 CanBeSimplified = Amt1 == Amt->getOperand(i);
16938 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16939 CanBeSimplified = Amt2 == Amt->getOperand(j);
16943 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16944 isa<ConstantSDNode>(Amt2)) {
16945 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16946 EVT CastVT = MVT::v4i32;
16948 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
16949 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16951 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
16952 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16953 if (TargetOpcode == X86ISD::MOVSD)
16954 CastVT = MVT::v2i64;
16955 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
16956 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
16957 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16959 return DAG.getBitcast(VT, Result);
16963 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16964 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
16965 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, dl, VT));
16967 SDValue VSelM = DAG.getConstant(0x80, dl, VT);
16968 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16969 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16971 // r = VSELECT(r, shl(r, 4), a);
16972 SDValue M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(4, dl, VT));
16973 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16976 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16977 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16978 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16980 // r = VSELECT(r, shl(r, 2), a);
16981 M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(2, dl, VT));
16982 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16985 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16986 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16987 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16989 // return VSELECT(r, r+r, a);
16990 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16991 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16995 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16996 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16997 // solution better.
16998 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16999 MVT ExtVT = MVT::v8i32;
17001 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17002 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
17003 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
17004 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17005 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
17008 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
17009 MVT ExtVT = MVT::v8i32;
17010 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17011 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
17012 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
17013 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
17014 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
17015 ALo = DAG.getBitcast(ExtVT, ALo);
17016 AHi = DAG.getBitcast(ExtVT, AHi);
17017 RLo = DAG.getBitcast(ExtVT, RLo);
17018 RHi = DAG.getBitcast(ExtVT, RHi);
17019 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
17020 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
17021 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
17022 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
17023 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
17026 // Decompose 256-bit shifts into smaller 128-bit shifts.
17027 if (VT.is256BitVector()) {
17028 unsigned NumElems = VT.getVectorNumElements();
17029 MVT EltVT = VT.getVectorElementType();
17030 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17032 // Extract the two vectors
17033 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17034 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17036 // Recreate the shift amount vectors
17037 SDValue Amt1, Amt2;
17038 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17039 // Constant shift amount
17040 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
17041 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
17042 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
17044 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17045 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17047 // Variable shift amount
17048 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17049 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17052 // Issue new vector shifts for the smaller types
17053 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17054 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17056 // Concatenate the result back
17057 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17063 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17064 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17065 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17066 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17067 // has only one use.
17068 SDNode *N = Op.getNode();
17069 SDValue LHS = N->getOperand(0);
17070 SDValue RHS = N->getOperand(1);
17071 unsigned BaseOp = 0;
17074 switch (Op.getOpcode()) {
17075 default: llvm_unreachable("Unknown ovf instruction!");
17077 // A subtract of one will be selected as a INC. Note that INC doesn't
17078 // set CF, so we can't do this for UADDO.
17079 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17081 BaseOp = X86ISD::INC;
17082 Cond = X86::COND_O;
17085 BaseOp = X86ISD::ADD;
17086 Cond = X86::COND_O;
17089 BaseOp = X86ISD::ADD;
17090 Cond = X86::COND_B;
17093 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17094 // set CF, so we can't do this for USUBO.
17095 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17097 BaseOp = X86ISD::DEC;
17098 Cond = X86::COND_O;
17101 BaseOp = X86ISD::SUB;
17102 Cond = X86::COND_O;
17105 BaseOp = X86ISD::SUB;
17106 Cond = X86::COND_B;
17109 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
17110 Cond = X86::COND_O;
17112 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
17113 if (N->getValueType(0) == MVT::i8) {
17114 BaseOp = X86ISD::UMUL8;
17115 Cond = X86::COND_O;
17118 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
17120 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
17123 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
17124 DAG.getConstant(X86::COND_O, DL, MVT::i32),
17125 SDValue(Sum.getNode(), 2));
17127 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17131 // Also sets EFLAGS.
17132 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
17133 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
17136 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17137 DAG.getConstant(Cond, DL, MVT::i32),
17138 SDValue(Sum.getNode(), 1));
17140 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17143 /// Returns true if the operand type is exactly twice the native width, and
17144 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
17145 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
17146 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
17147 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
17148 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
17151 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
17152 else if (OpWidth == 128)
17153 return Subtarget->hasCmpxchg16b();
17158 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
17159 return needsCmpXchgNb(SI->getValueOperand()->getType());
17162 // Note: this turns large loads into lock cmpxchg8b/16b.
17163 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
17164 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
17165 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
17166 return needsCmpXchgNb(PTy->getElementType());
17169 TargetLoweringBase::AtomicRMWExpansionKind
17170 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
17171 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17172 const Type *MemType = AI->getType();
17174 // If the operand is too big, we must see if cmpxchg8/16b is available
17175 // and default to library calls otherwise.
17176 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
17177 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
17178 : AtomicRMWExpansionKind::None;
17181 AtomicRMWInst::BinOp Op = AI->getOperation();
17184 llvm_unreachable("Unknown atomic operation");
17185 case AtomicRMWInst::Xchg:
17186 case AtomicRMWInst::Add:
17187 case AtomicRMWInst::Sub:
17188 // It's better to use xadd, xsub or xchg for these in all cases.
17189 return AtomicRMWExpansionKind::None;
17190 case AtomicRMWInst::Or:
17191 case AtomicRMWInst::And:
17192 case AtomicRMWInst::Xor:
17193 // If the atomicrmw's result isn't actually used, we can just add a "lock"
17194 // prefix to a normal instruction for these operations.
17195 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
17196 : AtomicRMWExpansionKind::None;
17197 case AtomicRMWInst::Nand:
17198 case AtomicRMWInst::Max:
17199 case AtomicRMWInst::Min:
17200 case AtomicRMWInst::UMax:
17201 case AtomicRMWInst::UMin:
17202 // These always require a non-trivial set of data operations on x86. We must
17203 // use a cmpxchg loop.
17204 return AtomicRMWExpansionKind::CmpXChg;
17208 static bool hasMFENCE(const X86Subtarget& Subtarget) {
17209 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
17210 // no-sse2). There isn't any reason to disable it if the target processor
17212 return Subtarget.hasSSE2() || Subtarget.is64Bit();
17216 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
17217 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17218 const Type *MemType = AI->getType();
17219 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
17220 // there is no benefit in turning such RMWs into loads, and it is actually
17221 // harmful as it introduces a mfence.
17222 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
17225 auto Builder = IRBuilder<>(AI);
17226 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
17227 auto SynchScope = AI->getSynchScope();
17228 // We must restrict the ordering to avoid generating loads with Release or
17229 // ReleaseAcquire orderings.
17230 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
17231 auto Ptr = AI->getPointerOperand();
17233 // Before the load we need a fence. Here is an example lifted from
17234 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
17237 // x.store(1, relaxed);
17238 // r1 = y.fetch_add(0, release);
17240 // y.fetch_add(42, acquire);
17241 // r2 = x.load(relaxed);
17242 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
17243 // lowered to just a load without a fence. A mfence flushes the store buffer,
17244 // making the optimization clearly correct.
17245 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
17246 // otherwise, we might be able to be more agressive on relaxed idempotent
17247 // rmw. In practice, they do not look useful, so we don't try to be
17248 // especially clever.
17249 if (SynchScope == SingleThread)
17250 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
17251 // the IR level, so we must wrap it in an intrinsic.
17254 if (!hasMFENCE(*Subtarget))
17255 // FIXME: it might make sense to use a locked operation here but on a
17256 // different cache-line to prevent cache-line bouncing. In practice it
17257 // is probably a small win, and x86 processors without mfence are rare
17258 // enough that we do not bother.
17262 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
17263 Builder.CreateCall(MFence, {});
17265 // Finally we can emit the atomic load.
17266 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
17267 AI->getType()->getPrimitiveSizeInBits());
17268 Loaded->setAtomic(Order, SynchScope);
17269 AI->replaceAllUsesWith(Loaded);
17270 AI->eraseFromParent();
17274 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
17275 SelectionDAG &DAG) {
17277 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
17278 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
17279 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
17280 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
17282 // The only fence that needs an instruction is a sequentially-consistent
17283 // cross-thread fence.
17284 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
17285 if (hasMFENCE(*Subtarget))
17286 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
17288 SDValue Chain = Op.getOperand(0);
17289 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
17291 DAG.getRegister(X86::ESP, MVT::i32), // Base
17292 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
17293 DAG.getRegister(0, MVT::i32), // Index
17294 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
17295 DAG.getRegister(0, MVT::i32), // Segment.
17299 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
17300 return SDValue(Res, 0);
17303 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
17304 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17307 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17308 SelectionDAG &DAG) {
17309 MVT T = Op.getSimpleValueType();
17313 switch(T.SimpleTy) {
17314 default: llvm_unreachable("Invalid value type!");
17315 case MVT::i8: Reg = X86::AL; size = 1; break;
17316 case MVT::i16: Reg = X86::AX; size = 2; break;
17317 case MVT::i32: Reg = X86::EAX; size = 4; break;
17319 assert(Subtarget->is64Bit() && "Node not type legal!");
17320 Reg = X86::RAX; size = 8;
17323 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17324 Op.getOperand(2), SDValue());
17325 SDValue Ops[] = { cpIn.getValue(0),
17328 DAG.getTargetConstant(size, DL, MVT::i8),
17329 cpIn.getValue(1) };
17330 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17331 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17332 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17336 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17337 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17338 MVT::i32, cpOut.getValue(2));
17339 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17340 DAG.getConstant(X86::COND_E, DL, MVT::i8),
17343 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17344 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17345 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17349 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17350 SelectionDAG &DAG) {
17351 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17352 MVT DstVT = Op.getSimpleValueType();
17354 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17355 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17356 if (DstVT != MVT::f64)
17357 // This conversion needs to be expanded.
17360 SDValue InVec = Op->getOperand(0);
17362 unsigned NumElts = SrcVT.getVectorNumElements();
17363 EVT SVT = SrcVT.getVectorElementType();
17365 // Widen the vector in input in the case of MVT::v2i32.
17366 // Example: from MVT::v2i32 to MVT::v4i32.
17367 SmallVector<SDValue, 16> Elts;
17368 for (unsigned i = 0, e = NumElts; i != e; ++i)
17369 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17370 DAG.getIntPtrConstant(i, dl)));
17372 // Explicitly mark the extra elements as Undef.
17373 Elts.append(NumElts, DAG.getUNDEF(SVT));
17375 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17376 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17377 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
17378 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17379 DAG.getIntPtrConstant(0, dl));
17382 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17383 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17384 assert((DstVT == MVT::i64 ||
17385 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17386 "Unexpected custom BITCAST");
17387 // i64 <=> MMX conversions are Legal.
17388 if (SrcVT==MVT::i64 && DstVT.isVector())
17390 if (DstVT==MVT::i64 && SrcVT.isVector())
17392 // MMX <=> MMX conversions are Legal.
17393 if (SrcVT.isVector() && DstVT.isVector())
17395 // All other conversions need to be expanded.
17399 /// Compute the horizontal sum of bytes in V for the elements of VT.
17401 /// Requires V to be a byte vector and VT to be an integer vector type with
17402 /// wider elements than V's type. The width of the elements of VT determines
17403 /// how many bytes of V are summed horizontally to produce each element of the
17405 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
17406 const X86Subtarget *Subtarget,
17407 SelectionDAG &DAG) {
17409 MVT ByteVecVT = V.getSimpleValueType();
17410 MVT EltVT = VT.getVectorElementType();
17411 int NumElts = VT.getVectorNumElements();
17412 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
17413 "Expected value to have byte element type.");
17414 assert(EltVT != MVT::i8 &&
17415 "Horizontal byte sum only makes sense for wider elements!");
17416 unsigned VecSize = VT.getSizeInBits();
17417 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
17419 // PSADBW instruction horizontally add all bytes and leave the result in i64
17420 // chunks, thus directly computes the pop count for v2i64 and v4i64.
17421 if (EltVT == MVT::i64) {
17422 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
17423 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
17424 return DAG.getBitcast(VT, V);
17427 if (EltVT == MVT::i32) {
17428 // We unpack the low half and high half into i32s interleaved with zeros so
17429 // that we can use PSADBW to horizontally sum them. The most useful part of
17430 // this is that it lines up the results of two PSADBW instructions to be
17431 // two v2i64 vectors which concatenated are the 4 population counts. We can
17432 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
17433 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
17434 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
17435 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
17437 // Do the horizontal sums into two v2i64s.
17438 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
17439 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
17440 DAG.getBitcast(ByteVecVT, Low), Zeros);
17441 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
17442 DAG.getBitcast(ByteVecVT, High), Zeros);
17444 // Merge them together.
17445 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
17446 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
17447 DAG.getBitcast(ShortVecVT, Low),
17448 DAG.getBitcast(ShortVecVT, High));
17450 return DAG.getBitcast(VT, V);
17453 // The only element type left is i16.
17454 assert(EltVT == MVT::i16 && "Unknown how to handle type");
17456 // To obtain pop count for each i16 element starting from the pop count for
17457 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
17458 // right by 8. It is important to shift as i16s as i8 vector shift isn't
17459 // directly supported.
17460 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
17461 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
17462 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
17463 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
17464 DAG.getBitcast(ByteVecVT, V));
17465 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
17468 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
17469 const X86Subtarget *Subtarget,
17470 SelectionDAG &DAG) {
17471 MVT VT = Op.getSimpleValueType();
17472 MVT EltVT = VT.getVectorElementType();
17473 unsigned VecSize = VT.getSizeInBits();
17475 // Implement a lookup table in register by using an algorithm based on:
17476 // http://wm.ite.pl/articles/sse-popcount.html
17478 // The general idea is that every lower byte nibble in the input vector is an
17479 // index into a in-register pre-computed pop count table. We then split up the
17480 // input vector in two new ones: (1) a vector with only the shifted-right
17481 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
17482 // masked out higher ones) for each byte. PSHUB is used separately with both
17483 // to index the in-register table. Next, both are added and the result is a
17484 // i8 vector where each element contains the pop count for input byte.
17486 // To obtain the pop count for elements != i8, we follow up with the same
17487 // approach and use additional tricks as described below.
17489 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
17490 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
17491 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
17492 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
17494 int NumByteElts = VecSize / 8;
17495 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
17496 SDValue In = DAG.getBitcast(ByteVecVT, Op);
17497 SmallVector<SDValue, 16> LUTVec;
17498 for (int i = 0; i < NumByteElts; ++i)
17499 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
17500 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
17501 SmallVector<SDValue, 16> Mask0F(NumByteElts,
17502 DAG.getConstant(0x0F, DL, MVT::i8));
17503 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
17506 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
17507 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
17508 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
17511 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
17513 // The input vector is used as the shuffle mask that index elements into the
17514 // LUT. After counting low and high nibbles, add the vector to obtain the
17515 // final pop count per i8 element.
17516 SDValue HighPopCnt =
17517 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
17518 SDValue LowPopCnt =
17519 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
17520 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
17522 if (EltVT == MVT::i8)
17525 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
17528 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
17529 const X86Subtarget *Subtarget,
17530 SelectionDAG &DAG) {
17531 MVT VT = Op.getSimpleValueType();
17532 assert(VT.is128BitVector() &&
17533 "Only 128-bit vector bitmath lowering supported.");
17535 int VecSize = VT.getSizeInBits();
17536 MVT EltVT = VT.getVectorElementType();
17537 int Len = EltVT.getSizeInBits();
17539 // This is the vectorized version of the "best" algorithm from
17540 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
17541 // with a minor tweak to use a series of adds + shifts instead of vector
17542 // multiplications. Implemented for all integer vector types. We only use
17543 // this when we don't have SSSE3 which allows a LUT-based lowering that is
17544 // much faster, even faster than using native popcnt instructions.
17546 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
17547 MVT VT = V.getSimpleValueType();
17548 SmallVector<SDValue, 32> Shifters(
17549 VT.getVectorNumElements(),
17550 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
17551 return DAG.getNode(OpCode, DL, VT, V,
17552 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
17554 auto GetMask = [&](SDValue V, APInt Mask) {
17555 MVT VT = V.getSimpleValueType();
17556 SmallVector<SDValue, 32> Masks(
17557 VT.getVectorNumElements(),
17558 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
17559 return DAG.getNode(ISD::AND, DL, VT, V,
17560 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
17563 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
17564 // x86, so set the SRL type to have elements at least i16 wide. This is
17565 // correct because all of our SRLs are followed immediately by a mask anyways
17566 // that handles any bits that sneak into the high bits of the byte elements.
17567 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
17571 // v = v - ((v >> 1) & 0x55555555...)
17573 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
17574 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
17575 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
17577 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
17578 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
17579 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
17580 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
17581 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
17583 // v = (v + (v >> 4)) & 0x0F0F0F0F...
17584 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
17585 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
17586 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
17588 // At this point, V contains the byte-wise population count, and we are
17589 // merely doing a horizontal sum if necessary to get the wider element
17591 if (EltVT == MVT::i8)
17594 return LowerHorizontalByteSum(
17595 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
17599 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17600 SelectionDAG &DAG) {
17601 MVT VT = Op.getSimpleValueType();
17602 // FIXME: Need to add AVX-512 support here!
17603 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17604 "Unknown CTPOP type to handle");
17605 SDLoc DL(Op.getNode());
17606 SDValue Op0 = Op.getOperand(0);
17608 if (!Subtarget->hasSSSE3()) {
17609 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
17610 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
17611 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
17614 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
17615 unsigned NumElems = VT.getVectorNumElements();
17617 // Extract each 128-bit vector, compute pop count and concat the result.
17618 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
17619 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
17621 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
17622 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
17623 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
17626 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
17629 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17630 SelectionDAG &DAG) {
17631 assert(Op.getValueType().isVector() &&
17632 "We only do custom lowering for vector population count.");
17633 return LowerVectorCTPOP(Op, Subtarget, DAG);
17636 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
17637 SDNode *Node = Op.getNode();
17639 EVT T = Node->getValueType(0);
17640 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
17641 DAG.getConstant(0, dl, T), Node->getOperand(2));
17642 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
17643 cast<AtomicSDNode>(Node)->getMemoryVT(),
17644 Node->getOperand(0),
17645 Node->getOperand(1), negOp,
17646 cast<AtomicSDNode>(Node)->getMemOperand(),
17647 cast<AtomicSDNode>(Node)->getOrdering(),
17648 cast<AtomicSDNode>(Node)->getSynchScope());
17651 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
17652 SDNode *Node = Op.getNode();
17654 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
17656 // Convert seq_cst store -> xchg
17657 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
17658 // FIXME: On 32-bit, store -> fist or movq would be more efficient
17659 // (The only way to get a 16-byte store is cmpxchg16b)
17660 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
17661 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
17662 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17663 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
17664 cast<AtomicSDNode>(Node)->getMemoryVT(),
17665 Node->getOperand(0),
17666 Node->getOperand(1), Node->getOperand(2),
17667 cast<AtomicSDNode>(Node)->getMemOperand(),
17668 cast<AtomicSDNode>(Node)->getOrdering(),
17669 cast<AtomicSDNode>(Node)->getSynchScope());
17670 return Swap.getValue(1);
17672 // Other atomic stores have a simple pattern.
17676 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
17677 EVT VT = Op.getNode()->getSimpleValueType(0);
17679 // Let legalize expand this if it isn't a legal type yet.
17680 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17683 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17686 bool ExtraOp = false;
17687 switch (Op.getOpcode()) {
17688 default: llvm_unreachable("Invalid code");
17689 case ISD::ADDC: Opc = X86ISD::ADD; break;
17690 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
17691 case ISD::SUBC: Opc = X86ISD::SUB; break;
17692 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
17696 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17698 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17699 Op.getOperand(1), Op.getOperand(2));
17702 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
17703 SelectionDAG &DAG) {
17704 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
17706 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
17707 // which returns the values as { float, float } (in XMM0) or
17708 // { double, double } (which is returned in XMM0, XMM1).
17710 SDValue Arg = Op.getOperand(0);
17711 EVT ArgVT = Arg.getValueType();
17712 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17714 TargetLowering::ArgListTy Args;
17715 TargetLowering::ArgListEntry Entry;
17719 Entry.isSExt = false;
17720 Entry.isZExt = false;
17721 Args.push_back(Entry);
17723 bool isF64 = ArgVT == MVT::f64;
17724 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
17725 // the small struct {f32, f32} is returned in (eax, edx). For f64,
17726 // the results are returned via SRet in memory.
17727 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
17728 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17729 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
17731 Type *RetTy = isF64
17732 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
17733 : (Type*)VectorType::get(ArgTy, 4);
17735 TargetLowering::CallLoweringInfo CLI(DAG);
17736 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
17737 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
17739 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
17742 // Returned in xmm0 and xmm1.
17743 return CallResult.first;
17745 // Returned in bits 0:31 and 32:64 xmm0.
17746 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17747 CallResult.first, DAG.getIntPtrConstant(0, dl));
17748 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17749 CallResult.first, DAG.getIntPtrConstant(1, dl));
17750 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
17751 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
17754 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
17755 SelectionDAG &DAG) {
17756 assert(Subtarget->hasAVX512() &&
17757 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17759 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
17760 EVT VT = N->getValue().getValueType();
17761 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
17764 // X86 scatter kills mask register, so its type should be added to
17765 // the list of return values
17766 if (N->getNumValues() == 1) {
17767 SDValue Index = N->getIndex();
17768 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17769 !Index.getValueType().is512BitVector())
17770 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17772 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
17773 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17774 N->getOperand(3), Index };
17776 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
17777 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
17778 return SDValue(NewScatter.getNode(), 0);
17783 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
17784 SelectionDAG &DAG) {
17785 assert(Subtarget->hasAVX512() &&
17786 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17788 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
17789 EVT VT = Op.getValueType();
17790 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
17793 SDValue Index = N->getIndex();
17794 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17795 !Index.getValueType().is512BitVector()) {
17796 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17797 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17798 N->getOperand(3), Index };
17799 DAG.UpdateNodeOperands(N, Ops);
17804 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
17805 SelectionDAG &DAG) const {
17806 // TODO: Eventually, the lowering of these nodes should be informed by or
17807 // deferred to the GC strategy for the function in which they appear. For
17808 // now, however, they must be lowered to something. Since they are logically
17809 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17810 // require special handling for these nodes), lower them as literal NOOPs for
17812 SmallVector<SDValue, 2> Ops;
17814 Ops.push_back(Op.getOperand(0));
17815 if (Op->getGluedNode())
17816 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17819 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17820 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17825 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
17826 SelectionDAG &DAG) const {
17827 // TODO: Eventually, the lowering of these nodes should be informed by or
17828 // deferred to the GC strategy for the function in which they appear. For
17829 // now, however, they must be lowered to something. Since they are logically
17830 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17831 // require special handling for these nodes), lower them as literal NOOPs for
17833 SmallVector<SDValue, 2> Ops;
17835 Ops.push_back(Op.getOperand(0));
17836 if (Op->getGluedNode())
17837 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17840 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17841 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17846 /// LowerOperation - Provide custom lowering hooks for some operations.
17848 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
17849 switch (Op.getOpcode()) {
17850 default: llvm_unreachable("Should not custom lower this!");
17851 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
17852 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
17853 return LowerCMP_SWAP(Op, Subtarget, DAG);
17854 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
17855 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
17856 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
17857 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
17858 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
17859 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
17860 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
17861 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
17862 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
17863 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
17864 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
17865 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
17866 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
17867 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
17868 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
17869 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
17870 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
17871 case ISD::SHL_PARTS:
17872 case ISD::SRA_PARTS:
17873 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
17874 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
17875 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
17876 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
17877 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
17878 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
17879 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
17880 case ISD::SIGN_EXTEND_VECTOR_INREG:
17881 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
17882 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
17883 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
17884 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
17885 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
17887 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
17888 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
17889 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
17890 case ISD::SETCC: return LowerSETCC(Op, DAG);
17891 case ISD::SELECT: return LowerSELECT(Op, DAG);
17892 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
17893 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
17894 case ISD::VASTART: return LowerVASTART(Op, DAG);
17895 case ISD::VAARG: return LowerVAARG(Op, DAG);
17896 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
17897 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
17898 case ISD::INTRINSIC_VOID:
17899 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
17900 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
17901 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
17902 case ISD::FRAME_TO_ARGS_OFFSET:
17903 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
17904 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
17905 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
17906 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
17907 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
17908 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
17909 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
17910 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
17911 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
17912 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
17913 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
17914 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
17915 case ISD::UMUL_LOHI:
17916 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
17919 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
17925 case ISD::UMULO: return LowerXALUO(Op, DAG);
17926 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
17927 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
17931 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
17932 case ISD::ADD: return LowerADD(Op, DAG);
17933 case ISD::SUB: return LowerSUB(Op, DAG);
17934 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
17935 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
17936 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
17937 case ISD::GC_TRANSITION_START:
17938 return LowerGC_TRANSITION_START(Op, DAG);
17939 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
17943 /// ReplaceNodeResults - Replace a node with an illegal result type
17944 /// with a new node built out of custom code.
17945 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
17946 SmallVectorImpl<SDValue>&Results,
17947 SelectionDAG &DAG) const {
17949 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17950 switch (N->getOpcode()) {
17952 llvm_unreachable("Do not know how to custom type legalize this operation!");
17953 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17954 case X86ISD::FMINC:
17956 case X86ISD::FMAXC:
17957 case X86ISD::FMAX: {
17958 EVT VT = N->getValueType(0);
17959 if (VT != MVT::v2f32)
17960 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17961 SDValue UNDEF = DAG.getUNDEF(VT);
17962 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17963 N->getOperand(0), UNDEF);
17964 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17965 N->getOperand(1), UNDEF);
17966 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17969 case ISD::SIGN_EXTEND_INREG:
17974 // We don't want to expand or promote these.
17981 case ISD::UDIVREM: {
17982 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17983 Results.push_back(V);
17986 case ISD::FP_TO_SINT:
17987 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
17988 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
17989 if (N->getOperand(0).getValueType() == MVT::f16)
17992 case ISD::FP_TO_UINT: {
17993 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17995 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17998 std::pair<SDValue,SDValue> Vals =
17999 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18000 SDValue FIST = Vals.first, StackSlot = Vals.second;
18001 if (FIST.getNode()) {
18002 EVT VT = N->getValueType(0);
18003 // Return a load from the stack slot.
18004 if (StackSlot.getNode())
18005 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18006 MachinePointerInfo(),
18007 false, false, false, 0));
18009 Results.push_back(FIST);
18013 case ISD::UINT_TO_FP: {
18014 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18015 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18016 N->getValueType(0) != MVT::v2f32)
18018 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18020 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
18022 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18023 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18024 DAG.getBitcast(MVT::v2i64, VBias));
18025 Or = DAG.getBitcast(MVT::v2f64, Or);
18026 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18027 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18030 case ISD::FP_ROUND: {
18031 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18033 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18034 Results.push_back(V);
18037 case ISD::FP_EXTEND: {
18038 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
18039 // No other ValueType for FP_EXTEND should reach this point.
