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 // Only provide customized ctpop vector bit twiddling for vector types we
846 // know to perform better than using the popcnt instructions on each vector
847 // element. If popcnt isn't supported, always provide the custom version.
848 if (!Subtarget->hasPOPCNT()) {
849 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
850 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
853 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
854 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
855 MVT VT = (MVT::SimpleValueType)i;
856 // Do not attempt to custom lower non-power-of-2 vectors
857 if (!isPowerOf2_32(VT.getVectorNumElements()))
859 // Do not attempt to custom lower non-128-bit vectors
860 if (!VT.is128BitVector())
862 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
863 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
864 setOperationAction(ISD::VSELECT, VT, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
868 // We support custom legalizing of sext and anyext loads for specific
869 // memory vector types which we can load as a scalar (or sequence of
870 // scalars) and extend in-register to a legal 128-bit vector type. For sext
871 // loads these must work with a single scalar load.
872 for (MVT VT : MVT::integer_vector_valuetypes()) {
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
874 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
875 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
881 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
888 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
889 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
893 if (Subtarget->is64Bit()) {
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
895 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
898 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
899 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
900 MVT VT = (MVT::SimpleValueType)i;
902 // Do not attempt to promote non-128-bit vectors
903 if (!VT.is128BitVector())
906 setOperationAction(ISD::AND, VT, Promote);
907 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
908 setOperationAction(ISD::OR, VT, Promote);
909 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
910 setOperationAction(ISD::XOR, VT, Promote);
911 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
912 setOperationAction(ISD::LOAD, VT, Promote);
913 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
914 setOperationAction(ISD::SELECT, VT, Promote);
915 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
918 // Custom lower v2i64 and v2f64 selects.
919 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
920 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
921 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
922 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
924 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
925 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
927 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
928 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
929 // As there is no 64-bit GPR available, we need build a special custom
930 // sequence to convert from v2i32 to v2f32.
931 if (!Subtarget->is64Bit())
932 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
934 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
935 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
937 for (MVT VT : MVT::fp_vector_valuetypes())
938 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
940 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
941 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
942 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
945 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
946 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
947 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
948 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
949 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
950 setOperationAction(ISD::FRINT, RoundedTy, Legal);
951 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
954 // FIXME: Do we need to handle scalar-to-vector here?
955 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
957 // We directly match byte blends in the backend as they match the VSELECT
959 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
961 // SSE41 brings specific instructions for doing vector sign extend even in
962 // cases where we don't have SRA.
963 for (MVT VT : MVT::integer_vector_valuetypes()) {
964 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
965 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
966 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
969 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
970 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
971 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
972 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
973 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
974 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
977 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
978 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
979 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
980 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
981 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
984 // i8 and i16 vectors are custom because the source register and source
985 // source memory operand types are not the same width. f32 vectors are
986 // custom since the immediate controlling the insert encodes additional
988 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
989 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
990 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
991 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
993 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
994 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
995 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
998 // FIXME: these should be Legal, but that's only for the case where
999 // the index is constant. For now custom expand to deal with that.
1000 if (Subtarget->is64Bit()) {
1001 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1002 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1006 if (Subtarget->hasSSE2()) {
1007 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1008 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1010 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1011 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1013 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1014 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1016 // In the customized shift lowering, the legal cases in AVX2 will be
1018 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1019 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1021 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1022 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1024 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1027 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1028 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1029 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1030 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1031 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1032 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1033 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1035 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1036 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1037 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1039 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1040 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1041 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1042 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1043 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1044 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1045 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1046 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1047 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1048 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1049 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1050 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1052 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1053 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1054 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1055 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1056 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1057 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1058 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1059 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1060 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1063 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1065 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1066 // even though v8i16 is a legal type.
1067 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1068 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1069 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1071 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1072 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1073 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1075 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1076 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1078 for (MVT VT : MVT::fp_vector_valuetypes())
1079 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1081 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1082 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1084 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1085 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1087 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1088 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1090 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1091 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1092 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1093 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1095 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1096 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1097 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1099 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1100 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1101 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1102 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1103 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1104 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1105 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1106 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1107 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1108 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1109 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1110 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1112 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1113 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1114 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1115 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1116 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1117 setOperationAction(ISD::FMA, MVT::f32, Legal);
1118 setOperationAction(ISD::FMA, MVT::f64, Legal);
1121 if (Subtarget->hasInt256()) {
1122 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1123 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1124 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1125 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1127 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1128 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1129 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1130 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1132 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1133 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1134 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1135 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1137 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1138 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1139 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1140 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1142 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1143 // when we have a 256bit-wide blend with immediate.
1144 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1146 // Only provide customized ctpop vector bit twiddling for vector types we
1147 // know to perform better than using the popcnt instructions on each
1148 // vector element. If popcnt isn't supported, always provide the custom
1150 if (!Subtarget->hasPOPCNT())
1151 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1153 // Custom CTPOP always performs better on natively supported v8i32
1154 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1156 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1157 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1158 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1159 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1160 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1161 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1162 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1164 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1165 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1166 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1167 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1168 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1169 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1171 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1172 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1173 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1174 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1176 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1177 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1178 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1179 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1181 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1182 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1183 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1184 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1187 // In the customized shift lowering, the legal cases in AVX2 will be
1189 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1190 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1192 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1193 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1195 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1197 // Custom lower several nodes for 256-bit types.
1198 for (MVT VT : MVT::vector_valuetypes()) {
1199 if (VT.getScalarSizeInBits() >= 32) {
1200 setOperationAction(ISD::MLOAD, VT, Legal);
1201 setOperationAction(ISD::MSTORE, VT, Legal);
1203 // Extract subvector is special because the value type
1204 // (result) is 128-bit but the source is 256-bit wide.
1205 if (VT.is128BitVector()) {
1206 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1208 // Do not attempt to custom lower other non-256-bit vectors
1209 if (!VT.is256BitVector())
1212 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1213 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1214 setOperationAction(ISD::VSELECT, VT, Custom);
1215 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1216 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1217 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1218 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1219 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1222 if (Subtarget->hasInt256())
1223 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1226 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1227 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1228 MVT VT = (MVT::SimpleValueType)i;
1230 // Do not attempt to promote non-256-bit vectors
1231 if (!VT.is256BitVector())
1234 setOperationAction(ISD::AND, VT, Promote);
1235 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1236 setOperationAction(ISD::OR, VT, Promote);
1237 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1238 setOperationAction(ISD::XOR, VT, Promote);
1239 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1240 setOperationAction(ISD::LOAD, VT, Promote);
1241 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1242 setOperationAction(ISD::SELECT, VT, Promote);
1243 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1247 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1248 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1249 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1250 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1251 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1253 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1254 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1255 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1257 for (MVT VT : MVT::fp_vector_valuetypes())
1258 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1260 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1261 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1262 setOperationAction(ISD::XOR, MVT::i1, Legal);
1263 setOperationAction(ISD::OR, MVT::i1, Legal);
1264 setOperationAction(ISD::AND, MVT::i1, Legal);
1265 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1266 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1267 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1268 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1269 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1271 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1272 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1273 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1274 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1275 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1276 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1278 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1279 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1280 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1281 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1282 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1283 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1284 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1285 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1287 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1288 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1289 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1290 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1291 if (Subtarget->is64Bit()) {
1292 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1293 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1294 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1295 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1297 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1298 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1299 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1300 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1301 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1302 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1303 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1304 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1305 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1306 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1307 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1308 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1309 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1310 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1311 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1312 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1314 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1315 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1316 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1317 if (Subtarget->hasDQI()) {
1318 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1319 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1321 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1322 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1323 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1324 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1325 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1326 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1327 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1328 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1329 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1330 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1331 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1332 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1333 if (Subtarget->hasDQI()) {
1334 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1335 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1337 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1338 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1339 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1340 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1341 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1342 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1343 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1344 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1345 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1346 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1348 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1349 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1350 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1351 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1352 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1354 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1355 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1357 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1359 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1360 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1361 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1362 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1363 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1364 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1365 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1366 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1367 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1368 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1369 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1371 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1372 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1374 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1375 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1377 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1379 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1380 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1382 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1383 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1385 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1386 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1388 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1389 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1390 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1391 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1392 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1393 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1395 if (Subtarget->hasCDI()) {
1396 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1397 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1399 if (Subtarget->hasDQI()) {
1400 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1401 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1402 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1404 // Custom lower several nodes.
1405 for (MVT VT : MVT::vector_valuetypes()) {
1406 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1408 setOperationAction(ISD::AND, VT, Legal);
1409 setOperationAction(ISD::OR, VT, Legal);
1410 setOperationAction(ISD::XOR, VT, Legal);
1412 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1413 setOperationAction(ISD::MGATHER, VT, Custom);
1414 setOperationAction(ISD::MSCATTER, VT, Custom);
1416 // Extract subvector is special because the value type
1417 // (result) is 256/128-bit but the source is 512-bit wide.
1418 if (VT.is128BitVector() || VT.is256BitVector()) {
1419 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1421 if (VT.getVectorElementType() == MVT::i1)
1422 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1424 // Do not attempt to custom lower other non-512-bit vectors
1425 if (!VT.is512BitVector())
1428 if (EltSize >= 32) {
1429 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1430 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1431 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1432 setOperationAction(ISD::VSELECT, VT, Legal);
1433 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1434 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1435 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1436 setOperationAction(ISD::MLOAD, VT, Legal);
1437 setOperationAction(ISD::MSTORE, VT, Legal);
1440 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1441 MVT VT = (MVT::SimpleValueType)i;
1443 // Do not attempt to promote non-512-bit vectors.
1444 if (!VT.is512BitVector())
1447 setOperationAction(ISD::SELECT, VT, Promote);
1448 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1452 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1453 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1454 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1456 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1457 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1459 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1460 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1461 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1462 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1463 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1464 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1465 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1466 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1467 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1468 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1470 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1471 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1472 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1473 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1474 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1475 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1476 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1477 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1479 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1480 const MVT VT = (MVT::SimpleValueType)i;
1482 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1484 // Do not attempt to promote non-512-bit vectors.
1485 if (!VT.is512BitVector())
1489 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1490 setOperationAction(ISD::VSELECT, VT, Legal);
1495 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1496 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1497 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1499 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1500 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1501 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1502 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1503 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1504 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1505 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1506 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1507 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1508 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1510 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1511 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1512 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1513 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1514 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1515 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1516 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1517 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1520 // We want to custom lower some of our intrinsics.
1521 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1522 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1523 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1524 if (!Subtarget->is64Bit())
1525 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1527 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1528 // handle type legalization for these operations here.
1530 // FIXME: We really should do custom legalization for addition and
1531 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1532 // than generic legalization for 64-bit multiplication-with-overflow, though.
1533 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1534 // Add/Sub/Mul with overflow operations are custom lowered.
1536 setOperationAction(ISD::SADDO, VT, Custom);
1537 setOperationAction(ISD::UADDO, VT, Custom);
1538 setOperationAction(ISD::SSUBO, VT, Custom);
1539 setOperationAction(ISD::USUBO, VT, Custom);
1540 setOperationAction(ISD::SMULO, VT, Custom);
1541 setOperationAction(ISD::UMULO, VT, Custom);
1545 if (!Subtarget->is64Bit()) {
1546 // These libcalls are not available in 32-bit.
1547 setLibcallName(RTLIB::SHL_I128, nullptr);
1548 setLibcallName(RTLIB::SRL_I128, nullptr);
1549 setLibcallName(RTLIB::SRA_I128, nullptr);
1552 // Combine sin / cos into one node or libcall if possible.
1553 if (Subtarget->hasSinCos()) {
1554 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1555 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1556 if (Subtarget->isTargetDarwin()) {
1557 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1558 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1559 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1560 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1564 if (Subtarget->isTargetWin64()) {
1565 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1566 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1567 setOperationAction(ISD::SREM, MVT::i128, Custom);
1568 setOperationAction(ISD::UREM, MVT::i128, Custom);
1569 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1570 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1573 // We have target-specific dag combine patterns for the following nodes:
1574 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1575 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1576 setTargetDAGCombine(ISD::BITCAST);
1577 setTargetDAGCombine(ISD::VSELECT);
1578 setTargetDAGCombine(ISD::SELECT);
1579 setTargetDAGCombine(ISD::SHL);
1580 setTargetDAGCombine(ISD::SRA);
1581 setTargetDAGCombine(ISD::SRL);
1582 setTargetDAGCombine(ISD::OR);
1583 setTargetDAGCombine(ISD::AND);
1584 setTargetDAGCombine(ISD::ADD);
1585 setTargetDAGCombine(ISD::FADD);
1586 setTargetDAGCombine(ISD::FSUB);
1587 setTargetDAGCombine(ISD::FMA);
1588 setTargetDAGCombine(ISD::SUB);
1589 setTargetDAGCombine(ISD::LOAD);
1590 setTargetDAGCombine(ISD::MLOAD);
1591 setTargetDAGCombine(ISD::STORE);
1592 setTargetDAGCombine(ISD::MSTORE);
1593 setTargetDAGCombine(ISD::ZERO_EXTEND);
1594 setTargetDAGCombine(ISD::ANY_EXTEND);
1595 setTargetDAGCombine(ISD::SIGN_EXTEND);
1596 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1597 setTargetDAGCombine(ISD::SINT_TO_FP);
1598 setTargetDAGCombine(ISD::SETCC);
1599 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1600 setTargetDAGCombine(ISD::BUILD_VECTOR);
1601 setTargetDAGCombine(ISD::MUL);
1602 setTargetDAGCombine(ISD::XOR);
1604 computeRegisterProperties(Subtarget->getRegisterInfo());
1606 // On Darwin, -Os means optimize for size without hurting performance,
1607 // do not reduce the limit.
1608 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1609 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1610 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1611 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1612 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1613 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1614 setPrefLoopAlignment(4); // 2^4 bytes.
1616 // Predictable cmov don't hurt on atom because it's in-order.
1617 PredictableSelectIsExpensive = !Subtarget->isAtom();
1618 EnableExtLdPromotion = true;
1619 setPrefFunctionAlignment(4); // 2^4 bytes.
1621 verifyIntrinsicTables();
1624 // This has so far only been implemented for 64-bit MachO.
1625 bool X86TargetLowering::useLoadStackGuardNode() const {
1626 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1629 TargetLoweringBase::LegalizeTypeAction
1630 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1631 if (ExperimentalVectorWideningLegalization &&
1632 VT.getVectorNumElements() != 1 &&
1633 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1634 return TypeWidenVector;
1636 return TargetLoweringBase::getPreferredVectorAction(VT);
1639 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1641 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1643 const unsigned NumElts = VT.getVectorNumElements();
1644 const EVT EltVT = VT.getVectorElementType();
1645 if (VT.is512BitVector()) {
1646 if (Subtarget->hasAVX512())
1647 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1648 EltVT == MVT::f32 || EltVT == MVT::f64)
1650 case 8: return MVT::v8i1;
1651 case 16: return MVT::v16i1;
1653 if (Subtarget->hasBWI())
1654 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1656 case 32: return MVT::v32i1;
1657 case 64: return MVT::v64i1;
1661 if (VT.is256BitVector() || VT.is128BitVector()) {
1662 if (Subtarget->hasVLX())
1663 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1664 EltVT == MVT::f32 || EltVT == MVT::f64)
1666 case 2: return MVT::v2i1;
1667 case 4: return MVT::v4i1;
1668 case 8: return MVT::v8i1;
1670 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1671 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1673 case 8: return MVT::v8i1;
1674 case 16: return MVT::v16i1;
1675 case 32: return MVT::v32i1;
1679 return VT.changeVectorElementTypeToInteger();
1682 /// Helper for getByValTypeAlignment to determine
1683 /// the desired ByVal argument alignment.