18040 assert(N->getValueType(0) == MVT::v2f32 &&
18041 "Do not know how to legalize this Node");
18044 case ISD::INTRINSIC_W_CHAIN: {
18045 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18047 default : llvm_unreachable("Do not know how to custom type "
18048 "legalize this intrinsic operation!");
18049 case Intrinsic::x86_rdtsc:
18050 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18052 case Intrinsic::x86_rdtscp:
18053 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18055 case Intrinsic::x86_rdpmc:
18056 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18059 case ISD::READCYCLECOUNTER: {
18060 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18063 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18064 EVT T = N->getValueType(0);
18065 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18066 bool Regs64bit = T == MVT::i128;
18067 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18068 SDValue cpInL, cpInH;
18069 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18070 DAG.getConstant(0, dl, HalfT));
18071 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18072 DAG.getConstant(1, dl, HalfT));
18073 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18074 Regs64bit ? X86::RAX : X86::EAX,
18076 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18077 Regs64bit ? X86::RDX : X86::EDX,
18078 cpInH, cpInL.getValue(1));
18079 SDValue swapInL, swapInH;
18080 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18081 DAG.getConstant(0, dl, HalfT));
18082 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18083 DAG.getConstant(1, dl, HalfT));
18084 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18085 Regs64bit ? X86::RBX : X86::EBX,
18086 swapInL, cpInH.getValue(1));
18087 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18088 Regs64bit ? X86::RCX : X86::ECX,
18089 swapInH, swapInL.getValue(1));
18090 SDValue Ops[] = { swapInH.getValue(0),
18092 swapInH.getValue(1) };
18093 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18094 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18095 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18096 X86ISD::LCMPXCHG8_DAG;
18097 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18098 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
18099 Regs64bit ? X86::RAX : X86::EAX,
18100 HalfT, Result.getValue(1));
18101 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
18102 Regs64bit ? X86::RDX : X86::EDX,
18103 HalfT, cpOutL.getValue(2));
18104 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
18106 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
18107 MVT::i32, cpOutH.getValue(2));
18109 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18110 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
18111 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
18113 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
18114 Results.push_back(Success);
18115 Results.push_back(EFLAGS.getValue(1));
18118 case ISD::ATOMIC_SWAP:
18119 case ISD::ATOMIC_LOAD_ADD:
18120 case ISD::ATOMIC_LOAD_SUB:
18121 case ISD::ATOMIC_LOAD_AND:
18122 case ISD::ATOMIC_LOAD_OR:
18123 case ISD::ATOMIC_LOAD_XOR:
18124 case ISD::ATOMIC_LOAD_NAND:
18125 case ISD::ATOMIC_LOAD_MIN:
18126 case ISD::ATOMIC_LOAD_MAX:
18127 case ISD::ATOMIC_LOAD_UMIN:
18128 case ISD::ATOMIC_LOAD_UMAX:
18129 case ISD::ATOMIC_LOAD: {
18130 // Delegate to generic TypeLegalization. Situations we can really handle
18131 // should have already been dealt with by AtomicExpandPass.cpp.
18134 case ISD::BITCAST: {
18135 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18136 EVT DstVT = N->getValueType(0);
18137 EVT SrcVT = N->getOperand(0)->getValueType(0);
18139 if (SrcVT != MVT::f64 ||
18140 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
18143 unsigned NumElts = DstVT.getVectorNumElements();
18144 EVT SVT = DstVT.getVectorElementType();
18145 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18146 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
18147 MVT::v2f64, N->getOperand(0));
18148 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
18150 if (ExperimentalVectorWideningLegalization) {
18151 // If we are legalizing vectors by widening, we already have the desired
18152 // legal vector type, just return it.
18153 Results.push_back(ToVecInt);
18157 SmallVector<SDValue, 8> Elts;
18158 for (unsigned i = 0, e = NumElts; i != e; ++i)
18159 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
18160 ToVecInt, DAG.getIntPtrConstant(i, dl)));
18162 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
18167 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
18168 switch ((X86ISD::NodeType)Opcode) {
18169 case X86ISD::FIRST_NUMBER: break;
18170 case X86ISD::BSF: return "X86ISD::BSF";
18171 case X86ISD::BSR: return "X86ISD::BSR";
18172 case X86ISD::SHLD: return "X86ISD::SHLD";
18173 case X86ISD::SHRD: return "X86ISD::SHRD";
18174 case X86ISD::FAND: return "X86ISD::FAND";
18175 case X86ISD::FANDN: return "X86ISD::FANDN";
18176 case X86ISD::FOR: return "X86ISD::FOR";
18177 case X86ISD::FXOR: return "X86ISD::FXOR";
18178 case X86ISD::FILD: return "X86ISD::FILD";
18179 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
18180 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
18181 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
18182 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
18183 case X86ISD::FLD: return "X86ISD::FLD";
18184 case X86ISD::FST: return "X86ISD::FST";
18185 case X86ISD::CALL: return "X86ISD::CALL";
18186 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
18187 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
18188 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
18189 case X86ISD::BT: return "X86ISD::BT";
18190 case X86ISD::CMP: return "X86ISD::CMP";
18191 case X86ISD::COMI: return "X86ISD::COMI";
18192 case X86ISD::UCOMI: return "X86ISD::UCOMI";
18193 case X86ISD::CMPM: return "X86ISD::CMPM";
18194 case X86ISD::CMPMU: return "X86ISD::CMPMU";
18195 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
18196 case X86ISD::SETCC: return "X86ISD::SETCC";
18197 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
18198 case X86ISD::FSETCC: return "X86ISD::FSETCC";
18199 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
18200 case X86ISD::CMOV: return "X86ISD::CMOV";
18201 case X86ISD::BRCOND: return "X86ISD::BRCOND";
18202 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
18203 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
18204 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
18205 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
18206 case X86ISD::Wrapper: return "X86ISD::Wrapper";
18207 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
18208 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
18209 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
18210 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
18211 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
18212 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
18213 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
18214 case X86ISD::PINSRB: return "X86ISD::PINSRB";
18215 case X86ISD::PINSRW: return "X86ISD::PINSRW";
18216 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
18217 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
18218 case X86ISD::ANDNP: return "X86ISD::ANDNP";
18219 case X86ISD::PSIGN: return "X86ISD::PSIGN";
18220 case X86ISD::BLENDI: return "X86ISD::BLENDI";
18221 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
18222 case X86ISD::ADDUS: return "X86ISD::ADDUS";
18223 case X86ISD::SUBUS: return "X86ISD::SUBUS";
18224 case X86ISD::HADD: return "X86ISD::HADD";
18225 case X86ISD::HSUB: return "X86ISD::HSUB";
18226 case X86ISD::FHADD: return "X86ISD::FHADD";
18227 case X86ISD::FHSUB: return "X86ISD::FHSUB";
18228 case X86ISD::UMAX: return "X86ISD::UMAX";
18229 case X86ISD::UMIN: return "X86ISD::UMIN";
18230 case X86ISD::SMAX: return "X86ISD::SMAX";
18231 case X86ISD::SMIN: return "X86ISD::SMIN";
18232 case X86ISD::FMAX: return "X86ISD::FMAX";
18233 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
18234 case X86ISD::FMIN: return "X86ISD::FMIN";
18235 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
18236 case X86ISD::FMAXC: return "X86ISD::FMAXC";
18237 case X86ISD::FMINC: return "X86ISD::FMINC";
18238 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
18239 case X86ISD::FRCP: return "X86ISD::FRCP";
18240 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
18241 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
18242 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
18243 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
18244 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
18245 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
18246 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
18247 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
18248 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
18249 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
18250 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
18251 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
18252 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
18253 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
18254 case X86ISD::VZEXT: return "X86ISD::VZEXT";
18255 case X86ISD::VSEXT: return "X86ISD::VSEXT";
18256 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
18257 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
18258 case X86ISD::VINSERT: return "X86ISD::VINSERT";
18259 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
18260 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
18261 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
18262 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
18263 case X86ISD::VSHL: return "X86ISD::VSHL";
18264 case X86ISD::VSRL: return "X86ISD::VSRL";
18265 case X86ISD::VSRA: return "X86ISD::VSRA";
18266 case X86ISD::VSHLI: return "X86ISD::VSHLI";
18267 case X86ISD::VSRLI: return "X86ISD::VSRLI";
18268 case X86ISD::VSRAI: return "X86ISD::VSRAI";
18269 case X86ISD::CMPP: return "X86ISD::CMPP";
18270 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
18271 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
18272 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
18273 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
18274 case X86ISD::ADD: return "X86ISD::ADD";
18275 case X86ISD::SUB: return "X86ISD::SUB";
18276 case X86ISD::ADC: return "X86ISD::ADC";
18277 case X86ISD::SBB: return "X86ISD::SBB";
18278 case X86ISD::SMUL: return "X86ISD::SMUL";
18279 case X86ISD::UMUL: return "X86ISD::UMUL";
18280 case X86ISD::SMUL8: return "X86ISD::SMUL8";
18281 case X86ISD::UMUL8: return "X86ISD::UMUL8";
18282 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
18283 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
18284 case X86ISD::INC: return "X86ISD::INC";
18285 case X86ISD::DEC: return "X86ISD::DEC";
18286 case X86ISD::OR: return "X86ISD::OR";
18287 case X86ISD::XOR: return "X86ISD::XOR";
18288 case X86ISD::AND: return "X86ISD::AND";
18289 case X86ISD::BEXTR: return "X86ISD::BEXTR";
18290 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
18291 case X86ISD::PTEST: return "X86ISD::PTEST";
18292 case X86ISD::TESTP: return "X86ISD::TESTP";
18293 case X86ISD::TESTM: return "X86ISD::TESTM";
18294 case X86ISD::TESTNM: return "X86ISD::TESTNM";
18295 case X86ISD::KORTEST: return "X86ISD::KORTEST";
18296 case X86ISD::PACKSS: return "X86ISD::PACKSS";
18297 case X86ISD::PACKUS: return "X86ISD::PACKUS";
18298 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
18299 case X86ISD::VALIGN: return "X86ISD::VALIGN";
18300 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
18301 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
18302 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
18303 case X86ISD::SHUFP: return "X86ISD::SHUFP";
18304 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
18305 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
18306 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
18307 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
18308 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
18309 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
18310 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
18311 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
18312 case X86ISD::MOVSD: return "X86ISD::MOVSD";
18313 case X86ISD::MOVSS: return "X86ISD::MOVSS";
18314 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
18315 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
18316 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
18317 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
18318 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
18319 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
18320 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
18321 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
18322 case X86ISD::VPERMV: return "X86ISD::VPERMV";
18323 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
18324 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
18325 case X86ISD::VPERMI: return "X86ISD::VPERMI";
18326 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
18327 case X86ISD::VRANGE: return "X86ISD::VRANGE";
18328 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
18329 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
18330 case X86ISD::PSADBW: return "X86ISD::PSADBW";
18331 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
18332 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
18333 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
18334 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
18335 case X86ISD::MFENCE: return "X86ISD::MFENCE";
18336 case X86ISD::SFENCE: return "X86ISD::SFENCE";
18337 case X86ISD::LFENCE: return "X86ISD::LFENCE";
18338 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
18339 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
18340 case X86ISD::SAHF: return "X86ISD::SAHF";
18341 case X86ISD::RDRAND: return "X86ISD::RDRAND";
18342 case X86ISD::RDSEED: return "X86ISD::RDSEED";
18343 case X86ISD::FMADD: return "X86ISD::FMADD";
18344 case X86ISD::FMSUB: return "X86ISD::FMSUB";
18345 case X86ISD::FNMADD: return "X86ISD::FNMADD";
18346 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
18347 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
18348 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
18349 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
18350 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
18351 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
18352 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
18353 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
18354 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
18355 case X86ISD::RNDSCALE: return "X86ISD::RNDSCALE";
18356 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
18357 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
18358 case X86ISD::XTEST: return "X86ISD::XTEST";
18359 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
18360 case X86ISD::EXPAND: return "X86ISD::EXPAND";
18361 case X86ISD::SELECT: return "X86ISD::SELECT";
18362 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
18363 case X86ISD::RCP28: return "X86ISD::RCP28";
18364 case X86ISD::EXP2: return "X86ISD::EXP2";
18365 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
18366 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
18367 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
18368 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
18369 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
18370 case X86ISD::ADDS: return "X86ISD::ADDS";
18371 case X86ISD::SUBS: return "X86ISD::SUBS";
18376 // isLegalAddressingMode - Return true if the addressing mode represented
18377 // by AM is legal for this target, for a load/store of the specified type.
18378 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
18380 unsigned AS) const {
18381 // X86 supports extremely general addressing modes.
18382 CodeModel::Model M = getTargetMachine().getCodeModel();
18383 Reloc::Model R = getTargetMachine().getRelocationModel();
18385 // X86 allows a sign-extended 32-bit immediate field as a displacement.
18386 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
18391 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
18393 // If a reference to this global requires an extra load, we can't fold it.
18394 if (isGlobalStubReference(GVFlags))
18397 // If BaseGV requires a register for the PIC base, we cannot also have a
18398 // BaseReg specified.
18399 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
18402 // If lower 4G is not available, then we must use rip-relative addressing.
18403 if ((M != CodeModel::Small || R != Reloc::Static) &&
18404 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
18408 switch (AM.Scale) {
18414 // These scales always work.
18419 // These scales are formed with basereg+scalereg. Only accept if there is
18424 default: // Other stuff never works.
18431 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
18432 unsigned Bits = Ty->getScalarSizeInBits();
18434 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
18435 // particularly cheaper than those without.
18439 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
18440 // variable shifts just as cheap as scalar ones.
18441 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
18444 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
18445 // fully general vector.
18449 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
18450 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18452 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
18453 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
18454 return NumBits1 > NumBits2;
18457 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
18458 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18461 if (!isTypeLegal(EVT::getEVT(Ty1)))
18464 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
18466 // Assuming the caller doesn't have a zeroext or signext return parameter,
18467 // truncation all the way down to i1 is valid.
18471 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
18472 return isInt<32>(Imm);
18475 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
18476 // Can also use sub to handle negated immediates.
18477 return isInt<32>(Imm);
18480 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
18481 if (!VT1.isInteger() || !VT2.isInteger())
18483 unsigned NumBits1 = VT1.getSizeInBits();
18484 unsigned NumBits2 = VT2.getSizeInBits();
18485 return NumBits1 > NumBits2;
18488 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
18489 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18490 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
18493 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
18494 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18495 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
18498 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
18499 EVT VT1 = Val.getValueType();
18500 if (isZExtFree(VT1, VT2))
18503 if (Val.getOpcode() != ISD::LOAD)
18506 if (!VT1.isSimple() || !VT1.isInteger() ||
18507 !VT2.isSimple() || !VT2.isInteger())
18510 switch (VT1.getSimpleVT().SimpleTy) {
18515 // X86 has 8, 16, and 32-bit zero-extending loads.
18522 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
18525 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
18526 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
18529 VT = VT.getScalarType();
18531 if (!VT.isSimple())
18534 switch (VT.getSimpleVT().SimpleTy) {
18545 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
18546 // i16 instructions are longer (0x66 prefix) and potentially slower.
18547 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18550 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18551 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18552 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18553 /// are assumed to be legal.
18555 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18557 if (!VT.isSimple())
18560 // Not for i1 vectors
18561 if (VT.getScalarType() == MVT::i1)
18564 // Very little shuffling can be done for 64-bit vectors right now.
18565 if (VT.getSizeInBits() == 64)
18568 // We only care that the types being shuffled are legal. The lowering can
18569 // handle any possible shuffle mask that results.
18570 return isTypeLegal(VT.getSimpleVT());
18574 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18576 // Just delegate to the generic legality, clear masks aren't special.
18577 return isShuffleMaskLegal(Mask, VT);
18580 //===----------------------------------------------------------------------===//
18581 // X86 Scheduler Hooks
18582 //===----------------------------------------------------------------------===//
18584 /// Utility function to emit xbegin specifying the start of an RTM region.
18585 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18586 const TargetInstrInfo *TII) {
18587 DebugLoc DL = MI->getDebugLoc();
18589 const BasicBlock *BB = MBB->getBasicBlock();
18590 MachineFunction::iterator I = MBB;
18593 // For the v = xbegin(), we generate
18604 MachineBasicBlock *thisMBB = MBB;
18605 MachineFunction *MF = MBB->getParent();
18606 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18607 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18608 MF->insert(I, mainMBB);
18609 MF->insert(I, sinkMBB);
18611 // Transfer the remainder of BB and its successor edges to sinkMBB.
18612 sinkMBB->splice(sinkMBB->begin(), MBB,
18613 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18614 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18618 // # fallthrough to mainMBB
18619 // # abortion to sinkMBB
18620 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18621 thisMBB->addSuccessor(mainMBB);
18622 thisMBB->addSuccessor(sinkMBB);
18626 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18627 mainMBB->addSuccessor(sinkMBB);
18630 // EAX is live into the sinkMBB
18631 sinkMBB->addLiveIn(X86::EAX);
18632 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18633 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18636 MI->eraseFromParent();
18640 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18641 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18642 // in the .td file.
18643 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18644 const TargetInstrInfo *TII) {
18646 switch (MI->getOpcode()) {
18647 default: llvm_unreachable("illegal opcode!");
18648 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18649 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18650 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18651 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18652 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18653 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18654 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18655 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18658 DebugLoc dl = MI->getDebugLoc();
18659 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18661 unsigned NumArgs = MI->getNumOperands();
18662 for (unsigned i = 1; i < NumArgs; ++i) {
18663 MachineOperand &Op = MI->getOperand(i);
18664 if (!(Op.isReg() && Op.isImplicit()))
18665 MIB.addOperand(Op);
18667 if (MI->hasOneMemOperand())
18668 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18670 BuildMI(*BB, MI, dl,
18671 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18672 .addReg(X86::XMM0);
18674 MI->eraseFromParent();
18678 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18679 // defs in an instruction pattern
18680 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18681 const TargetInstrInfo *TII) {
18683 switch (MI->getOpcode()) {
18684 default: llvm_unreachable("illegal opcode!");
18685 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18686 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18687 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18688 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18689 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18690 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18691 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18692 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18695 DebugLoc dl = MI->getDebugLoc();
18696 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18698 unsigned NumArgs = MI->getNumOperands(); // remove the results
18699 for (unsigned i = 1; i < NumArgs; ++i) {
18700 MachineOperand &Op = MI->getOperand(i);
18701 if (!(Op.isReg() && Op.isImplicit()))
18702 MIB.addOperand(Op);
18704 if (MI->hasOneMemOperand())
18705 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18707 BuildMI(*BB, MI, dl,
18708 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18711 MI->eraseFromParent();
18715 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18716 const X86Subtarget *Subtarget) {
18717 DebugLoc dl = MI->getDebugLoc();
18718 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18719 // Address into RAX/EAX, other two args into ECX, EDX.
18720 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18721 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18722 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18723 for (int i = 0; i < X86::AddrNumOperands; ++i)
18724 MIB.addOperand(MI->getOperand(i));
18726 unsigned ValOps = X86::AddrNumOperands;
18727 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18728 .addReg(MI->getOperand(ValOps).getReg());
18729 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18730 .addReg(MI->getOperand(ValOps+1).getReg());
18732 // The instruction doesn't actually take any operands though.
18733 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18735 MI->eraseFromParent(); // The pseudo is gone now.
18739 MachineBasicBlock *
18740 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
18741 MachineBasicBlock *MBB) const {
18742 // Emit va_arg instruction on X86-64.
18744 // Operands to this pseudo-instruction:
18745 // 0 ) Output : destination address (reg)
18746 // 1-5) Input : va_list address (addr, i64mem)
18747 // 6 ) ArgSize : Size (in bytes) of vararg type
18748 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18749 // 8 ) Align : Alignment of type
18750 // 9 ) EFLAGS (implicit-def)
18752 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18753 static_assert(X86::AddrNumOperands == 5,
18754 "VAARG_64 assumes 5 address operands");
18756 unsigned DestReg = MI->getOperand(0).getReg();
18757 MachineOperand &Base = MI->getOperand(1);
18758 MachineOperand &Scale = MI->getOperand(2);
18759 MachineOperand &Index = MI->getOperand(3);
18760 MachineOperand &Disp = MI->getOperand(4);
18761 MachineOperand &Segment = MI->getOperand(5);
18762 unsigned ArgSize = MI->getOperand(6).getImm();
18763 unsigned ArgMode = MI->getOperand(7).getImm();
18764 unsigned Align = MI->getOperand(8).getImm();
18766 // Memory Reference
18767 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
18768 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18769 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18771 // Machine Information
18772 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18773 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
18774 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
18775 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
18776 DebugLoc DL = MI->getDebugLoc();
18778 // struct va_list {
18781 // i64 overflow_area (address)
18782 // i64 reg_save_area (address)
18784 // sizeof(va_list) = 24
18785 // alignment(va_list) = 8
18787 unsigned TotalNumIntRegs = 6;
18788 unsigned TotalNumXMMRegs = 8;
18789 bool UseGPOffset = (ArgMode == 1);
18790 bool UseFPOffset = (ArgMode == 2);
18791 unsigned MaxOffset = TotalNumIntRegs * 8 +
18792 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
18794 /* Align ArgSize to a multiple of 8 */
18795 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
18796 bool NeedsAlign = (Align > 8);
18798 MachineBasicBlock *thisMBB = MBB;
18799 MachineBasicBlock *overflowMBB;
18800 MachineBasicBlock *offsetMBB;
18801 MachineBasicBlock *endMBB;
18803 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
18804 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
18805 unsigned OffsetReg = 0;
18807 if (!UseGPOffset && !UseFPOffset) {
18808 // If we only pull from the overflow region, we don't create a branch.
18809 // We don't need to alter control flow.
18810 OffsetDestReg = 0; // unused
18811 OverflowDestReg = DestReg;
18813 offsetMBB = nullptr;
18814 overflowMBB = thisMBB;
18817 // First emit code to check if gp_offset (or fp_offset) is below the bound.
18818 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
18819 // If not, pull from overflow_area. (branch to overflowMBB)
18824 // offsetMBB overflowMBB
18829 // Registers for the PHI in endMBB
18830 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
18831 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
18833 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18834 MachineFunction *MF = MBB->getParent();
18835 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18836 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18837 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18839 MachineFunction::iterator MBBIter = MBB;
18842 // Insert the new basic blocks
18843 MF->insert(MBBIter, offsetMBB);
18844 MF->insert(MBBIter, overflowMBB);
18845 MF->insert(MBBIter, endMBB);
18847 // Transfer the remainder of MBB and its successor edges to endMBB.
18848 endMBB->splice(endMBB->begin(), thisMBB,
18849 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
18850 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
18852 // Make offsetMBB and overflowMBB successors of thisMBB
18853 thisMBB->addSuccessor(offsetMBB);
18854 thisMBB->addSuccessor(overflowMBB);
18856 // endMBB is a successor of both offsetMBB and overflowMBB
18857 offsetMBB->addSuccessor(endMBB);
18858 overflowMBB->addSuccessor(endMBB);
18860 // Load the offset value into a register
18861 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18862 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
18866 .addDisp(Disp, UseFPOffset ? 4 : 0)
18867 .addOperand(Segment)
18868 .setMemRefs(MMOBegin, MMOEnd);
18870 // Check if there is enough room left to pull this argument.
18871 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
18873 .addImm(MaxOffset + 8 - ArgSizeA8);
18875 // Branch to "overflowMBB" if offset >= max
18876 // Fall through to "offsetMBB" otherwise
18877 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
18878 .addMBB(overflowMBB);
18881 // In offsetMBB, emit code to use the reg_save_area.
18883 assert(OffsetReg != 0);
18885 // Read the reg_save_area address.
18886 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
18887 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
18892 .addOperand(Segment)
18893 .setMemRefs(MMOBegin, MMOEnd);
18895 // Zero-extend the offset
18896 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
18897 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
18900 .addImm(X86::sub_32bit);
18902 // Add the offset to the reg_save_area to get the final address.
18903 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
18904 .addReg(OffsetReg64)
18905 .addReg(RegSaveReg);
18907 // Compute the offset for the next argument
18908 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18909 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
18911 .addImm(UseFPOffset ? 16 : 8);
18913 // Store it back into the va_list.
18914 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
18918 .addDisp(Disp, UseFPOffset ? 4 : 0)
18919 .addOperand(Segment)
18920 .addReg(NextOffsetReg)
18921 .setMemRefs(MMOBegin, MMOEnd);
18924 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
18929 // Emit code to use overflow area
18932 // Load the overflow_area address into a register.
18933 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
18934 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
18939 .addOperand(Segment)
18940 .setMemRefs(MMOBegin, MMOEnd);
18942 // If we need to align it, do so. Otherwise, just copy the address
18943 // to OverflowDestReg.
18945 // Align the overflow address
18946 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
18947 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
18949 // aligned_addr = (addr + (align-1)) & ~(align-1)
18950 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
18951 .addReg(OverflowAddrReg)
18954 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
18956 .addImm(~(uint64_t)(Align-1));
18958 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
18959 .addReg(OverflowAddrReg);
18962 // Compute the next overflow address after this argument.
18963 // (the overflow address should be kept 8-byte aligned)
18964 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
18965 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
18966 .addReg(OverflowDestReg)
18967 .addImm(ArgSizeA8);
18969 // Store the new overflow address.
18970 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
18975 .addOperand(Segment)
18976 .addReg(NextAddrReg)
18977 .setMemRefs(MMOBegin, MMOEnd);
18979 // If we branched, emit the PHI to the front of endMBB.
18981 BuildMI(*endMBB, endMBB->begin(), DL,
18982 TII->get(X86::PHI), DestReg)
18983 .addReg(OffsetDestReg).addMBB(offsetMBB)
18984 .addReg(OverflowDestReg).addMBB(overflowMBB);
18987 // Erase the pseudo instruction
18988 MI->eraseFromParent();
18993 MachineBasicBlock *
18994 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18996 MachineBasicBlock *MBB) const {
18997 // Emit code to save XMM registers to the stack. The ABI says that the
18998 // number of registers to save is given in %al, so it's theoretically
18999 // possible to do an indirect jump trick to avoid saving all of them,
19000 // however this code takes a simpler approach and just executes all
19001 // of the stores if %al is non-zero. It's less code, and it's probably
19002 // easier on the hardware branch predictor, and stores aren't all that
19003 // expensive anyway.
19005 // Create the new basic blocks. One block contains all the XMM stores,
19006 // and one block is the final destination regardless of whether any
19007 // stores were performed.
19008 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19009 MachineFunction *F = MBB->getParent();
19010 MachineFunction::iterator MBBIter = MBB;
19012 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19013 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19014 F->insert(MBBIter, XMMSaveMBB);
19015 F->insert(MBBIter, EndMBB);
19017 // Transfer the remainder of MBB and its successor edges to EndMBB.
19018 EndMBB->splice(EndMBB->begin(), MBB,
19019 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19020 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19022 // The original block will now fall through to the XMM save block.
19023 MBB->addSuccessor(XMMSaveMBB);
19024 // The XMMSaveMBB will fall through to the end block.
19025 XMMSaveMBB->addSuccessor(EndMBB);
19027 // Now add the instructions.
19028 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19029 DebugLoc DL = MI->getDebugLoc();
19031 unsigned CountReg = MI->getOperand(0).getReg();
19032 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19033 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19035 if (!Subtarget->isTargetWin64()) {
19036 // If %al is 0, branch around the XMM save block.
19037 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19038 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
19039 MBB->addSuccessor(EndMBB);
19042 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19043 // that was just emitted, but clearly shouldn't be "saved".
19044 assert((MI->getNumOperands() <= 3 ||
19045 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19046 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19047 && "Expected last argument to be EFLAGS");
19048 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19049 // In the XMM save block, save all the XMM argument registers.
19050 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19051 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19052 MachineMemOperand *MMO =
19053 F->getMachineMemOperand(
19054 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
19055 MachineMemOperand::MOStore,
19056 /*Size=*/16, /*Align=*/16);
19057 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19058 .addFrameIndex(RegSaveFrameIndex)
19059 .addImm(/*Scale=*/1)
19060 .addReg(/*IndexReg=*/0)
19061 .addImm(/*Disp=*/Offset)
19062 .addReg(/*Segment=*/0)
19063 .addReg(MI->getOperand(i).getReg())
19064 .addMemOperand(MMO);
19067 MI->eraseFromParent(); // The pseudo instruction is gone now.
19072 // The EFLAGS operand of SelectItr might be missing a kill marker
19073 // because there were multiple uses of EFLAGS, and ISel didn't know
19074 // which to mark. Figure out whether SelectItr should have had a
19075 // kill marker, and set it if it should. Returns the correct kill
19077 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19078 MachineBasicBlock* BB,
19079 const TargetRegisterInfo* TRI) {
19080 // Scan forward through BB for a use/def of EFLAGS.
19081 MachineBasicBlock::iterator miI(std::next(SelectItr));
19082 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19083 const MachineInstr& mi = *miI;
19084 if (mi.readsRegister(X86::EFLAGS))
19086 if (mi.definesRegister(X86::EFLAGS))
19087 break; // Should have kill-flag - update below.
19090 // If we hit the end of the block, check whether EFLAGS is live into a
19092 if (miI == BB->end()) {
19093 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
19094 sEnd = BB->succ_end();
19095 sItr != sEnd; ++sItr) {
19096 MachineBasicBlock* succ = *sItr;
19097 if (succ->isLiveIn(X86::EFLAGS))
19102 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
19103 // out. SelectMI should have a kill flag on EFLAGS.
19104 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
19108 MachineBasicBlock *
19109 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
19110 MachineBasicBlock *BB) const {
19111 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19112 DebugLoc DL = MI->getDebugLoc();
19114 // To "insert" a SELECT_CC instruction, we actually have to insert the
19115 // diamond control-flow pattern. The incoming instruction knows the
19116 // destination vreg to set, the condition code register to branch on, the
19117 // true/false values to select between, and a branch opcode to use.
19118 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19119 MachineFunction::iterator It = BB;
19125 // cmpTY ccX, r1, r2
19127 // fallthrough --> copy0MBB
19128 MachineBasicBlock *thisMBB = BB;
19129 MachineFunction *F = BB->getParent();
19131 // We also lower double CMOVs:
19132 // (CMOV (CMOV F, T, cc1), T, cc2)
19133 // to two successives branches. For that, we look for another CMOV as the
19134 // following instruction.
19136 // Without this, we would add a PHI between the two jumps, which ends up
19137 // creating a few copies all around. For instance, for
19139 // (sitofp (zext (fcmp une)))
19141 // we would generate:
19143 // ucomiss %xmm1, %xmm0
19144 // movss <1.0f>, %xmm0
19145 // movaps %xmm0, %xmm1
19147 // xorps %xmm1, %xmm1
19150 // movaps %xmm1, %xmm0
19154 // because this custom-inserter would have generated:
19166 // A: X = ...; Y = ...
19168 // C: Z = PHI [X, A], [Y, B]
19170 // E: PHI [X, C], [Z, D]
19172 // If we lower both CMOVs in a single step, we can instead generate:
19184 // A: X = ...; Y = ...
19186 // E: PHI [X, A], [X, C], [Y, D]
19188 // Which, in our sitofp/fcmp example, gives us something like:
19190 // ucomiss %xmm1, %xmm0
19191 // movss <1.0f>, %xmm0
19194 // xorps %xmm0, %xmm0
19198 MachineInstr *NextCMOV = nullptr;
19199 MachineBasicBlock::iterator NextMIIt =
19200 std::next(MachineBasicBlock::iterator(MI));
19201 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
19202 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
19203 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
19204 NextCMOV = &*NextMIIt;
19206 MachineBasicBlock *jcc1MBB = nullptr;
19208 // If we have a double CMOV, we lower it to two successive branches to
19209 // the same block. EFLAGS is used by both, so mark it as live in the second.
19211 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
19212 F->insert(It, jcc1MBB);
19213 jcc1MBB->addLiveIn(X86::EFLAGS);
19216 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
19217 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
19218 F->insert(It, copy0MBB);
19219 F->insert(It, sinkMBB);
19221 // If the EFLAGS register isn't dead in the terminator, then claim that it's
19222 // live into the sink and copy blocks.
19223 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
19225 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
19226 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
19227 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
19228 copy0MBB->addLiveIn(X86::EFLAGS);
19229 sinkMBB->addLiveIn(X86::EFLAGS);
19232 // Transfer the remainder of BB and its successor edges to sinkMBB.
19233 sinkMBB->splice(sinkMBB->begin(), BB,
19234 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19235 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
19237 // Add the true and fallthrough blocks as its successors.
19239 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
19240 BB->addSuccessor(jcc1MBB);
19242 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
19243 // jump to the sinkMBB.
19244 jcc1MBB->addSuccessor(copy0MBB);
19245 jcc1MBB->addSuccessor(sinkMBB);
19247 BB->addSuccessor(copy0MBB);
19250 // The true block target of the first (or only) branch is always sinkMBB.
19251 BB->addSuccessor(sinkMBB);
19253 // Create the conditional branch instruction.
19255 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
19256 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
19259 unsigned Opc2 = X86::GetCondBranchFromCond(
19260 (X86::CondCode)NextCMOV->getOperand(3).getImm());
19261 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
19265 // %FalseValue = ...
19266 // # fallthrough to sinkMBB
19267 copy0MBB->addSuccessor(sinkMBB);
19270 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
19272 MachineInstrBuilder MIB =
19273 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
19274 MI->getOperand(0).getReg())
19275 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
19276 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
19278 // If we have a double CMOV, the second Jcc provides the same incoming
19279 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
19281 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
19282 // Copy the PHI result to the register defined by the second CMOV.
19283 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
19284 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
19285 .addReg(MI->getOperand(0).getReg());
19286 NextCMOV->eraseFromParent();
19289 MI->eraseFromParent(); // The pseudo instruction is gone now.
19293 MachineBasicBlock *
19294 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
19295 MachineBasicBlock *BB) const {
19296 MachineFunction *MF = BB->getParent();
19297 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19298 DebugLoc DL = MI->getDebugLoc();
19299 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19301 assert(MF->shouldSplitStack());
19303 const bool Is64Bit = Subtarget->is64Bit();
19304 const bool IsLP64 = Subtarget->isTarget64BitLP64();
19306 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
19307 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
19310 // ... [Till the alloca]
19311 // If stacklet is not large enough, jump to mallocMBB
19314 // Allocate by subtracting from RSP
19315 // Jump to continueMBB
19318 // Allocate by call to runtime
19322 // [rest of original BB]
19325 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19326 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19327 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19329 MachineRegisterInfo &MRI = MF->getRegInfo();
19330 const TargetRegisterClass *AddrRegClass =
19331 getRegClassFor(getPointerTy());
19333 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19334 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19335 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
19336 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
19337 sizeVReg = MI->getOperand(1).getReg(),
19338 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
19340 MachineFunction::iterator MBBIter = BB;
19343 MF->insert(MBBIter, bumpMBB);
19344 MF->insert(MBBIter, mallocMBB);
19345 MF->insert(MBBIter, continueMBB);
19347 continueMBB->splice(continueMBB->begin(), BB,
19348 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19349 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
19351 // Add code to the main basic block to check if the stack limit has been hit,
19352 // and if so, jump to mallocMBB otherwise to bumpMBB.
19353 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
19354 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
19355 .addReg(tmpSPVReg).addReg(sizeVReg);
19356 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
19357 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
19358 .addReg(SPLimitVReg);
19359 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
19361 // bumpMBB simply decreases the stack pointer, since we know the current
19362 // stacklet has enough space.
19363 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
19364 .addReg(SPLimitVReg);
19365 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
19366 .addReg(SPLimitVReg);
19367 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19369 // Calls into a routine in libgcc to allocate more space from the heap.
19370 const uint32_t *RegMask =
19371 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
19373 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
19375 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19376 .addExternalSymbol("__morestack_allocate_stack_space")
19377 .addRegMask(RegMask)
19378 .addReg(X86::RDI, RegState::Implicit)
19379 .addReg(X86::RAX, RegState::ImplicitDefine);
19380 } else if (Is64Bit) {
19381 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
19383 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19384 .addExternalSymbol("__morestack_allocate_stack_space")
19385 .addRegMask(RegMask)
19386 .addReg(X86::EDI, RegState::Implicit)
19387 .addReg(X86::EAX, RegState::ImplicitDefine);
19389 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
19391 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
19392 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
19393 .addExternalSymbol("__morestack_allocate_stack_space")
19394 .addRegMask(RegMask)
19395 .addReg(X86::EAX, RegState::ImplicitDefine);
19399 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
19402 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
19403 .addReg(IsLP64 ? X86::RAX : X86::EAX);
19404 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19406 // Set up the CFG correctly.
19407 BB->addSuccessor(bumpMBB);
19408 BB->addSuccessor(mallocMBB);
19409 mallocMBB->addSuccessor(continueMBB);
19410 bumpMBB->addSuccessor(continueMBB);
19412 // Take care of the PHI nodes.
19413 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
19414 MI->getOperand(0).getReg())
19415 .addReg(mallocPtrVReg).addMBB(mallocMBB)
19416 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
19418 // Delete the original pseudo instruction.
19419 MI->eraseFromParent();
19422 return continueMBB;
19425 MachineBasicBlock *
19426 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
19427 MachineBasicBlock *BB) const {
19428 DebugLoc DL = MI->getDebugLoc();
19430 assert(!Subtarget->isTargetMachO());
19432 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
19434 MI->eraseFromParent(); // The pseudo instruction is gone now.
19438 MachineBasicBlock *
19439 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
19440 MachineBasicBlock *BB) const {
19441 // This is pretty easy. We're taking the value that we received from
19442 // our load from the relocation, sticking it in either RDI (x86-64)
19443 // or EAX and doing an indirect call. The return value will then
19444 // be in the normal return register.
19445 MachineFunction *F = BB->getParent();
19446 const X86InstrInfo *TII = Subtarget->getInstrInfo();
19447 DebugLoc DL = MI->getDebugLoc();
19449 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
19450 assert(MI->getOperand(3).isGlobal() && "This should be a global");
19452 // Get a register mask for the lowered call.
19453 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
19454 // proper register mask.
19455 const uint32_t *RegMask =
19456 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
19457 if (Subtarget->is64Bit()) {
19458 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19459 TII->get(X86::MOV64rm), X86::RDI)
19461 .addImm(0).addReg(0)
19462 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19463 MI->getOperand(3).getTargetFlags())
19465 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
19466 addDirectMem(MIB, X86::RDI);
19467 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
19468 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
19469 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19470 TII->get(X86::MOV32rm), X86::EAX)
19472 .addImm(0).addReg(0)
19473 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19474 MI->getOperand(3).getTargetFlags())
19476 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19477 addDirectMem(MIB, X86::EAX);
19478 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19480 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19481 TII->get(X86::MOV32rm), X86::EAX)
19482 .addReg(TII->getGlobalBaseReg(F))
19483 .addImm(0).addReg(0)
19484 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19485 MI->getOperand(3).getTargetFlags())
19487 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19488 addDirectMem(MIB, X86::EAX);
19489 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19492 MI->eraseFromParent(); // The pseudo instruction is gone now.
19496 MachineBasicBlock *
19497 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
19498 MachineBasicBlock *MBB) const {
19499 DebugLoc DL = MI->getDebugLoc();
19500 MachineFunction *MF = MBB->getParent();
19501 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19502 MachineRegisterInfo &MRI = MF->getRegInfo();
19504 const BasicBlock *BB = MBB->getBasicBlock();
19505 MachineFunction::iterator I = MBB;
19508 // Memory Reference
19509 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19510 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19513 unsigned MemOpndSlot = 0;
19515 unsigned CurOp = 0;
19517 DstReg = MI->getOperand(CurOp++).getReg();
19518 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
19519 assert(RC->hasType(MVT::i32) && "Invalid destination!");
19520 unsigned mainDstReg = MRI.createVirtualRegister(RC);
19521 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
19523 MemOpndSlot = CurOp;
19525 MVT PVT = getPointerTy();
19526 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19527 "Invalid Pointer Size!");
19529 // For v = setjmp(buf), we generate
19532 // buf[LabelOffset] = restoreMBB
19533 // SjLjSetup restoreMBB
19539 // v = phi(main, restore)
19542 // if base pointer being used, load it from frame
19545 MachineBasicBlock *thisMBB = MBB;
19546 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19547 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19548 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
19549 MF->insert(I, mainMBB);
19550 MF->insert(I, sinkMBB);
19551 MF->push_back(restoreMBB);
19553 MachineInstrBuilder MIB;
19555 // Transfer the remainder of BB and its successor edges to sinkMBB.
19556 sinkMBB->splice(sinkMBB->begin(), MBB,
19557 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19558 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19561 unsigned PtrStoreOpc = 0;
19562 unsigned LabelReg = 0;
19563 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19564 Reloc::Model RM = MF->getTarget().getRelocationModel();
19565 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19566 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19568 // Prepare IP either in reg or imm.
19569 if (!UseImmLabel) {
19570 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19571 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19572 LabelReg = MRI.createVirtualRegister(PtrRC);
19573 if (Subtarget->is64Bit()) {
19574 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19578 .addMBB(restoreMBB)
19581 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19582 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19583 .addReg(XII->getGlobalBaseReg(MF))
19586 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19590 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19592 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19593 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19594 if (i == X86::AddrDisp)
19595 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19597 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19600 MIB.addReg(LabelReg);
19602 MIB.addMBB(restoreMBB);
19603 MIB.setMemRefs(MMOBegin, MMOEnd);
19605 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19606 .addMBB(restoreMBB);
19608 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19609 MIB.addRegMask(RegInfo->getNoPreservedMask());
19610 thisMBB->addSuccessor(mainMBB);
19611 thisMBB->addSuccessor(restoreMBB);
19615 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19616 mainMBB->addSuccessor(sinkMBB);
19619 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19620 TII->get(X86::PHI), DstReg)
19621 .addReg(mainDstReg).addMBB(mainMBB)
19622 .addReg(restoreDstReg).addMBB(restoreMBB);
19625 if (RegInfo->hasBasePointer(*MF)) {
19626 const bool Uses64BitFramePtr =
19627 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
19628 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
19629 X86FI->setRestoreBasePointer(MF);
19630 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
19631 unsigned BasePtr = RegInfo->getBaseRegister();
19632 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
19633 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
19634 FramePtr, true, X86FI->getRestoreBasePointerOffset())
19635 .setMIFlag(MachineInstr::FrameSetup);
19637 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19638 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
19639 restoreMBB->addSuccessor(sinkMBB);
19641 MI->eraseFromParent();
19645 MachineBasicBlock *
19646 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19647 MachineBasicBlock *MBB) const {
19648 DebugLoc DL = MI->getDebugLoc();
19649 MachineFunction *MF = MBB->getParent();
19650 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19651 MachineRegisterInfo &MRI = MF->getRegInfo();
19653 // Memory Reference
19654 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19655 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19657 MVT PVT = getPointerTy();
19658 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19659 "Invalid Pointer Size!");
19661 const TargetRegisterClass *RC =
19662 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19663 unsigned Tmp = MRI.createVirtualRegister(RC);
19664 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19665 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19666 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19667 unsigned SP = RegInfo->getStackRegister();
19669 MachineInstrBuilder MIB;
19671 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19672 const int64_t SPOffset = 2 * PVT.getStoreSize();
19674 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19675 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19678 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19679 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19680 MIB.addOperand(MI->getOperand(i));
19681 MIB.setMemRefs(MMOBegin, MMOEnd);
19683 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19684 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19685 if (i == X86::AddrDisp)
19686 MIB.addDisp(MI->getOperand(i), LabelOffset);
19688 MIB.addOperand(MI->getOperand(i));
19690 MIB.setMemRefs(MMOBegin, MMOEnd);
19692 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19693 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19694 if (i == X86::AddrDisp)
19695 MIB.addDisp(MI->getOperand(i), SPOffset);
19697 MIB.addOperand(MI->getOperand(i));
19699 MIB.setMemRefs(MMOBegin, MMOEnd);
19701 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19703 MI->eraseFromParent();
19707 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19708 // accumulator loops. Writing back to the accumulator allows the coalescer
19709 // to remove extra copies in the loop.
19710 MachineBasicBlock *
19711 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19712 MachineBasicBlock *MBB) const {
19713 MachineOperand &AddendOp = MI->getOperand(3);
19715 // Bail out early if the addend isn't a register - we can't switch these.
19716 if (!AddendOp.isReg())
19719 MachineFunction &MF = *MBB->getParent();
19720 MachineRegisterInfo &MRI = MF.getRegInfo();
19722 // Check whether the addend is defined by a PHI:
19723 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19724 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19725 if (!AddendDef.isPHI())
19728 // Look for the following pattern:
19730 // %addend = phi [%entry, 0], [%loop, %result]
19732 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19736 // %addend = phi [%entry, 0], [%loop, %result]
19738 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19740 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19741 assert(AddendDef.getOperand(i).isReg());
19742 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19743 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19744 if (&PHISrcInst == MI) {
19745 // Found a matching instruction.
19746 unsigned NewFMAOpc = 0;
19747 switch (MI->getOpcode()) {
19748 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19749 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19750 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19751 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19752 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19753 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19754 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19755 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19756 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19757 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19758 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19759 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19760 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19761 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19762 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19763 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19764 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
19765 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
19766 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
19767 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
19769 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19770 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19771 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19772 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19773 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19774 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19775 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19776 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19777 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
19778 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
19779 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
19780 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
19781 default: llvm_unreachable("Unrecognized FMA variant.");
19784 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
19785 MachineInstrBuilder MIB =
19786 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19787 .addOperand(MI->getOperand(0))
19788 .addOperand(MI->getOperand(3))
19789 .addOperand(MI->getOperand(2))
19790 .addOperand(MI->getOperand(1));
19791 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19792 MI->eraseFromParent();
19799 MachineBasicBlock *
19800 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19801 MachineBasicBlock *BB) const {
19802 switch (MI->getOpcode()) {
19803 default: llvm_unreachable("Unexpected instr type to insert");
19804 case X86::TAILJMPd64:
19805 case X86::TAILJMPr64:
19806 case X86::TAILJMPm64:
19807 case X86::TAILJMPd64_REX:
19808 case X86::TAILJMPr64_REX:
19809 case X86::TAILJMPm64_REX:
19810 llvm_unreachable("TAILJMP64 would not be touched here.");
19811 case X86::TCRETURNdi64:
19812 case X86::TCRETURNri64:
19813 case X86::TCRETURNmi64:
19815 case X86::WIN_ALLOCA:
19816 return EmitLoweredWinAlloca(MI, BB);
19817 case X86::SEG_ALLOCA_32:
19818 case X86::SEG_ALLOCA_64:
19819 return EmitLoweredSegAlloca(MI, BB);
19820 case X86::TLSCall_32:
19821 case X86::TLSCall_64:
19822 return EmitLoweredTLSCall(MI, BB);
19823 case X86::CMOV_GR8:
19824 case X86::CMOV_FR32:
19825 case X86::CMOV_FR64:
19826 case X86::CMOV_V4F32:
19827 case X86::CMOV_V2F64:
19828 case X86::CMOV_V2I64:
19829 case X86::CMOV_V8F32:
19830 case X86::CMOV_V4F64:
19831 case X86::CMOV_V4I64:
19832 case X86::CMOV_V16F32:
19833 case X86::CMOV_V8F64:
19834 case X86::CMOV_V8I64:
19835 case X86::CMOV_GR16:
19836 case X86::CMOV_GR32:
19837 case X86::CMOV_RFP32:
19838 case X86::CMOV_RFP64:
19839 case X86::CMOV_RFP80:
19840 case X86::CMOV_V8I1:
19841 case X86::CMOV_V16I1:
19842 case X86::CMOV_V32I1:
19843 case X86::CMOV_V64I1:
19844 return EmitLoweredSelect(MI, BB);
19846 case X86::FP32_TO_INT16_IN_MEM:
19847 case X86::FP32_TO_INT32_IN_MEM:
19848 case X86::FP32_TO_INT64_IN_MEM:
19849 case X86::FP64_TO_INT16_IN_MEM:
19850 case X86::FP64_TO_INT32_IN_MEM:
19851 case X86::FP64_TO_INT64_IN_MEM:
19852 case X86::FP80_TO_INT16_IN_MEM:
19853 case X86::FP80_TO_INT32_IN_MEM:
19854 case X86::FP80_TO_INT64_IN_MEM: {
19855 MachineFunction *F = BB->getParent();
19856 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19857 DebugLoc DL = MI->getDebugLoc();
19859 // Change the floating point control register to use "round towards zero"
19860 // mode when truncating to an integer value.
19861 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19862 addFrameReference(BuildMI(*BB, MI, DL,
19863 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19865 // Load the old value of the high byte of the control word...
19867 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
19868 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
19871 // Set the high part to be round to zero...
19872 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
19875 // Reload the modified control word now...
19876 addFrameReference(BuildMI(*BB, MI, DL,
19877 TII->get(X86::FLDCW16m)), CWFrameIdx);
19879 // Restore the memory image of control word to original value
19880 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
19883 // Get the X86 opcode to use.
19885 switch (MI->getOpcode()) {
19886 default: llvm_unreachable("illegal opcode!");
19887 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
19888 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
19889 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
19890 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
19891 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
19892 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
19893 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
19894 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
19895 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
19899 MachineOperand &Op = MI->getOperand(0);
19901 AM.BaseType = X86AddressMode::RegBase;
19902 AM.Base.Reg = Op.getReg();
19904 AM.BaseType = X86AddressMode::FrameIndexBase;
19905 AM.Base.FrameIndex = Op.getIndex();
19907 Op = MI->getOperand(1);
19909 AM.Scale = Op.getImm();
19910 Op = MI->getOperand(2);
19912 AM.IndexReg = Op.getImm();
19913 Op = MI->getOperand(3);
19914 if (Op.isGlobal()) {
19915 AM.GV = Op.getGlobal();
19917 AM.Disp = Op.getImm();
19919 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
19920 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
19922 // Reload the original control word now.
19923 addFrameReference(BuildMI(*BB, MI, DL,
19924 TII->get(X86::FLDCW16m)), CWFrameIdx);
19926 MI->eraseFromParent(); // The pseudo instruction is gone now.
19929 // String/text processing lowering.
19930 case X86::PCMPISTRM128REG:
19931 case X86::VPCMPISTRM128REG:
19932 case X86::PCMPISTRM128MEM:
19933 case X86::VPCMPISTRM128MEM:
19934 case X86::PCMPESTRM128REG:
19935 case X86::VPCMPESTRM128REG:
19936 case X86::PCMPESTRM128MEM:
19937 case X86::VPCMPESTRM128MEM:
19938 assert(Subtarget->hasSSE42() &&
19939 "Target must have SSE4.2 or AVX features enabled");
19940 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
19942 // String/text processing lowering.
19943 case X86::PCMPISTRIREG:
19944 case X86::VPCMPISTRIREG:
19945 case X86::PCMPISTRIMEM:
19946 case X86::VPCMPISTRIMEM:
19947 case X86::PCMPESTRIREG:
19948 case X86::VPCMPESTRIREG:
19949 case X86::PCMPESTRIMEM:
19950 case X86::VPCMPESTRIMEM:
19951 assert(Subtarget->hasSSE42() &&
19952 "Target must have SSE4.2 or AVX features enabled");
19953 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
19955 // Thread synchronization.
19957 return EmitMonitor(MI, BB, Subtarget);
19961 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
19963 case X86::VASTART_SAVE_XMM_REGS:
19964 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
19966 case X86::VAARG_64:
19967 return EmitVAARG64WithCustomInserter(MI, BB);
19969 case X86::EH_SjLj_SetJmp32:
19970 case X86::EH_SjLj_SetJmp64:
19971 return emitEHSjLjSetJmp(MI, BB);
19973 case X86::EH_SjLj_LongJmp32:
19974 case X86::EH_SjLj_LongJmp64:
19975 return emitEHSjLjLongJmp(MI, BB);
19977 case TargetOpcode::STATEPOINT:
19978 // As an implementation detail, STATEPOINT shares the STACKMAP format at
19979 // this point in the process. We diverge later.
19980 return emitPatchPoint(MI, BB);
19982 case TargetOpcode::STACKMAP:
19983 case TargetOpcode::PATCHPOINT:
19984 return emitPatchPoint(MI, BB);
19986 case X86::VFMADDPDr213r:
19987 case X86::VFMADDPSr213r:
19988 case X86::VFMADDSDr213r:
19989 case X86::VFMADDSSr213r:
19990 case X86::VFMSUBPDr213r:
19991 case X86::VFMSUBPSr213r:
19992 case X86::VFMSUBSDr213r:
19993 case X86::VFMSUBSSr213r:
19994 case X86::VFNMADDPDr213r:
19995 case X86::VFNMADDPSr213r:
19996 case X86::VFNMADDSDr213r:
19997 case X86::VFNMADDSSr213r:
19998 case X86::VFNMSUBPDr213r:
19999 case X86::VFNMSUBPSr213r:
20000 case X86::VFNMSUBSDr213r:
20001 case X86::VFNMSUBSSr213r:
20002 case X86::VFMADDSUBPDr213r:
20003 case X86::VFMADDSUBPSr213r:
20004 case X86::VFMSUBADDPDr213r:
20005 case X86::VFMSUBADDPSr213r:
20006 case X86::VFMADDPDr213rY:
20007 case X86::VFMADDPSr213rY:
20008 case X86::VFMSUBPDr213rY:
20009 case X86::VFMSUBPSr213rY:
20010 case X86::VFNMADDPDr213rY:
20011 case X86::VFNMADDPSr213rY:
20012 case X86::VFNMSUBPDr213rY:
20013 case X86::VFNMSUBPSr213rY:
20014 case X86::VFMADDSUBPDr213rY:
20015 case X86::VFMADDSUBPSr213rY:
20016 case X86::VFMSUBADDPDr213rY:
20017 case X86::VFMSUBADDPSr213rY:
20018 return emitFMA3Instr(MI, BB);
20022 //===----------------------------------------------------------------------===//
20023 // X86 Optimization Hooks
20024 //===----------------------------------------------------------------------===//
20026 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
20029 const SelectionDAG &DAG,
20030 unsigned Depth) const {
20031 unsigned BitWidth = KnownZero.getBitWidth();
20032 unsigned Opc = Op.getOpcode();
20033 assert((Opc >= ISD::BUILTIN_OP_END ||
20034 Opc == ISD::INTRINSIC_WO_CHAIN ||
20035 Opc == ISD::INTRINSIC_W_CHAIN ||
20036 Opc == ISD::INTRINSIC_VOID) &&
20037 "Should use MaskedValueIsZero if you don't know whether Op"
20038 " is a target node!");
20040 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
20054 // These nodes' second result is a boolean.
20055 if (Op.getResNo() == 0)
20058 case X86ISD::SETCC:
20059 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
20061 case ISD::INTRINSIC_WO_CHAIN: {
20062 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
20063 unsigned NumLoBits = 0;
20066 case Intrinsic::x86_sse_movmsk_ps:
20067 case Intrinsic::x86_avx_movmsk_ps_256:
20068 case Intrinsic::x86_sse2_movmsk_pd:
20069 case Intrinsic::x86_avx_movmsk_pd_256:
20070 case Intrinsic::x86_mmx_pmovmskb:
20071 case Intrinsic::x86_sse2_pmovmskb_128:
20072 case Intrinsic::x86_avx2_pmovmskb: {
20073 // High bits of movmskp{s|d}, pmovmskb are known zero.
20075 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
20076 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
20077 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
20078 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
20079 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
20080 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
20081 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
20082 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
20084 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
20093 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
20095 const SelectionDAG &,
20096 unsigned Depth) const {
20097 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
20098 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
20099 return Op.getValueType().getScalarType().getSizeInBits();
20105 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
20106 /// node is a GlobalAddress + offset.