1684 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1687 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1688 if (VTy->getBitWidth() == 128)
1690 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1691 unsigned EltAlign = 0;
1692 getMaxByValAlign(ATy->getElementType(), EltAlign);
1693 if (EltAlign > MaxAlign)
1694 MaxAlign = EltAlign;
1695 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1696 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1697 unsigned EltAlign = 0;
1698 getMaxByValAlign(STy->getElementType(i), EltAlign);
1699 if (EltAlign > MaxAlign)
1700 MaxAlign = EltAlign;
1707 /// Return the desired alignment for ByVal aggregate
1708 /// function arguments in the caller parameter area. For X86, aggregates
1709 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1710 /// are at 4-byte boundaries.
1711 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1712 if (Subtarget->is64Bit()) {
1713 // Max of 8 and alignment of type.
1714 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1721 if (Subtarget->hasSSE1())
1722 getMaxByValAlign(Ty, Align);
1726 /// Returns the target specific optimal type for load
1727 /// and store operations as a result of memset, memcpy, and memmove
1728 /// lowering. If DstAlign is zero that means it's safe to destination
1729 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1730 /// means there isn't a need to check it against alignment requirement,
1731 /// probably because the source does not need to be loaded. If 'IsMemset' is
1732 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1733 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1734 /// source is constant so it does not need to be loaded.
1735 /// It returns EVT::Other if the type should be determined using generic
1736 /// target-independent logic.
1738 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1739 unsigned DstAlign, unsigned SrcAlign,
1740 bool IsMemset, bool ZeroMemset,
1742 MachineFunction &MF) const {
1743 const Function *F = MF.getFunction();
1744 if ((!IsMemset || ZeroMemset) &&
1745 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1747 (Subtarget->isUnalignedMemAccessFast() ||
1748 ((DstAlign == 0 || DstAlign >= 16) &&
1749 (SrcAlign == 0 || SrcAlign >= 16)))) {
1751 if (Subtarget->hasInt256())
1753 if (Subtarget->hasFp256())
1756 if (Subtarget->hasSSE2())
1758 if (Subtarget->hasSSE1())
1760 } else if (!MemcpyStrSrc && Size >= 8 &&
1761 !Subtarget->is64Bit() &&
1762 Subtarget->hasSSE2()) {
1763 // Do not use f64 to lower memcpy if source is string constant. It's
1764 // better to use i32 to avoid the loads.
1768 if (Subtarget->is64Bit() && Size >= 8)
1773 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1775 return X86ScalarSSEf32;
1776 else if (VT == MVT::f64)
1777 return X86ScalarSSEf64;
1782 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1787 *Fast = Subtarget->isUnalignedMemAccessFast();
1791 /// Return the entry encoding for a jump table in the
1792 /// current function. The returned value is a member of the
1793 /// MachineJumpTableInfo::JTEntryKind enum.
1794 unsigned X86TargetLowering::getJumpTableEncoding() const {
1795 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1797 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1798 Subtarget->isPICStyleGOT())
1799 return MachineJumpTableInfo::EK_Custom32;
1801 // Otherwise, use the normal jump table encoding heuristics.
1802 return TargetLowering::getJumpTableEncoding();
1805 bool X86TargetLowering::useSoftFloat() const {
1806 return Subtarget->useSoftFloat();
1810 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1811 const MachineBasicBlock *MBB,
1812 unsigned uid,MCContext &Ctx) const{
1813 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1814 Subtarget->isPICStyleGOT());
1815 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1817 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1818 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1821 /// Returns relocation base for the given PIC jumptable.
1822 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1823 SelectionDAG &DAG) const {
1824 if (!Subtarget->is64Bit())
1825 // This doesn't have SDLoc associated with it, but is not really the
1826 // same as a Register.
1827 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1831 /// This returns the relocation base for the given PIC jumptable,
1832 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1833 const MCExpr *X86TargetLowering::
1834 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1835 MCContext &Ctx) const {
1836 // X86-64 uses RIP relative addressing based on the jump table label.
1837 if (Subtarget->isPICStyleRIPRel())
1838 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1840 // Otherwise, the reference is relative to the PIC base.
1841 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1844 std::pair<const TargetRegisterClass *, uint8_t>
1845 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1847 const TargetRegisterClass *RRC = nullptr;
1849 switch (VT.SimpleTy) {
1851 return TargetLowering::findRepresentativeClass(TRI, VT);
1852 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1853 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1856 RRC = &X86::VR64RegClass;
1858 case MVT::f32: case MVT::f64:
1859 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1860 case MVT::v4f32: case MVT::v2f64:
1861 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1863 RRC = &X86::VR128RegClass;
1866 return std::make_pair(RRC, Cost);
1869 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1870 unsigned &Offset) const {
1871 if (!Subtarget->isTargetLinux())
1874 if (Subtarget->is64Bit()) {
1875 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1877 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1889 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1890 unsigned DestAS) const {
1891 assert(SrcAS != DestAS && "Expected different address spaces!");
1893 return SrcAS < 256 && DestAS < 256;
1896 //===----------------------------------------------------------------------===//
1897 // Return Value Calling Convention Implementation
1898 //===----------------------------------------------------------------------===//
1900 #include "X86GenCallingConv.inc"
1903 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1904 MachineFunction &MF, bool isVarArg,
1905 const SmallVectorImpl<ISD::OutputArg> &Outs,
1906 LLVMContext &Context) const {
1907 SmallVector<CCValAssign, 16> RVLocs;
1908 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1909 return CCInfo.CheckReturn(Outs, RetCC_X86);
1912 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1913 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1918 X86TargetLowering::LowerReturn(SDValue Chain,
1919 CallingConv::ID CallConv, bool isVarArg,
1920 const SmallVectorImpl<ISD::OutputArg> &Outs,
1921 const SmallVectorImpl<SDValue> &OutVals,
1922 SDLoc dl, SelectionDAG &DAG) const {
1923 MachineFunction &MF = DAG.getMachineFunction();
1924 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1926 SmallVector<CCValAssign, 16> RVLocs;
1927 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1928 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1931 SmallVector<SDValue, 6> RetOps;
1932 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1933 // Operand #1 = Bytes To Pop
1934 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
1937 // Copy the result values into the output registers.
1938 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1939 CCValAssign &VA = RVLocs[i];
1940 assert(VA.isRegLoc() && "Can only return in registers!");
1941 SDValue ValToCopy = OutVals[i];
1942 EVT ValVT = ValToCopy.getValueType();
1944 // Promote values to the appropriate types.
1945 if (VA.getLocInfo() == CCValAssign::SExt)
1946 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1947 else if (VA.getLocInfo() == CCValAssign::ZExt)
1948 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1949 else if (VA.getLocInfo() == CCValAssign::AExt) {
1950 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
1951 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1953 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1955 else if (VA.getLocInfo() == CCValAssign::BCvt)
1956 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1958 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1959 "Unexpected FP-extend for return value.");
1961 // If this is x86-64, and we disabled SSE, we can't return FP values,
1962 // or SSE or MMX vectors.
1963 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1964 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1965 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1966 report_fatal_error("SSE register return with SSE disabled");
1968 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1969 // llvm-gcc has never done it right and no one has noticed, so this
1970 // should be OK for now.
1971 if (ValVT == MVT::f64 &&
1972 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1973 report_fatal_error("SSE2 register return with SSE2 disabled");
1975 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1976 // the RET instruction and handled by the FP Stackifier.
1977 if (VA.getLocReg() == X86::FP0 ||
1978 VA.getLocReg() == X86::FP1) {
1979 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1980 // change the value to the FP stack register class.
1981 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1982 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1983 RetOps.push_back(ValToCopy);
1984 // Don't emit a copytoreg.
1988 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1989 // which is returned in RAX / RDX.
1990 if (Subtarget->is64Bit()) {
1991 if (ValVT == MVT::x86mmx) {
1992 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1993 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1994 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1996 // If we don't have SSE2 available, convert to v4f32 so the generated
1997 // register is legal.
1998 if (!Subtarget->hasSSE2())
1999 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2004 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2005 Flag = Chain.getValue(1);
2006 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2009 // All x86 ABIs require that for returning structs by value we copy
2010 // the sret argument into %rax/%eax (depending on ABI) for the return.
2011 // We saved the argument into a virtual register in the entry block,
2012 // so now we copy the value out and into %rax/%eax.
2014 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2015 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2016 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2017 // either case FuncInfo->setSRetReturnReg() will have been called.
2018 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2019 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2022 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2023 X86::RAX : X86::EAX;
2024 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2025 Flag = Chain.getValue(1);
2027 // RAX/EAX now acts like a return value.
2028 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2031 RetOps[0] = Chain; // Update chain.
2033 // Add the flag if we have it.
2035 RetOps.push_back(Flag);
2037 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2040 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2041 if (N->getNumValues() != 1)
2043 if (!N->hasNUsesOfValue(1, 0))
2046 SDValue TCChain = Chain;
2047 SDNode *Copy = *N->use_begin();
2048 if (Copy->getOpcode() == ISD::CopyToReg) {
2049 // If the copy has a glue operand, we conservatively assume it isn't safe to
2050 // perform a tail call.
2051 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2053 TCChain = Copy->getOperand(0);
2054 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2057 bool HasRet = false;
2058 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2060 if (UI->getOpcode() != X86ISD::RET_FLAG)
2062 // If we are returning more than one value, we can definitely
2063 // not make a tail call see PR19530
2064 if (UI->getNumOperands() > 4)
2066 if (UI->getNumOperands() == 4 &&
2067 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2080 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2081 ISD::NodeType ExtendKind) const {
2083 // TODO: Is this also valid on 32-bit?
2084 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2085 ReturnMVT = MVT::i8;
2087 ReturnMVT = MVT::i32;
2089 EVT MinVT = getRegisterType(Context, ReturnMVT);
2090 return VT.bitsLT(MinVT) ? MinVT : VT;
2093 /// Lower the result values of a call into the
2094 /// appropriate copies out of appropriate physical registers.
2097 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2098 CallingConv::ID CallConv, bool isVarArg,
2099 const SmallVectorImpl<ISD::InputArg> &Ins,
2100 SDLoc dl, SelectionDAG &DAG,
2101 SmallVectorImpl<SDValue> &InVals) const {
2103 // Assign locations to each value returned by this call.
2104 SmallVector<CCValAssign, 16> RVLocs;
2105 bool Is64Bit = Subtarget->is64Bit();
2106 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2108 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2110 // Copy all of the result registers out of their specified physreg.
2111 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2112 CCValAssign &VA = RVLocs[i];
2113 EVT CopyVT = VA.getLocVT();
2115 // If this is x86-64, and we disabled SSE, we can't return FP values
2116 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2117 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2118 report_fatal_error("SSE register return with SSE disabled");
2121 // If we prefer to use the value in xmm registers, copy it out as f80 and
2122 // use a truncate to move it from fp stack reg to xmm reg.
2123 bool RoundAfterCopy = false;
2124 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2125 isScalarFPTypeInSSEReg(VA.getValVT())) {
2127 RoundAfterCopy = (CopyVT != VA.getLocVT());
2130 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2131 CopyVT, InFlag).getValue(1);
2132 SDValue Val = Chain.getValue(0);
2135 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2136 // This truncation won't change the value.
2137 DAG.getIntPtrConstant(1, dl));
2139 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2140 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2142 InFlag = Chain.getValue(2);
2143 InVals.push_back(Val);
2149 //===----------------------------------------------------------------------===//
2150 // C & StdCall & Fast Calling Convention implementation
2151 //===----------------------------------------------------------------------===//
2152 // StdCall calling convention seems to be standard for many Windows' API
2153 // routines and around. It differs from C calling convention just a little:
2154 // callee should clean up the stack, not caller. Symbols should be also
2155 // decorated in some fancy way :) It doesn't support any vector arguments.
2156 // For info on fast calling convention see Fast Calling Convention (tail call)
2157 // implementation LowerX86_32FastCCCallTo.
2159 /// CallIsStructReturn - Determines whether a call uses struct return
2161 enum StructReturnType {
2166 static StructReturnType
2167 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2169 return NotStructReturn;
2171 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2172 if (!Flags.isSRet())
2173 return NotStructReturn;
2174 if (Flags.isInReg())
2175 return RegStructReturn;
2176 return StackStructReturn;
2179 /// Determines whether a function uses struct return semantics.
2180 static StructReturnType
2181 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2183 return NotStructReturn;
2185 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2186 if (!Flags.isSRet())
2187 return NotStructReturn;
2188 if (Flags.isInReg())
2189 return RegStructReturn;
2190 return StackStructReturn;
2193 /// Make a copy of an aggregate at address specified by "Src" to address
2194 /// "Dst" with size and alignment information specified by the specific
2195 /// parameter attribute. The copy will be passed as a byval function parameter.
2197 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2198 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2200 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2202 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2203 /*isVolatile*/false, /*AlwaysInline=*/true,
2204 /*isTailCall*/false,
2205 MachinePointerInfo(), MachinePointerInfo());
2208 /// Return true if the calling convention is one that
2209 /// supports tail call optimization.
2210 static bool IsTailCallConvention(CallingConv::ID CC) {
2211 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2212 CC == CallingConv::HiPE);
2215 /// \brief Return true if the calling convention is a C calling convention.
2216 static bool IsCCallConvention(CallingConv::ID CC) {
2217 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2218 CC == CallingConv::X86_64_SysV);
2221 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2222 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2226 CallingConv::ID CalleeCC = CS.getCallingConv();
2227 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2233 /// Return true if the function is being made into
2234 /// a tailcall target by changing its ABI.
2235 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2236 bool GuaranteedTailCallOpt) {
2237 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2241 X86TargetLowering::LowerMemArgument(SDValue Chain,
2242 CallingConv::ID CallConv,
2243 const SmallVectorImpl<ISD::InputArg> &Ins,
2244 SDLoc dl, SelectionDAG &DAG,
2245 const CCValAssign &VA,
2246 MachineFrameInfo *MFI,
2248 // Create the nodes corresponding to a load from this parameter slot.
2249 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2250 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2251 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2252 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2255 // If value is passed by pointer we have address passed instead of the value
2257 bool ExtendedInMem = VA.isExtInLoc() &&
2258 VA.getValVT().getScalarType() == MVT::i1;
2260 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2261 ValVT = VA.getLocVT();
2263 ValVT = VA.getValVT();
2265 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2266 // changed with more analysis.
2267 // In case of tail call optimization mark all arguments mutable. Since they
2268 // could be overwritten by lowering of arguments in case of a tail call.
2269 if (Flags.isByVal()) {
2270 unsigned Bytes = Flags.getByValSize();
2271 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2272 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2273 return DAG.getFrameIndex(FI, getPointerTy());
2275 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2276 VA.getLocMemOffset(), isImmutable);
2277 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2278 SDValue Val = DAG.getLoad(ValVT, dl, Chain, FIN,
2279 MachinePointerInfo::getFixedStack(FI),
2280 false, false, false, 0);
2281 return ExtendedInMem ?
2282 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2286 // FIXME: Get this from tablegen.
2287 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2288 const X86Subtarget *Subtarget) {
2289 assert(Subtarget->is64Bit());
2291 if (Subtarget->isCallingConvWin64(CallConv)) {
2292 static const MCPhysReg GPR64ArgRegsWin64[] = {
2293 X86::RCX, X86::RDX, X86::R8, X86::R9
2295 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2298 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2299 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2301 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2304 // FIXME: Get this from tablegen.
2305 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2306 CallingConv::ID CallConv,
2307 const X86Subtarget *Subtarget) {
2308 assert(Subtarget->is64Bit());
2309 if (Subtarget->isCallingConvWin64(CallConv)) {
2310 // The XMM registers which might contain var arg parameters are shadowed
2311 // in their paired GPR. So we only need to save the GPR to their home
2313 // TODO: __vectorcall will change this.