20107 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
20108 const GlobalValue* &GA,
20109 int64_t &Offset) const {
20110 if (N->getOpcode() == X86ISD::Wrapper) {
20111 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
20112 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
20113 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
20117 return TargetLowering::isGAPlusOffset(N, GA, Offset);
20120 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
20121 /// same as extracting the high 128-bit part of 256-bit vector and then
20122 /// inserting the result into the low part of a new 256-bit vector
20123 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
20124 EVT VT = SVOp->getValueType(0);
20125 unsigned NumElems = VT.getVectorNumElements();
20127 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20128 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
20129 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20130 SVOp->getMaskElt(j) >= 0)
20136 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
20137 /// same as extracting the low 128-bit part of 256-bit vector and then
20138 /// inserting the result into the high part of a new 256-bit vector
20139 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
20140 EVT VT = SVOp->getValueType(0);
20141 unsigned NumElems = VT.getVectorNumElements();
20143 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20144 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
20145 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20146 SVOp->getMaskElt(j) >= 0)
20152 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
20153 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
20154 TargetLowering::DAGCombinerInfo &DCI,
20155 const X86Subtarget* Subtarget) {
20157 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20158 SDValue V1 = SVOp->getOperand(0);
20159 SDValue V2 = SVOp->getOperand(1);
20160 EVT VT = SVOp->getValueType(0);
20161 unsigned NumElems = VT.getVectorNumElements();
20163 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
20164 V2.getOpcode() == ISD::CONCAT_VECTORS) {
20168 // V UNDEF BUILD_VECTOR UNDEF
20170 // CONCAT_VECTOR CONCAT_VECTOR
20173 // RESULT: V + zero extended
20175 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
20176 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
20177 V1.getOperand(1).getOpcode() != ISD::UNDEF)
20180 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
20183 // To match the shuffle mask, the first half of the mask should
20184 // be exactly the first vector, and all the rest a splat with the
20185 // first element of the second one.
20186 for (unsigned i = 0; i != NumElems/2; ++i)
20187 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
20188 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
20191 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
20192 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
20193 if (Ld->hasNUsesOfValue(1, 0)) {
20194 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
20195 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
20197 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
20199 Ld->getPointerInfo(),
20200 Ld->getAlignment(),
20201 false/*isVolatile*/, true/*ReadMem*/,
20202 false/*WriteMem*/);
20204 // Make sure the newly-created LOAD is in the same position as Ld in
20205 // terms of dependency. We create a TokenFactor for Ld and ResNode,
20206 // and update uses of Ld's output chain to use the TokenFactor.
20207 if (Ld->hasAnyUseOfValue(1)) {
20208 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
20209 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
20210 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
20211 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
20212 SDValue(ResNode.getNode(), 1));
20215 return DAG.getBitcast(VT, ResNode);
20219 // Emit a zeroed vector and insert the desired subvector on its
20221 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
20222 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
20223 return DCI.CombineTo(N, InsV);
20226 //===--------------------------------------------------------------------===//
20227 // Combine some shuffles into subvector extracts and inserts:
20230 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20231 if (isShuffleHigh128VectorInsertLow(SVOp)) {
20232 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
20233 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
20234 return DCI.CombineTo(N, InsV);
20237 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20238 if (isShuffleLow128VectorInsertHigh(SVOp)) {
20239 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
20240 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
20241 return DCI.CombineTo(N, InsV);
20247 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
20250 /// This is the leaf of the recursive combinine below. When we have found some
20251 /// chain of single-use x86 shuffle instructions and accumulated the combined
20252 /// shuffle mask represented by them, this will try to pattern match that mask
20253 /// into either a single instruction if there is a special purpose instruction
20254 /// for this operation, or into a PSHUFB instruction which is a fully general
20255 /// instruction but should only be used to replace chains over a certain depth.
20256 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
20257 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
20258 TargetLowering::DAGCombinerInfo &DCI,
20259 const X86Subtarget *Subtarget) {
20260 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
20262 // Find the operand that enters the chain. Note that multiple uses are OK
20263 // here, we're not going to remove the operand we find.
20264 SDValue Input = Op.getOperand(0);
20265 while (Input.getOpcode() == ISD::BITCAST)
20266 Input = Input.getOperand(0);
20268 MVT VT = Input.getSimpleValueType();
20269 MVT RootVT = Root.getSimpleValueType();
20272 // Just remove no-op shuffle masks.
20273 if (Mask.size() == 1) {
20274 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
20279 // Use the float domain if the operand type is a floating point type.
20280 bool FloatDomain = VT.isFloatingPoint();
20282 // For floating point shuffles, we don't have free copies in the shuffle
20283 // instructions or the ability to load as part of the instruction, so
20284 // canonicalize their shuffles to UNPCK or MOV variants.
20286 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
20287 // vectors because it can have a load folded into it that UNPCK cannot. This
20288 // doesn't preclude something switching to the shorter encoding post-RA.
20290 // FIXME: Should teach these routines about AVX vector widths.
20291 if (FloatDomain && VT.getSizeInBits() == 128) {
20292 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
20293 bool Lo = Mask.equals({0, 0});
20296 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
20297 // is no slower than UNPCKLPD but has the option to fold the input operand
20298 // into even an unaligned memory load.
20299 if (Lo && Subtarget->hasSSE3()) {
20300 Shuffle = X86ISD::MOVDDUP;
20301 ShuffleVT = MVT::v2f64;
20303 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
20304 // than the UNPCK variants.
20305 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
20306 ShuffleVT = MVT::v4f32;
20308 if (Depth == 1 && Root->getOpcode() == Shuffle)
20309 return false; // Nothing to do!
20310 Op = DAG.getBitcast(ShuffleVT, Input);
20311 DCI.AddToWorklist(Op.getNode());
20312 if (Shuffle == X86ISD::MOVDDUP)
20313 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20315 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20316 DCI.AddToWorklist(Op.getNode());
20317 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
20321 if (Subtarget->hasSSE3() &&
20322 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
20323 bool Lo = Mask.equals({0, 0, 2, 2});
20324 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
20325 MVT ShuffleVT = MVT::v4f32;
20326 if (Depth == 1 && Root->getOpcode() == Shuffle)
20327 return false; // Nothing to do!
20328 Op = DAG.getBitcast(ShuffleVT, Input);
20329 DCI.AddToWorklist(Op.getNode());
20330 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20331 DCI.AddToWorklist(Op.getNode());
20332 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
20336 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
20337 bool Lo = Mask.equals({0, 0, 1, 1});
20338 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20339 MVT ShuffleVT = MVT::v4f32;
20340 if (Depth == 1 && Root->getOpcode() == Shuffle)
20341 return false; // Nothing to do!
20342 Op = DAG.getBitcast(ShuffleVT, Input);
20343 DCI.AddToWorklist(Op.getNode());
20344 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20345 DCI.AddToWorklist(Op.getNode());
20346 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
20352 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
20353 // variants as none of these have single-instruction variants that are
20354 // superior to the UNPCK formulation.
20355 if (!FloatDomain && VT.getSizeInBits() == 128 &&
20356 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20357 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
20358 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
20360 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
20361 bool Lo = Mask[0] == 0;
20362 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20363 if (Depth == 1 && Root->getOpcode() == Shuffle)
20364 return false; // Nothing to do!
20366 switch (Mask.size()) {
20368 ShuffleVT = MVT::v8i16;
20371 ShuffleVT = MVT::v16i8;
20374 llvm_unreachable("Impossible mask size!");
20376 Op = DAG.getBitcast(ShuffleVT, Input);
20377 DCI.AddToWorklist(Op.getNode());
20378 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20379 DCI.AddToWorklist(Op.getNode());
20380 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
20385 // Don't try to re-form single instruction chains under any circumstances now
20386 // that we've done encoding canonicalization for them.
20390 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
20391 // can replace them with a single PSHUFB instruction profitably. Intel's
20392 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
20393 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
20394 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
20395 SmallVector<SDValue, 16> PSHUFBMask;
20396 int NumBytes = VT.getSizeInBits() / 8;
20397 int Ratio = NumBytes / Mask.size();
20398 for (int i = 0; i < NumBytes; ++i) {
20399 if (Mask[i / Ratio] == SM_SentinelUndef) {
20400 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
20403 int M = Mask[i / Ratio] != SM_SentinelZero
20404 ? Ratio * Mask[i / Ratio] + i % Ratio
20406 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
20408 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
20409 Op = DAG.getBitcast(ByteVT, Input);
20410 DCI.AddToWorklist(Op.getNode());
20411 SDValue PSHUFBMaskOp =
20412 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
20413 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
20414 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
20415 DCI.AddToWorklist(Op.getNode());
20416 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
20421 // Failed to find any combines.
20425 /// \brief Fully generic combining of x86 shuffle instructions.
20427 /// This should be the last combine run over the x86 shuffle instructions. Once
20428 /// they have been fully optimized, this will recursively consider all chains
20429 /// of single-use shuffle instructions, build a generic model of the cumulative
20430 /// shuffle operation, and check for simpler instructions which implement this
20431 /// operation. We use this primarily for two purposes:
20433 /// 1) Collapse generic shuffles to specialized single instructions when
20434 /// equivalent. In most cases, this is just an encoding size win, but
20435 /// sometimes we will collapse multiple generic shuffles into a single
20436 /// special-purpose shuffle.
20437 /// 2) Look for sequences of shuffle instructions with 3 or more total
20438 /// instructions, and replace them with the slightly more expensive SSSE3
20439 /// PSHUFB instruction if available. We do this as the last combining step
20440 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
20441 /// a suitable short sequence of other instructions. The PHUFB will either
20442 /// use a register or have to read from memory and so is slightly (but only
20443 /// slightly) more expensive than the other shuffle instructions.
20445 /// Because this is inherently a quadratic operation (for each shuffle in
20446 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
20447 /// This should never be an issue in practice as the shuffle lowering doesn't
20448 /// produce sequences of more than 8 instructions.
20450 /// FIXME: We will currently miss some cases where the redundant shuffling
20451 /// would simplify under the threshold for PSHUFB formation because of
20452 /// combine-ordering. To fix this, we should do the redundant instruction
20453 /// combining in this recursive walk.
20454 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
20455 ArrayRef<int> RootMask,
20456 int Depth, bool HasPSHUFB,
20458 TargetLowering::DAGCombinerInfo &DCI,
20459 const X86Subtarget *Subtarget) {
20460 // Bound the depth of our recursive combine because this is ultimately
20461 // quadratic in nature.
20465 // Directly rip through bitcasts to find the underlying operand.
20466 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
20467 Op = Op.getOperand(0);
20469 MVT VT = Op.getSimpleValueType();
20470 if (!VT.isVector())
20471 return false; // Bail if we hit a non-vector.
20473 assert(Root.getSimpleValueType().isVector() &&
20474 "Shuffles operate on vector types!");
20475 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
20476 "Can only combine shuffles of the same vector register size.");
20478 if (!isTargetShuffle(Op.getOpcode()))
20480 SmallVector<int, 16> OpMask;
20482 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
20483 // We only can combine unary shuffles which we can decode the mask for.
20484 if (!HaveMask || !IsUnary)
20487 assert(VT.getVectorNumElements() == OpMask.size() &&
20488 "Different mask size from vector size!");
20489 assert(((RootMask.size() > OpMask.size() &&
20490 RootMask.size() % OpMask.size() == 0) ||
20491 (OpMask.size() > RootMask.size() &&
20492 OpMask.size() % RootMask.size() == 0) ||
20493 OpMask.size() == RootMask.size()) &&
20494 "The smaller number of elements must divide the larger.");
20495 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
20496 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
20497 assert(((RootRatio == 1 && OpRatio == 1) ||
20498 (RootRatio == 1) != (OpRatio == 1)) &&
20499 "Must not have a ratio for both incoming and op masks!");
20501 SmallVector<int, 16> Mask;
20502 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
20504 // Merge this shuffle operation's mask into our accumulated mask. Note that
20505 // this shuffle's mask will be the first applied to the input, followed by the
20506 // root mask to get us all the way to the root value arrangement. The reason
20507 // for this order is that we are recursing up the operation chain.
20508 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
20509 int RootIdx = i / RootRatio;
20510 if (RootMask[RootIdx] < 0) {
20511 // This is a zero or undef lane, we're done.
20512 Mask.push_back(RootMask[RootIdx]);
20516 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
20517 int OpIdx = RootMaskedIdx / OpRatio;
20518 if (OpMask[OpIdx] < 0) {
20519 // The incoming lanes are zero or undef, it doesn't matter which ones we
20521 Mask.push_back(OpMask[OpIdx]);
20525 // Ok, we have non-zero lanes, map them through.
20526 Mask.push_back(OpMask[OpIdx] * OpRatio +
20527 RootMaskedIdx % OpRatio);
20530 // See if we can recurse into the operand to combine more things.
20531 switch (Op.getOpcode()) {
20532 case X86ISD::PSHUFB:
20534 case X86ISD::PSHUFD:
20535 case X86ISD::PSHUFHW:
20536 case X86ISD::PSHUFLW:
20537 if (Op.getOperand(0).hasOneUse() &&
20538 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20539 HasPSHUFB, DAG, DCI, Subtarget))
20543 case X86ISD::UNPCKL:
20544 case X86ISD::UNPCKH:
20545 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
20546 // We can't check for single use, we have to check that this shuffle is the only user.
20547 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
20548 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20549 HasPSHUFB, DAG, DCI, Subtarget))
20554 // Minor canonicalization of the accumulated shuffle mask to make it easier
20555 // to match below. All this does is detect masks with squential pairs of
20556 // elements, and shrink them to the half-width mask. It does this in a loop
20557 // so it will reduce the size of the mask to the minimal width mask which
20558 // performs an equivalent shuffle.
20559 SmallVector<int, 16> WidenedMask;
20560 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
20561 Mask = std::move(WidenedMask);
20562 WidenedMask.clear();
20565 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20569 /// \brief Get the PSHUF-style mask from PSHUF node.
20571 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20572 /// PSHUF-style masks that can be reused with such instructions.
20573 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20574 MVT VT = N.getSimpleValueType();
20575 SmallVector<int, 4> Mask;
20577 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
20581 // If we have more than 128-bits, only the low 128-bits of shuffle mask
20582 // matter. Check that the upper masks are repeats and remove them.
20583 if (VT.getSizeInBits() > 128) {
20584 int LaneElts = 128 / VT.getScalarSizeInBits();
20586 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
20587 for (int j = 0; j < LaneElts; ++j)
20588 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
20589 "Mask doesn't repeat in high 128-bit lanes!");
20591 Mask.resize(LaneElts);
20594 switch (N.getOpcode()) {
20595 case X86ISD::PSHUFD:
20597 case X86ISD::PSHUFLW:
20600 case X86ISD::PSHUFHW:
20601 Mask.erase(Mask.begin(), Mask.begin() + 4);
20602 for (int &M : Mask)
20606 llvm_unreachable("No valid shuffle instruction found!");
20610 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20612 /// We walk up the chain and look for a combinable shuffle, skipping over
20613 /// shuffles that we could hoist this shuffle's transformation past without
20614 /// altering anything.
20616 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20618 TargetLowering::DAGCombinerInfo &DCI) {
20619 assert(N.getOpcode() == X86ISD::PSHUFD &&
20620 "Called with something other than an x86 128-bit half shuffle!");
20623 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20624 // of the shuffles in the chain so that we can form a fresh chain to replace
20626 SmallVector<SDValue, 8> Chain;
20627 SDValue V = N.getOperand(0);
20628 for (; V.hasOneUse(); V = V.getOperand(0)) {
20629 switch (V.getOpcode()) {
20631 return SDValue(); // Nothing combined!
20634 // Skip bitcasts as we always know the type for the target specific
20638 case X86ISD::PSHUFD:
20639 // Found another dword shuffle.
20642 case X86ISD::PSHUFLW:
20643 // Check that the low words (being shuffled) are the identity in the
20644 // dword shuffle, and the high words are self-contained.
20645 if (Mask[0] != 0 || Mask[1] != 1 ||
20646 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20649 Chain.push_back(V);
20652 case X86ISD::PSHUFHW:
20653 // Check that the high words (being shuffled) are the identity in the
20654 // dword shuffle, and the low words are self-contained.
20655 if (Mask[2] != 2 || Mask[3] != 3 ||
20656 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20659 Chain.push_back(V);
20662 case X86ISD::UNPCKL:
20663 case X86ISD::UNPCKH:
20664 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20665 // shuffle into a preceding word shuffle.
20666 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
20667 V.getSimpleValueType().getScalarType() != MVT::i16)
20670 // Search for a half-shuffle which we can combine with.
20671 unsigned CombineOp =
20672 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20673 if (V.getOperand(0) != V.getOperand(1) ||
20674 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20676 Chain.push_back(V);
20677 V = V.getOperand(0);
20679 switch (V.getOpcode()) {
20681 return SDValue(); // Nothing to combine.
20683 case X86ISD::PSHUFLW:
20684 case X86ISD::PSHUFHW:
20685 if (V.getOpcode() == CombineOp)
20688 Chain.push_back(V);
20692 V = V.getOperand(0);
20696 } while (V.hasOneUse());
20699 // Break out of the loop if we break out of the switch.
20703 if (!V.hasOneUse())
20704 // We fell out of the loop without finding a viable combining instruction.
20707 // Merge this node's mask and our incoming mask.
20708 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20709 for (int &M : Mask)
20711 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20712 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20714 // Rebuild the chain around this new shuffle.
20715 while (!Chain.empty()) {
20716 SDValue W = Chain.pop_back_val();
20718 if (V.getValueType() != W.getOperand(0).getValueType())
20719 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
20721 switch (W.getOpcode()) {
20723 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20725 case X86ISD::UNPCKL:
20726 case X86ISD::UNPCKH:
20727 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20730 case X86ISD::PSHUFD:
20731 case X86ISD::PSHUFLW:
20732 case X86ISD::PSHUFHW:
20733 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20737 if (V.getValueType() != N.getValueType())
20738 V = DAG.getBitcast(N.getValueType(), V);
20740 // Return the new chain to replace N.
20744 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20746 /// We walk up the chain, skipping shuffles of the other half and looking
20747 /// through shuffles which switch halves trying to find a shuffle of the same
20748 /// pair of dwords.
20749 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20751 TargetLowering::DAGCombinerInfo &DCI) {
20753 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20754 "Called with something other than an x86 128-bit half shuffle!");
20756 unsigned CombineOpcode = N.getOpcode();
20758 // Walk up a single-use chain looking for a combinable shuffle.
20759 SDValue V = N.getOperand(0);
20760 for (; V.hasOneUse(); V = V.getOperand(0)) {
20761 switch (V.getOpcode()) {
20763 return false; // Nothing combined!
20766 // Skip bitcasts as we always know the type for the target specific
20770 case X86ISD::PSHUFLW:
20771 case X86ISD::PSHUFHW:
20772 if (V.getOpcode() == CombineOpcode)
20775 // Other-half shuffles are no-ops.
20778 // Break out of the loop if we break out of the switch.
20782 if (!V.hasOneUse())
20783 // We fell out of the loop without finding a viable combining instruction.
20786 // Combine away the bottom node as its shuffle will be accumulated into
20787 // a preceding shuffle.
20788 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20790 // Record the old value.
20793 // Merge this node's mask and our incoming mask (adjusted to account for all
20794 // the pshufd instructions encountered).
20795 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20796 for (int &M : Mask)
20798 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20799 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20801 // Check that the shuffles didn't cancel each other out. If not, we need to
20802 // combine to the new one.
20804 // Replace the combinable shuffle with the combined one, updating all users
20805 // so that we re-evaluate the chain here.
20806 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20811 /// \brief Try to combine x86 target specific shuffles.
20812 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20813 TargetLowering::DAGCombinerInfo &DCI,
20814 const X86Subtarget *Subtarget) {
20816 MVT VT = N.getSimpleValueType();
20817 SmallVector<int, 4> Mask;
20819 switch (N.getOpcode()) {
20820 case X86ISD::PSHUFD:
20821 case X86ISD::PSHUFLW:
20822 case X86ISD::PSHUFHW:
20823 Mask = getPSHUFShuffleMask(N);
20824 assert(Mask.size() == 4);
20830 // Nuke no-op shuffles that show up after combining.
20831 if (isNoopShuffleMask(Mask))
20832 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20834 // Look for simplifications involving one or two shuffle instructions.
20835 SDValue V = N.getOperand(0);
20836 switch (N.getOpcode()) {
20839 case X86ISD::PSHUFLW:
20840 case X86ISD::PSHUFHW:
20841 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
20843 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20844 return SDValue(); // We combined away this shuffle, so we're done.
20846 // See if this reduces to a PSHUFD which is no more expensive and can
20847 // combine with more operations. Note that it has to at least flip the
20848 // dwords as otherwise it would have been removed as a no-op.
20849 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
20850 int DMask[] = {0, 1, 2, 3};
20851 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20852 DMask[DOffset + 0] = DOffset + 1;
20853 DMask[DOffset + 1] = DOffset + 0;
20854 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
20855 V = DAG.getBitcast(DVT, V);
20856 DCI.AddToWorklist(V.getNode());
20857 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
20858 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
20859 DCI.AddToWorklist(V.getNode());
20860 return DAG.getBitcast(VT, V);
20863 // Look for shuffle patterns which can be implemented as a single unpack.
20864 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20865 // only works when we have a PSHUFD followed by two half-shuffles.
20866 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20867 (V.getOpcode() == X86ISD::PSHUFLW ||
20868 V.getOpcode() == X86ISD::PSHUFHW) &&
20869 V.getOpcode() != N.getOpcode() &&
20871 SDValue D = V.getOperand(0);
20872 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20873 D = D.getOperand(0);
20874 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20875 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20876 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20877 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20878 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20880 for (int i = 0; i < 4; ++i) {
20881 WordMask[i + NOffset] = Mask[i] + NOffset;
20882 WordMask[i + VOffset] = VMask[i] + VOffset;
20884 // Map the word mask through the DWord mask.
20886 for (int i = 0; i < 8; ++i)
20887 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20888 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20889 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
20890 // We can replace all three shuffles with an unpack.
20891 V = DAG.getBitcast(VT, D.getOperand(0));
20892 DCI.AddToWorklist(V.getNode());
20893 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
20902 case X86ISD::PSHUFD:
20903 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
20912 /// \brief Try to combine a shuffle into a target-specific add-sub node.
20914 /// We combine this directly on the abstract vector shuffle nodes so it is
20915 /// easier to generically match. We also insert dummy vector shuffle nodes for
20916 /// the operands which explicitly discard the lanes which are unused by this
20917 /// operation to try to flow through the rest of the combiner the fact that
20918 /// they're unused.
20919 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
20921 EVT VT = N->getValueType(0);
20923 // We only handle target-independent shuffles.
20924 // FIXME: It would be easy and harmless to use the target shuffle mask
20925 // extraction tool to support more.
20926 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
20929 auto *SVN = cast<ShuffleVectorSDNode>(N);
20930 ArrayRef<int> Mask = SVN->getMask();
20931 SDValue V1 = N->getOperand(0);
20932 SDValue V2 = N->getOperand(1);
20934 // We require the first shuffle operand to be the SUB node, and the second to
20935 // be the ADD node.
20936 // FIXME: We should support the commuted patterns.
20937 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
20940 // If there are other uses of these operations we can't fold them.
20941 if (!V1->hasOneUse() || !V2->hasOneUse())
20944 // Ensure that both operations have the same operands. Note that we can
20945 // commute the FADD operands.
20946 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
20947 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
20948 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
20951 // We're looking for blends between FADD and FSUB nodes. We insist on these
20952 // nodes being lined up in a specific expected pattern.
20953 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
20954 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
20955 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
20958 // Only specific types are legal at this point, assert so we notice if and
20959 // when these change.
20960 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
20961 VT == MVT::v4f64) &&
20962 "Unknown vector type encountered!");
20964 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
20967 /// PerformShuffleCombine - Performs several different shuffle combines.
20968 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
20969 TargetLowering::DAGCombinerInfo &DCI,
20970 const X86Subtarget *Subtarget) {
20972 SDValue N0 = N->getOperand(0);
20973 SDValue N1 = N->getOperand(1);
20974 EVT VT = N->getValueType(0);
20976 // Don't create instructions with illegal types after legalize types has run.
20977 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20978 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
20981 // If we have legalized the vector types, look for blends of FADD and FSUB
20982 // nodes that we can fuse into an ADDSUB node.
20983 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
20984 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
20987 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
20988 if (Subtarget->hasFp256() && VT.is256BitVector() &&
20989 N->getOpcode() == ISD::VECTOR_SHUFFLE)
20990 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
20992 // During Type Legalization, when promoting illegal vector types,
20993 // the backend might introduce new shuffle dag nodes and bitcasts.
20995 // This code performs the following transformation:
20996 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
20997 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
20999 // We do this only if both the bitcast and the BINOP dag nodes have
21000 // one use. Also, perform this transformation only if the new binary
21001 // operation is legal. This is to avoid introducing dag nodes that
21002 // potentially need to be further expanded (or custom lowered) into a
21003 // less optimal sequence of dag nodes.
21004 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
21005 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
21006 N0.getOpcode() == ISD::BITCAST) {
21007 SDValue BC0 = N0.getOperand(0);
21008 EVT SVT = BC0.getValueType();
21009 unsigned Opcode = BC0.getOpcode();
21010 unsigned NumElts = VT.getVectorNumElements();
21012 if (BC0.hasOneUse() && SVT.isVector() &&
21013 SVT.getVectorNumElements() * 2 == NumElts &&
21014 TLI.isOperationLegal(Opcode, VT)) {
21015 bool CanFold = false;
21027 unsigned SVTNumElts = SVT.getVectorNumElements();
21028 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21029 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
21030 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
21031 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
21032 CanFold = SVOp->getMaskElt(i) < 0;
21035 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
21036 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
21037 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
21038 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
21043 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
21044 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
21045 // consecutive, non-overlapping, and in the right order.
21046 SmallVector<SDValue, 16> Elts;
21047 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
21048 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
21050 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
21054 if (isTargetShuffle(N->getOpcode())) {
21056 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
21057 if (Shuffle.getNode())
21060 // Try recursively combining arbitrary sequences of x86 shuffle
21061 // instructions into higher-order shuffles. We do this after combining
21062 // specific PSHUF instruction sequences into their minimal form so that we
21063 // can evaluate how many specialized shuffle instructions are involved in
21064 // a particular chain.
21065 SmallVector<int, 1> NonceMask; // Just a placeholder.
21066 NonceMask.push_back(0);
21067 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
21068 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
21070 return SDValue(); // This routine will use CombineTo to replace N.
21076 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
21077 /// specific shuffle of a load can be folded into a single element load.
21078 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
21079 /// shuffles have been custom lowered so we need to handle those here.
21080 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
21081 TargetLowering::DAGCombinerInfo &DCI) {
21082 if (DCI.isBeforeLegalizeOps())
21085 SDValue InVec = N->getOperand(0);
21086 SDValue EltNo = N->getOperand(1);
21088 if (!isa<ConstantSDNode>(EltNo))
21091 EVT OriginalVT = InVec.getValueType();
21093 if (InVec.getOpcode() == ISD::BITCAST) {
21094 // Don't duplicate a load with other uses.
21095 if (!InVec.hasOneUse())
21097 EVT BCVT = InVec.getOperand(0).getValueType();
21098 if (!BCVT.isVector() ||
21099 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
21101 InVec = InVec.getOperand(0);
21104 EVT CurrentVT = InVec.getValueType();
21106 if (!isTargetShuffle(InVec.getOpcode()))
21109 // Don't duplicate a load with other uses.
21110 if (!InVec.hasOneUse())
21113 SmallVector<int, 16> ShuffleMask;
21115 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
21116 ShuffleMask, UnaryShuffle))
21119 // Select the input vector, guarding against out of range extract vector.
21120 unsigned NumElems = CurrentVT.getVectorNumElements();
21121 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
21122 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
21123 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
21124 : InVec.getOperand(1);
21126 // If inputs to shuffle are the same for both ops, then allow 2 uses
21127 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
21128 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
21130 if (LdNode.getOpcode() == ISD::BITCAST) {
21131 // Don't duplicate a load with other uses.
21132 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
21135 AllowedUses = 1; // only allow 1 load use if we have a bitcast
21136 LdNode = LdNode.getOperand(0);
21139 if (!ISD::isNormalLoad(LdNode.getNode()))
21142 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
21144 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
21147 EVT EltVT = N->getValueType(0);
21148 // If there's a bitcast before the shuffle, check if the load type and
21149 // alignment is valid.