2317 const Function *Fn = MF.getFunction();
2318 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2319 bool isSoftFloat = Subtarget->useSoftFloat();
2320 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2321 "SSE register cannot be used when SSE is disabled!");
2322 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2323 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2327 static const MCPhysReg XMMArgRegs64Bit[] = {
2328 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2329 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2331 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2335 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2336 CallingConv::ID CallConv,
2338 const SmallVectorImpl<ISD::InputArg> &Ins,
2341 SmallVectorImpl<SDValue> &InVals)
2343 MachineFunction &MF = DAG.getMachineFunction();
2344 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2345 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2347 const Function* Fn = MF.getFunction();
2348 if (Fn->hasExternalLinkage() &&
2349 Subtarget->isTargetCygMing() &&
2350 Fn->getName() == "main")
2351 FuncInfo->setForceFramePointer(true);
2353 MachineFrameInfo *MFI = MF.getFrameInfo();
2354 bool Is64Bit = Subtarget->is64Bit();
2355 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2357 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2358 "Var args not supported with calling convention fastcc, ghc or hipe");
2360 // Assign locations to all of the incoming arguments.
2361 SmallVector<CCValAssign, 16> ArgLocs;
2362 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2364 // Allocate shadow area for Win64
2366 CCInfo.AllocateStack(32, 8);
2368 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2370 unsigned LastVal = ~0U;
2372 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2373 CCValAssign &VA = ArgLocs[i];
2374 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2376 assert(VA.getValNo() != LastVal &&
2377 "Don't support value assigned to multiple locs yet");
2379 LastVal = VA.getValNo();
2381 if (VA.isRegLoc()) {
2382 EVT RegVT = VA.getLocVT();
2383 const TargetRegisterClass *RC;
2384 if (RegVT == MVT::i32)
2385 RC = &X86::GR32RegClass;
2386 else if (Is64Bit && RegVT == MVT::i64)
2387 RC = &X86::GR64RegClass;
2388 else if (RegVT == MVT::f32)
2389 RC = &X86::FR32RegClass;
2390 else if (RegVT == MVT::f64)
2391 RC = &X86::FR64RegClass;
2392 else if (RegVT.is512BitVector())
2393 RC = &X86::VR512RegClass;
2394 else if (RegVT.is256BitVector())
2395 RC = &X86::VR256RegClass;
2396 else if (RegVT.is128BitVector())
2397 RC = &X86::VR128RegClass;
2398 else if (RegVT == MVT::x86mmx)
2399 RC = &X86::VR64RegClass;
2400 else if (RegVT == MVT::i1)
2401 RC = &X86::VK1RegClass;
2402 else if (RegVT == MVT::v8i1)
2403 RC = &X86::VK8RegClass;
2404 else if (RegVT == MVT::v16i1)
2405 RC = &X86::VK16RegClass;
2406 else if (RegVT == MVT::v32i1)
2407 RC = &X86::VK32RegClass;
2408 else if (RegVT == MVT::v64i1)
2409 RC = &X86::VK64RegClass;
2411 llvm_unreachable("Unknown argument type!");
2413 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2414 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2416 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2417 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2419 if (VA.getLocInfo() == CCValAssign::SExt)
2420 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2421 DAG.getValueType(VA.getValVT()));
2422 else if (VA.getLocInfo() == CCValAssign::ZExt)
2423 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2424 DAG.getValueType(VA.getValVT()));
2425 else if (VA.getLocInfo() == CCValAssign::BCvt)
2426 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2428 if (VA.isExtInLoc()) {
2429 // Handle MMX values passed in XMM regs.
2430 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2431 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2433 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2436 assert(VA.isMemLoc());
2437 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2440 // If value is passed via pointer - do a load.
2441 if (VA.getLocInfo() == CCValAssign::Indirect)
2442 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2443 MachinePointerInfo(), false, false, false, 0);
2445 InVals.push_back(ArgValue);
2448 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2449 // All x86 ABIs require that for returning structs by value we copy the
2450 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2451 // the argument into a virtual register so that we can access it from the
2453 if (Ins[i].Flags.isSRet()) {
2454 unsigned Reg = FuncInfo->getSRetReturnReg();
2456 MVT PtrTy = getPointerTy();
2457 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2458 FuncInfo->setSRetReturnReg(Reg);
2460 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2461 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2466 unsigned StackSize = CCInfo.getNextStackOffset();
2467 // Align stack specially for tail calls.
2468 if (FuncIsMadeTailCallSafe(CallConv,
2469 MF.getTarget().Options.GuaranteedTailCallOpt))
2470 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2472 // If the function takes variable number of arguments, make a frame index for
2473 // the start of the first vararg value... for expansion of llvm.va_start. We
2474 // can skip this if there are no va_start calls.
2475 if (MFI->hasVAStart() &&
2476 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2477 CallConv != CallingConv::X86_ThisCall))) {
2478 FuncInfo->setVarArgsFrameIndex(
2479 MFI->CreateFixedObject(1, StackSize, true));
2482 MachineModuleInfo &MMI = MF.getMMI();
2483 const Function *WinEHParent = nullptr;
2484 if (IsWin64 && MMI.hasWinEHFuncInfo(Fn))
2485 WinEHParent = MMI.getWinEHParent(Fn);
2486 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2487 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2489 // Figure out if XMM registers are in use.
2490 assert(!(Subtarget->useSoftFloat() &&
2491 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2492 "SSE register cannot be used when SSE is disabled!");
2494 // 64-bit calling conventions support varargs and register parameters, so we
2495 // have to do extra work to spill them in the prologue.
2496 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2497 // Find the first unallocated argument registers.
2498 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2499 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2500 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2501 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2502 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2503 "SSE register cannot be used when SSE is disabled!");
2505 // Gather all the live in physical registers.
2506 SmallVector<SDValue, 6> LiveGPRs;
2507 SmallVector<SDValue, 8> LiveXMMRegs;
2509 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2510 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2512 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2514 if (!ArgXMMs.empty()) {
2515 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2516 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2517 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2518 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2519 LiveXMMRegs.push_back(
2520 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2525 // Get to the caller-allocated home save location. Add 8 to account
2526 // for the return address.
2527 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2528 FuncInfo->setRegSaveFrameIndex(
2529 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2530 // Fixup to set vararg frame on shadow area (4 x i64).
2532 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2534 // For X86-64, if there are vararg parameters that are passed via
2535 // registers, then we must store them to their spots on the stack so
2536 // they may be loaded by deferencing the result of va_next.
2537 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2538 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2539 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2540 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2543 // Store the integer parameter registers.
2544 SmallVector<SDValue, 8> MemOps;
2545 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2547 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2548 for (SDValue Val : LiveGPRs) {
2549 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2550 DAG.getIntPtrConstant(Offset, dl));
2552 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2553 MachinePointerInfo::getFixedStack(
2554 FuncInfo->getRegSaveFrameIndex(), Offset),
2556 MemOps.push_back(Store);
2560 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2561 // Now store the XMM (fp + vector) parameter registers.
2562 SmallVector<SDValue, 12> SaveXMMOps;
2563 SaveXMMOps.push_back(Chain);
2564 SaveXMMOps.push_back(ALVal);
2565 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2566 FuncInfo->getRegSaveFrameIndex(), dl));
2567 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2568 FuncInfo->getVarArgsFPOffset(), dl));
2569 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2571 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2572 MVT::Other, SaveXMMOps));
2575 if (!MemOps.empty())
2576 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2577 } else if (IsWinEHOutlined) {
2578 // Get to the caller-allocated home save location. Add 8 to account
2579 // for the return address.
2580 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2581 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2582 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2584 MMI.getWinEHFuncInfo(Fn)
2585 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2586 FuncInfo->getRegSaveFrameIndex();
2588 // Store the second integer parameter (rdx) into rsp+16 relative to the
2589 // stack pointer at the entry of the function.
2591 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), getPointerTy());
2592 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2593 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2594 Chain = DAG.getStore(
2595 Val.getValue(1), dl, Val, RSFIN,
2596 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2597 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2600 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2601 // Find the largest legal vector type.
2602 MVT VecVT = MVT::Other;
2603 // FIXME: Only some x86_32 calling conventions support AVX512.
2604 if (Subtarget->hasAVX512() &&
2605 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2606 CallConv == CallingConv::Intel_OCL_BI)))
2607 VecVT = MVT::v16f32;
2608 else if (Subtarget->hasAVX())
2610 else if (Subtarget->hasSSE2())
2613 // We forward some GPRs and some vector types.
2614 SmallVector<MVT, 2> RegParmTypes;
2615 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2616 RegParmTypes.push_back(IntVT);
2617 if (VecVT != MVT::Other)
2618 RegParmTypes.push_back(VecVT);
2620 // Compute the set of forwarded registers. The rest are scratch.
2621 SmallVectorImpl<ForwardedRegister> &Forwards =
2622 FuncInfo->getForwardedMustTailRegParms();
2623 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2625 // Conservatively forward AL on x86_64, since it might be used for varargs.
2626 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2627 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2628 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2631 // Copy all forwards from physical to virtual registers.
2632 for (ForwardedRegister &F : Forwards) {
2633 // FIXME: Can we use a less constrained schedule?
2634 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2635 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2636 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2640 // Some CCs need callee pop.
2641 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2642 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2643 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2645 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2646 // If this is an sret function, the return should pop the hidden pointer.
2647 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2648 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2649 argsAreStructReturn(Ins) == StackStructReturn)
2650 FuncInfo->setBytesToPopOnReturn(4);
2654 // RegSaveFrameIndex is X86-64 only.
2655 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2656 if (CallConv == CallingConv::X86_FastCall ||
2657 CallConv == CallingConv::X86_ThisCall)
2658 // fastcc functions can't have varargs.
2659 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2662 FuncInfo->setArgumentStackSize(StackSize);
2664 if (IsWinEHParent) {
2665 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2666 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2667 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2668 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2669 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2670 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2671 /*isVolatile=*/true,
2672 /*isNonTemporal=*/false, /*Alignment=*/0);
2679 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2680 SDValue StackPtr, SDValue Arg,
2681 SDLoc dl, SelectionDAG &DAG,
2682 const CCValAssign &VA,
2683 ISD::ArgFlagsTy Flags) const {
2684 unsigned LocMemOffset = VA.getLocMemOffset();
2685 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2686 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2687 if (Flags.isByVal())
2688 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2690 return DAG.getStore(Chain, dl, Arg, PtrOff,
2691 MachinePointerInfo::getStack(LocMemOffset),
2695 /// Emit a load of return address if tail call
2696 /// optimization is performed and it is required.
2698 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2699 SDValue &OutRetAddr, SDValue Chain,
2700 bool IsTailCall, bool Is64Bit,
2701 int FPDiff, SDLoc dl) const {
2702 // Adjust the Return address stack slot.
2703 EVT VT = getPointerTy();
2704 OutRetAddr = getReturnAddressFrameIndex(DAG);
2706 // Load the "old" Return address.
2707 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2708 false, false, false, 0);
2709 return SDValue(OutRetAddr.getNode(), 1);
2712 /// Emit a store of the return address if tail call
2713 /// optimization is performed and it is required (FPDiff!=0).
2714 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2715 SDValue Chain, SDValue RetAddrFrIdx,
2716 EVT PtrVT, unsigned SlotSize,
2717 int FPDiff, SDLoc dl) {
2718 // Store the return address to the appropriate stack slot.
2719 if (!FPDiff) return Chain;
2720 // Calculate the new stack slot for the return address.
2721 int NewReturnAddrFI =
2722 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2724 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2725 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2726 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2732 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2733 SmallVectorImpl<SDValue> &InVals) const {
2734 SelectionDAG &DAG = CLI.DAG;
2736 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2737 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2738 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2739 SDValue Chain = CLI.Chain;
2740 SDValue Callee = CLI.Callee;
2741 CallingConv::ID CallConv = CLI.CallConv;
2742 bool &isTailCall = CLI.IsTailCall;
2743 bool isVarArg = CLI.IsVarArg;
2745 MachineFunction &MF = DAG.getMachineFunction();
2746 bool Is64Bit = Subtarget->is64Bit();
2747 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2748 StructReturnType SR = callIsStructReturn(Outs);
2749 bool IsSibcall = false;
2750 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2752 if (MF.getTarget().Options.DisableTailCalls)
2755 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2757 // Force this to be a tail call. The verifier rules are enough to ensure
2758 // that we can lower this successfully without moving the return address
2761 } else if (isTailCall) {
2762 // Check if it's really possible to do a tail call.
2763 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2764 isVarArg, SR != NotStructReturn,
2765 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2766 Outs, OutVals, Ins, DAG);
2768 // Sibcalls are automatically detected tailcalls which do not require
2770 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2777 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2778 "Var args not supported with calling convention fastcc, ghc or hipe");
2780 // Analyze operands of the call, assigning locations to each operand.
2781 SmallVector<CCValAssign, 16> ArgLocs;
2782 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2784 // Allocate shadow area for Win64
2786 CCInfo.AllocateStack(32, 8);
2788 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2790 // Get a count of how many bytes are to be pushed on the stack.
2791 unsigned NumBytes = CCInfo.getNextStackOffset();
2793 // This is a sibcall. The memory operands are available in caller's
2794 // own caller's stack.
2796 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2797 IsTailCallConvention(CallConv))
2798 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2801 if (isTailCall && !IsSibcall && !IsMustTail) {
2802 // Lower arguments at fp - stackoffset + fpdiff.
2803 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2805 FPDiff = NumBytesCallerPushed - NumBytes;
2807 // Set the delta of movement of the returnaddr stackslot.
2808 // But only set if delta is greater than previous delta.
2809 if (FPDiff < X86Info->getTCReturnAddrDelta())
2810 X86Info->setTCReturnAddrDelta(FPDiff);
2813 unsigned NumBytesToPush = NumBytes;
2814 unsigned NumBytesToPop = NumBytes;
2816 // If we have an inalloca argument, all stack space has already been allocated
2817 // for us and be right at the top of the stack. We don't support multiple
2818 // arguments passed in memory when using inalloca.
2819 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2821 if (!ArgLocs.back().isMemLoc())
2822 report_fatal_error("cannot use inalloca attribute on a register "
2824 if (ArgLocs.back().getLocMemOffset() != 0)
2825 report_fatal_error("any parameter with the inalloca attribute must be "
2826 "the only memory argument");
2830 Chain = DAG.getCALLSEQ_START(
2831 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
2833 SDValue RetAddrFrIdx;
2834 // Load return address for tail calls.
2835 if (isTailCall && FPDiff)
2836 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2837 Is64Bit, FPDiff, dl);
2839 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2840 SmallVector<SDValue, 8> MemOpChains;
2843 // Walk the register/memloc assignments, inserting copies/loads. In the case
2844 // of tail call optimization arguments are handle later.
2845 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2846 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2847 // Skip inalloca arguments, they have already been written.
2848 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2849 if (Flags.isInAlloca())
2852 CCValAssign &VA = ArgLocs[i];
2853 EVT RegVT = VA.getLocVT();
2854 SDValue Arg = OutVals[i];
2855 bool isByVal = Flags.isByVal();
2857 // Promote the value if needed.
2858 switch (VA.getLocInfo()) {
2859 default: llvm_unreachable("Unknown loc info!");
2860 case CCValAssign::Full: break;
2861 case CCValAssign::SExt:
2862 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2864 case CCValAssign::ZExt:
2865 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2867 case CCValAssign::AExt:
2868 if (Arg.getValueType().isVector() &&
2869 Arg.getValueType().getScalarType() == MVT::i1)
2870 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2871 else if (RegVT.is128BitVector()) {
2872 // Special case: passing MMX values in XMM registers.
2873 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2874 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2875 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2877 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2879 case CCValAssign::BCvt:
2880 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2882 case CCValAssign::Indirect: {
2883 // Store the argument.
2884 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2885 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2886 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2887 MachinePointerInfo::getFixedStack(FI),
2894 if (VA.isRegLoc()) {
2895 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2896 if (isVarArg && IsWin64) {
2897 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2898 // shadow reg if callee is a varargs function.
2899 unsigned ShadowReg = 0;
2900 switch (VA.getLocReg()) {
2901 case X86::XMM0: ShadowReg = X86::RCX; break;
2902 case X86::XMM1: ShadowReg = X86::RDX; break;
2903 case X86::XMM2: ShadowReg = X86::R8; break;
2904 case X86::XMM3: ShadowReg = X86::R9; break;
2907 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2909 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2910 assert(VA.isMemLoc());
2911 if (!StackPtr.getNode())
2912 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2914 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2915 dl, DAG, VA, Flags));
2919 if (!MemOpChains.empty())
2920 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2922 if (Subtarget->isPICStyleGOT()) {
2923 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2926 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2927 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2929 // If we are tail calling and generating PIC/GOT style code load the
2930 // address of the callee into ECX. The value in ecx is used as target of
2931 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2932 // for tail calls on PIC/GOT architectures. Normally we would just put the
2933 // address of GOT into ebx and then call target@PLT. But for tail calls
2934 // ebx would be restored (since ebx is callee saved) before jumping to the
2937 // Note: The actual moving to ECX is done further down.
2938 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2939 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2940 !G->getGlobal()->hasProtectedVisibility())
2941 Callee = LowerGlobalAddress(Callee, DAG);
2942 else if (isa<ExternalSymbolSDNode>(Callee))
2943 Callee = LowerExternalSymbol(Callee, DAG);
2947 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2948 // From AMD64 ABI document:
2949 // For calls that may call functions that use varargs or stdargs
2950 // (prototype-less calls or calls to functions containing ellipsis (...) in
2951 // the declaration) %al is used as hidden argument to specify the number
2952 // of SSE registers used. The contents of %al do not need to match exactly
2953 // the number of registers, but must be an ubound on the number of SSE
2954 // registers used and is in the range 0 - 8 inclusive.
2956 // Count the number of XMM registers allocated.
2957 static const MCPhysReg XMMArgRegs[] = {
2958 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2959 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2961 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2962 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2963 && "SSE registers cannot be used when SSE is disabled");
2965 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2966 DAG.getConstant(NumXMMRegs, dl,
2970 if (isVarArg && IsMustTail) {
2971 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
2972 for (const auto &F : Forwards) {
2973 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2974 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
2978 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2979 // don't need this because the eligibility check rejects calls that require
2980 // shuffling arguments passed in memory.
2981 if (!IsSibcall && isTailCall) {
2982 // Force all the incoming stack arguments to be loaded from the stack
2983 // before any new outgoing arguments are stored to the stack, because the
2984 // outgoing stack slots may alias the incoming argument stack slots, and
2985 // the alias isn't otherwise explicit. This is slightly more conservative
2986 // than necessary, because it means that each store effectively depends
2987 // on every argument instead of just those arguments it would clobber.
2988 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2990 SmallVector<SDValue, 8> MemOpChains2;
2993 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2994 CCValAssign &VA = ArgLocs[i];
2997 assert(VA.isMemLoc());
2998 SDValue Arg = OutVals[i];
2999 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3000 // Skip inalloca arguments. They don't require any work.
3001 if (Flags.isInAlloca())
3003 // Create frame index.
3004 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3005 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3006 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3007 FIN = DAG.getFrameIndex(FI, getPointerTy());
3009 if (Flags.isByVal()) {
3010 // Copy relative to framepointer.
3011 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3012 if (!StackPtr.getNode())
3013 StackPtr = DAG.getCopyFromReg(Chain, dl,
3014 RegInfo->getStackRegister(),
3016 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3018 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3022 // Store relative to framepointer.
3023 MemOpChains2.push_back(
3024 DAG.getStore(ArgChain, dl, Arg, FIN,
3025 MachinePointerInfo::getFixedStack(FI),
3030 if (!MemOpChains2.empty())
3031 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3033 // Store the return address to the appropriate stack slot.
3034 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3035 getPointerTy(), RegInfo->getSlotSize(),
3039 // Build a sequence of copy-to-reg nodes chained together with token chain
3040 // and flag operands which copy the outgoing args into registers.
3042 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3043 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3044 RegsToPass[i].second, InFlag);
3045 InFlag = Chain.getValue(1);
3048 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3049 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3050 // In the 64-bit large code model, we have to make all calls
3051 // through a register, since the call instruction's 32-bit
3052 // pc-relative offset may not be large enough to hold the whole
3054 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3055 // If the callee is a GlobalAddress node (quite common, every direct call
3056 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3058 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3060 // We should use extra load for direct calls to dllimported functions in
3062 const GlobalValue *GV = G->getGlobal();
3063 if (!GV->hasDLLImportStorageClass()) {
3064 unsigned char OpFlags = 0;
3065 bool ExtraLoad = false;
3066 unsigned WrapperKind = ISD::DELETED_NODE;
3068 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3069 // external symbols most go through the PLT in PIC mode. If the symbol
3070 // has hidden or protected visibility, or if it is static or local, then
3071 // we don't need to use the PLT - we can directly call it.
3072 if (Subtarget->isTargetELF() &&
3073 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3074 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3075 OpFlags = X86II::MO_PLT;
3076 } else if (Subtarget->isPICStyleStubAny() &&
3077 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3078 (!Subtarget->getTargetTriple().isMacOSX() ||
3079 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3080 // PC-relative references to external symbols should go through $stub,
3081 // unless we're building with the leopard linker or later, which
3082 // automatically synthesizes these stubs.
3083 OpFlags = X86II::MO_DARWIN_STUB;
3084 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3085 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3086 // If the function is marked as non-lazy, generate an indirect call
3087 // which loads from the GOT directly. This avoids runtime overhead
3088 // at the cost of eager binding (and one extra byte of encoding).
3089 OpFlags = X86II::MO_GOTPCREL;
3090 WrapperKind = X86ISD::WrapperRIP;
3094 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3095 G->getOffset(), OpFlags);
3097 // Add a wrapper if needed.
3098 if (WrapperKind != ISD::DELETED_NODE)
3099 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3100 // Add extra indirection if needed.
3102 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3103 MachinePointerInfo::getGOT(),
3104 false, false, false, 0);
3106 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3107 unsigned char OpFlags = 0;
3109 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3110 // external symbols should go through the PLT.
3111 if (Subtarget->isTargetELF() &&
3112 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3113 OpFlags = X86II::MO_PLT;
3114 } else if (Subtarget->isPICStyleStubAny() &&
3115 (!Subtarget->getTargetTriple().isMacOSX() ||
3116 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3117 // PC-relative references to external symbols should go through $stub,
3118 // unless we're building with the leopard linker or later, which
3119 // automatically synthesizes these stubs.
3120 OpFlags = X86II::MO_DARWIN_STUB;
3123 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3125 } else if (Subtarget->isTarget64BitILP32() &&
3126 Callee->getValueType(0) == MVT::i32) {
3127 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3128 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3131 // Returns a chain & a flag for retval copy to use.
3132 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3133 SmallVector<SDValue, 8> Ops;
3135 if (!IsSibcall && isTailCall) {
3136 Chain = DAG.getCALLSEQ_END(Chain,
3137 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3138 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3139 InFlag = Chain.getValue(1);
3142 Ops.push_back(Chain);
3143 Ops.push_back(Callee);
3146 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3148 // Add argument registers to the end of the list so that they are known live
3150 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3151 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3152 RegsToPass[i].second.getValueType()));
3154 // Add a register mask operand representing the call-preserved registers.
3155 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3156 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
3157 assert(Mask && "Missing call preserved mask for calling convention");
3158 Ops.push_back(DAG.getRegisterMask(Mask));
3160 if (InFlag.getNode())
3161 Ops.push_back(InFlag);
3165 //// If this is the first return lowered for this function, add the regs
3166 //// to the liveout set for the function.
3167 // This isn't right, although it's probably harmless on x86; liveouts
3168 // should be computed from returns not tail calls. Consider a void
3169 // function making a tail call to a function returning int.
3170 MF.getFrameInfo()->setHasTailCall();
3171 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3174 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3175 InFlag = Chain.getValue(1);
3177 // Create the CALLSEQ_END node.
3178 unsigned NumBytesForCalleeToPop;
3179 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3180 DAG.getTarget().Options.GuaranteedTailCallOpt))
3181 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3182 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3183 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3184 SR == StackStructReturn)
3185 // If this is a call to a struct-return function, the callee
3186 // pops the hidden struct pointer, so we have to push it back.
3187 // This is common for Darwin/X86, Linux & Mingw32 targets.
3188 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3189 NumBytesForCalleeToPop = 4;
3191 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3193 // Returns a flag for retval copy to use.
3195 Chain = DAG.getCALLSEQ_END(Chain,
3196 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3197 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3200 InFlag = Chain.getValue(1);
3203 // Handle result values, copying them out of physregs into vregs that we
3205 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3206 Ins, dl, DAG, InVals);
3209 //===----------------------------------------------------------------------===//
3210 // Fast Calling Convention (tail call) implementation
3211 //===----------------------------------------------------------------------===//
3213 // Like std call, callee cleans arguments, convention except that ECX is
3214 // reserved for storing the tail called function address. Only 2 registers are
3215 // free for argument passing (inreg). Tail call optimization is performed
3217 // * tailcallopt is enabled
3218 // * caller/callee are fastcc
3219 // On X86_64 architecture with GOT-style position independent code only local
3220 // (within module) calls are supported at the moment.
3221 // To keep the stack aligned according to platform abi the function
3222 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3223 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3224 // If a tail called function callee has more arguments than the caller the
3225 // caller needs to make sure that there is room to move the RETADDR to. This is
3226 // achieved by reserving an area the size of the argument delta right after the
3227 // original RETADDR, but before the saved framepointer or the spilled registers
3228 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3240 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3241 /// for a 16 byte align requirement.
3243 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3244 SelectionDAG& DAG) const {
3245 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3246 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3247 unsigned StackAlignment = TFI.getStackAlignment();
3248 uint64_t AlignMask = StackAlignment - 1;
3249 int64_t Offset = StackSize;
3250 unsigned SlotSize = RegInfo->getSlotSize();
3251 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3252 // Number smaller than 12 so just add the difference.
3253 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3255 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3256 Offset = ((~AlignMask) & Offset) + StackAlignment +
3257 (StackAlignment-SlotSize);
3262 /// MatchingStackOffset - Return true if the given stack call argument is
3263 /// already available in the same position (relatively) of the caller's
3264 /// incoming argument stack.
3266 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3267 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3268 const X86InstrInfo *TII) {
3269 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3271 if (Arg.getOpcode() == ISD::CopyFromReg) {
3272 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3273 if (!TargetRegisterInfo::isVirtualRegister(VR))
3275 MachineInstr *Def = MRI->getVRegDef(VR);
3278 if (!Flags.isByVal()) {
3279 if (!TII->isLoadFromStackSlot(Def, FI))
3282 unsigned Opcode = Def->getOpcode();
3283 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3284 Opcode == X86::LEA64_32r) &&
3285 Def->getOperand(1).isFI()) {
3286 FI = Def->getOperand(1).getIndex();
3287 Bytes = Flags.getByValSize();
3291 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3292 if (Flags.isByVal())
3293 // ByVal argument is passed in as a pointer but it's now being
3294 // dereferenced. e.g.
3295 // define @foo(%struct.X* %A) {
3296 // tail call @bar(%struct.X* byval %A)
3299 SDValue Ptr = Ld->getBasePtr();
3300 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3303 FI = FINode->getIndex();
3304 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3305 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3306 FI = FINode->getIndex();
3307 Bytes = Flags.getByValSize();
3311 assert(FI != INT_MAX);
3312 if (!MFI->isFixedObjectIndex(FI))
3314 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3317 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3318 /// for tail call optimization. Targets which want to do tail call
3319 /// optimization should implement this function.
3321 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3322 CallingConv::ID CalleeCC,
3324 bool isCalleeStructRet,
3325 bool isCallerStructRet,
3327 const SmallVectorImpl<ISD::OutputArg> &Outs,
3328 const SmallVectorImpl<SDValue> &OutVals,
3329 const SmallVectorImpl<ISD::InputArg> &Ins,
3330 SelectionDAG &DAG) const {
3331 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3334 // If -tailcallopt is specified, make fastcc functions tail-callable.
3335 const MachineFunction &MF = DAG.getMachineFunction();
3336 const Function *CallerF = MF.getFunction();
3338 // If the function return type is x86_fp80 and the callee return type is not,
3339 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3340 // perform a tailcall optimization here.
3341 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3344 CallingConv::ID CallerCC = CallerF->getCallingConv();
3345 bool CCMatch = CallerCC == CalleeCC;
3346 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3347 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3349 // Win64 functions have extra shadow space for argument homing. Don't do the
3350 // sibcall if the caller and callee have mismatched expectations for this
3352 if (IsCalleeWin64 != IsCallerWin64)
3355 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3356 if (IsTailCallConvention(CalleeCC) && CCMatch)
3361 // Look for obvious safe cases to perform tail call optimization that do not
3362 // require ABI changes. This is what gcc calls sibcall.
3364 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3365 // emit a special epilogue.
3366 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3367 if (RegInfo->needsStackRealignment(MF))
3370 // Also avoid sibcall optimization if either caller or callee uses struct
3371 // return semantics.
3372 if (isCalleeStructRet || isCallerStructRet)
3375 // An stdcall/thiscall caller is expected to clean up its arguments; the
3376 // callee isn't going to do that.
3377 // FIXME: this is more restrictive than needed. We could produce a tailcall
3378 // when the stack adjustment matches. For example, with a thiscall that takes
3379 // only one argument.
3380 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3381 CallerCC == CallingConv::X86_ThisCall))
3384 // Do not sibcall optimize vararg calls unless all arguments are passed via
3386 if (isVarArg && !Outs.empty()) {
3388 // Optimizing for varargs on Win64 is unlikely to be safe without
3389 // additional testing.
3390 if (IsCalleeWin64 || IsCallerWin64)
3393 SmallVector<CCValAssign, 16> ArgLocs;
3394 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3397 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3398 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3399 if (!ArgLocs[i].isRegLoc())
3403 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3404 // stack. Therefore, if it's not used by the call it is not safe to optimize
3405 // this into a sibcall.
3406 bool Unused = false;
3407 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3414 SmallVector<CCValAssign, 16> RVLocs;
3415 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3417 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3418 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3419 CCValAssign &VA = RVLocs[i];
3420 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3425 // If the calling conventions do not match, then we'd better make sure the
3426 // results are returned in the same way as what the caller expects.
3428 SmallVector<CCValAssign, 16> RVLocs1;
3429 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3431 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3433 SmallVector<CCValAssign, 16> RVLocs2;
3434 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3436 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3438 if (RVLocs1.size() != RVLocs2.size())
3440 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3441 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3443 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3445 if (RVLocs1[i].isRegLoc()) {
3446 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3449 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3455 // If the callee takes no arguments then go on to check the results of the
3457 if (!Outs.empty()) {
3458 // Check if stack adjustment is needed. For now, do not do this if any
3459 // argument is passed on the stack.