21150 unsigned Align = LN0->getAlignment();
21151 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21152 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
21153 EltVT.getTypeForEVT(*DAG.getContext()));
21155 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
21158 // All checks match so transform back to vector_shuffle so that DAG combiner
21159 // can finish the job
21162 // Create shuffle node taking into account the case that its a unary shuffle
21163 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
21164 : InVec.getOperand(1);
21165 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
21166 InVec.getOperand(0), Shuffle,
21168 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
21169 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
21173 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
21174 /// special and don't usually play with other vector types, it's better to
21175 /// handle them early to be sure we emit efficient code by avoiding
21176 /// store-load conversions.
21177 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
21178 if (N->getValueType(0) != MVT::x86mmx ||
21179 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
21180 N->getOperand(0)->getValueType(0) != MVT::v2i32)
21183 SDValue V = N->getOperand(0);
21184 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
21185 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
21186 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
21187 N->getValueType(0), V.getOperand(0));
21192 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
21193 /// generation and convert it from being a bunch of shuffles and extracts
21194 /// into a somewhat faster sequence. For i686, the best sequence is apparently
21195 /// storing the value and loading scalars back, while for x64 we should
21196 /// use 64-bit extracts and shifts.
21197 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
21198 TargetLowering::DAGCombinerInfo &DCI) {
21199 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
21200 if (NewOp.getNode())
21203 SDValue InputVector = N->getOperand(0);
21204 SDLoc dl(InputVector);
21205 // Detect mmx to i32 conversion through a v2i32 elt extract.
21206 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
21207 N->getValueType(0) == MVT::i32 &&
21208 InputVector.getValueType() == MVT::v2i32) {
21210 // The bitcast source is a direct mmx result.
21211 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
21212 if (MMXSrc.getValueType() == MVT::x86mmx)
21213 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
21214 N->getValueType(0),
21215 InputVector.getNode()->getOperand(0));
21217 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
21218 SDValue MMXSrcOp = MMXSrc.getOperand(0);
21219 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
21220 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
21221 MMXSrcOp.getOpcode() == ISD::BITCAST &&
21222 MMXSrcOp.getValueType() == MVT::v1i64 &&
21223 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
21224 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
21225 N->getValueType(0),
21226 MMXSrcOp.getOperand(0));
21229 EVT VT = N->getValueType(0);
21231 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
21232 InputVector.getOpcode() == ISD::BITCAST &&
21233 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
21234 uint64_t ExtractedElt =
21235 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
21236 uint64_t InputValue =
21237 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
21238 uint64_t Res = (InputValue >> ExtractedElt) & 1;
21239 return DAG.getConstant(Res, dl, MVT::i1);
21241 // Only operate on vectors of 4 elements, where the alternative shuffling
21242 // gets to be more expensive.
21243 if (InputVector.getValueType() != MVT::v4i32)
21246 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
21247 // single use which is a sign-extend or zero-extend, and all elements are
21249 SmallVector<SDNode *, 4> Uses;
21250 unsigned ExtractedElements = 0;
21251 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
21252 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
21253 if (UI.getUse().getResNo() != InputVector.getResNo())
21256 SDNode *Extract = *UI;
21257 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
21260 if (Extract->getValueType(0) != MVT::i32)
21262 if (!Extract->hasOneUse())
21264 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
21265 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
21267 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
21270 // Record which element was extracted.
21271 ExtractedElements |=
21272 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
21274 Uses.push_back(Extract);
21277 // If not all the elements were used, this may not be worthwhile.
21278 if (ExtractedElements != 15)
21281 // Ok, we've now decided to do the transformation.
21282 // If 64-bit shifts are legal, use the extract-shift sequence,
21283 // otherwise bounce the vector off the cache.
21284 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21287 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
21288 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
21289 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
21290 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
21291 DAG.getConstant(0, dl, VecIdxTy));
21292 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
21293 DAG.getConstant(1, dl, VecIdxTy));
21295 SDValue ShAmt = DAG.getConstant(32, dl,
21296 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
21297 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
21298 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
21299 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
21300 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
21301 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
21302 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
21304 // Store the value to a temporary stack slot.
21305 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
21306 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
21307 MachinePointerInfo(), false, false, 0);
21309 EVT ElementType = InputVector.getValueType().getVectorElementType();
21310 unsigned EltSize = ElementType.getSizeInBits() / 8;
21312 // Replace each use (extract) with a load of the appropriate element.
21313 for (unsigned i = 0; i < 4; ++i) {
21314 uint64_t Offset = EltSize * i;
21315 SDValue OffsetVal = DAG.getConstant(Offset, dl, TLI.getPointerTy());
21317 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
21318 StackPtr, OffsetVal);
21320 // Load the scalar.
21321 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
21322 ScalarAddr, MachinePointerInfo(),
21323 false, false, false, 0);
21328 // Replace the extracts
21329 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
21330 UE = Uses.end(); UI != UE; ++UI) {
21331 SDNode *Extract = *UI;
21333 SDValue Idx = Extract->getOperand(1);
21334 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
21335 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
21338 // The replacement was made in place; don't return anything.
21342 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
21343 static std::pair<unsigned, bool>
21344 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
21345 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
21346 if (!VT.isVector())
21347 return std::make_pair(0, false);
21349 bool NeedSplit = false;
21350 switch (VT.getSimpleVT().SimpleTy) {
21351 default: return std::make_pair(0, false);
21354 if (!Subtarget->hasVLX())
21355 return std::make_pair(0, false);
21359 if (!Subtarget->hasBWI())
21360 return std::make_pair(0, false);
21364 if (!Subtarget->hasAVX512())
21365 return std::make_pair(0, false);
21370 if (!Subtarget->hasAVX2())
21372 if (!Subtarget->hasAVX())
21373 return std::make_pair(0, false);
21378 if (!Subtarget->hasSSE2())
21379 return std::make_pair(0, false);
21382 // SSE2 has only a small subset of the operations.
21383 bool hasUnsigned = Subtarget->hasSSE41() ||
21384 (Subtarget->hasSSE2() && VT == MVT::v16i8);
21385 bool hasSigned = Subtarget->hasSSE41() ||
21386 (Subtarget->hasSSE2() && VT == MVT::v8i16);
21388 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21391 // Check for x CC y ? x : y.
21392 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21393 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21398 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21401 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21404 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21407 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21409 // Check for x CC y ? y : x -- a min/max with reversed arms.
21410 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21411 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21416 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21419 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21422 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21425 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21429 return std::make_pair(Opc, NeedSplit);
21433 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
21434 const X86Subtarget *Subtarget) {
21436 SDValue Cond = N->getOperand(0);
21437 SDValue LHS = N->getOperand(1);
21438 SDValue RHS = N->getOperand(2);
21440 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
21441 SDValue CondSrc = Cond->getOperand(0);
21442 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
21443 Cond = CondSrc->getOperand(0);
21446 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
21449 // A vselect where all conditions and data are constants can be optimized into
21450 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
21451 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
21452 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
21455 unsigned MaskValue = 0;
21456 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
21459 MVT VT = N->getSimpleValueType(0);
21460 unsigned NumElems = VT.getVectorNumElements();
21461 SmallVector<int, 8> ShuffleMask(NumElems, -1);
21462 for (unsigned i = 0; i < NumElems; ++i) {
21463 // Be sure we emit undef where we can.
21464 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
21465 ShuffleMask[i] = -1;
21467 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
21470 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21471 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
21473 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
21476 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
21478 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
21479 TargetLowering::DAGCombinerInfo &DCI,
21480 const X86Subtarget *Subtarget) {
21482 SDValue Cond = N->getOperand(0);
21483 // Get the LHS/RHS of the select.
21484 SDValue LHS = N->getOperand(1);
21485 SDValue RHS = N->getOperand(2);
21486 EVT VT = LHS.getValueType();
21487 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21489 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
21490 // instructions match the semantics of the common C idiom x<y?x:y but not
21491 // x<=y?x:y, because of how they handle negative zero (which can be
21492 // ignored in unsafe-math mode).
21493 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
21494 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
21495 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
21496 (Subtarget->hasSSE2() ||
21497 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
21498 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21500 unsigned Opcode = 0;
21501 // Check for x CC y ? x : y.
21502 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21503 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21507 // Converting this to a min would handle NaNs incorrectly, and swapping
21508 // the operands would cause it to handle comparisons between positive
21509 // and negative zero incorrectly.
21510 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21511 if (!DAG.getTarget().Options.UnsafeFPMath &&
21512 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21514 std::swap(LHS, RHS);
21516 Opcode = X86ISD::FMIN;
21519 // Converting this to a min would handle comparisons between positive
21520 // and negative zero incorrectly.
21521 if (!DAG.getTarget().Options.UnsafeFPMath &&
21522 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21524 Opcode = X86ISD::FMIN;
21527 // Converting this to a min would handle both negative zeros and NaNs
21528 // incorrectly, but we can swap the operands to fix both.
21529 std::swap(LHS, RHS);
21533 Opcode = X86ISD::FMIN;
21537 // Converting this to a max would handle comparisons between positive
21538 // and negative zero incorrectly.
21539 if (!DAG.getTarget().Options.UnsafeFPMath &&
21540 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21542 Opcode = X86ISD::FMAX;
21545 // Converting this to a max would handle NaNs incorrectly, and swapping
21546 // the operands would cause it to handle comparisons between positive
21547 // and negative zero incorrectly.
21548 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21549 if (!DAG.getTarget().Options.UnsafeFPMath &&
21550 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21552 std::swap(LHS, RHS);
21554 Opcode = X86ISD::FMAX;
21557 // Converting this to a max would handle both negative zeros and NaNs
21558 // incorrectly, but we can swap the operands to fix both.
21559 std::swap(LHS, RHS);
21563 Opcode = X86ISD::FMAX;
21566 // Check for x CC y ? y : x -- a min/max with reversed arms.
21567 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21568 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21572 // Converting this to a min would handle comparisons between positive
21573 // and negative zero incorrectly, and swapping the operands would
21574 // cause it to handle NaNs incorrectly.
21575 if (!DAG.getTarget().Options.UnsafeFPMath &&
21576 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21577 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21579 std::swap(LHS, RHS);
21581 Opcode = X86ISD::FMIN;
21584 // Converting this to a min would handle NaNs incorrectly.
21585 if (!DAG.getTarget().Options.UnsafeFPMath &&
21586 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21588 Opcode = X86ISD::FMIN;
21591 // Converting this to a min would handle both negative zeros and NaNs
21592 // incorrectly, but we can swap the operands to fix both.
21593 std::swap(LHS, RHS);
21597 Opcode = X86ISD::FMIN;
21601 // Converting this to a max would handle NaNs incorrectly.
21602 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21604 Opcode = X86ISD::FMAX;
21607 // Converting this to a max would handle comparisons between positive
21608 // and negative zero incorrectly, and swapping the operands would
21609 // cause it to handle NaNs incorrectly.
21610 if (!DAG.getTarget().Options.UnsafeFPMath &&
21611 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21612 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21614 std::swap(LHS, RHS);
21616 Opcode = X86ISD::FMAX;
21619 // Converting this to a max would handle both negative zeros and NaNs
21620 // incorrectly, but we can swap the operands to fix both.
21621 std::swap(LHS, RHS);
21625 Opcode = X86ISD::FMAX;
21631 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21634 EVT CondVT = Cond.getValueType();
21635 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21636 CondVT.getVectorElementType() == MVT::i1) {
21637 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21638 // lowering on KNL. In this case we convert it to
21639 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21640 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21641 // Since SKX these selects have a proper lowering.
21642 EVT OpVT = LHS.getValueType();
21643 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21644 (OpVT.getVectorElementType() == MVT::i8 ||
21645 OpVT.getVectorElementType() == MVT::i16) &&
21646 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21647 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21648 DCI.AddToWorklist(Cond.getNode());
21649 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21652 // If this is a select between two integer constants, try to do some
21654 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21655 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21656 // Don't do this for crazy integer types.
21657 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21658 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21659 // so that TrueC (the true value) is larger than FalseC.
21660 bool NeedsCondInvert = false;
21662 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21663 // Efficiently invertible.
21664 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21665 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21666 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21667 NeedsCondInvert = true;
21668 std::swap(TrueC, FalseC);
21671 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21672 if (FalseC->getAPIntValue() == 0 &&
21673 TrueC->getAPIntValue().isPowerOf2()) {
21674 if (NeedsCondInvert) // Invert the condition if needed.
21675 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21676 DAG.getConstant(1, DL, Cond.getValueType()));
21678 // Zero extend the condition if needed.
21679 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21681 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21682 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21683 DAG.getConstant(ShAmt, DL, MVT::i8));
21686 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21687 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21688 if (NeedsCondInvert) // Invert the condition if needed.
21689 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21690 DAG.getConstant(1, DL, Cond.getValueType()));
21692 // Zero extend the condition if needed.
21693 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21694 FalseC->getValueType(0), Cond);
21695 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21696 SDValue(FalseC, 0));
21699 // Optimize cases that will turn into an LEA instruction. This requires
21700 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21701 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21702 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21703 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21705 bool isFastMultiplier = false;
21707 switch ((unsigned char)Diff) {
21709 case 1: // result = add base, cond
21710 case 2: // result = lea base( , cond*2)
21711 case 3: // result = lea base(cond, cond*2)
21712 case 4: // result = lea base( , cond*4)
21713 case 5: // result = lea base(cond, cond*4)
21714 case 8: // result = lea base( , cond*8)
21715 case 9: // result = lea base(cond, cond*8)
21716 isFastMultiplier = true;
21721 if (isFastMultiplier) {
21722 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21723 if (NeedsCondInvert) // Invert the condition if needed.
21724 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21725 DAG.getConstant(1, DL, Cond.getValueType()));
21727 // Zero extend the condition if needed.
21728 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21730 // Scale the condition by the difference.
21732 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21733 DAG.getConstant(Diff, DL,
21734 Cond.getValueType()));
21736 // Add the base if non-zero.
21737 if (FalseC->getAPIntValue() != 0)
21738 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21739 SDValue(FalseC, 0));
21746 // Canonicalize max and min:
21747 // (x > y) ? x : y -> (x >= y) ? x : y
21748 // (x < y) ? x : y -> (x <= y) ? x : y
21749 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21750 // the need for an extra compare
21751 // against zero. e.g.
21752 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21754 // testl %edi, %edi
21756 // cmovgl %edi, %eax
21760 // cmovsl %eax, %edi
21761 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21762 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21763 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21764 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21769 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21770 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21771 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21772 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21777 // Early exit check
21778 if (!TLI.isTypeLegal(VT))
21781 // Match VSELECTs into subs with unsigned saturation.
21782 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21783 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21784 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21785 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21786 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21788 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21789 // left side invert the predicate to simplify logic below.
21791 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21793 CC = ISD::getSetCCInverse(CC, true);
21794 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21798 if (Other.getNode() && Other->getNumOperands() == 2 &&
21799 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21800 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21801 SDValue CondRHS = Cond->getOperand(1);
21803 // Look for a general sub with unsigned saturation first.
21804 // x >= y ? x-y : 0 --> subus x, y
21805 // x > y ? x-y : 0 --> subus x, y
21806 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21807 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21808 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21810 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21811 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21812 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21813 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21814 // If the RHS is a constant we have to reverse the const
21815 // canonicalization.
21816 // x > C-1 ? x+-C : 0 --> subus x, C
21817 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21818 CondRHSConst->getAPIntValue() ==
21819 (-OpRHSConst->getAPIntValue() - 1))
21820 return DAG.getNode(
21821 X86ISD::SUBUS, DL, VT, OpLHS,
21822 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
21824 // Another special case: If C was a sign bit, the sub has been
21825 // canonicalized into a xor.
21826 // FIXME: Would it be better to use computeKnownBits to determine
21827 // whether it's safe to decanonicalize the xor?
21828 // x s< 0 ? x^C : 0 --> subus x, C
21829 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21830 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21831 OpRHSConst->getAPIntValue().isSignBit())
21832 // Note that we have to rebuild the RHS constant here to ensure we
21833 // don't rely on particular values of undef lanes.
21834 return DAG.getNode(
21835 X86ISD::SUBUS, DL, VT, OpLHS,
21836 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
21841 // Try to match a min/max vector operation.
21842 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21843 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21844 unsigned Opc = ret.first;
21845 bool NeedSplit = ret.second;
21847 if (Opc && NeedSplit) {
21848 unsigned NumElems = VT.getVectorNumElements();
21849 // Extract the LHS vectors
21850 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21851 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21853 // Extract the RHS vectors
21854 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21855 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21857 // Create min/max for each subvector
21858 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21859 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21861 // Merge the result
21862 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21864 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21867 // Simplify vector selection if condition value type matches vselect
21869 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
21870 assert(Cond.getValueType().isVector() &&
21871 "vector select expects a vector selector!");
21873 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21874 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21876 // Try invert the condition if true value is not all 1s and false value
21878 if (!TValIsAllOnes && !FValIsAllZeros &&
21879 // Check if the selector will be produced by CMPP*/PCMP*
21880 Cond.getOpcode() == ISD::SETCC &&
21881 // Check if SETCC has already been promoted
21882 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
21883 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21884 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21886 if (TValIsAllZeros || FValIsAllOnes) {
21887 SDValue CC = Cond.getOperand(2);
21888 ISD::CondCode NewCC =
21889 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21890 Cond.getOperand(0).getValueType().isInteger());
21891 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21892 std::swap(LHS, RHS);
21893 TValIsAllOnes = FValIsAllOnes;
21894 FValIsAllZeros = TValIsAllZeros;
21898 if (TValIsAllOnes || FValIsAllZeros) {
21901 if (TValIsAllOnes && FValIsAllZeros)
21903 else if (TValIsAllOnes)
21905 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
21906 else if (FValIsAllZeros)
21907 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21908 DAG.getBitcast(CondVT, LHS));
21910 return DAG.getBitcast(VT, Ret);
21914 // We should generate an X86ISD::BLENDI from a vselect if its argument
21915 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
21916 // constants. This specific pattern gets generated when we split a
21917 // selector for a 512 bit vector in a machine without AVX512 (but with
21918 // 256-bit vectors), during legalization:
21920 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
21922 // Iff we find this pattern and the build_vectors are built from
21923 // constants, we translate the vselect into a shuffle_vector that we
21924 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
21925 if ((N->getOpcode() == ISD::VSELECT ||
21926 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
21927 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
21928 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
21929 if (Shuffle.getNode())
21933 // If this is a *dynamic* select (non-constant condition) and we can match
21934 // this node with one of the variable blend instructions, restructure the
21935 // condition so that the blends can use the high bit of each element and use
21936 // SimplifyDemandedBits to simplify the condition operand.
21937 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
21938 !DCI.isBeforeLegalize() &&
21939 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
21940 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
21942 // Don't optimize vector selects that map to mask-registers.
21946 // We can only handle the cases where VSELECT is directly legal on the
21947 // subtarget. We custom lower VSELECT nodes with constant conditions and
21948 // this makes it hard to see whether a dynamic VSELECT will correctly
21949 // lower, so we both check the operation's status and explicitly handle the
21950 // cases where a *dynamic* blend will fail even though a constant-condition
21951 // blend could be custom lowered.
21952 // FIXME: We should find a better way to handle this class of problems.
21953 // Potentially, we should combine constant-condition vselect nodes
21954 // pre-legalization into shuffles and not mark as many types as custom
21956 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
21958 // FIXME: We don't support i16-element blends currently. We could and
21959 // should support them by making *all* the bits in the condition be set
21960 // rather than just the high bit and using an i8-element blend.
21961 if (VT.getScalarType() == MVT::i16)
21963 // Dynamic blending was only available from SSE4.1 onward.
21964 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
21966 // Byte blends are only available in AVX2
21967 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
21968 !Subtarget->hasAVX2())
21971 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
21972 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
21974 APInt KnownZero, KnownOne;
21975 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
21976 DCI.isBeforeLegalizeOps());
21977 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
21978 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
21980 // If we changed the computation somewhere in the DAG, this change
21981 // will affect all users of Cond.
21982 // Make sure it is fine and update all the nodes so that we do not
21983 // use the generic VSELECT anymore. Otherwise, we may perform
21984 // wrong optimizations as we messed up with the actual expectation
21985 // for the vector boolean values.
21986 if (Cond != TLO.Old) {
21987 // Check all uses of that condition operand to check whether it will be
21988 // consumed by non-BLEND instructions, which may depend on all bits are
21990 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21992 if (I->getOpcode() != ISD::VSELECT)
21993 // TODO: Add other opcodes eventually lowered into BLEND.
21996 // Update all the users of the condition, before committing the change,
21997 // so that the VSELECT optimizations that expect the correct vector
21998 // boolean value will not be triggered.
21999 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22001 DAG.ReplaceAllUsesOfValueWith(
22003 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
22004 Cond, I->getOperand(1), I->getOperand(2)));
22005 DCI.CommitTargetLoweringOpt(TLO);
22008 // At this point, only Cond is changed. Change the condition
22009 // just for N to keep the opportunity to optimize all other
22010 // users their own way.
22011 DAG.ReplaceAllUsesOfValueWith(
22013 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
22014 TLO.New, N->getOperand(1), N->getOperand(2)));
22022 // Check whether a boolean test is testing a boolean value generated by
22023 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22026 // Simplify the following patterns:
22027 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22028 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
22029 // to (Op EFLAGS Cond)
22031 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
22032 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
22033 // to (Op EFLAGS !Cond)
22035 // where Op could be BRCOND or CMOV.
22037 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
22038 // Quit if not CMP and SUB with its value result used.
22039 if (Cmp.getOpcode() != X86ISD::CMP &&
22040 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
22043 // Quit if not used as a boolean value.
22044 if (CC != X86::COND_E && CC != X86::COND_NE)
22047 // Check CMP operands. One of them should be 0 or 1 and the other should be
22048 // an SetCC or extended from it.
22049 SDValue Op1 = Cmp.getOperand(0);
22050 SDValue Op2 = Cmp.getOperand(1);
22053 const ConstantSDNode* C = nullptr;
22054 bool needOppositeCond = (CC == X86::COND_E);
22055 bool checkAgainstTrue = false; // Is it a comparison against 1?
22057 if ((C = dyn_cast<ConstantSDNode>(Op1)))
22059 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
22061 else // Quit if all operands are not constants.
22064 if (C->getZExtValue() == 1) {
22065 needOppositeCond = !needOppositeCond;
22066 checkAgainstTrue = true;
22067 } else if (C->getZExtValue() != 0)
22068 // Quit if the constant is neither 0 or 1.
22071 bool truncatedToBoolWithAnd = false;
22072 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
22073 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
22074 SetCC.getOpcode() == ISD::TRUNCATE ||
22075 SetCC.getOpcode() == ISD::AND) {
22076 if (SetCC.getOpcode() == ISD::AND) {
22078 ConstantSDNode *CS;
22079 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
22080 CS->getZExtValue() == 1)
22082 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
22083 CS->getZExtValue() == 1)
22087 SetCC = SetCC.getOperand(OpIdx);
22088 truncatedToBoolWithAnd = true;
22090 SetCC = SetCC.getOperand(0);
22093 switch (SetCC.getOpcode()) {
22094 case X86ISD::SETCC_CARRY:
22095 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
22096 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
22097 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
22098 // truncated to i1 using 'and'.
22099 if (checkAgainstTrue && !truncatedToBoolWithAnd)
22101 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
22102 "Invalid use of SETCC_CARRY!");
22104 case X86ISD::SETCC:
22105 // Set the condition code or opposite one if necessary.
22106 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
22107 if (needOppositeCond)
22108 CC = X86::GetOppositeBranchCondition(CC);
22109 return SetCC.getOperand(1);
22110 case X86ISD::CMOV: {
22111 // Check whether false/true value has canonical one, i.e. 0 or 1.
22112 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
22113 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
22114 // Quit if true value is not a constant.
22117 // Quit if false value is not a constant.
22119 SDValue Op = SetCC.getOperand(0);
22120 // Skip 'zext' or 'trunc' node.
22121 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
22122 Op.getOpcode() == ISD::TRUNCATE)
22123 Op = Op.getOperand(0);
22124 // A special case for rdrand/rdseed, where 0 is set if false cond is
22126 if ((Op.getOpcode() != X86ISD::RDRAND &&
22127 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
22130 // Quit if false value is not the constant 0 or 1.
22131 bool FValIsFalse = true;
22132 if (FVal && FVal->getZExtValue() != 0) {
22133 if (FVal->getZExtValue() != 1)
22135 // If FVal is 1, opposite cond is needed.
22136 needOppositeCond = !needOppositeCond;
22137 FValIsFalse = false;
22139 // Quit if TVal is not the constant opposite of FVal.
22140 if (FValIsFalse && TVal->getZExtValue() != 1)
22142 if (!FValIsFalse && TVal->getZExtValue() != 0)
22144 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
22145 if (needOppositeCond)
22146 CC = X86::GetOppositeBranchCondition(CC);
22147 return SetCC.getOperand(3);
22154 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
22156 /// (X86or (X86setcc) (X86setcc))
22157 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
22158 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
22159 X86::CondCode &CC1, SDValue &Flags,
22161 if (Cond->getOpcode() == X86ISD::CMP) {
22162 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
22163 if (!CondOp1C || !CondOp1C->isNullValue())
22166 Cond = Cond->getOperand(0);
22171 SDValue SetCC0, SetCC1;
22172 switch (Cond->getOpcode()) {
22173 default: return false;
22180 SetCC0 = Cond->getOperand(0);
22181 SetCC1 = Cond->getOperand(1);
22185 // Make sure we have SETCC nodes, using the same flags value.
22186 if (SetCC0.getOpcode() != X86ISD::SETCC ||
22187 SetCC1.getOpcode() != X86ISD::SETCC ||
22188 SetCC0->getOperand(1) != SetCC1->getOperand(1))
22191 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
22192 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
22193 Flags = SetCC0->getOperand(1);
22197 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
22198 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
22199 TargetLowering::DAGCombinerInfo &DCI,
22200 const X86Subtarget *Subtarget) {
22203 // If the flag operand isn't dead, don't touch this CMOV.
22204 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
22207 SDValue FalseOp = N->getOperand(0);
22208 SDValue TrueOp = N->getOperand(1);
22209 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
22210 SDValue Cond = N->getOperand(3);
22212 if (CC == X86::COND_E || CC == X86::COND_NE) {
22213 switch (Cond.getOpcode()) {
22217 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
22218 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
22219 return (CC == X86::COND_E) ? FalseOp : TrueOp;
22225 Flags = checkBoolTestSetCCCombine(Cond, CC);
22226 if (Flags.getNode() &&
22227 // Extra check as FCMOV only supports a subset of X86 cond.
22228 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
22229 SDValue Ops[] = { FalseOp, TrueOp,
22230 DAG.getConstant(CC, DL, MVT::i8), Flags };
22231 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
22234 // If this is a select between two integer constants, try to do some
22235 // optimizations. Note that the operands are ordered the opposite of SELECT
22237 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
22238 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
22239 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
22240 // larger than FalseC (the false value).
22241 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
22242 CC = X86::GetOppositeBranchCondition(CC);
22243 std::swap(TrueC, FalseC);
22244 std::swap(TrueOp, FalseOp);
22247 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
22248 // This is efficient for any integer data type (including i8/i16) and
22250 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
22251 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22252 DAG.getConstant(CC, DL, MVT::i8), Cond);
22254 // Zero extend the condition if needed.
22255 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
22257 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22258 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
22259 DAG.getConstant(ShAmt, DL, MVT::i8));
22260 if (N->getNumValues() == 2) // Dead flag value?
22261 return DCI.CombineTo(N, Cond, SDValue());
22265 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
22266 // for any integer data type, including i8/i16.
22267 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22268 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22269 DAG.getConstant(CC, DL, MVT::i8), Cond);
22271 // Zero extend the condition if needed.
22272 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22273 FalseC->getValueType(0), Cond);
22274 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22275 SDValue(FalseC, 0));
22277 if (N->getNumValues() == 2) // Dead flag value?