3460 SmallVector<CCValAssign, 16> ArgLocs;
3461 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3464 // Allocate shadow area for Win64
3466 CCInfo.AllocateStack(32, 8);
3468 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3469 if (CCInfo.getNextStackOffset()) {
3470 MachineFunction &MF = DAG.getMachineFunction();
3471 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3474 // Check if the arguments are already laid out in the right way as
3475 // the caller's fixed stack objects.
3476 MachineFrameInfo *MFI = MF.getFrameInfo();
3477 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3478 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3479 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3480 CCValAssign &VA = ArgLocs[i];
3481 SDValue Arg = OutVals[i];
3482 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3483 if (VA.getLocInfo() == CCValAssign::Indirect)
3485 if (!VA.isRegLoc()) {
3486 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3493 // If the tailcall address may be in a register, then make sure it's
3494 // possible to register allocate for it. In 32-bit, the call address can
3495 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3496 // callee-saved registers are restored. These happen to be the same
3497 // registers used to pass 'inreg' arguments so watch out for those.
3498 if (!Subtarget->is64Bit() &&
3499 ((!isa<GlobalAddressSDNode>(Callee) &&
3500 !isa<ExternalSymbolSDNode>(Callee)) ||
3501 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3502 unsigned NumInRegs = 0;
3503 // In PIC we need an extra register to formulate the address computation
3505 unsigned MaxInRegs =
3506 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3508 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3509 CCValAssign &VA = ArgLocs[i];
3512 unsigned Reg = VA.getLocReg();
3515 case X86::EAX: case X86::EDX: case X86::ECX:
3516 if (++NumInRegs == MaxInRegs)
3528 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3529 const TargetLibraryInfo *libInfo) const {
3530 return X86::createFastISel(funcInfo, libInfo);
3533 //===----------------------------------------------------------------------===//
3534 // Other Lowering Hooks
3535 //===----------------------------------------------------------------------===//
3537 static bool MayFoldLoad(SDValue Op) {
3538 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3541 static bool MayFoldIntoStore(SDValue Op) {
3542 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3545 static bool isTargetShuffle(unsigned Opcode) {
3547 default: return false;
3548 case X86ISD::BLENDI:
3549 case X86ISD::PSHUFB:
3550 case X86ISD::PSHUFD:
3551 case X86ISD::PSHUFHW:
3552 case X86ISD::PSHUFLW:
3554 case X86ISD::PALIGNR:
3555 case X86ISD::MOVLHPS:
3556 case X86ISD::MOVLHPD:
3557 case X86ISD::MOVHLPS:
3558 case X86ISD::MOVLPS:
3559 case X86ISD::MOVLPD:
3560 case X86ISD::MOVSHDUP:
3561 case X86ISD::MOVSLDUP:
3562 case X86ISD::MOVDDUP:
3565 case X86ISD::UNPCKL:
3566 case X86ISD::UNPCKH:
3567 case X86ISD::VPERMILPI:
3568 case X86ISD::VPERM2X128:
3569 case X86ISD::VPERMI:
3574 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3575 SDValue V1, unsigned TargetMask,
3576 SelectionDAG &DAG) {
3578 default: llvm_unreachable("Unknown x86 shuffle node");
3579 case X86ISD::PSHUFD:
3580 case X86ISD::PSHUFHW:
3581 case X86ISD::PSHUFLW:
3582 case X86ISD::VPERMILPI:
3583 case X86ISD::VPERMI:
3584 return DAG.getNode(Opc, dl, VT, V1,
3585 DAG.getConstant(TargetMask, dl, MVT::i8));
3589 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3590 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3592 default: llvm_unreachable("Unknown x86 shuffle node");
3593 case X86ISD::MOVLHPS:
3594 case X86ISD::MOVLHPD:
3595 case X86ISD::MOVHLPS:
3596 case X86ISD::MOVLPS:
3597 case X86ISD::MOVLPD:
3600 case X86ISD::UNPCKL:
3601 case X86ISD::UNPCKH:
3602 return DAG.getNode(Opc, dl, VT, V1, V2);
3606 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3607 MachineFunction &MF = DAG.getMachineFunction();
3608 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3609 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3610 int ReturnAddrIndex = FuncInfo->getRAIndex();
3612 if (ReturnAddrIndex == 0) {
3613 // Set up a frame object for the return address.
3614 unsigned SlotSize = RegInfo->getSlotSize();
3615 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3618 FuncInfo->setRAIndex(ReturnAddrIndex);
3621 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3624 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3625 bool hasSymbolicDisplacement) {
3626 // Offset should fit into 32 bit immediate field.
3627 if (!isInt<32>(Offset))
3630 // If we don't have a symbolic displacement - we don't have any extra
3632 if (!hasSymbolicDisplacement)
3635 // FIXME: Some tweaks might be needed for medium code model.
3636 if (M != CodeModel::Small && M != CodeModel::Kernel)
3639 // For small code model we assume that latest object is 16MB before end of 31
3640 // bits boundary. We may also accept pretty large negative constants knowing
3641 // that all objects are in the positive half of address space.
3642 if (M == CodeModel::Small && Offset < 16*1024*1024)
3645 // For kernel code model we know that all object resist in the negative half
3646 // of 32bits address space. We may not accept negative offsets, since they may
3647 // be just off and we may accept pretty large positive ones.
3648 if (M == CodeModel::Kernel && Offset >= 0)
3654 /// isCalleePop - Determines whether the callee is required to pop its
3655 /// own arguments. Callee pop is necessary to support tail calls.
3656 bool X86::isCalleePop(CallingConv::ID CallingConv,
3657 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3658 switch (CallingConv) {
3661 case CallingConv::X86_StdCall:
3662 case CallingConv::X86_FastCall:
3663 case CallingConv::X86_ThisCall:
3665 case CallingConv::Fast:
3666 case CallingConv::GHC:
3667 case CallingConv::HiPE:
3674 /// \brief Return true if the condition is an unsigned comparison operation.
3675 static bool isX86CCUnsigned(unsigned X86CC) {
3677 default: llvm_unreachable("Invalid integer condition!");
3678 case X86::COND_E: return true;
3679 case X86::COND_G: return false;
3680 case X86::COND_GE: return false;
3681 case X86::COND_L: return false;
3682 case X86::COND_LE: return false;
3683 case X86::COND_NE: return true;
3684 case X86::COND_B: return true;
3685 case X86::COND_A: return true;
3686 case X86::COND_BE: return true;
3687 case X86::COND_AE: return true;
3689 llvm_unreachable("covered switch fell through?!");
3692 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3693 /// specific condition code, returning the condition code and the LHS/RHS of the
3694 /// comparison to make.
3695 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3696 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3698 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3699 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3700 // X > -1 -> X == 0, jump !sign.
3701 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3702 return X86::COND_NS;
3704 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3705 // X < 0 -> X == 0, jump on sign.
3708 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3710 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3711 return X86::COND_LE;
3715 switch (SetCCOpcode) {
3716 default: llvm_unreachable("Invalid integer condition!");
3717 case ISD::SETEQ: return X86::COND_E;
3718 case ISD::SETGT: return X86::COND_G;
3719 case ISD::SETGE: return X86::COND_GE;
3720 case ISD::SETLT: return X86::COND_L;
3721 case ISD::SETLE: return X86::COND_LE;
3722 case ISD::SETNE: return X86::COND_NE;
3723 case ISD::SETULT: return X86::COND_B;
3724 case ISD::SETUGT: return X86::COND_A;
3725 case ISD::SETULE: return X86::COND_BE;
3726 case ISD::SETUGE: return X86::COND_AE;
3730 // First determine if it is required or is profitable to flip the operands.
3732 // If LHS is a foldable load, but RHS is not, flip the condition.
3733 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3734 !ISD::isNON_EXTLoad(RHS.getNode())) {
3735 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3736 std::swap(LHS, RHS);
3739 switch (SetCCOpcode) {
3745 std::swap(LHS, RHS);
3749 // On a floating point condition, the flags are set as follows:
3751 // 0 | 0 | 0 | X > Y
3752 // 0 | 0 | 1 | X < Y
3753 // 1 | 0 | 0 | X == Y
3754 // 1 | 1 | 1 | unordered
3755 switch (SetCCOpcode) {
3756 default: llvm_unreachable("Condcode should be pre-legalized away");
3758 case ISD::SETEQ: return X86::COND_E;
3759 case ISD::SETOLT: // flipped
3761 case ISD::SETGT: return X86::COND_A;
3762 case ISD::SETOLE: // flipped
3764 case ISD::SETGE: return X86::COND_AE;
3765 case ISD::SETUGT: // flipped
3767 case ISD::SETLT: return X86::COND_B;
3768 case ISD::SETUGE: // flipped
3770 case ISD::SETLE: return X86::COND_BE;
3772 case ISD::SETNE: return X86::COND_NE;
3773 case ISD::SETUO: return X86::COND_P;
3774 case ISD::SETO: return X86::COND_NP;
3776 case ISD::SETUNE: return X86::COND_INVALID;
3780 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3781 /// code. Current x86 isa includes the following FP cmov instructions:
3782 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3783 static bool hasFPCMov(unsigned X86CC) {
3799 /// isFPImmLegal - Returns true if the target can instruction select the
3800 /// specified FP immediate natively. If false, the legalizer will
3801 /// materialize the FP immediate as a load from a constant pool.
3802 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3803 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3804 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3810 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3811 ISD::LoadExtType ExtTy,
3813 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3814 // relocation target a movq or addq instruction: don't let the load shrink.
3815 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3816 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3817 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3818 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3822 /// \brief Returns true if it is beneficial to convert a load of a constant
3823 /// to just the constant itself.
3824 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3826 assert(Ty->isIntegerTy());
3828 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3829 if (BitSize == 0 || BitSize > 64)
3834 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3835 unsigned Index) const {
3836 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3839 return (Index == 0 || Index == ResVT.getVectorNumElements());
3842 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3843 // Speculate cttz only if we can directly use TZCNT.
3844 return Subtarget->hasBMI();
3847 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3848 // Speculate ctlz only if we can directly use LZCNT.
3849 return Subtarget->hasLZCNT();
3852 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3853 /// the specified range (L, H].
3854 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3855 return (Val < 0) || (Val >= Low && Val < Hi);
3858 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3859 /// specified value.
3860 static bool isUndefOrEqual(int Val, int CmpVal) {
3861 return (Val < 0 || Val == CmpVal);
3864 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3865 /// from position Pos and ending in Pos+Size, falls within the specified
3866 /// sequential range (Low, Low+Size]. or is undef.
3867 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3868 unsigned Pos, unsigned Size, int Low) {
3869 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3870 if (!isUndefOrEqual(Mask[i], Low))
3875 /// isVEXTRACTIndex - Return true if the specified
3876 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3877 /// suitable for instruction that extract 128 or 256 bit vectors
3878 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3879 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3880 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3883 // The index should be aligned on a vecWidth-bit boundary.
3885 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3887 MVT VT = N->getSimpleValueType(0);
3888 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3889 bool Result = (Index * ElSize) % vecWidth == 0;
3894 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3895 /// operand specifies a subvector insert that is suitable for input to
3896 /// insertion of 128 or 256-bit subvectors
3897 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3898 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3899 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3901 // The index should be aligned on a vecWidth-bit boundary.
3903 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3905 MVT VT = N->getSimpleValueType(0);
3906 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3907 bool Result = (Index * ElSize) % vecWidth == 0;
3912 bool X86::isVINSERT128Index(SDNode *N) {
3913 return isVINSERTIndex(N, 128);
3916 bool X86::isVINSERT256Index(SDNode *N) {
3917 return isVINSERTIndex(N, 256);
3920 bool X86::isVEXTRACT128Index(SDNode *N) {
3921 return isVEXTRACTIndex(N, 128);
3924 bool X86::isVEXTRACT256Index(SDNode *N) {
3925 return isVEXTRACTIndex(N, 256);
3928 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3929 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3930 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3931 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3934 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3936 MVT VecVT = N->getOperand(0).getSimpleValueType();
3937 MVT ElVT = VecVT.getVectorElementType();
3939 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3940 return Index / NumElemsPerChunk;
3943 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3944 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3945 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3946 llvm_unreachable("Illegal insert subvector for VINSERT");
3949 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3951 MVT VecVT = N->getSimpleValueType(0);
3952 MVT ElVT = VecVT.getVectorElementType();
3954 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3955 return Index / NumElemsPerChunk;
3958 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
3959 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3960 /// and VINSERTI128 instructions.
3961 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
3962 return getExtractVEXTRACTImmediate(N, 128);
3965 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
3966 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
3967 /// and VINSERTI64x4 instructions.
3968 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
3969 return getExtractVEXTRACTImmediate(N, 256);
3972 /// getInsertVINSERT128Immediate - Return the appropriate immediate
3973 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3974 /// and VINSERTI128 instructions.
3975 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
3976 return getInsertVINSERTImmediate(N, 128);
3979 /// getInsertVINSERT256Immediate - Return the appropriate immediate
3980 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
3981 /// and VINSERTI64x4 instructions.
3982 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
3983 return getInsertVINSERTImmediate(N, 256);
3986 /// isZero - Returns true if Elt is a constant integer zero
3987 static bool isZero(SDValue V) {
3988 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
3989 return C && C->isNullValue();
3992 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3994 bool X86::isZeroNode(SDValue Elt) {
3997 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
3998 return CFP->getValueAPF().isPosZero();
4002 /// getZeroVector - Returns a vector of specified type with all zero elements.
4004 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4005 SelectionDAG &DAG, SDLoc dl) {
4006 assert(VT.isVector() && "Expected a vector type");
4008 // Always build SSE zero vectors as <4 x i32> bitcasted
4009 // to their dest type. This ensures they get CSE'd.
4011 if (VT.is128BitVector()) { // SSE
4012 if (Subtarget->hasSSE2()) { // SSE2
4013 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4014 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4016 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4017 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4019 } else if (VT.is256BitVector()) { // AVX
4020 if (Subtarget->hasInt256()) { // AVX2
4021 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4022 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4023 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4025 // 256-bit logic and arithmetic instructions in AVX are all
4026 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4027 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4028 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4029 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4031 } else if (VT.is512BitVector()) { // AVX-512
4032 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4033 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4034 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4035 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4036 } else if (VT.getScalarType() == MVT::i1) {
4038 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4039 && "Unexpected vector type");
4040 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4041 && "Unexpected vector type");
4042 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4043 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4044 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4046 llvm_unreachable("Unexpected vector type");
4048 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4051 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4052 SelectionDAG &DAG, SDLoc dl,
4053 unsigned vectorWidth) {
4054 assert((vectorWidth == 128 || vectorWidth == 256) &&
4055 "Unsupported vector width");
4056 EVT VT = Vec.getValueType();
4057 EVT ElVT = VT.getVectorElementType();
4058 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4059 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4060 VT.getVectorNumElements()/Factor);
4062 // Extract from UNDEF is UNDEF.
4063 if (Vec.getOpcode() == ISD::UNDEF)
4064 return DAG.getUNDEF(ResultVT);
4066 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4067 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4069 // This is the index of the first element of the vectorWidth-bit chunk
4071 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4074 // If the input is a buildvector just emit a smaller one.
4075 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4076 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4077 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4080 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4081 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4084 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4085 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4086 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4087 /// instructions or a simple subregister reference. Idx is an index in the
4088 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4089 /// lowering EXTRACT_VECTOR_ELT operations easier.
4090 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4091 SelectionDAG &DAG, SDLoc dl) {
4092 assert((Vec.getValueType().is256BitVector() ||
4093 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4094 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4097 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4098 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4099 SelectionDAG &DAG, SDLoc dl) {
4100 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4101 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4104 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4105 unsigned IdxVal, SelectionDAG &DAG,
4106 SDLoc dl, unsigned vectorWidth) {
4107 assert((vectorWidth == 128 || vectorWidth == 256) &&
4108 "Unsupported vector width");
4109 // Inserting UNDEF is Result
4110 if (Vec.getOpcode() == ISD::UNDEF)
4112 EVT VT = Vec.getValueType();
4113 EVT ElVT = VT.getVectorElementType();
4114 EVT ResultVT = Result.getValueType();
4116 // Insert the relevant vectorWidth bits.
4117 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4119 // This is the index of the first element of the vectorWidth-bit chunk
4121 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4124 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4125 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4128 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4129 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4130 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4131 /// simple superregister reference. Idx is an index in the 128 bits
4132 /// we want. It need not be aligned to a 128-bit boundary. That makes
4133 /// lowering INSERT_VECTOR_ELT operations easier.
4134 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4135 SelectionDAG &DAG, SDLoc dl) {
4136 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4138 // For insertion into the zero index (low half) of a 256-bit vector, it is
4139 // more efficient to generate a blend with immediate instead of an insert*128.
4140 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4141 // extend the subvector to the size of the result vector. Make sure that
4142 // we are not recursing on that node by checking for undef here.
4143 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4144 Result.getOpcode() != ISD::UNDEF) {
4145 EVT ResultVT = Result.getValueType();
4146 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4147 SDValue Undef = DAG.getUNDEF(ResultVT);
4148 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4151 // The blend instruction, and therefore its mask, depend on the data type.
4152 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4153 if (ScalarType.isFloatingPoint()) {
4154 // Choose either vblendps (float) or vblendpd (double).
4155 unsigned ScalarSize = ScalarType.getSizeInBits();
4156 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4157 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4158 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4159 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4162 const X86Subtarget &Subtarget =
4163 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4165 // AVX2 is needed for 256-bit integer blend support.
4166 // Integers must be cast to 32-bit because there is only vpblendd;
4167 // vpblendw can't be used for this because it has a handicapped mask.
4169 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4170 // is still more efficient than using the wrong domain vinsertf128 that
4171 // will be created by InsertSubVector().
4172 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4174 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4175 Vec256 = DAG.getNode(ISD::BITCAST, dl, CastVT, Vec256);
4176 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4177 return DAG.getNode(ISD::BITCAST, dl, ResultVT, Vec256);
4180 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4183 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4184 SelectionDAG &DAG, SDLoc dl) {
4185 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4186 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4189 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4190 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4191 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4192 /// large BUILD_VECTORS.
4193 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4194 unsigned NumElems, SelectionDAG &DAG,
4196 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4197 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4200 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4201 unsigned NumElems, SelectionDAG &DAG,
4203 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4204 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4207 /// getOnesVector - Returns a vector of specified type with all bits set.
4208 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4209 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4210 /// Then bitcast to their original type, ensuring they get CSE'd.
4211 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4213 assert(VT.isVector() && "Expected a vector type");
4215 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4217 if (VT.is256BitVector()) {
4218 if (HasInt256) { // AVX2
4219 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4220 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4222 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4223 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4225 } else if (VT.is128BitVector()) {
4226 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4228 llvm_unreachable("Unexpected vector type");
4230 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4233 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4234 /// operation of specified width.
4235 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4237 unsigned NumElems = VT.getVectorNumElements();
4238 SmallVector<int, 8> Mask;
4239 Mask.push_back(NumElems);
4240 for (unsigned i = 1; i != NumElems; ++i)
4242 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4245 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4246 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4248 unsigned NumElems = VT.getVectorNumElements();
4249 SmallVector<int, 8> Mask;
4250 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4252 Mask.push_back(i + NumElems);
4254 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4257 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4258 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4260 unsigned NumElems = VT.getVectorNumElements();
4261 SmallVector<int, 8> Mask;
4262 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4263 Mask.push_back(i + Half);
4264 Mask.push_back(i + NumElems + Half);
4266 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4269 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4270 /// vector of zero or undef vector. This produces a shuffle where the low
4271 /// element of V2 is swizzled into the zero/undef vector, landing at element
4272 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4273 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4275 const X86Subtarget *Subtarget,
4276 SelectionDAG &DAG) {
4277 MVT VT = V2.getSimpleValueType();
4279 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4280 unsigned NumElems = VT.getVectorNumElements();
4281 SmallVector<int, 16> MaskVec;
4282 for (unsigned i = 0; i != NumElems; ++i)
4283 // If this is the insertion idx, put the low elt of V2 here.
4284 MaskVec.push_back(i == Idx ? NumElems : i);
4285 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4288 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4289 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4290 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4291 /// shuffles which use a single input multiple times, and in those cases it will
4292 /// adjust the mask to only have indices within that single input.
4293 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4294 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4295 unsigned NumElems = VT.getVectorNumElements();
4299 bool IsFakeUnary = false;
4300 switch(N->getOpcode()) {
4301 case X86ISD::BLENDI:
4302 ImmN = N->getOperand(N->getNumOperands()-1);
4303 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4306 ImmN = N->getOperand(N->getNumOperands()-1);
4307 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4308 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4310 case X86ISD::UNPCKH:
4311 DecodeUNPCKHMask(VT, Mask);
4312 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4314 case X86ISD::UNPCKL:
4315 DecodeUNPCKLMask(VT, Mask);
4316 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4318 case X86ISD::MOVHLPS:
4319 DecodeMOVHLPSMask(NumElems, Mask);
4320 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4322 case X86ISD::MOVLHPS:
4323 DecodeMOVLHPSMask(NumElems, Mask);
4324 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4326 case X86ISD::PALIGNR:
4327 ImmN = N->getOperand(N->getNumOperands()-1);
4328 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4330 case X86ISD::PSHUFD:
4331 case X86ISD::VPERMILPI:
4332 ImmN = N->getOperand(N->getNumOperands()-1);
4333 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4336 case X86ISD::PSHUFHW:
4337 ImmN = N->getOperand(N->getNumOperands()-1);
4338 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4341 case X86ISD::PSHUFLW:
4342 ImmN = N->getOperand(N->getNumOperands()-1);
4343 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4346 case X86ISD::PSHUFB: {
4348 SDValue MaskNode = N->getOperand(1);
4349 while (MaskNode->getOpcode() == ISD::BITCAST)
4350 MaskNode = MaskNode->getOperand(0);
4352 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4353 // If we have a build-vector, then things are easy.
4354 EVT VT = MaskNode.getValueType();
4355 assert(VT.isVector() &&
4356 "Can't produce a non-vector with a build_vector!");
4357 if (!VT.isInteger())
4360 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4362 SmallVector<uint64_t, 32> RawMask;
4363 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4364 SDValue Op = MaskNode->getOperand(i);
4365 if (Op->getOpcode() == ISD::UNDEF) {
4366 RawMask.push_back((uint64_t)SM_SentinelUndef);
4369 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4372 APInt MaskElement = CN->getAPIntValue();
4374 // We now have to decode the element which could be any integer size and
4375 // extract each byte of it.
4376 for (int j = 0; j < NumBytesPerElement; ++j) {
4377 // Note that this is x86 and so always little endian: the low byte is
4378 // the first byte of the mask.
4379 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4380 MaskElement = MaskElement.lshr(8);
4383 DecodePSHUFBMask(RawMask, Mask);
4387 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4391 SDValue Ptr = MaskLoad->getBasePtr();
4392 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4393 Ptr->getOpcode() == X86ISD::WrapperRIP)
4394 Ptr = Ptr->getOperand(0);
4396 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4397 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4400 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4401 DecodePSHUFBMask(C, Mask);
4409 case X86ISD::VPERMI:
4410 ImmN = N->getOperand(N->getNumOperands()-1);
4411 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4416 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4418 case X86ISD::VPERM2X128:
4419 ImmN = N->getOperand(N->getNumOperands()-1);
4420 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4421 if (Mask.empty()) return false;
4423 case X86ISD::MOVSLDUP:
4424 DecodeMOVSLDUPMask(VT, Mask);
4427 case X86ISD::MOVSHDUP:
4428 DecodeMOVSHDUPMask(VT, Mask);
4431 case X86ISD::MOVDDUP:
4432 DecodeMOVDDUPMask(VT, Mask);
4435 case X86ISD::MOVLHPD:
4436 case X86ISD::MOVLPD:
4437 case X86ISD::MOVLPS:
4438 // Not yet implemented
4440 default: llvm_unreachable("unknown target shuffle node");
4443 // If we have a fake unary shuffle, the shuffle mask is spread across two
4444 // inputs that are actually the same node. Re-map the mask to always point
4445 // into the first input.
4448 if (M >= (int)Mask.size())
4454 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4455 /// element of the result of the vector shuffle.
4456 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4459 return SDValue(); // Limit search depth.
4461 SDValue V = SDValue(N, 0);
4462 EVT VT = V.getValueType();
4463 unsigned Opcode = V.getOpcode();
4465 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4466 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4467 int Elt = SV->getMaskElt(Index);
4470 return DAG.getUNDEF(VT.getVectorElementType());
4472 unsigned NumElems = VT.getVectorNumElements();
4473 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4474 : SV->getOperand(1);
4475 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4478 // Recurse into target specific vector shuffles to find scalars.
4479 if (isTargetShuffle(Opcode)) {
4480 MVT ShufVT = V.getSimpleValueType();
4481 unsigned NumElems = ShufVT.getVectorNumElements();
4482 SmallVector<int, 16> ShuffleMask;
4485 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4488 int Elt = ShuffleMask[Index];
4490 return DAG.getUNDEF(ShufVT.getVectorElementType());
4492 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4494 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4498 // Actual nodes that may contain scalar elements
4499 if (Opcode == ISD::BITCAST) {
4500 V = V.getOperand(0);
4501 EVT SrcVT = V.getValueType();
4502 unsigned NumElems = VT.getVectorNumElements();
4504 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4508 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4509 return (Index == 0) ? V.getOperand(0)
4510 : DAG.getUNDEF(VT.getVectorElementType());
4512 if (V.getOpcode() == ISD::BUILD_VECTOR)
4513 return V.getOperand(Index);
4518 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4520 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4521 unsigned NumNonZero, unsigned NumZero,
4523 const X86Subtarget* Subtarget,
4524 const TargetLowering &TLI) {
4532 // SSE4.1 - use PINSRB to insert each byte directly.
4533 if (Subtarget->hasSSE41()) {
4534 for (unsigned i = 0; i < 16; ++i) {
4535 bool isNonZero = (NonZeros & (1 << i)) != 0;
4539 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4541 V = DAG.getUNDEF(MVT::v16i8);
4544 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4545 MVT::v16i8, V, Op.getOperand(i),
4546 DAG.getIntPtrConstant(i, dl));
4553 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4554 for (unsigned i = 0; i < 16; ++i) {
4555 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4556 if (ThisIsNonZero && First) {
4558 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4560 V = DAG.getUNDEF(MVT::v8i16);
4565 SDValue ThisElt, LastElt;
4566 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4567 if (LastIsNonZero) {
4568 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4569 MVT::i16, Op.getOperand(i-1));
4571 if (ThisIsNonZero) {
4572 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4573 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4574 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4576 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4580 if (ThisElt.getNode())
4581 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4582 DAG.getIntPtrConstant(i/2, dl));
4586 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4589 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4591 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4592 unsigned NumNonZero, unsigned NumZero,
4594 const X86Subtarget* Subtarget,
4595 const TargetLowering &TLI) {
4602 for (unsigned i = 0; i < 8; ++i) {
4603 bool isNonZero = (NonZeros & (1 << i)) != 0;
4607 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4609 V = DAG.getUNDEF(MVT::v8i16);
4612 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4613 MVT::v8i16, V, Op.getOperand(i),
4614 DAG.getIntPtrConstant(i, dl));
4621 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4622 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4623 const X86Subtarget *Subtarget,
4624 const TargetLowering &TLI) {
4625 // Find all zeroable elements.
4626 std::bitset<4> Zeroable;
4627 for (int i=0; i < 4; ++i) {
4628 SDValue Elt = Op->getOperand(i);
4629 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4631 assert(Zeroable.size() - Zeroable.count() > 1 &&
4632 "We expect at least two non-zero elements!");
4634 // We only know how to deal with build_vector nodes where elements are either
4635 // zeroable or extract_vector_elt with constant index.
4636 SDValue FirstNonZero;
4637 unsigned FirstNonZeroIdx;
4638 for (unsigned i=0; i < 4; ++i) {
4641 SDValue Elt = Op->getOperand(i);
4642 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4643 !isa<ConstantSDNode>(Elt.getOperand(1)))
4645 // Make sure that this node is extracting from a 128-bit vector.
4646 MVT VT = Elt.getOperand(0).getSimpleValueType();
4647 if (!VT.is128BitVector())
4649 if (!FirstNonZero.getNode()) {
4651 FirstNonZeroIdx = i;
4655 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4656 SDValue V1 = FirstNonZero.getOperand(0);
4657 MVT VT = V1.getSimpleValueType();
4659 // See if this build_vector can be lowered as a blend with zero.
4661 unsigned EltMaskIdx, EltIdx;
4663 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4664 if (Zeroable[EltIdx]) {
4665 // The zero vector will be on the right hand side.
4666 Mask[EltIdx] = EltIdx+4;
4670 Elt = Op->getOperand(EltIdx);
4671 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4672 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4673 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4675 Mask[EltIdx] = EltIdx;
4679 // Let the shuffle legalizer deal with blend operations.
4680 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4681 if (V1.getSimpleValueType() != VT)
4682 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4683 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4686 // See if we can lower this build_vector to a INSERTPS.
4687 if (!Subtarget->hasSSE41())
4690 SDValue V2 = Elt.getOperand(0);
4691 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4694 bool CanFold = true;
4695 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4699 SDValue Current = Op->getOperand(i);
4700 SDValue SrcVector = Current->getOperand(0);
4703 CanFold = SrcVector == V1 &&
4704 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4710 assert(V1.getNode() && "Expected at least two non-zero elements!");
4711 if (V1.getSimpleValueType() != MVT::v4f32)
4712 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4713 if (V2.getSimpleValueType() != MVT::v4f32)
4714 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4716 // Ok, we can emit an INSERTPS instruction.
4717 unsigned ZMask = Zeroable.to_ulong();
4719 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4720 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4722 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4723 DAG.getIntPtrConstant(InsertPSMask, DL));
4724 return DAG.getNode(ISD::BITCAST, DL, VT, Result);
4727 /// Return a vector logical shift node.
4728 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4729 unsigned NumBits, SelectionDAG &DAG,
4730 const TargetLowering &TLI, SDLoc dl) {
4731 assert(VT.is128BitVector() && "Unknown type for VShift");
4732 MVT ShVT = MVT::v2i64;
4733 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4734 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4735 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4736 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4737 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4738 return DAG.getNode(ISD::BITCAST, dl, VT,
4739 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4743 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4745 // Check if the scalar load can be widened into a vector load. And if
4746 // the address is "base + cst" see if the cst can be "absorbed" into
4747 // the shuffle mask.
4748 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4749 SDValue Ptr = LD->getBasePtr();
4750 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4752 EVT PVT = LD->getValueType(0);
4753 if (PVT != MVT::i32 && PVT != MVT::f32)
4758 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4759 FI = FINode->getIndex();
4761 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4762 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4763 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4764 Offset = Ptr.getConstantOperandVal(1);
4765 Ptr = Ptr.getOperand(0);
4770 // FIXME: 256-bit vector instructions don't require a strict alignment,
4771 // improve this code to support it better.
4772 unsigned RequiredAlign = VT.getSizeInBits()/8;
4773 SDValue Chain = LD->getChain();
4774 // Make sure the stack object alignment is at least 16 or 32.
4775 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4776 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4777 if (MFI->isFixedObjectIndex(FI)) {
4778 // Can't change the alignment. FIXME: It's possible to compute
4779 // the exact stack offset and reference FI + adjust offset instead.