22278 return DCI.CombineTo(N, Cond, SDValue());
22282 // Optimize cases that will turn into an LEA instruction. This requires
22283 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22284 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22285 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22286 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22288 bool isFastMultiplier = false;
22290 switch ((unsigned char)Diff) {
22292 case 1: // result = add base, cond
22293 case 2: // result = lea base( , cond*2)
22294 case 3: // result = lea base(cond, cond*2)
22295 case 4: // result = lea base( , cond*4)
22296 case 5: // result = lea base(cond, cond*4)
22297 case 8: // result = lea base( , cond*8)
22298 case 9: // result = lea base(cond, cond*8)
22299 isFastMultiplier = true;
22304 if (isFastMultiplier) {
22305 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22306 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22307 DAG.getConstant(CC, DL, MVT::i8), Cond);
22308 // Zero extend the condition if needed.
22309 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22311 // Scale the condition by the difference.
22313 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22314 DAG.getConstant(Diff, DL, Cond.getValueType()));
22316 // Add the base if non-zero.
22317 if (FalseC->getAPIntValue() != 0)
22318 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22319 SDValue(FalseC, 0));
22320 if (N->getNumValues() == 2) // Dead flag value?
22321 return DCI.CombineTo(N, Cond, SDValue());
22328 // Handle these cases:
22329 // (select (x != c), e, c) -> select (x != c), e, x),
22330 // (select (x == c), c, e) -> select (x == c), x, e)
22331 // where the c is an integer constant, and the "select" is the combination
22332 // of CMOV and CMP.
22334 // The rationale for this change is that the conditional-move from a constant
22335 // needs two instructions, however, conditional-move from a register needs
22336 // only one instruction.
22338 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
22339 // some instruction-combining opportunities. This opt needs to be
22340 // postponed as late as possible.
22342 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
22343 // the DCI.xxxx conditions are provided to postpone the optimization as
22344 // late as possible.
22346 ConstantSDNode *CmpAgainst = nullptr;
22347 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
22348 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
22349 !isa<ConstantSDNode>(Cond.getOperand(0))) {
22351 if (CC == X86::COND_NE &&
22352 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
22353 CC = X86::GetOppositeBranchCondition(CC);
22354 std::swap(TrueOp, FalseOp);
22357 if (CC == X86::COND_E &&
22358 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
22359 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
22360 DAG.getConstant(CC, DL, MVT::i8), Cond };
22361 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
22366 // Fold and/or of setcc's to double CMOV:
22367 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
22368 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
22370 // This combine lets us generate:
22371 // cmovcc1 (jcc1 if we don't have CMOV)
22377 // cmovne (jne if we don't have CMOV)
22378 // When we can't use the CMOV instruction, it might increase branch
22380 // When we can use CMOV, or when there is no mispredict, this improves
22381 // throughput and reduces register pressure.
22383 if (CC == X86::COND_NE) {
22385 X86::CondCode CC0, CC1;
22387 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
22389 std::swap(FalseOp, TrueOp);
22390 CC0 = X86::GetOppositeBranchCondition(CC0);
22391 CC1 = X86::GetOppositeBranchCondition(CC1);
22394 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
22396 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
22397 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
22398 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
22399 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
22407 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
22408 const X86Subtarget *Subtarget) {
22409 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
22411 default: return SDValue();
22412 // SSE/AVX/AVX2 blend intrinsics.
22413 case Intrinsic::x86_avx2_pblendvb:
22414 // Don't try to simplify this intrinsic if we don't have AVX2.
22415 if (!Subtarget->hasAVX2())
22418 case Intrinsic::x86_avx_blendv_pd_256:
22419 case Intrinsic::x86_avx_blendv_ps_256:
22420 // Don't try to simplify this intrinsic if we don't have AVX.
22421 if (!Subtarget->hasAVX())
22424 case Intrinsic::x86_sse41_blendvps:
22425 case Intrinsic::x86_sse41_blendvpd:
22426 case Intrinsic::x86_sse41_pblendvb: {
22427 SDValue Op0 = N->getOperand(1);
22428 SDValue Op1 = N->getOperand(2);
22429 SDValue Mask = N->getOperand(3);
22431 // Don't try to simplify this intrinsic if we don't have SSE4.1.
22432 if (!Subtarget->hasSSE41())
22435 // fold (blend A, A, Mask) -> A
22438 // fold (blend A, B, allZeros) -> A
22439 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
22441 // fold (blend A, B, allOnes) -> B
22442 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
22445 // Simplify the case where the mask is a constant i32 value.
22446 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
22447 if (C->isNullValue())
22449 if (C->isAllOnesValue())
22456 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
22457 case Intrinsic::x86_sse2_psrai_w:
22458 case Intrinsic::x86_sse2_psrai_d:
22459 case Intrinsic::x86_avx2_psrai_w:
22460 case Intrinsic::x86_avx2_psrai_d:
22461 case Intrinsic::x86_sse2_psra_w:
22462 case Intrinsic::x86_sse2_psra_d:
22463 case Intrinsic::x86_avx2_psra_w:
22464 case Intrinsic::x86_avx2_psra_d: {
22465 SDValue Op0 = N->getOperand(1);
22466 SDValue Op1 = N->getOperand(2);
22467 EVT VT = Op0.getValueType();
22468 assert(VT.isVector() && "Expected a vector type!");
22470 if (isa<BuildVectorSDNode>(Op1))
22471 Op1 = Op1.getOperand(0);
22473 if (!isa<ConstantSDNode>(Op1))
22476 EVT SVT = VT.getVectorElementType();
22477 unsigned SVTBits = SVT.getSizeInBits();
22479 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
22480 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
22481 uint64_t ShAmt = C.getZExtValue();
22483 // Don't try to convert this shift into a ISD::SRA if the shift
22484 // count is bigger than or equal to the element size.
22485 if (ShAmt >= SVTBits)
22488 // Trivial case: if the shift count is zero, then fold this
22489 // into the first operand.
22493 // Replace this packed shift intrinsic with a target independent
22496 SDValue Splat = DAG.getConstant(C, DL, VT);
22497 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
22502 /// PerformMulCombine - Optimize a single multiply with constant into two
22503 /// in order to implement it with two cheaper instructions, e.g.
22504 /// LEA + SHL, LEA + LEA.
22505 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
22506 TargetLowering::DAGCombinerInfo &DCI) {
22507 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
22510 EVT VT = N->getValueType(0);
22511 if (VT != MVT::i64 && VT != MVT::i32)
22514 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
22517 uint64_t MulAmt = C->getZExtValue();
22518 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
22521 uint64_t MulAmt1 = 0;
22522 uint64_t MulAmt2 = 0;
22523 if ((MulAmt % 9) == 0) {
22525 MulAmt2 = MulAmt / 9;
22526 } else if ((MulAmt % 5) == 0) {
22528 MulAmt2 = MulAmt / 5;
22529 } else if ((MulAmt % 3) == 0) {
22531 MulAmt2 = MulAmt / 3;
22534 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
22537 if (isPowerOf2_64(MulAmt2) &&
22538 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
22539 // If second multiplifer is pow2, issue it first. We want the multiply by
22540 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
22542 std::swap(MulAmt1, MulAmt2);
22545 if (isPowerOf2_64(MulAmt1))
22546 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
22547 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
22549 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
22550 DAG.getConstant(MulAmt1, DL, VT));
22552 if (isPowerOf2_64(MulAmt2))
22553 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
22554 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
22556 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
22557 DAG.getConstant(MulAmt2, DL, VT));
22559 // Do not add new nodes to DAG combiner worklist.
22560 DCI.CombineTo(N, NewMul, false);
22565 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22566 SDValue N0 = N->getOperand(0);
22567 SDValue N1 = N->getOperand(1);
22568 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22569 EVT VT = N0.getValueType();
22571 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22572 // since the result of setcc_c is all zero's or all ones.
22573 if (VT.isInteger() && !VT.isVector() &&
22574 N1C && N0.getOpcode() == ISD::AND &&
22575 N0.getOperand(1).getOpcode() == ISD::Constant) {
22576 SDValue N00 = N0.getOperand(0);
22577 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22578 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22579 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22580 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22581 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22582 APInt ShAmt = N1C->getAPIntValue();
22583 Mask = Mask.shl(ShAmt);
22586 return DAG.getNode(ISD::AND, DL, VT,
22587 N00, DAG.getConstant(Mask, DL, VT));
22592 // Hardware support for vector shifts is sparse which makes us scalarize the
22593 // vector operations in many cases. Also, on sandybridge ADD is faster than
22595 // (shl V, 1) -> add V,V
22596 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22597 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22598 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22599 // We shift all of the values by one. In many cases we do not have
22600 // hardware support for this operation. This is better expressed as an ADD
22602 if (N1SplatC->getZExtValue() == 1)
22603 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22609 /// \brief Returns a vector of 0s if the node in input is a vector logical
22610 /// shift by a constant amount which is known to be bigger than or equal
22611 /// to the vector element size in bits.
22612 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22613 const X86Subtarget *Subtarget) {
22614 EVT VT = N->getValueType(0);
22616 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22617 (!Subtarget->hasInt256() ||
22618 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22621 SDValue Amt = N->getOperand(1);
22623 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22624 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22625 APInt ShiftAmt = AmtSplat->getAPIntValue();
22626 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22628 // SSE2/AVX2 logical shifts always return a vector of 0s
22629 // if the shift amount is bigger than or equal to
22630 // the element size. The constant shift amount will be
22631 // encoded as a 8-bit immediate.
22632 if (ShiftAmt.trunc(8).uge(MaxAmount))
22633 return getZeroVector(VT, Subtarget, DAG, DL);
22639 /// PerformShiftCombine - Combine shifts.
22640 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22641 TargetLowering::DAGCombinerInfo &DCI,
22642 const X86Subtarget *Subtarget) {
22643 if (N->getOpcode() == ISD::SHL) {
22644 SDValue V = PerformSHLCombine(N, DAG);
22645 if (V.getNode()) return V;
22648 if (N->getOpcode() != ISD::SRA) {
22649 // Try to fold this logical shift into a zero vector.
22650 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22651 if (V.getNode()) return V;
22657 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22658 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22659 // and friends. Likewise for OR -> CMPNEQSS.
22660 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22661 TargetLowering::DAGCombinerInfo &DCI,
22662 const X86Subtarget *Subtarget) {
22665 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22666 // we're requiring SSE2 for both.
22667 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22668 SDValue N0 = N->getOperand(0);
22669 SDValue N1 = N->getOperand(1);
22670 SDValue CMP0 = N0->getOperand(1);
22671 SDValue CMP1 = N1->getOperand(1);
22674 // The SETCCs should both refer to the same CMP.
22675 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22678 SDValue CMP00 = CMP0->getOperand(0);
22679 SDValue CMP01 = CMP0->getOperand(1);
22680 EVT VT = CMP00.getValueType();
22682 if (VT == MVT::f32 || VT == MVT::f64) {
22683 bool ExpectingFlags = false;
22684 // Check for any users that want flags:
22685 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22686 !ExpectingFlags && UI != UE; ++UI)
22687 switch (UI->getOpcode()) {
22692 ExpectingFlags = true;
22694 case ISD::CopyToReg:
22695 case ISD::SIGN_EXTEND:
22696 case ISD::ZERO_EXTEND:
22697 case ISD::ANY_EXTEND:
22701 if (!ExpectingFlags) {
22702 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22703 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22705 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22706 X86::CondCode tmp = cc0;
22711 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22712 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22713 // FIXME: need symbolic constants for these magic numbers.
22714 // See X86ATTInstPrinter.cpp:printSSECC().
22715 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22716 if (Subtarget->hasAVX512()) {
22717 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22719 DAG.getConstant(x86cc, DL, MVT::i8));
22720 if (N->getValueType(0) != MVT::i1)
22721 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22725 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22726 CMP00.getValueType(), CMP00, CMP01,
22727 DAG.getConstant(x86cc, DL,
22730 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22731 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22733 if (is64BitFP && !Subtarget->is64Bit()) {
22734 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22735 // 64-bit integer, since that's not a legal type. Since
22736 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22737 // bits, but can do this little dance to extract the lowest 32 bits
22738 // and work with those going forward.
22739 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22741 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
22742 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22743 Vector32, DAG.getIntPtrConstant(0, DL));
22747 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
22748 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22749 DAG.getConstant(1, DL, IntVT));
22750 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
22752 return OneBitOfTruth;
22760 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22761 /// so it can be folded inside ANDNP.
22762 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22763 EVT VT = N->getValueType(0);
22765 // Match direct AllOnes for 128 and 256-bit vectors
22766 if (ISD::isBuildVectorAllOnes(N))
22769 // Look through a bit convert.
22770 if (N->getOpcode() == ISD::BITCAST)
22771 N = N->getOperand(0).getNode();
22773 // Sometimes the operand may come from a insert_subvector building a 256-bit
22775 if (VT.is256BitVector() &&
22776 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22777 SDValue V1 = N->getOperand(0);
22778 SDValue V2 = N->getOperand(1);
22780 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22781 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22782 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22783 ISD::isBuildVectorAllOnes(V2.getNode()))
22790 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22791 // register. In most cases we actually compare or select YMM-sized registers
22792 // and mixing the two types creates horrible code. This method optimizes
22793 // some of the transition sequences.
22794 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22795 TargetLowering::DAGCombinerInfo &DCI,
22796 const X86Subtarget *Subtarget) {
22797 EVT VT = N->getValueType(0);
22798 if (!VT.is256BitVector())
22801 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22802 N->getOpcode() == ISD::ZERO_EXTEND ||
22803 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22805 SDValue Narrow = N->getOperand(0);
22806 EVT NarrowVT = Narrow->getValueType(0);
22807 if (!NarrowVT.is128BitVector())
22810 if (Narrow->getOpcode() != ISD::XOR &&
22811 Narrow->getOpcode() != ISD::AND &&
22812 Narrow->getOpcode() != ISD::OR)
22815 SDValue N0 = Narrow->getOperand(0);
22816 SDValue N1 = Narrow->getOperand(1);
22819 // The Left side has to be a trunc.
22820 if (N0.getOpcode() != ISD::TRUNCATE)
22823 // The type of the truncated inputs.
22824 EVT WideVT = N0->getOperand(0)->getValueType(0);
22828 // The right side has to be a 'trunc' or a constant vector.
22829 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22830 ConstantSDNode *RHSConstSplat = nullptr;
22831 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22832 RHSConstSplat = RHSBV->getConstantSplatNode();
22833 if (!RHSTrunc && !RHSConstSplat)
22836 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22838 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22841 // Set N0 and N1 to hold the inputs to the new wide operation.
22842 N0 = N0->getOperand(0);
22843 if (RHSConstSplat) {
22844 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22845 SDValue(RHSConstSplat, 0));
22846 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22847 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22848 } else if (RHSTrunc) {
22849 N1 = N1->getOperand(0);
22852 // Generate the wide operation.
22853 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22854 unsigned Opcode = N->getOpcode();
22856 case ISD::ANY_EXTEND:
22858 case ISD::ZERO_EXTEND: {
22859 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22860 APInt Mask = APInt::getAllOnesValue(InBits);
22861 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22862 return DAG.getNode(ISD::AND, DL, VT,
22863 Op, DAG.getConstant(Mask, DL, VT));
22865 case ISD::SIGN_EXTEND:
22866 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22867 Op, DAG.getValueType(NarrowVT));
22869 llvm_unreachable("Unexpected opcode");
22873 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
22874 TargetLowering::DAGCombinerInfo &DCI,
22875 const X86Subtarget *Subtarget) {
22876 SDValue N0 = N->getOperand(0);
22877 SDValue N1 = N->getOperand(1);
22880 // A vector zext_in_reg may be represented as a shuffle,
22881 // feeding into a bitcast (this represents anyext) feeding into
22882 // an and with a mask.
22883 // We'd like to try to combine that into a shuffle with zero
22884 // plus a bitcast, removing the and.
22885 if (N0.getOpcode() != ISD::BITCAST ||
22886 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
22889 // The other side of the AND should be a splat of 2^C, where C
22890 // is the number of bits in the source type.
22891 if (N1.getOpcode() == ISD::BITCAST)
22892 N1 = N1.getOperand(0);
22893 if (N1.getOpcode() != ISD::BUILD_VECTOR)
22895 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
22897 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
22898 EVT SrcType = Shuffle->getValueType(0);
22900 // We expect a single-source shuffle
22901 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
22904 unsigned SrcSize = SrcType.getScalarSizeInBits();
22906 APInt SplatValue, SplatUndef;
22907 unsigned SplatBitSize;
22909 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
22910 SplatBitSize, HasAnyUndefs))
22913 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
22914 // Make sure the splat matches the mask we expect
22915 if (SplatBitSize > ResSize ||
22916 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
22919 // Make sure the input and output size make sense
22920 if (SrcSize >= ResSize || ResSize % SrcSize)
22923 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
22924 // The number of u's between each two values depends on the ratio between
22925 // the source and dest type.
22926 unsigned ZextRatio = ResSize / SrcSize;
22927 bool IsZext = true;
22928 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
22929 if (i % ZextRatio) {
22930 if (Shuffle->getMaskElt(i) > 0) {
22936 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
22937 // Expected element number
22947 // Ok, perform the transformation - replace the shuffle with
22948 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
22949 // (instead of undef) where the k elements come from the zero vector.
22950 SmallVector<int, 8> Mask;
22951 unsigned NumElems = SrcType.getVectorNumElements();
22952 for (unsigned i = 0; i < NumElems; ++i)
22954 Mask.push_back(NumElems);
22956 Mask.push_back(i / ZextRatio);
22958 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
22959 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
22960 return DAG.getBitcast(N0.getValueType(), NewShuffle);
22963 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22964 TargetLowering::DAGCombinerInfo &DCI,
22965 const X86Subtarget *Subtarget) {
22966 if (DCI.isBeforeLegalizeOps())
22969 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
22972 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
22975 EVT VT = N->getValueType(0);
22976 SDValue N0 = N->getOperand(0);
22977 SDValue N1 = N->getOperand(1);
22980 // Create BEXTR instructions
22981 // BEXTR is ((X >> imm) & (2**size-1))
22982 if (VT == MVT::i32 || VT == MVT::i64) {
22983 // Check for BEXTR.
22984 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22985 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22986 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22987 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22988 if (MaskNode && ShiftNode) {
22989 uint64_t Mask = MaskNode->getZExtValue();
22990 uint64_t Shift = ShiftNode->getZExtValue();
22991 if (isMask_64(Mask)) {
22992 uint64_t MaskSize = countPopulation(Mask);
22993 if (Shift + MaskSize <= VT.getSizeInBits())
22994 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22995 DAG.getConstant(Shift | (MaskSize << 8), DL,
23004 // Want to form ANDNP nodes:
23005 // 1) In the hopes of then easily combining them with OR and AND nodes
23006 // to form PBLEND/PSIGN.
23007 // 2) To match ANDN packed intrinsics
23008 if (VT != MVT::v2i64 && VT != MVT::v4i64)
23011 // Check LHS for vnot
23012 if (N0.getOpcode() == ISD::XOR &&
23013 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
23014 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
23015 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
23017 // Check RHS for vnot
23018 if (N1.getOpcode() == ISD::XOR &&
23019 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
23020 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
23021 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
23026 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
23027 TargetLowering::DAGCombinerInfo &DCI,
23028 const X86Subtarget *Subtarget) {
23029 if (DCI.isBeforeLegalizeOps())
23032 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
23036 SDValue N0 = N->getOperand(0);
23037 SDValue N1 = N->getOperand(1);
23038 EVT VT = N->getValueType(0);
23040 // look for psign/blend
23041 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
23042 if (!Subtarget->hasSSSE3() ||
23043 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
23046 // Canonicalize pandn to RHS
23047 if (N0.getOpcode() == X86ISD::ANDNP)
23049 // or (and (m, y), (pandn m, x))
23050 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
23051 SDValue Mask = N1.getOperand(0);
23052 SDValue X = N1.getOperand(1);
23054 if (N0.getOperand(0) == Mask)
23055 Y = N0.getOperand(1);
23056 if (N0.getOperand(1) == Mask)
23057 Y = N0.getOperand(0);
23059 // Check to see if the mask appeared in both the AND and ANDNP and
23063 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
23064 // Look through mask bitcast.
23065 if (Mask.getOpcode() == ISD::BITCAST)
23066 Mask = Mask.getOperand(0);
23067 if (X.getOpcode() == ISD::BITCAST)
23068 X = X.getOperand(0);
23069 if (Y.getOpcode() == ISD::BITCAST)
23070 Y = Y.getOperand(0);
23072 EVT MaskVT = Mask.getValueType();
23074 // Validate that the Mask operand is a vector sra node.
23075 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
23076 // there is no psrai.b
23077 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
23078 unsigned SraAmt = ~0;
23079 if (Mask.getOpcode() == ISD::SRA) {
23080 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
23081 if (auto *AmtConst = AmtBV->getConstantSplatNode())
23082 SraAmt = AmtConst->getZExtValue();
23083 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
23084 SDValue SraC = Mask.getOperand(1);
23085 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
23087 if ((SraAmt + 1) != EltBits)
23092 // Now we know we at least have a plendvb with the mask val. See if
23093 // we can form a psignb/w/d.
23094 // psign = x.type == y.type == mask.type && y = sub(0, x);
23095 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
23096 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
23097 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
23098 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
23099 "Unsupported VT for PSIGN");
23100 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
23101 return DAG.getBitcast(VT, Mask);
23103 // PBLENDVB only available on SSE 4.1
23104 if (!Subtarget->hasSSE41())
23107 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
23109 X = DAG.getBitcast(BlendVT, X);
23110 Y = DAG.getBitcast(BlendVT, Y);
23111 Mask = DAG.getBitcast(BlendVT, Mask);
23112 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
23113 return DAG.getBitcast(VT, Mask);
23117 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
23120 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
23121 MachineFunction &MF = DAG.getMachineFunction();
23123 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
23125 // SHLD/SHRD instructions have lower register pressure, but on some
23126 // platforms they have higher latency than the equivalent
23127 // series of shifts/or that would otherwise be generated.
23128 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
23129 // have higher latencies and we are not optimizing for size.
23130 if (!OptForSize && Subtarget->isSHLDSlow())
23133 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
23135 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
23137 if (!N0.hasOneUse() || !N1.hasOneUse())
23140 SDValue ShAmt0 = N0.getOperand(1);
23141 if (ShAmt0.getValueType() != MVT::i8)
23143 SDValue ShAmt1 = N1.getOperand(1);
23144 if (ShAmt1.getValueType() != MVT::i8)
23146 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
23147 ShAmt0 = ShAmt0.getOperand(0);
23148 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
23149 ShAmt1 = ShAmt1.getOperand(0);
23152 unsigned Opc = X86ISD::SHLD;
23153 SDValue Op0 = N0.getOperand(0);
23154 SDValue Op1 = N1.getOperand(0);
23155 if (ShAmt0.getOpcode() == ISD::SUB) {
23156 Opc = X86ISD::SHRD;
23157 std::swap(Op0, Op1);
23158 std::swap(ShAmt0, ShAmt1);
23161 unsigned Bits = VT.getSizeInBits();
23162 if (ShAmt1.getOpcode() == ISD::SUB) {
23163 SDValue Sum = ShAmt1.getOperand(0);
23164 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
23165 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
23166 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
23167 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
23168 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
23169 return DAG.getNode(Opc, DL, VT,
23171 DAG.getNode(ISD::TRUNCATE, DL,
23174 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
23175 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
23177 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
23178 return DAG.getNode(Opc, DL, VT,
23179 N0.getOperand(0), N1.getOperand(0),
23180 DAG.getNode(ISD::TRUNCATE, DL,
23187 // Generate NEG and CMOV for integer abs.
23188 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
23189 EVT VT = N->getValueType(0);
23191 // Since X86 does not have CMOV for 8-bit integer, we don't convert
23192 // 8-bit integer abs to NEG and CMOV.
23193 if (VT.isInteger() && VT.getSizeInBits() == 8)
23196 SDValue N0 = N->getOperand(0);
23197 SDValue N1 = N->getOperand(1);
23200 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
23201 // and change it to SUB and CMOV.
23202 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
23203 N0.getOpcode() == ISD::ADD &&
23204 N0.getOperand(1) == N1 &&
23205 N1.getOpcode() == ISD::SRA &&
23206 N1.getOperand(0) == N0.getOperand(0))
23207 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
23208 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
23209 // Generate SUB & CMOV.
23210 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
23211 DAG.getConstant(0, DL, VT), N0.getOperand(0));
23213 SDValue Ops[] = { N0.getOperand(0), Neg,
23214 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
23215 SDValue(Neg.getNode(), 1) };
23216 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
23221 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
23222 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
23223 TargetLowering::DAGCombinerInfo &DCI,
23224 const X86Subtarget *Subtarget) {
23225 if (DCI.isBeforeLegalizeOps())
23228 if (Subtarget->hasCMov()) {
23229 SDValue RV = performIntegerAbsCombine(N, DAG);
23237 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
23238 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
23239 TargetLowering::DAGCombinerInfo &DCI,
23240 const X86Subtarget *Subtarget) {
23241 LoadSDNode *Ld = cast<LoadSDNode>(N);
23242 EVT RegVT = Ld->getValueType(0);
23243 EVT MemVT = Ld->getMemoryVT();
23245 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23247 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
23248 // into two 16-byte operations.
23249 ISD::LoadExtType Ext = Ld->getExtensionType();
23250 unsigned Alignment = Ld->getAlignment();
23251 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
23252 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
23253 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
23254 unsigned NumElems = RegVT.getVectorNumElements();
23258 SDValue Ptr = Ld->getBasePtr();
23259 SDValue Increment = DAG.getConstant(16, dl, TLI.getPointerTy());
23261 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
23263 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23264 Ld->getPointerInfo(), Ld->isVolatile(),
23265 Ld->isNonTemporal(), Ld->isInvariant(),
23267 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23268 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23269 Ld->getPointerInfo(), Ld->isVolatile(),
23270 Ld->isNonTemporal(), Ld->isInvariant(),
23271 std::min(16U, Alignment));
23272 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
23274 Load2.getValue(1));
23276 SDValue NewVec = DAG.getUNDEF(RegVT);
23277 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
23278 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
23279 return DCI.CombineTo(N, NewVec, TF, true);
23285 /// PerformMLOADCombine - Resolve extending loads
23286 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
23287 TargetLowering::DAGCombinerInfo &DCI,
23288 const X86Subtarget *Subtarget) {
23289 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
23290 if (Mld->getExtensionType() != ISD::SEXTLOAD)
23293 EVT VT = Mld->getValueType(0);
23294 unsigned NumElems = VT.getVectorNumElements();
23295 EVT LdVT = Mld->getMemoryVT();
23298 assert(LdVT != VT && "Cannot extend to the same type");
23299 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
23300 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
23301 // From, To sizes and ElemCount must be pow of two
23302 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23303 "Unexpected size for extending masked load");
23305 unsigned SizeRatio = ToSz / FromSz;
23306 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
23308 // Create a type on which we perform the shuffle
23309 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23310 LdVT.getScalarType(), NumElems*SizeRatio);
23311 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23313 // Convert Src0 value
23314 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
23315 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
23316 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23317 for (unsigned i = 0; i != NumElems; ++i)
23318 ShuffleVec[i] = i * SizeRatio;
23320 // Can't shuffle using an illegal type.
23321 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23322 && "WideVecVT should be legal");
23323 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
23324 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
23326 // Prepare the new mask
23328 SDValue Mask = Mld->getMask();
23329 if (Mask.getValueType() == VT) {
23330 // Mask and original value have the same type
23331 NewMask = DAG.getBitcast(WideVecVT, Mask);
23332 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23333 for (unsigned i = 0; i != NumElems; ++i)
23334 ShuffleVec[i] = i * SizeRatio;
23335 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23336 ShuffleVec[i] = NumElems*SizeRatio;
23337 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23338 DAG.getConstant(0, dl, WideVecVT),
23342 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23343 unsigned WidenNumElts = NumElems*SizeRatio;
23344 unsigned MaskNumElts = VT.getVectorNumElements();
23345 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23348 unsigned NumConcat = WidenNumElts / MaskNumElts;
23349 SmallVector<SDValue, 16> Ops(NumConcat);
23350 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23352 for (unsigned i = 1; i != NumConcat; ++i)
23355 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23358 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
23359 Mld->getBasePtr(), NewMask, WideSrc0,
23360 Mld->getMemoryVT(), Mld->getMemOperand(),
23362 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
23363 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
23366 /// PerformMSTORECombine - Resolve truncating stores
23367 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
23368 const X86Subtarget *Subtarget) {
23369 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
23370 if (!Mst->isTruncatingStore())
23373 EVT VT = Mst->getValue().getValueType();
23374 unsigned NumElems = VT.getVectorNumElements();
23375 EVT StVT = Mst->getMemoryVT();
23378 assert(StVT != VT && "Cannot truncate to the same type");
23379 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23380 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23382 // From, To sizes and ElemCount must be pow of two
23383 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23384 "Unexpected size for truncating masked store");
23385 // We are going to use the original vector elt for storing.
23386 // Accumulated smaller vector elements must be a multiple of the store size.
23387 assert (((NumElems * FromSz) % ToSz) == 0 &&
23388 "Unexpected ratio for truncating masked store");
23390 unsigned SizeRatio = FromSz / ToSz;
23391 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23393 // Create a type on which we perform the shuffle
23394 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23395 StVT.getScalarType(), NumElems*SizeRatio);
23397 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23399 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
23400 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23401 for (unsigned i = 0; i != NumElems; ++i)
23402 ShuffleVec[i] = i * SizeRatio;
23404 // Can't shuffle using an illegal type.
23405 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23406 && "WideVecVT should be legal");
23408 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23409 DAG.getUNDEF(WideVecVT),
23413 SDValue Mask = Mst->getMask();
23414 if (Mask.getValueType() == VT) {
23415 // Mask and original value have the same type
23416 NewMask = DAG.getBitcast(WideVecVT, Mask);
23417 for (unsigned i = 0; i != NumElems; ++i)
23418 ShuffleVec[i] = i * SizeRatio;
23419 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23420 ShuffleVec[i] = NumElems*SizeRatio;
23421 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23422 DAG.getConstant(0, dl, WideVecVT),
23426 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23427 unsigned WidenNumElts = NumElems*SizeRatio;
23428 unsigned MaskNumElts = VT.getVectorNumElements();
23429 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23432 unsigned NumConcat = WidenNumElts / MaskNumElts;
23433 SmallVector<SDValue, 16> Ops(NumConcat);
23434 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23436 for (unsigned i = 1; i != NumConcat; ++i)
23439 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23442 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
23443 NewMask, StVT, Mst->getMemOperand(), false);
23445 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
23446 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
23447 const X86Subtarget *Subtarget) {
23448 StoreSDNode *St = cast<StoreSDNode>(N);
23449 EVT VT = St->getValue().getValueType();
23450 EVT StVT = St->getMemoryVT();
23452 SDValue StoredVal = St->getOperand(1);
23453 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23455 // If we are saving a concatenation of two XMM registers and 32-byte stores
23456 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
23457 unsigned Alignment = St->getAlignment();
23458 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
23459 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
23460 StVT == VT && !IsAligned) {
23461 unsigned NumElems = VT.getVectorNumElements();
23465 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
23466 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
23468 SDValue Stride = DAG.getConstant(16, dl, TLI.getPointerTy());
23469 SDValue Ptr0 = St->getBasePtr();
23470 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
23472 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
23473 St->getPointerInfo(), St->isVolatile(),
23474 St->isNonTemporal(), Alignment);
23475 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
23476 St->getPointerInfo(), St->isVolatile(),
23477 St->isNonTemporal(),
23478 std::min(16U, Alignment));
23479 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
23482 // Optimize trunc store (of multiple scalars) to shuffle and store.
23483 // First, pack all of the elements in one place. Next, store to memory
23484 // in fewer chunks.
23485 if (St->isTruncatingStore() && VT.isVector()) {
23486 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23487 unsigned NumElems = VT.getVectorNumElements();
23488 assert(StVT != VT && "Cannot truncate to the same type");
23489 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23490 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23492 // From, To sizes and ElemCount must be pow of two
23493 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
23494 // We are going to use the original vector elt for storing.
23495 // Accumulated smaller vector elements must be a multiple of the store size.
23496 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
23498 unsigned SizeRatio = FromSz / ToSz;
23500 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23502 // Create a type on which we perform the shuffle
23503 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23504 StVT.getScalarType(), NumElems*SizeRatio);
23506 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23508 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
23509 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
23510 for (unsigned i = 0; i != NumElems; ++i)
23511 ShuffleVec[i] = i * SizeRatio;
23513 // Can't shuffle using an illegal type.
23514 if (!TLI.isTypeLegal(WideVecVT))
23517 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23518 DAG.getUNDEF(WideVecVT),
23520 // At this point all of the data is stored at the bottom of the
23521 // register. We now need to save it to mem.
23523 // Find the largest store unit
23524 MVT StoreType = MVT::i8;
23525 for (MVT Tp : MVT::integer_valuetypes()) {
23526 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
23530 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
23531 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
23532 (64 <= NumElems * ToSz))
23533 StoreType = MVT::f64;
23535 // Bitcast the original vector into a vector of store-size units
23536 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
23537 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
23538 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
23539 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
23540 SmallVector<SDValue, 8> Chains;
23541 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, dl,
23542 TLI.getPointerTy());
23543 SDValue Ptr = St->getBasePtr();
23545 // Perform one or more big stores into memory.
23546 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
23547 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
23548 StoreType, ShuffWide,
23549 DAG.getIntPtrConstant(i, dl));
23550 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
23551 St->getPointerInfo(), St->isVolatile(),
23552 St->isNonTemporal(), St->getAlignment());
23553 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23554 Chains.push_back(Ch);
23557 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
23560 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
23561 // the FP state in cases where an emms may be missing.
23562 // A preferable solution to the general problem is to figure out the right
23563 // places to insert EMMS. This qualifies as a quick hack.
23565 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
23566 if (VT.getSizeInBits() != 64)
23569 const Function *F = DAG.getMachineFunction().getFunction();
23570 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
23572 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
23573 if ((VT.isVector() ||
23574 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23575 isa<LoadSDNode>(St->getValue()) &&
23576 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23577 St->getChain().hasOneUse() && !St->isVolatile()) {
23578 SDNode* LdVal = St->getValue().getNode();
23579 LoadSDNode *Ld = nullptr;
23580 int TokenFactorIndex = -1;
23581 SmallVector<SDValue, 8> Ops;
23582 SDNode* ChainVal = St->getChain().getNode();
23583 // Must be a store of a load. We currently handle two cases: the load
23584 // is a direct child, and it's under an intervening TokenFactor. It is
23585 // possible to dig deeper under nested TokenFactors.
23586 if (ChainVal == LdVal)
23587 Ld = cast<LoadSDNode>(St->getChain());
23588 else if (St->getValue().hasOneUse() &&
23589 ChainVal->getOpcode() == ISD::TokenFactor) {
23590 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23591 if (ChainVal->getOperand(i).getNode() == LdVal) {
23592 TokenFactorIndex = i;
23593 Ld = cast<LoadSDNode>(St->getValue());
23595 Ops.push_back(ChainVal->getOperand(i));
23599 if (!Ld || !ISD::isNormalLoad(Ld))
23602 // If this is not the MMX case, i.e. we are just turning i64 load/store
23603 // into f64 load/store, avoid the transformation if there are multiple
23604 // uses of the loaded value.
23605 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23610 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23611 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23613 if (Subtarget->is64Bit() || F64IsLegal) {
23614 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23615 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23616 Ld->getPointerInfo(), Ld->isVolatile(),
23617 Ld->isNonTemporal(), Ld->isInvariant(),
23618 Ld->getAlignment());
23619 SDValue NewChain = NewLd.getValue(1);
23620 if (TokenFactorIndex != -1) {
23621 Ops.push_back(NewChain);
23622 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23624 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23625 St->getPointerInfo(),
23626 St->isVolatile(), St->isNonTemporal(),
23627 St->getAlignment());
23630 // Otherwise, lower to two pairs of 32-bit loads / stores.
23631 SDValue LoAddr = Ld->getBasePtr();
23632 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23633 DAG.getConstant(4, LdDL, MVT::i32));
23635 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23636 Ld->getPointerInfo(),
23637 Ld->isVolatile(), Ld->isNonTemporal(),
23638 Ld->isInvariant(), Ld->getAlignment());
23639 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23640 Ld->getPointerInfo().getWithOffset(4),
23641 Ld->isVolatile(), Ld->isNonTemporal(),
23643 MinAlign(Ld->getAlignment(), 4));
23645 SDValue NewChain = LoLd.getValue(1);
23646 if (TokenFactorIndex != -1) {
23647 Ops.push_back(LoLd);
23648 Ops.push_back(HiLd);
23649 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23652 LoAddr = St->getBasePtr();
23653 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23654 DAG.getConstant(4, StDL, MVT::i32));
23656 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23657 St->getPointerInfo(),
23658 St->isVolatile(), St->isNonTemporal(),
23659 St->getAlignment());
23660 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23661 St->getPointerInfo().getWithOffset(4),
23663 St->isNonTemporal(),
23664 MinAlign(St->getAlignment(), 4));
23665 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23668 // This is similar to the above case, but here we handle a scalar 64-bit
23669 // integer store that is extracted from a vector on a 32-bit target.
23670 // If we have SSE2, then we can treat it like a floating-point double
23671 // to get past legalization. The execution dependencies fixup pass will
23672 // choose the optimal machine instruction for the store if this really is
23673 // an integer or v2f32 rather than an f64.
23674 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
23675 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
23676 SDValue OldExtract = St->getOperand(1);
23677 SDValue ExtOp0 = OldExtract.getOperand(0);
23678 unsigned VecSize = ExtOp0.getValueSizeInBits();
23679 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
23680 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
23681 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
23682 BitCast, OldExtract.getOperand(1));
23683 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
23684 St->getPointerInfo(), St->isVolatile(),
23685 St->isNonTemporal(), St->getAlignment());
23691 /// Return 'true' if this vector operation is "horizontal"
23692 /// and return the operands for the horizontal operation in LHS and RHS. A
23693 /// horizontal operation performs the binary operation on successive elements
23694 /// of its first operand, then on successive elements of its second operand,
23695 /// returning the resulting values in a vector. For example, if
23696 /// A = < float a0, float a1, float a2, float a3 >
23698 /// B = < float b0, float b1, float b2, float b3 >
23699 /// then the result of doing a horizontal operation on A and B is
23700 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23701 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23702 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23703 /// set to A, RHS to B, and the routine returns 'true'.
23704 /// Note that the binary operation should have the property that if one of the
23705 /// operands is UNDEF then the result is UNDEF.
23706 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23707 // Look for the following pattern: if
23708 // A = < float a0, float a1, float a2, float a3 >
23709 // B = < float b0, float b1, float b2, float b3 >
23711 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23712 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23713 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23714 // which is A horizontal-op B.
23716 // At least one of the operands should be a vector shuffle.
23717 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23718 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23721 MVT VT = LHS.getSimpleValueType();
23723 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23724 "Unsupported vector type for horizontal add/sub");
23726 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23727 // operate independently on 128-bit lanes.
23728 unsigned NumElts = VT.getVectorNumElements();
23729 unsigned NumLanes = VT.getSizeInBits()/128;
23730 unsigned NumLaneElts = NumElts / NumLanes;
23731 assert((NumLaneElts % 2 == 0) &&
23732 "Vector type should have an even number of elements in each lane");
23733 unsigned HalfLaneElts = NumLaneElts/2;
23735 // View LHS in the form
23736 // LHS = VECTOR_SHUFFLE A, B, LMask
23737 // If LHS is not a shuffle then pretend it is the shuffle
23738 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23739 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23742 SmallVector<int, 16> LMask(NumElts);
23743 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23744 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23745 A = LHS.getOperand(0);
23746 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23747 B = LHS.getOperand(1);
23748 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23749 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23751 if (LHS.getOpcode() != ISD::UNDEF)
23753 for (unsigned i = 0; i != NumElts; ++i)
23757 // Likewise, view RHS in the form
23758 // RHS = VECTOR_SHUFFLE C, D, RMask
23760 SmallVector<int, 16> RMask(NumElts);
23761 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23762 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23763 C = RHS.getOperand(0);
23764 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23765 D = RHS.getOperand(1);
23766 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23767 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23769 if (RHS.getOpcode() != ISD::UNDEF)
23771 for (unsigned i = 0; i != NumElts; ++i)
23775 // Check that the shuffles are both shuffling the same vectors.
23776 if (!(A == C && B == D) && !(A == D && B == C))
23779 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23780 if (!A.getNode() && !B.getNode())
23783 // If A and B occur in reverse order in RHS, then "swap" them (which means
23784 // rewriting the mask).
23786 ShuffleVectorSDNode::commuteMask(RMask);
23788 // At this point LHS and RHS are equivalent to
23789 // LHS = VECTOR_SHUFFLE A, B, LMask
23790 // RHS = VECTOR_SHUFFLE A, B, RMask
23791 // Check that the masks correspond to performing a horizontal operation.
23792 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23793 for (unsigned i = 0; i != NumLaneElts; ++i) {
23794 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23796 // Ignore any UNDEF components.
23797 if (LIdx < 0 || RIdx < 0 ||
23798 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23799 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23802 // Check that successive elements are being operated on. If not, this is
23803 // not a horizontal operation.
23804 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23805 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23806 if (!(LIdx == Index && RIdx == Index + 1) &&
23807 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23812 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23813 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23817 /// Do target-specific dag combines on floating point adds.
23818 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23819 const X86Subtarget *Subtarget) {
23820 EVT VT = N->getValueType(0);
23821 SDValue LHS = N->getOperand(0);
23822 SDValue RHS = N->getOperand(1);
23824 // Try to synthesize horizontal adds from adds of shuffles.
23825 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23826 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23827 isHorizontalBinOp(LHS, RHS, true))
23828 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23832 /// Do target-specific dag combines on floating point subs.
23833 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23834 const X86Subtarget *Subtarget) {
23835 EVT VT = N->getValueType(0);
23836 SDValue LHS = N->getOperand(0);
23837 SDValue RHS = N->getOperand(1);
23839 // Try to synthesize horizontal subs from subs of shuffles.
23840 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23841 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23842 isHorizontalBinOp(LHS, RHS, false))
23843 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23847 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
23848 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23849 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23851 // F[X]OR(0.0, x) -> x
23852 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23853 if (C->getValueAPF().isPosZero())
23854 return N->getOperand(1);
23856 // F[X]OR(x, 0.0) -> x
23857 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23858 if (C->getValueAPF().isPosZero())
23859 return N->getOperand(0);
23863 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
23864 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23865 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23867 // Only perform optimizations if UnsafeMath is used.
23868 if (!DAG.getTarget().Options.UnsafeFPMath)
23871 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23872 // into FMINC and FMAXC, which are Commutative operations.
23873 unsigned NewOp = 0;
23874 switch (N->getOpcode()) {
23875 default: llvm_unreachable("unknown opcode");
23876 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23877 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23880 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23881 N->getOperand(0), N->getOperand(1));
23884 /// Do target-specific dag combines on X86ISD::FAND nodes.
23885 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23886 // FAND(0.0, x) -> 0.0
23887 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23888 if (C->getValueAPF().isPosZero())
23889 return N->getOperand(0);
23891 // FAND(x, 0.0) -> 0.0
23892 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23893 if (C->getValueAPF().isPosZero())
23894 return N->getOperand(1);
23899 /// Do target-specific dag combines on X86ISD::FANDN nodes
23900 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23901 // FANDN(0.0, x) -> x
23902 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23903 if (C->getValueAPF().isPosZero())
23904 return N->getOperand(1);
23906 // FANDN(x, 0.0) -> 0.0
23907 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23908 if (C->getValueAPF().isPosZero())
23909 return N->getOperand(1);
23914 static SDValue PerformBTCombine(SDNode *N,
23916 TargetLowering::DAGCombinerInfo &DCI) {
23917 // BT ignores high bits in the bit index operand.
23918 SDValue Op1 = N->getOperand(1);
23919 if (Op1.hasOneUse()) {
23920 unsigned BitWidth = Op1.getValueSizeInBits();
23921 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23922 APInt KnownZero, KnownOne;
23923 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23924 !DCI.isBeforeLegalizeOps());
23925 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23926 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23927 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23928 DCI.CommitTargetLoweringOpt(TLO);
23933 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23934 SDValue Op = N->getOperand(0);
23935 if (Op.getOpcode() == ISD::BITCAST)
23936 Op = Op.getOperand(0);
23937 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23938 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23939 VT.getVectorElementType().getSizeInBits() ==
23940 OpVT.getVectorElementType().getSizeInBits()) {
23941 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23946 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23947 const X86Subtarget *Subtarget) {
23948 EVT VT = N->getValueType(0);
23949 if (!VT.isVector())
23952 SDValue N0 = N->getOperand(0);
23953 SDValue N1 = N->getOperand(1);
23954 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23957 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23958 // both SSE and AVX2 since there is no sign-extended shift right
23959 // operation on a vector with 64-bit elements.
23960 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23961 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23962 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23963 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23964 SDValue N00 = N0.getOperand(0);
23966 // EXTLOAD has a better solution on AVX2,
23967 // it may be replaced with X86ISD::VSEXT node.
23968 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23969 if (!ISD::isNormalLoad(N00.getNode()))
23972 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23973 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23975 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23981 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23982 TargetLowering::DAGCombinerInfo &DCI,
23983 const X86Subtarget *Subtarget) {
23984 SDValue N0 = N->getOperand(0);
23985 EVT VT = N->getValueType(0);
23986 EVT SVT = VT.getScalarType();
23987 EVT InVT = N0->getValueType(0);
23988 EVT InSVT = InVT.getScalarType();
23991 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
23992 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
23993 // This exposes the sext to the sdivrem lowering, so that it directly extends
23994 // from AH (which we otherwise need to do contortions to access).
23995 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
23996 InVT == MVT::i8 && VT == MVT::i32) {
23997 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23998 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
23999 N0.getOperand(0), N0.getOperand(1));
24000 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24001 return R.getValue(1);
24004 if (!DCI.isBeforeLegalizeOps()) {
24005 if (N0.getValueType() == MVT::i1) {
24006 SDValue Zero = DAG.getConstant(0, DL, VT);
24008 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
24009 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
24014 if (VT.isVector()) {
24015 auto ExtendToVec128 = [&DAG](SDLoc DL, SDValue N) {
24016 EVT InVT = N->getValueType(0);
24017 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
24018 128 / InVT.getScalarSizeInBits());
24019 SmallVector<SDValue, 8> Opnds(128 / InVT.getSizeInBits(),
24020 DAG.getUNDEF(InVT));
24022 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
24025 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
24026 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
24027 if (VT.getSizeInBits() == 128 &&
24028 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24029 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24030 SDValue ExOp = ExtendToVec128(DL, N0);
24031 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
24034 // On pre-AVX2 targets, split into 128-bit nodes of
24035 // ISD::SIGN_EXTEND_VECTOR_INREG.
24036 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
24037 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24038 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24039 unsigned NumVecs = VT.getSizeInBits() / 128;
24040 unsigned NumSubElts = 128 / SVT.getSizeInBits();
24041 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
24042 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
24044 SmallVector<SDValue, 8> Opnds;
24045 for (unsigned i = 0, Offset = 0; i != NumVecs;
24046 ++i, Offset += NumSubElts) {
24047 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
24048 DAG.getIntPtrConstant(Offset, DL));
24049 SrcVec = ExtendToVec128(DL, SrcVec);
24050 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
24051 Opnds.push_back(SrcVec);
24053 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
24057 if (!Subtarget->hasFp256())
24060 if (VT.isVector() && VT.getSizeInBits() == 256) {
24061 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
24069 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
24070 const X86Subtarget* Subtarget) {
24072 EVT VT = N->getValueType(0);
24074 // Let legalize expand this if it isn't a legal type yet.
24075 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
24078 EVT ScalarVT = VT.getScalarType();
24079 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
24080 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
24083 SDValue A = N->getOperand(0);
24084 SDValue B = N->getOperand(1);
24085 SDValue C = N->getOperand(2);
24087 bool NegA = (A.getOpcode() == ISD::FNEG);
24088 bool NegB = (B.getOpcode() == ISD::FNEG);
24089 bool NegC = (C.getOpcode() == ISD::FNEG);
24091 // Negative multiplication when NegA xor NegB
24092 bool NegMul = (NegA != NegB);
24094 A = A.getOperand(0);
24096 B = B.getOperand(0);
24098 C = C.getOperand(0);
24102 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
24104 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
24106 return DAG.getNode(Opcode, dl, VT, A, B, C);
24109 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
24110 TargetLowering::DAGCombinerInfo &DCI,
24111 const X86Subtarget *Subtarget) {
24112 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
24113 // (and (i32 x86isd::setcc_carry), 1)
24114 // This eliminates the zext. This transformation is necessary because
24115 // ISD::SETCC is always legalized to i8.
24117 SDValue N0 = N->getOperand(0);
24118 EVT VT = N->getValueType(0);
24120 if (N0.getOpcode() == ISD::AND &&
24122 N0.getOperand(0).hasOneUse()) {
24123 SDValue N00 = N0.getOperand(0);
24124 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24125 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24126 if (!C || C->getZExtValue() != 1)
24128 return DAG.getNode(ISD::AND, dl, VT,
24129 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24130 N00.getOperand(0), N00.getOperand(1)),
24131 DAG.getConstant(1, dl, VT));
24135 if (N0.getOpcode() == ISD::TRUNCATE &&
24137 N0.getOperand(0).hasOneUse()) {
24138 SDValue N00 = N0.getOperand(0);
24139 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24140 return DAG.getNode(ISD::AND, dl, VT,
24141 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24142 N00.getOperand(0), N00.getOperand(1)),
24143 DAG.getConstant(1, dl, VT));
24146 if (VT.is256BitVector()) {
24147 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
24152 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
24153 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
24154 // This exposes the zext to the udivrem lowering, so that it directly extends
24155 // from AH (which we otherwise need to do contortions to access).
24156 if (N0.getOpcode() == ISD::UDIVREM &&
24157 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
24158 (VT == MVT::i32 || VT == MVT::i64)) {
24159 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24160 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
24161 N0.getOperand(0), N0.getOperand(1));
24162 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24163 return R.getValue(1);
24169 // Optimize x == -y --> x+y == 0
24170 // x != -y --> x+y != 0
24171 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
24172 const X86Subtarget* Subtarget) {
24173 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
24174 SDValue LHS = N->getOperand(0);
24175 SDValue RHS = N->getOperand(1);
24176 EVT VT = N->getValueType(0);
24179 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
24180 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
24181 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
24182 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
24183 LHS.getOperand(1));
24184 return DAG.getSetCC(DL, N->getValueType(0), addV,
24185 DAG.getConstant(0, DL, addV.getValueType()), CC);
24187 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
24188 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
24189 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
24190 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
24191 RHS.getOperand(1));
24192 return DAG.getSetCC(DL, N->getValueType(0), addV,
24193 DAG.getConstant(0, DL, addV.getValueType()), CC);
24196 if (VT.getScalarType() == MVT::i1 &&
24197 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
24199 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
24200 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
24201 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
24203 if (!IsSEXT0 || !IsVZero1) {
24204 // Swap the operands and update the condition code.
24205 std::swap(LHS, RHS);
24206 CC = ISD::getSetCCSwappedOperands(CC);
24208 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
24209 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
24210 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
24213 if (IsSEXT0 && IsVZero1) {
24214 assert(VT == LHS.getOperand(0).getValueType() &&
24215 "Uexpected operand type");
24216 if (CC == ISD::SETGT)
24217 return DAG.getConstant(0, DL, VT);
24218 if (CC == ISD::SETLE)
24219 return DAG.getConstant(1, DL, VT);
24220 if (CC == ISD::SETEQ || CC == ISD::SETGE)
24221 return DAG.getNOT(DL, LHS.getOperand(0), VT);
24223 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
24224 "Unexpected condition code!");
24225 return LHS.getOperand(0);
24232 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
24233 SelectionDAG &DAG) {
24235 MVT VT = Load->getSimpleValueType(0);
24236 MVT EVT = VT.getVectorElementType();
24237 SDValue Addr = Load->getOperand(1);
24238 SDValue NewAddr = DAG.getNode(
24239 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
24240 DAG.getConstant(Index * EVT.getStoreSize(), dl,
24241 Addr.getSimpleValueType()));
24244 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
24245 DAG.getMachineFunction().getMachineMemOperand(
24246 Load->getMemOperand(), 0, EVT.getStoreSize()));
24250 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
24251 const X86Subtarget *Subtarget) {
24253 MVT VT = N->getOperand(1)->getSimpleValueType(0);
24254 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
24255 "X86insertps is only defined for v4x32");
24257 SDValue Ld = N->getOperand(1);
24258 if (MayFoldLoad(Ld)) {
24259 // Extract the countS bits from the immediate so we can get the proper
24260 // address when narrowing the vector load to a specific element.
24261 // When the second source op is a memory address, insertps doesn't use
24262 // countS and just gets an f32 from that address.
24263 unsigned DestIndex =
24264 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
24266 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
24268 // Create this as a scalar to vector to match the instruction pattern.
24269 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
24270 // countS bits are ignored when loading from memory on insertps, which
24271 // means we don't need to explicitly set them to 0.
24272 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
24273 LoadScalarToVector, N->getOperand(2));
24278 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
24279 SDValue V0 = N->getOperand(0);
24280 SDValue V1 = N->getOperand(1);
24282 EVT VT = N->getValueType(0);
24284 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
24285 // operands and changing the mask to 1. This saves us a bunch of
24286 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
24287 // x86InstrInfo knows how to commute this back after instruction selection
24288 // if it would help register allocation.
24290 // TODO: If optimizing for size or a processor that doesn't suffer from
24291 // partial register update stalls, this should be transformed into a MOVSD
24292 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
24294 if (VT == MVT::v2f64)
24295 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
24296 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
24297 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
24298 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
24304 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
24305 // as "sbb reg,reg", since it can be extended without zext and produces
24306 // an all-ones bit which is more useful than 0/1 in some cases.
24307 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
24310 return DAG.getNode(ISD::AND, DL, VT,
24311 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
24312 DAG.getConstant(X86::COND_B, DL, MVT::i8),
24314 DAG.getConstant(1, DL, VT));
24315 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
24316 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
24317 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
24318 DAG.getConstant(X86::COND_B, DL, MVT::i8),
24322 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
24323 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
24324 TargetLowering::DAGCombinerInfo &DCI,
24325 const X86Subtarget *Subtarget) {
24327 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
24328 SDValue EFLAGS = N->getOperand(1);
24330 if (CC == X86::COND_A) {
24331 // Try to convert COND_A into COND_B in an attempt to facilitate
24332 // materializing "setb reg".
24334 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
24335 // cannot take an immediate as its first operand.
24337 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
24338 EFLAGS.getValueType().isInteger() &&
24339 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
24340 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
24341 EFLAGS.getNode()->getVTList(),
24342 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
24343 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
24344 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
24348 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
24349 // a zext and produces an all-ones bit which is more useful than 0/1 in some
24351 if (CC == X86::COND_B)
24352 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
24356 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24357 if (Flags.getNode()) {
24358 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24359 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
24365 // Optimize branch condition evaluation.
24367 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
24368 TargetLowering::DAGCombinerInfo &DCI,
24369 const X86Subtarget *Subtarget) {
24371 SDValue Chain = N->getOperand(0);
24372 SDValue Dest = N->getOperand(1);
24373 SDValue EFLAGS = N->getOperand(3);
24374 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
24378 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24379 if (Flags.getNode()) {
24380 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24381 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
24388 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
24389 SelectionDAG &DAG) {
24390 // Take advantage of vector comparisons producing 0 or -1 in each lane to
24391 // optimize away operation when it's from a constant.