4780 // If someone *really* cares about this. That's the way to implement it.
4783 MFI->setObjectAlignment(FI, RequiredAlign);
4787 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4788 // Ptr + (Offset & ~15).
4791 if ((Offset % RequiredAlign) & 3)
4793 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4796 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4797 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4800 int EltNo = (Offset - StartOffset) >> 2;
4801 unsigned NumElems = VT.getVectorNumElements();
4803 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4804 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4805 LD->getPointerInfo().getWithOffset(StartOffset),
4806 false, false, false, 0);
4808 SmallVector<int, 8> Mask(NumElems, EltNo);
4810 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4816 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4817 /// elements can be replaced by a single large load which has the same value as
4818 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4820 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4822 /// FIXME: we'd also like to handle the case where the last elements are zero
4823 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4824 /// There's even a handy isZeroNode for that purpose.
4825 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4826 SDLoc &DL, SelectionDAG &DAG,
4827 bool isAfterLegalize) {
4828 unsigned NumElems = Elts.size();
4830 LoadSDNode *LDBase = nullptr;
4831 unsigned LastLoadedElt = -1U;
4833 // For each element in the initializer, see if we've found a load or an undef.
4834 // If we don't find an initial load element, or later load elements are
4835 // non-consecutive, bail out.
4836 for (unsigned i = 0; i < NumElems; ++i) {
4837 SDValue Elt = Elts[i];
4838 // Look through a bitcast.
4839 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4840 Elt = Elt.getOperand(0);
4841 if (!Elt.getNode() ||
4842 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4845 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4847 LDBase = cast<LoadSDNode>(Elt.getNode());
4851 if (Elt.getOpcode() == ISD::UNDEF)
4854 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4855 EVT LdVT = Elt.getValueType();
4856 // Each loaded element must be the correct fractional portion of the
4857 // requested vector load.
4858 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4860 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4865 // If we have found an entire vector of loads and undefs, then return a large
4866 // load of the entire vector width starting at the base pointer. If we found
4867 // consecutive loads for the low half, generate a vzext_load node.
4868 if (LastLoadedElt == NumElems - 1) {
4869 assert(LDBase && "Did not find base load for merging consecutive loads");
4870 EVT EltVT = LDBase->getValueType(0);
4871 // Ensure that the input vector size for the merged loads matches the
4872 // cumulative size of the input elements.
4873 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4876 if (isAfterLegalize &&
4877 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4880 SDValue NewLd = SDValue();
4882 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4883 LDBase->getPointerInfo(), LDBase->isVolatile(),
4884 LDBase->isNonTemporal(), LDBase->isInvariant(),
4885 LDBase->getAlignment());
4887 if (LDBase->hasAnyUseOfValue(1)) {
4888 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4890 SDValue(NewLd.getNode(), 1));
4891 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4892 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4893 SDValue(NewLd.getNode(), 1));
4899 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4900 //of a v4i32 / v4f32. It's probably worth generalizing.
4901 EVT EltVT = VT.getVectorElementType();
4902 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4903 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4904 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4905 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4907 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4908 LDBase->getPointerInfo(),
4909 LDBase->getAlignment(),
4910 false/*isVolatile*/, true/*ReadMem*/,
4913 // Make sure the newly-created LOAD is in the same position as LDBase in
4914 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4915 // update uses of LDBase's output chain to use the TokenFactor.
4916 if (LDBase->hasAnyUseOfValue(1)) {
4917 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4918 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4919 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4920 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4921 SDValue(ResNode.getNode(), 1));
4924 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4929 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4930 /// to generate a splat value for the following cases:
4931 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4932 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4933 /// a scalar load, or a constant.
4934 /// The VBROADCAST node is returned when a pattern is found,
4935 /// or SDValue() otherwise.
4936 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4937 SelectionDAG &DAG) {
4938 // VBROADCAST requires AVX.
4939 // TODO: Splats could be generated for non-AVX CPUs using SSE
4940 // instructions, but there's less potential gain for only 128-bit vectors.
4941 if (!Subtarget->hasAVX())
4944 MVT VT = Op.getSimpleValueType();
4947 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4948 "Unsupported vector type for broadcast.");
4953 switch (Op.getOpcode()) {
4955 // Unknown pattern found.
4958 case ISD::BUILD_VECTOR: {
4959 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4960 BitVector UndefElements;
4961 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4963 // We need a splat of a single value to use broadcast, and it doesn't
4964 // make any sense if the value is only in one element of the vector.
4965 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4969 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4970 Ld.getOpcode() == ISD::ConstantFP);
4972 // Make sure that all of the users of a non-constant load are from the
4973 // BUILD_VECTOR node.
4974 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4979 case ISD::VECTOR_SHUFFLE: {
4980 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4982 // Shuffles must have a splat mask where the first element is
4984 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4987 SDValue Sc = Op.getOperand(0);
4988 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4989 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4991 if (!Subtarget->hasInt256())
4994 // Use the register form of the broadcast instruction available on AVX2.
4995 if (VT.getSizeInBits() >= 256)
4996 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4997 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5000 Ld = Sc.getOperand(0);
5001 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5002 Ld.getOpcode() == ISD::ConstantFP);
5004 // The scalar_to_vector node and the suspected
5005 // load node must have exactly one user.
5006 // Constants may have multiple users.
5008 // AVX-512 has register version of the broadcast
5009 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5010 Ld.getValueType().getSizeInBits() >= 32;
5011 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5018 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5019 bool IsGE256 = (VT.getSizeInBits() >= 256);
5021 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5022 // instruction to save 8 or more bytes of constant pool data.
5023 // TODO: If multiple splats are generated to load the same constant,
5024 // it may be detrimental to overall size. There needs to be a way to detect
5025 // that condition to know if this is truly a size win.
5026 const Function *F = DAG.getMachineFunction().getFunction();
5027 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
5029 // Handle broadcasting a single constant scalar from the constant pool
5031 // On Sandybridge (no AVX2), it is still better to load a constant vector
5032 // from the constant pool and not to broadcast it from a scalar.
5033 // But override that restriction when optimizing for size.
5034 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5035 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5036 EVT CVT = Ld.getValueType();
5037 assert(!CVT.isVector() && "Must not broadcast a vector type");
5039 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5040 // For size optimization, also splat v2f64 and v2i64, and for size opt
5041 // with AVX2, also splat i8 and i16.
5042 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5043 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5044 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5045 const Constant *C = nullptr;
5046 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5047 C = CI->getConstantIntValue();
5048 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5049 C = CF->getConstantFPValue();
5051 assert(C && "Invalid constant type");
5053 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5054 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5055 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5056 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5057 MachinePointerInfo::getConstantPool(),
5058 false, false, false, Alignment);
5060 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5064 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5066 // Handle AVX2 in-register broadcasts.
5067 if (!IsLoad && Subtarget->hasInt256() &&
5068 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5069 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5071 // The scalar source must be a normal load.
5075 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5076 (Subtarget->hasVLX() && ScalarSize == 64))
5077 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5079 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5080 // double since there is no vbroadcastsd xmm
5081 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5082 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5083 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5086 // Unsupported broadcast.
5090 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5091 /// underlying vector and index.
5093 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5095 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5097 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5098 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5101 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5103 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5105 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5106 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5109 // In this case the vector is the extract_subvector expression and the index
5110 // is 2, as specified by the shuffle.
5111 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5112 SDValue ShuffleVec = SVOp->getOperand(0);
5113 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5114 assert(ShuffleVecVT.getVectorElementType() ==
5115 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5117 int ShuffleIdx = SVOp->getMaskElt(Idx);
5118 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5119 ExtractedFromVec = ShuffleVec;
5125 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5126 MVT VT = Op.getSimpleValueType();
5128 // Skip if insert_vec_elt is not supported.
5129 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5130 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5134 unsigned NumElems = Op.getNumOperands();
5138 SmallVector<unsigned, 4> InsertIndices;
5139 SmallVector<int, 8> Mask(NumElems, -1);
5141 for (unsigned i = 0; i != NumElems; ++i) {
5142 unsigned Opc = Op.getOperand(i).getOpcode();
5144 if (Opc == ISD::UNDEF)
5147 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5148 // Quit if more than 1 elements need inserting.
5149 if (InsertIndices.size() > 1)
5152 InsertIndices.push_back(i);
5156 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5157 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5158 // Quit if non-constant index.
5159 if (!isa<ConstantSDNode>(ExtIdx))
5161 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5163 // Quit if extracted from vector of different type.
5164 if (ExtractedFromVec.getValueType() != VT)
5167 if (!VecIn1.getNode())
5168 VecIn1 = ExtractedFromVec;
5169 else if (VecIn1 != ExtractedFromVec) {
5170 if (!VecIn2.getNode())
5171 VecIn2 = ExtractedFromVec;
5172 else if (VecIn2 != ExtractedFromVec)
5173 // Quit if more than 2 vectors to shuffle
5177 if (ExtractedFromVec == VecIn1)
5179 else if (ExtractedFromVec == VecIn2)
5180 Mask[i] = Idx + NumElems;
5183 if (!VecIn1.getNode())
5186 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5187 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5188 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5189 unsigned Idx = InsertIndices[i];
5190 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5191 DAG.getIntPtrConstant(Idx, DL));
5197 static SDValue ConvertI1VectorToInterger(SDValue Op, SelectionDAG &DAG) {
5198 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5199 Op.getScalarValueSizeInBits() == 1 &&
5200 "Can not convert non-constant vector");
5201 uint64_t Immediate = 0;
5202 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5203 SDValue In = Op.getOperand(idx);
5204 if (In.getOpcode() != ISD::UNDEF)
5205 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5209 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5210 return DAG.getConstant(Immediate, dl, VT);
5212 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5214 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5216 MVT VT = Op.getSimpleValueType();
5217 assert((VT.getVectorElementType() == MVT::i1) &&
5218 "Unexpected type in LowerBUILD_VECTORvXi1!");
5221 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5222 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5223 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5224 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5227 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5228 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5229 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5230 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5233 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5234 SDValue Imm = ConvertI1VectorToInterger(Op, DAG);
5235 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5236 return DAG.getNode(ISD::BITCAST, dl, VT, Imm);
5237 SDValue ExtVec = DAG.getNode(ISD::BITCAST, dl, MVT::v8i1, Imm);
5238 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5239 DAG.getIntPtrConstant(0, dl));
5242 // Vector has one or more non-const elements
5243 uint64_t Immediate = 0;
5244 SmallVector<unsigned, 16> NonConstIdx;
5245 bool IsSplat = true;
5246 bool HasConstElts = false;
5248 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5249 SDValue In = Op.getOperand(idx);
5250 if (In.getOpcode() == ISD::UNDEF)
5252 if (!isa<ConstantSDNode>(In))
5253 NonConstIdx.push_back(idx);
5255 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5256 HasConstElts = true;
5260 else if (In != Op.getOperand(SplatIdx))
5264 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5266 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5267 DAG.getConstant(1, dl, VT),
5268 DAG.getConstant(0, dl, VT));
5270 // insert elements one by one
5274 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5275 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5277 else if (HasConstElts)
5278 Imm = DAG.getConstant(0, dl, VT);
5280 Imm = DAG.getUNDEF(VT);
5281 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5282 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, Imm);
5284 SDValue ExtVec = DAG.getNode(ISD::BITCAST, dl, MVT::v8i1, Imm);
5285 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5286 DAG.getIntPtrConstant(0, dl));
5289 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5290 unsigned InsertIdx = NonConstIdx[i];
5291 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5292 Op.getOperand(InsertIdx),
5293 DAG.getIntPtrConstant(InsertIdx, dl));
5298 /// \brief Return true if \p N implements a horizontal binop and return the
5299 /// operands for the horizontal binop into V0 and V1.
5301 /// This is a helper function of LowerToHorizontalOp().
5302 /// This function checks that the build_vector \p N in input implements a
5303 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5304 /// operation to match.
5305 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5306 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5307 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5310 /// This function only analyzes elements of \p N whose indices are
5311 /// in range [BaseIdx, LastIdx).
5312 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5314 unsigned BaseIdx, unsigned LastIdx,
5315 SDValue &V0, SDValue &V1) {
5316 EVT VT = N->getValueType(0);
5318 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5319 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5320 "Invalid Vector in input!");
5322 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5323 bool CanFold = true;
5324 unsigned ExpectedVExtractIdx = BaseIdx;
5325 unsigned NumElts = LastIdx - BaseIdx;
5326 V0 = DAG.getUNDEF(VT);
5327 V1 = DAG.getUNDEF(VT);
5329 // Check if N implements a horizontal binop.
5330 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5331 SDValue Op = N->getOperand(i + BaseIdx);
5334 if (Op->getOpcode() == ISD::UNDEF) {
5335 // Update the expected vector extract index.
5336 if (i * 2 == NumElts)
5337 ExpectedVExtractIdx = BaseIdx;
5338 ExpectedVExtractIdx += 2;
5342 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5347 SDValue Op0 = Op.getOperand(0);
5348 SDValue Op1 = Op.getOperand(1);
5350 // Try to match the following pattern:
5351 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5352 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5353 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5354 Op0.getOperand(0) == Op1.getOperand(0) &&
5355 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5356 isa<ConstantSDNode>(Op1.getOperand(1)));
5360 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5361 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5363 if (i * 2 < NumElts) {
5364 if (V0.getOpcode() == ISD::UNDEF) {
5365 V0 = Op0.getOperand(0);
5366 if (V0.getValueType() != VT)
5370 if (V1.getOpcode() == ISD::UNDEF) {
5371 V1 = Op0.getOperand(0);
5372 if (V1.getValueType() != VT)
5375 if (i * 2 == NumElts)
5376 ExpectedVExtractIdx = BaseIdx;
5379 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5380 if (I0 == ExpectedVExtractIdx)
5381 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5382 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5383 // Try to match the following dag sequence:
5384 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5385 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5389 ExpectedVExtractIdx += 2;
5395 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5396 /// a concat_vector.
5398 /// This is a helper function of LowerToHorizontalOp().
5399 /// This function expects two 256-bit vectors called V0 and V1.
5400 /// At first, each vector is split into two separate 128-bit vectors.
5401 /// Then, the resulting 128-bit vectors are used to implement two
5402 /// horizontal binary operations.
5404 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5406 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5407 /// the two new horizontal binop.
5408 /// When Mode is set, the first horizontal binop dag node would take as input
5409 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5410 /// horizontal binop dag node would take as input the lower 128-bit of V1
5411 /// and the upper 128-bit of V1.
5413 /// HADD V0_LO, V0_HI
5414 /// HADD V1_LO, V1_HI
5416 /// Otherwise, the first horizontal binop dag node takes as input the lower
5417 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5418 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5420 /// HADD V0_LO, V1_LO
5421 /// HADD V0_HI, V1_HI
5423 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5424 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5425 /// the upper 128-bits of the result.
5426 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5427 SDLoc DL, SelectionDAG &DAG,
5428 unsigned X86Opcode, bool Mode,
5429 bool isUndefLO, bool isUndefHI) {
5430 EVT VT = V0.getValueType();
5431 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5432 "Invalid nodes in input!");
5434 unsigned NumElts = VT.getVectorNumElements();
5435 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5436 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5437 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5438 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5439 EVT NewVT = V0_LO.getValueType();
5441 SDValue LO = DAG.getUNDEF(NewVT);
5442 SDValue HI = DAG.getUNDEF(NewVT);
5445 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5446 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5447 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5448 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5449 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5451 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5452 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5453 V1_LO->getOpcode() != ISD::UNDEF))
5454 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5456 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5457 V1_HI->getOpcode() != ISD::UNDEF))
5458 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5461 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5464 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5466 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5467 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5468 EVT VT = BV->getValueType(0);
5469 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5470 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5474 unsigned NumElts = VT.getVectorNumElements();
5475 SDValue InVec0 = DAG.getUNDEF(VT);
5476 SDValue InVec1 = DAG.getUNDEF(VT);
5478 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5479 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5481 // Odd-numbered elements in the input build vector are obtained from
5482 // adding two integer/float elements.