24393 // The general transformation is:
24394 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
24395 // AND(VECTOR_CMP(x,y), constant2)
24396 // constant2 = UNARYOP(constant)
24398 // Early exit if this isn't a vector operation, the operand of the
24399 // unary operation isn't a bitwise AND, or if the sizes of the operations
24400 // aren't the same.
24401 EVT VT = N->getValueType(0);
24402 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
24403 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
24404 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
24407 // Now check that the other operand of the AND is a constant. We could
24408 // make the transformation for non-constant splats as well, but it's unclear
24409 // that would be a benefit as it would not eliminate any operations, just
24410 // perform one more step in scalar code before moving to the vector unit.
24411 if (BuildVectorSDNode *BV =
24412 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
24413 // Bail out if the vector isn't a constant.
24414 if (!BV->isConstant())
24417 // Everything checks out. Build up the new and improved node.
24419 EVT IntVT = BV->getValueType(0);
24420 // Create a new constant of the appropriate type for the transformed
24422 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
24423 // The AND node needs bitcasts to/from an integer vector type around it.
24424 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
24425 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
24426 N->getOperand(0)->getOperand(0), MaskConst);
24427 SDValue Res = DAG.getBitcast(VT, NewAnd);
24434 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
24435 const X86Subtarget *Subtarget) {
24436 // First try to optimize away the conversion entirely when it's
24437 // conditionally from a constant. Vectors only.
24438 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
24439 if (Res != SDValue())
24442 // Now move on to more general possibilities.
24443 SDValue Op0 = N->getOperand(0);
24444 EVT InVT = Op0->getValueType(0);
24446 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
24447 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
24449 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
24450 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
24451 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
24454 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
24455 // a 32-bit target where SSE doesn't support i64->FP operations.
24456 if (Op0.getOpcode() == ISD::LOAD) {
24457 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
24458 EVT VT = Ld->getValueType(0);
24460 // This transformation is not supported if the result type is f16
24461 if (N->getValueType(0) == MVT::f16)
24464 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
24465 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
24466 !Subtarget->is64Bit() && VT == MVT::i64) {
24467 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
24468 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
24469 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
24476 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
24477 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
24478 X86TargetLowering::DAGCombinerInfo &DCI) {
24479 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
24480 // the result is either zero or one (depending on the input carry bit).
24481 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
24482 if (X86::isZeroNode(N->getOperand(0)) &&
24483 X86::isZeroNode(N->getOperand(1)) &&
24484 // We don't have a good way to replace an EFLAGS use, so only do this when
24486 SDValue(N, 1).use_empty()) {
24488 EVT VT = N->getValueType(0);
24489 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
24490 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
24491 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
24492 DAG.getConstant(X86::COND_B, DL,
24495 DAG.getConstant(1, DL, VT));
24496 return DCI.CombineTo(N, Res1, CarryOut);
24502 // fold (add Y, (sete X, 0)) -> adc 0, Y
24503 // (add Y, (setne X, 0)) -> sbb -1, Y
24504 // (sub (sete X, 0), Y) -> sbb 0, Y
24505 // (sub (setne X, 0), Y) -> adc -1, Y
24506 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
24509 // Look through ZExts.
24510 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
24511 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
24514 SDValue SetCC = Ext.getOperand(0);
24515 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
24518 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
24519 if (CC != X86::COND_E && CC != X86::COND_NE)
24522 SDValue Cmp = SetCC.getOperand(1);
24523 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
24524 !X86::isZeroNode(Cmp.getOperand(1)) ||
24525 !Cmp.getOperand(0).getValueType().isInteger())
24528 SDValue CmpOp0 = Cmp.getOperand(0);
24529 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
24530 DAG.getConstant(1, DL, CmpOp0.getValueType()));
24532 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
24533 if (CC == X86::COND_NE)
24534 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
24535 DL, OtherVal.getValueType(), OtherVal,
24536 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
24538 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
24539 DL, OtherVal.getValueType(), OtherVal,
24540 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
24543 /// PerformADDCombine - Do target-specific dag combines on integer adds.
24544 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
24545 const X86Subtarget *Subtarget) {
24546 EVT VT = N->getValueType(0);
24547 SDValue Op0 = N->getOperand(0);
24548 SDValue Op1 = N->getOperand(1);
24550 // Try to synthesize horizontal adds from adds of shuffles.
24551 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24552 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24553 isHorizontalBinOp(Op0, Op1, true))
24554 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
24556 return OptimizeConditionalInDecrement(N, DAG);
24559 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
24560 const X86Subtarget *Subtarget) {
24561 SDValue Op0 = N->getOperand(0);
24562 SDValue Op1 = N->getOperand(1);
24564 // X86 can't encode an immediate LHS of a sub. See if we can push the
24565 // negation into a preceding instruction.
24566 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
24567 // If the RHS of the sub is a XOR with one use and a constant, invert the
24568 // immediate. Then add one to the LHS of the sub so we can turn
24569 // X-Y -> X+~Y+1, saving one register.
24570 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
24571 isa<ConstantSDNode>(Op1.getOperand(1))) {
24572 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
24573 EVT VT = Op0.getValueType();
24574 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
24576 DAG.getConstant(~XorC, SDLoc(Op1), VT));
24577 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
24578 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
24582 // Try to synthesize horizontal adds from adds of shuffles.
24583 EVT VT = N->getValueType(0);
24584 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24585 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24586 isHorizontalBinOp(Op0, Op1, true))
24587 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
24589 return OptimizeConditionalInDecrement(N, DAG);
24592 /// performVZEXTCombine - Performs build vector combines
24593 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
24594 TargetLowering::DAGCombinerInfo &DCI,
24595 const X86Subtarget *Subtarget) {
24597 MVT VT = N->getSimpleValueType(0);
24598 SDValue Op = N->getOperand(0);
24599 MVT OpVT = Op.getSimpleValueType();
24600 MVT OpEltVT = OpVT.getVectorElementType();
24601 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
24603 // (vzext (bitcast (vzext (x)) -> (vzext x)
24605 while (V.getOpcode() == ISD::BITCAST)
24606 V = V.getOperand(0);
24608 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
24609 MVT InnerVT = V.getSimpleValueType();
24610 MVT InnerEltVT = InnerVT.getVectorElementType();
24612 // If the element sizes match exactly, we can just do one larger vzext. This
24613 // is always an exact type match as vzext operates on integer types.
24614 if (OpEltVT == InnerEltVT) {
24615 assert(OpVT == InnerVT && "Types must match for vzext!");
24616 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
24619 // The only other way we can combine them is if only a single element of the
24620 // inner vzext is used in the input to the outer vzext.
24621 if (InnerEltVT.getSizeInBits() < InputBits)
24624 // In this case, the inner vzext is completely dead because we're going to
24625 // only look at bits inside of the low element. Just do the outer vzext on
24626 // a bitcast of the input to the inner.
24627 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
24630 // Check if we can bypass extracting and re-inserting an element of an input
24631 // vector. Essentialy:
24632 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
24633 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
24634 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
24635 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
24636 SDValue ExtractedV = V.getOperand(0);
24637 SDValue OrigV = ExtractedV.getOperand(0);
24638 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
24639 if (ExtractIdx->getZExtValue() == 0) {
24640 MVT OrigVT = OrigV.getSimpleValueType();
24641 // Extract a subvector if necessary...
24642 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
24643 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
24644 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
24645 OrigVT.getVectorNumElements() / Ratio);
24646 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
24647 DAG.getIntPtrConstant(0, DL));
24649 Op = DAG.getBitcast(OpVT, OrigV);
24650 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
24657 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24658 DAGCombinerInfo &DCI) const {
24659 SelectionDAG &DAG = DCI.DAG;
24660 switch (N->getOpcode()) {
24662 case ISD::EXTRACT_VECTOR_ELT:
24663 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24666 case X86ISD::SHRUNKBLEND:
24667 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24668 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
24669 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24670 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24671 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24672 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24673 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24676 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24677 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24678 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24679 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24680 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24681 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
24682 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24683 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
24684 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
24685 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24686 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24688 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24690 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24691 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24692 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24693 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24694 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24695 case ISD::ANY_EXTEND:
24696 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24697 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24698 case ISD::SIGN_EXTEND_INREG:
24699 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24700 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24701 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24702 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24703 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24704 case X86ISD::SHUFP: // Handle all target specific shuffles
24705 case X86ISD::PALIGNR:
24706 case X86ISD::UNPCKH:
24707 case X86ISD::UNPCKL:
24708 case X86ISD::MOVHLPS:
24709 case X86ISD::MOVLHPS:
24710 case X86ISD::PSHUFB:
24711 case X86ISD::PSHUFD:
24712 case X86ISD::PSHUFHW:
24713 case X86ISD::PSHUFLW:
24714 case X86ISD::MOVSS:
24715 case X86ISD::MOVSD:
24716 case X86ISD::VPERMILPI:
24717 case X86ISD::VPERM2X128:
24718 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24719 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24720 case ISD::INTRINSIC_WO_CHAIN:
24721 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24722 case X86ISD::INSERTPS: {
24723 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
24724 return PerformINSERTPSCombine(N, DAG, Subtarget);
24727 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
24733 /// isTypeDesirableForOp - Return true if the target has native support for
24734 /// the specified value type and it is 'desirable' to use the type for the
24735 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24736 /// instruction encodings are longer and some i16 instructions are slow.
24737 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24738 if (!isTypeLegal(VT))
24740 if (VT != MVT::i16)
24747 case ISD::SIGN_EXTEND:
24748 case ISD::ZERO_EXTEND:
24749 case ISD::ANY_EXTEND:
24762 /// IsDesirableToPromoteOp - This method query the target whether it is
24763 /// beneficial for dag combiner to promote the specified node. If true, it
24764 /// should return the desired promotion type by reference.
24765 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24766 EVT VT = Op.getValueType();
24767 if (VT != MVT::i16)
24770 bool Promote = false;
24771 bool Commute = false;
24772 switch (Op.getOpcode()) {
24775 LoadSDNode *LD = cast<LoadSDNode>(Op);
24776 // If the non-extending load has a single use and it's not live out, then it
24777 // might be folded.
24778 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24779 Op.hasOneUse()*/) {
24780 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24781 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24782 // The only case where we'd want to promote LOAD (rather then it being
24783 // promoted as an operand is when it's only use is liveout.
24784 if (UI->getOpcode() != ISD::CopyToReg)
24791 case ISD::SIGN_EXTEND:
24792 case ISD::ZERO_EXTEND:
24793 case ISD::ANY_EXTEND:
24798 SDValue N0 = Op.getOperand(0);
24799 // Look out for (store (shl (load), x)).
24800 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24813 SDValue N0 = Op.getOperand(0);
24814 SDValue N1 = Op.getOperand(1);
24815 if (!Commute && MayFoldLoad(N1))
24817 // Avoid disabling potential load folding opportunities.
24818 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24820 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24830 //===----------------------------------------------------------------------===//
24831 // X86 Inline Assembly Support
24832 //===----------------------------------------------------------------------===//
24834 // Helper to match a string separated by whitespace.
24835 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
24836 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
24838 for (StringRef Piece : Pieces) {
24839 if (!S.startswith(Piece)) // Check if the piece matches.
24842 S = S.substr(Piece.size());
24843 StringRef::size_type Pos = S.find_first_not_of(" \t");
24844 if (Pos == 0) // We matched a prefix.
24853 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24855 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24856 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24857 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24858 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24860 if (AsmPieces.size() == 3)
24862 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24869 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24870 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24872 std::string AsmStr = IA->getAsmString();
24874 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24875 if (!Ty || Ty->getBitWidth() % 16 != 0)
24878 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24879 SmallVector<StringRef, 4> AsmPieces;
24880 SplitString(AsmStr, AsmPieces, ";\n");
24882 switch (AsmPieces.size()) {
24883 default: return false;
24885 // FIXME: this should verify that we are targeting a 486 or better. If not,
24886 // we will turn this bswap into something that will be lowered to logical
24887 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24888 // lower so don't worry about this.
24890 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
24891 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
24892 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
24893 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
24894 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
24895 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
24896 // No need to check constraints, nothing other than the equivalent of
24897 // "=r,0" would be valid here.
24898 return IntrinsicLowering::LowerToByteSwap(CI);
24901 // rorw $$8, ${0:w} --> llvm.bswap.i16
24902 if (CI->getType()->isIntegerTy(16) &&
24903 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24904 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
24905 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
24907 const std::string &ConstraintsStr = IA->getConstraintString();
24908 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24909 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24910 if (clobbersFlagRegisters(AsmPieces))
24911 return IntrinsicLowering::LowerToByteSwap(CI);
24915 if (CI->getType()->isIntegerTy(32) &&
24916 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24917 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
24918 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
24919 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
24921 const std::string &ConstraintsStr = IA->getConstraintString();
24922 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24923 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24924 if (clobbersFlagRegisters(AsmPieces))
24925 return IntrinsicLowering::LowerToByteSwap(CI);
24928 if (CI->getType()->isIntegerTy(64)) {
24929 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24930 if (Constraints.size() >= 2 &&
24931 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24932 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24933 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24934 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
24935 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
24936 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
24937 return IntrinsicLowering::LowerToByteSwap(CI);
24945 /// getConstraintType - Given a constraint letter, return the type of
24946 /// constraint it is for this target.
24947 X86TargetLowering::ConstraintType
24948 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24949 if (Constraint.size() == 1) {
24950 switch (Constraint[0]) {
24961 return C_RegisterClass;
24985 return TargetLowering::getConstraintType(Constraint);
24988 /// Examine constraint type and operand type and determine a weight value.
24989 /// This object must already have been set up with the operand type
24990 /// and the current alternative constraint selected.
24991 TargetLowering::ConstraintWeight
24992 X86TargetLowering::getSingleConstraintMatchWeight(
24993 AsmOperandInfo &info, const char *constraint) const {
24994 ConstraintWeight weight = CW_Invalid;
24995 Value *CallOperandVal = info.CallOperandVal;
24996 // If we don't have a value, we can't do a match,
24997 // but allow it at the lowest weight.
24998 if (!CallOperandVal)
25000 Type *type = CallOperandVal->getType();
25001 // Look at the constraint type.
25002 switch (*constraint) {
25004 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
25015 if (CallOperandVal->getType()->isIntegerTy())
25016 weight = CW_SpecificReg;
25021 if (type->isFloatingPointTy())
25022 weight = CW_SpecificReg;
25025 if (type->isX86_MMXTy() && Subtarget->hasMMX())
25026 weight = CW_SpecificReg;
25030 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
25031 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
25032 weight = CW_Register;
25035 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
25036 if (C->getZExtValue() <= 31)
25037 weight = CW_Constant;
25041 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25042 if (C->getZExtValue() <= 63)
25043 weight = CW_Constant;
25047 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25048 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
25049 weight = CW_Constant;
25053 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25054 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
25055 weight = CW_Constant;
25059 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25060 if (C->getZExtValue() <= 3)
25061 weight = CW_Constant;
25065 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25066 if (C->getZExtValue() <= 0xff)
25067 weight = CW_Constant;
25072 if (isa<ConstantFP>(CallOperandVal)) {
25073 weight = CW_Constant;
25077 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25078 if ((C->getSExtValue() >= -0x80000000LL) &&
25079 (C->getSExtValue() <= 0x7fffffffLL))
25080 weight = CW_Constant;
25084 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25085 if (C->getZExtValue() <= 0xffffffff)
25086 weight = CW_Constant;
25093 /// LowerXConstraint - try to replace an X constraint, which matches anything,
25094 /// with another that has more specific requirements based on the type of the
25095 /// corresponding operand.
25096 const char *X86TargetLowering::
25097 LowerXConstraint(EVT ConstraintVT) const {
25098 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
25099 // 'f' like normal targets.
25100 if (ConstraintVT.isFloatingPoint()) {
25101 if (Subtarget->hasSSE2())
25103 if (Subtarget->hasSSE1())
25107 return TargetLowering::LowerXConstraint(ConstraintVT);
25110 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
25111 /// vector. If it is invalid, don't add anything to Ops.
25112 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
25113 std::string &Constraint,
25114 std::vector<SDValue>&Ops,
25115 SelectionDAG &DAG) const {
25118 // Only support length 1 constraints for now.
25119 if (Constraint.length() > 1) return;
25121 char ConstraintLetter = Constraint[0];
25122 switch (ConstraintLetter) {
25125 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25126 if (C->getZExtValue() <= 31) {
25127 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25128 Op.getValueType());
25134 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25135 if (C->getZExtValue() <= 63) {
25136 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25137 Op.getValueType());
25143 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25144 if (isInt<8>(C->getSExtValue())) {
25145 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25146 Op.getValueType());
25152 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25153 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
25154 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
25155 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
25156 Op.getValueType());
25162 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25163 if (C->getZExtValue() <= 3) {
25164 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25165 Op.getValueType());
25171 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25172 if (C->getZExtValue() <= 255) {
25173 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25174 Op.getValueType());
25180 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25181 if (C->getZExtValue() <= 127) {
25182 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25183 Op.getValueType());
25189 // 32-bit signed value
25190 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25191 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
25192 C->getSExtValue())) {
25193 // Widen to 64 bits here to get it sign extended.
25194 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
25197 // FIXME gcc accepts some relocatable values here too, but only in certain
25198 // memory models; it's complicated.
25203 // 32-bit unsigned value
25204 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25205 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
25206 C->getZExtValue())) {
25207 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25208 Op.getValueType());
25212 // FIXME gcc accepts some relocatable values here too, but only in certain
25213 // memory models; it's complicated.
25217 // Literal immediates are always ok.
25218 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
25219 // Widen to 64 bits here to get it sign extended.
25220 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
25224 // In any sort of PIC mode addresses need to be computed at runtime by
25225 // adding in a register or some sort of table lookup. These can't
25226 // be used as immediates.
25227 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
25230 // If we are in non-pic codegen mode, we allow the address of a global (with
25231 // an optional displacement) to be used with 'i'.
25232 GlobalAddressSDNode *GA = nullptr;
25233 int64_t Offset = 0;
25235 // Match either (GA), (GA+C), (GA+C1+C2), etc.
25237 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
25238 Offset += GA->getOffset();
25240 } else if (Op.getOpcode() == ISD::ADD) {
25241 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
25242 Offset += C->getZExtValue();
25243 Op = Op.getOperand(0);
25246 } else if (Op.getOpcode() == ISD::SUB) {
25247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
25248 Offset += -C->getZExtValue();
25249 Op = Op.getOperand(0);
25254 // Otherwise, this isn't something we can handle, reject it.
25258 const GlobalValue *GV = GA->getGlobal();
25259 // If we require an extra load to get this address, as in PIC mode, we
25260 // can't accept it.
25261 if (isGlobalStubReference(
25262 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
25265 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
25266 GA->getValueType(0), Offset);
25271 if (Result.getNode()) {
25272 Ops.push_back(Result);
25275 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
25278 std::pair<unsigned, const TargetRegisterClass *>
25279 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
25280 const std::string &Constraint,
25282 // First, see if this is a constraint that directly corresponds to an LLVM
25284 if (Constraint.size() == 1) {
25285 // GCC Constraint Letters
25286 switch (Constraint[0]) {
25288 // TODO: Slight differences here in allocation order and leaving
25289 // RIP in the class. Do they matter any more here than they do
25290 // in the normal allocation?
25291 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
25292 if (Subtarget->is64Bit()) {
25293 if (VT == MVT::i32 || VT == MVT::f32)
25294 return std::make_pair(0U, &X86::GR32RegClass);
25295 if (VT == MVT::i16)
25296 return std::make_pair(0U, &X86::GR16RegClass);
25297 if (VT == MVT::i8 || VT == MVT::i1)
25298 return std::make_pair(0U, &X86::GR8RegClass);
25299 if (VT == MVT::i64 || VT == MVT::f64)
25300 return std::make_pair(0U, &X86::GR64RegClass);
25303 // 32-bit fallthrough
25304 case 'Q': // Q_REGS
25305 if (VT == MVT::i32 || VT == MVT::f32)
25306 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
25307 if (VT == MVT::i16)
25308 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
25309 if (VT == MVT::i8 || VT == MVT::i1)
25310 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
25311 if (VT == MVT::i64)
25312 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
25314 case 'r': // GENERAL_REGS
25315 case 'l': // INDEX_REGS
25316 if (VT == MVT::i8 || VT == MVT::i1)
25317 return std::make_pair(0U, &X86::GR8RegClass);
25318 if (VT == MVT::i16)
25319 return std::make_pair(0U, &X86::GR16RegClass);
25320 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
25321 return std::make_pair(0U, &X86::GR32RegClass);
25322 return std::make_pair(0U, &X86::GR64RegClass);
25323 case 'R': // LEGACY_REGS
25324 if (VT == MVT::i8 || VT == MVT::i1)
25325 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
25326 if (VT == MVT::i16)
25327 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
25328 if (VT == MVT::i32 || !Subtarget->is64Bit())
25329 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
25330 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
25331 case 'f': // FP Stack registers.
25332 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
25333 // value to the correct fpstack register class.
25334 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
25335 return std::make_pair(0U, &X86::RFP32RegClass);
25336 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
25337 return std::make_pair(0U, &X86::RFP64RegClass);
25338 return std::make_pair(0U, &X86::RFP80RegClass);
25339 case 'y': // MMX_REGS if MMX allowed.
25340 if (!Subtarget->hasMMX()) break;
25341 return std::make_pair(0U, &X86::VR64RegClass);
25342 case 'Y': // SSE_REGS if SSE2 allowed
25343 if (!Subtarget->hasSSE2()) break;
25345 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
25346 if (!Subtarget->hasSSE1()) break;
25348 switch (VT.SimpleTy) {
25350 // Scalar SSE types.
25353 return std::make_pair(0U, &X86::FR32RegClass);
25356 return std::make_pair(0U, &X86::FR64RegClass);
25364 return std::make_pair(0U, &X86::VR128RegClass);
25372 return std::make_pair(0U, &X86::VR256RegClass);
25377 return std::make_pair(0U, &X86::VR512RegClass);
25383 // Use the default implementation in TargetLowering to convert the register
25384 // constraint into a member of a register class.
25385 std::pair<unsigned, const TargetRegisterClass*> Res;
25386 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
25388 // Not found as a standard register?
25390 // Map st(0) -> st(7) -> ST0
25391 if (Constraint.size() == 7 && Constraint[0] == '{' &&
25392 tolower(Constraint[1]) == 's' &&
25393 tolower(Constraint[2]) == 't' &&
25394 Constraint[3] == '(' &&
25395 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
25396 Constraint[5] == ')' &&
25397 Constraint[6] == '}') {
25399 Res.first = X86::FP0+Constraint[4]-'0';
25400 Res.second = &X86::RFP80RegClass;
25404 // GCC allows "st(0)" to be called just plain "st".
25405 if (StringRef("{st}").equals_lower(Constraint)) {
25406 Res.first = X86::FP0;
25407 Res.second = &X86::RFP80RegClass;
25412 if (StringRef("{flags}").equals_lower(Constraint)) {
25413 Res.first = X86::EFLAGS;
25414 Res.second = &X86::CCRRegClass;
25418 // 'A' means EAX + EDX.
25419 if (Constraint == "A") {
25420 Res.first = X86::EAX;
25421 Res.second = &X86::GR32_ADRegClass;
25427 // Otherwise, check to see if this is a register class of the wrong value
25428 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
25429 // turn into {ax},{dx}.
25430 if (Res.second->hasType(VT))
25431 return Res; // Correct type already, nothing to do.
25433 // All of the single-register GCC register classes map their values onto
25434 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
25435 // really want an 8-bit or 32-bit register, map to the appropriate register
25436 // class and return the appropriate register.
25437 if (Res.second == &X86::GR16RegClass) {
25438 if (VT == MVT::i8 || VT == MVT::i1) {
25439 unsigned DestReg = 0;
25440 switch (Res.first) {
25442 case X86::AX: DestReg = X86::AL; break;
25443 case X86::DX: DestReg = X86::DL; break;
25444 case X86::CX: DestReg = X86::CL; break;
25445 case X86::BX: DestReg = X86::BL; break;
25448 Res.first = DestReg;
25449 Res.second = &X86::GR8RegClass;
25451 } else if (VT == MVT::i32 || VT == MVT::f32) {
25452 unsigned DestReg = 0;
25453 switch (Res.first) {
25455 case X86::AX: DestReg = X86::EAX; break;
25456 case X86::DX: DestReg = X86::EDX; break;
25457 case X86::CX: DestReg = X86::ECX; break;
25458 case X86::BX: DestReg = X86::EBX; break;
25459 case X86::SI: DestReg = X86::ESI; break;
25460 case X86::DI: DestReg = X86::EDI; break;
25461 case X86::BP: DestReg = X86::EBP; break;
25462 case X86::SP: DestReg = X86::ESP; break;
25465 Res.first = DestReg;
25466 Res.second = &X86::GR32RegClass;
25468 } else if (VT == MVT::i64 || VT == MVT::f64) {
25469 unsigned DestReg = 0;
25470 switch (Res.first) {
25472 case X86::AX: DestReg = X86::RAX; break;
25473 case X86::DX: DestReg = X86::RDX; break;
25474 case X86::CX: DestReg = X86::RCX; break;
25475 case X86::BX: DestReg = X86::RBX; break;
25476 case X86::SI: DestReg = X86::RSI; break;
25477 case X86::DI: DestReg = X86::RDI; break;
25478 case X86::BP: DestReg = X86::RBP; break;
25479 case X86::SP: DestReg = X86::RSP; break;
25482 Res.first = DestReg;
25483 Res.second = &X86::GR64RegClass;
25486 } else if (Res.second == &X86::FR32RegClass ||
25487 Res.second == &X86::FR64RegClass ||
25488 Res.second == &X86::VR128RegClass ||
25489 Res.second == &X86::VR256RegClass ||
25490 Res.second == &X86::FR32XRegClass ||
25491 Res.second == &X86::FR64XRegClass ||
25492 Res.second == &X86::VR128XRegClass ||
25493 Res.second == &X86::VR256XRegClass ||
25494 Res.second == &X86::VR512RegClass) {
25495 // Handle references to XMM physical registers that got mapped into the
25496 // wrong class. This can happen with constraints like {xmm0} where the
25497 // target independent register mapper will just pick the first match it can
25498 // find, ignoring the required type.
25500 if (VT == MVT::f32 || VT == MVT::i32)
25501 Res.second = &X86::FR32RegClass;
25502 else if (VT == MVT::f64 || VT == MVT::i64)
25503 Res.second = &X86::FR64RegClass;
25504 else if (X86::VR128RegClass.hasType(VT))
25505 Res.second = &X86::VR128RegClass;
25506 else if (X86::VR256RegClass.hasType(VT))
25507 Res.second = &X86::VR256RegClass;
25508 else if (X86::VR512RegClass.hasType(VT))
25509 Res.second = &X86::VR512RegClass;
25515 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
25517 unsigned AS) const {
25518 // Scaling factors are not free at all.
25519 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
25520 // will take 2 allocations in the out of order engine instead of 1
25521 // for plain addressing mode, i.e. inst (reg1).
25523 // vaddps (%rsi,%drx), %ymm0, %ymm1
25524 // Requires two allocations (one for the load, one for the computation)
25526 // vaddps (%rsi), %ymm0, %ymm1
25527 // Requires just 1 allocation, i.e., freeing allocations for other operations
25528 // and having less micro operations to execute.
25530 // For some X86 architectures, this is even worse because for instance for
25531 // stores, the complex addressing mode forces the instruction to use the
25532 // "load" ports instead of the dedicated "store" port.
25533 // E.g., on Haswell:
25534 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
25535 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
25536 if (isLegalAddressingMode(AM, Ty, AS))
25537 // Scale represents reg2 * scale, thus account for 1
25538 // as soon as we use a second register.
25539 return AM.Scale != 0;
25543 bool X86TargetLowering::isTargetFTOL() const {
25544 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();