5483 // Even-numbered elements in the input build vector are obtained from
5484 // subtracting two integer/float elements.
5485 unsigned ExpectedOpcode = ISD::FSUB;
5486 unsigned NextExpectedOpcode = ISD::FADD;
5487 bool AddFound = false;
5488 bool SubFound = false;
5490 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5491 SDValue Op = BV->getOperand(i);
5493 // Skip 'undef' values.
5494 unsigned Opcode = Op.getOpcode();
5495 if (Opcode == ISD::UNDEF) {
5496 std::swap(ExpectedOpcode, NextExpectedOpcode);
5500 // Early exit if we found an unexpected opcode.
5501 if (Opcode != ExpectedOpcode)
5504 SDValue Op0 = Op.getOperand(0);
5505 SDValue Op1 = Op.getOperand(1);
5507 // Try to match the following pattern:
5508 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5509 // Early exit if we cannot match that sequence.
5510 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5511 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5512 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5513 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5514 Op0.getOperand(1) != Op1.getOperand(1))
5517 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5521 // We found a valid add/sub node. Update the information accordingly.
5527 // Update InVec0 and InVec1.
5528 if (InVec0.getOpcode() == ISD::UNDEF) {
5529 InVec0 = Op0.getOperand(0);
5530 if (InVec0.getValueType() != VT)
5533 if (InVec1.getOpcode() == ISD::UNDEF) {
5534 InVec1 = Op1.getOperand(0);
5535 if (InVec1.getValueType() != VT)
5539 // Make sure that operands in input to each add/sub node always
5540 // come from a same pair of vectors.
5541 if (InVec0 != Op0.getOperand(0)) {
5542 if (ExpectedOpcode == ISD::FSUB)
5545 // FADD is commutable. Try to commute the operands
5546 // and then test again.
5547 std::swap(Op0, Op1);
5548 if (InVec0 != Op0.getOperand(0))
5552 if (InVec1 != Op1.getOperand(0))
5555 // Update the pair of expected opcodes.
5556 std::swap(ExpectedOpcode, NextExpectedOpcode);
5559 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5560 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5561 InVec1.getOpcode() != ISD::UNDEF)
5562 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5567 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5568 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5569 const X86Subtarget *Subtarget,
5570 SelectionDAG &DAG) {
5571 EVT VT = BV->getValueType(0);
5572 unsigned NumElts = VT.getVectorNumElements();
5573 unsigned NumUndefsLO = 0;
5574 unsigned NumUndefsHI = 0;
5575 unsigned Half = NumElts/2;
5577 // Count the number of UNDEF operands in the build_vector in input.
5578 for (unsigned i = 0, e = Half; i != e; ++i)
5579 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5582 for (unsigned i = Half, e = NumElts; i != e; ++i)
5583 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5586 // Early exit if this is either a build_vector of all UNDEFs or all the
5587 // operands but one are UNDEF.
5588 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5592 SDValue InVec0, InVec1;
5593 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5594 // Try to match an SSE3 float HADD/HSUB.
5595 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5596 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5598 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5599 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5600 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5601 // Try to match an SSSE3 integer HADD/HSUB.
5602 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5603 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5605 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5606 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5609 if (!Subtarget->hasAVX())
5612 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5613 // Try to match an AVX horizontal add/sub of packed single/double
5614 // precision floating point values from 256-bit vectors.
5615 SDValue InVec2, InVec3;
5616 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5617 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5618 ((InVec0.getOpcode() == ISD::UNDEF ||
5619 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5620 ((InVec1.getOpcode() == ISD::UNDEF ||
5621 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5622 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5624 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5625 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5626 ((InVec0.getOpcode() == ISD::UNDEF ||
5627 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5628 ((InVec1.getOpcode() == ISD::UNDEF ||
5629 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5630 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5631 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5632 // Try to match an AVX2 horizontal add/sub of signed integers.
5633 SDValue InVec2, InVec3;
5635 bool CanFold = true;
5637 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5638 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5639 ((InVec0.getOpcode() == ISD::UNDEF ||
5640 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5641 ((InVec1.getOpcode() == ISD::UNDEF ||
5642 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5643 X86Opcode = X86ISD::HADD;
5644 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5645 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5646 ((InVec0.getOpcode() == ISD::UNDEF ||
5647 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5648 ((InVec1.getOpcode() == ISD::UNDEF ||
5649 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5650 X86Opcode = X86ISD::HSUB;
5655 // Fold this build_vector into a single horizontal add/sub.
5656 // Do this only if the target has AVX2.
5657 if (Subtarget->hasAVX2())
5658 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5660 // Do not try to expand this build_vector into a pair of horizontal
5661 // add/sub if we can emit a pair of scalar add/sub.
5662 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5665 // Convert this build_vector into a pair of horizontal binop followed by
5667 bool isUndefLO = NumUndefsLO == Half;
5668 bool isUndefHI = NumUndefsHI == Half;
5669 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5670 isUndefLO, isUndefHI);
5674 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5675 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5677 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5678 X86Opcode = X86ISD::HADD;
5679 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5680 X86Opcode = X86ISD::HSUB;
5681 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5682 X86Opcode = X86ISD::FHADD;
5683 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5684 X86Opcode = X86ISD::FHSUB;
5688 // Don't try to expand this build_vector into a pair of horizontal add/sub
5689 // if we can simply emit a pair of scalar add/sub.
5690 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5693 // Convert this build_vector into two horizontal add/sub followed by
5695 bool isUndefLO = NumUndefsLO == Half;
5696 bool isUndefHI = NumUndefsHI == Half;
5697 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5698 isUndefLO, isUndefHI);
5705 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5708 MVT VT = Op.getSimpleValueType();
5709 MVT ExtVT = VT.getVectorElementType();
5710 unsigned NumElems = Op.getNumOperands();
5712 // Generate vectors for predicate vectors.
5713 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5714 return LowerBUILD_VECTORvXi1(Op, DAG);
5716 // Vectors containing all zeros can be matched by pxor and xorps later
5717 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5718 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5719 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5720 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5723 return getZeroVector(VT, Subtarget, DAG, dl);
5726 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5727 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5728 // vpcmpeqd on 256-bit vectors.
5729 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5730 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5733 if (!VT.is512BitVector())
5734 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5737 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5738 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5740 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5741 return HorizontalOp;
5742 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5745 unsigned EVTBits = ExtVT.getSizeInBits();
5747 unsigned NumZero = 0;
5748 unsigned NumNonZero = 0;
5749 unsigned NonZeros = 0;
5750 bool IsAllConstants = true;
5751 SmallSet<SDValue, 8> Values;
5752 for (unsigned i = 0; i < NumElems; ++i) {
5753 SDValue Elt = Op.getOperand(i);
5754 if (Elt.getOpcode() == ISD::UNDEF)
5757 if (Elt.getOpcode() != ISD::Constant &&
5758 Elt.getOpcode() != ISD::ConstantFP)
5759 IsAllConstants = false;
5760 if (X86::isZeroNode(Elt))
5763 NonZeros |= (1 << i);
5768 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5769 if (NumNonZero == 0)
5770 return DAG.getUNDEF(VT);
5772 // Special case for single non-zero, non-undef, element.
5773 if (NumNonZero == 1) {
5774 unsigned Idx = countTrailingZeros(NonZeros);
5775 SDValue Item = Op.getOperand(Idx);
5777 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5778 // the value are obviously zero, truncate the value to i32 and do the
5779 // insertion that way. Only do this if the value is non-constant or if the
5780 // value is a constant being inserted into element 0. It is cheaper to do
5781 // a constant pool load than it is to do a movd + shuffle.
5782 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5783 (!IsAllConstants || Idx == 0)) {
5784 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5786 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5787 EVT VecVT = MVT::v4i32;
5789 // Truncate the value (which may itself be a constant) to i32, and
5790 // convert it to a vector with movd (S2V+shuffle to zero extend).
5791 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5792 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5794 ISD::BITCAST, dl, VT,
5795 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5799 // If we have a constant or non-constant insertion into the low element of
5800 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5801 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5802 // depending on what the source datatype is.
5805 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5807 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5808 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5809 if (VT.is512BitVector()) {
5810 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5811 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5812 Item, DAG.getIntPtrConstant(0, dl));
5814 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5815 "Expected an SSE value type!");
5816 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5817 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5818 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5821 // We can't directly insert an i8 or i16 into a vector, so zero extend
5823 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5824 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5825 if (VT.is256BitVector()) {
5826 if (Subtarget->hasAVX()) {
5827 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5828 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5830 // Without AVX, we need to extend to a 128-bit vector and then
5831 // insert into the 256-bit vector.
5832 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5833 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5834 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5837 assert(VT.is128BitVector() && "Expected an SSE value type!");
5838 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5839 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5841 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5845 // Is it a vector logical left shift?
5846 if (NumElems == 2 && Idx == 1 &&
5847 X86::isZeroNode(Op.getOperand(0)) &&
5848 !X86::isZeroNode(Op.getOperand(1))) {
5849 unsigned NumBits = VT.getSizeInBits();
5850 return getVShift(true, VT,
5851 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5852 VT, Op.getOperand(1)),
5853 NumBits/2, DAG, *this, dl);
5856 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5859 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5860 // is a non-constant being inserted into an element other than the low one,
5861 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5862 // movd/movss) to move this into the low element, then shuffle it into
5864 if (EVTBits == 32) {
5865 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5866 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5870 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5871 if (Values.size() == 1) {
5872 if (EVTBits == 32) {
5873 // Instead of a shuffle like this:
5874 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5875 // Check if it's possible to issue this instead.
5876 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5877 unsigned Idx = countTrailingZeros(NonZeros);
5878 SDValue Item = Op.getOperand(Idx);
5879 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5880 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5885 // A vector full of immediates; various special cases are already
5886 // handled, so this is best done with a single constant-pool load.
5890 // For AVX-length vectors, see if we can use a vector load to get all of the
5891 // elements, otherwise build the individual 128-bit pieces and use
5892 // shuffles to put them in place.
5893 if (VT.is256BitVector() || VT.is512BitVector()) {
5894 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5896 // Check for a build vector of consecutive loads.
5897 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5900 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5902 // Build both the lower and upper subvector.
5903 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5904 makeArrayRef(&V[0], NumElems/2));
5905 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5906 makeArrayRef(&V[NumElems / 2], NumElems/2));
5908 // Recreate the wider vector with the lower and upper part.
5909 if (VT.is256BitVector())
5910 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5911 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5914 // Let legalizer expand 2-wide build_vectors.
5915 if (EVTBits == 64) {
5916 if (NumNonZero == 1) {
5917 // One half is zero or undef.
5918 unsigned Idx = countTrailingZeros(NonZeros);
5919 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5920 Op.getOperand(Idx));
5921 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5926 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5927 if (EVTBits == 8 && NumElems == 16)
5928 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5932 if (EVTBits == 16 && NumElems == 8)
5933 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5937 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5938 if (EVTBits == 32 && NumElems == 4)
5939 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
5942 // If element VT is == 32 bits, turn it into a number of shuffles.
5943 SmallVector<SDValue, 8> V(NumElems);
5944 if (NumElems == 4 && NumZero > 0) {
5945 for (unsigned i = 0; i < 4; ++i) {
5946 bool isZero = !(NonZeros & (1 << i));
5948 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5950 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5953 for (unsigned i = 0; i < 2; ++i) {
5954 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5957 V[i] = V[i*2]; // Must be a zero vector.
5960 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5963 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5966 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5971 bool Reverse1 = (NonZeros & 0x3) == 2;
5972 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5976 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5977 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5979 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5982 if (Values.size() > 1 && VT.is128BitVector()) {
5983 // Check for a build vector of consecutive loads.
5984 for (unsigned i = 0; i < NumElems; ++i)
5985 V[i] = Op.getOperand(i);
5987 // Check for elements which are consecutive loads.
5988 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5991 // Check for a build vector from mostly shuffle plus few inserting.
5992 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
5995 // For SSE 4.1, use insertps to put the high elements into the low element.
5996 if (Subtarget->hasSSE41()) {
5998 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5999 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6001 Result = DAG.getUNDEF(VT);
6003 for (unsigned i = 1; i < NumElems; ++i) {
6004 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6005 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6006 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6011 // Otherwise, expand into a number of unpckl*, start by extending each of
6012 // our (non-undef) elements to the full vector width with the element in the
6013 // bottom slot of the vector (which generates no code for SSE).
6014 for (unsigned i = 0; i < NumElems; ++i) {
6015 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6016 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6018 V[i] = DAG.getUNDEF(VT);
6021 // Next, we iteratively mix elements, e.g. for v4f32:
6022 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6023 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6024 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6025 unsigned EltStride = NumElems >> 1;
6026 while (EltStride != 0) {
6027 for (unsigned i = 0; i < EltStride; ++i) {
6028 // If V[i+EltStride] is undef and this is the first round of mixing,
6029 // then it is safe to just drop this shuffle: V[i] is already in the
6030 // right place, the one element (since it's the first round) being
6031 // inserted as undef can be dropped. This isn't safe for successive
6032 // rounds because they will permute elements within both vectors.
6033 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6034 EltStride == NumElems/2)
6037 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6046 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6047 // to create 256-bit vectors from two other 128-bit ones.
6048 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6050 MVT ResVT = Op.getSimpleValueType();
6052 assert((ResVT.is256BitVector() ||
6053 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6055 SDValue V1 = Op.getOperand(0);
6056 SDValue V2 = Op.getOperand(1);
6057 unsigned NumElems = ResVT.getVectorNumElements();
6058 if (ResVT.is256BitVector())
6059 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6061 if (Op.getNumOperands() == 4) {
6062 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6063 ResVT.getVectorNumElements()/2);
6064 SDValue V3 = Op.getOperand(2);
6065 SDValue V4 = Op.getOperand(3);
6066 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6067 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6069 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6072 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6073 const X86Subtarget *Subtarget,
6074 SelectionDAG & DAG) {
6076 MVT ResVT = Op.getSimpleValueType();
6077 unsigned NumOfOperands = Op.getNumOperands();
6079 assert(isPowerOf2_32(NumOfOperands) &&
6080 "Unexpected number of operands in CONCAT_VECTORS");
6082 if (NumOfOperands > 2) {
6083 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6084 ResVT.getVectorNumElements()/2);
6085 SmallVector<SDValue, 2> Ops;
6086 for (unsigned i = 0; i < NumOfOperands/2; i++)
6087 Ops.push_back(Op.getOperand(i));
6088 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6090 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6091 Ops.push_back(Op.getOperand(i));
6092 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6093 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6096 SDValue V1 = Op.getOperand(0);
6097 SDValue V2 = Op.getOperand(1);
6098 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6099 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6101 if (IsZeroV1 && IsZeroV2)
6102 return getZeroVector(ResVT, Subtarget, DAG, dl);
6104 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6105 SDValue Undef = DAG.getUNDEF(ResVT);
6106 unsigned NumElems = ResVT.getVectorNumElements();
6107 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6109 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6110 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6114 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6115 // Zero the upper bits of V1
6116 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6117 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6120 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6123 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6124 const X86Subtarget *Subtarget,
6125 SelectionDAG &DAG) {
6126 MVT VT = Op.getSimpleValueType();
6127 if (VT.getVectorElementType() == MVT::i1)
6128 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6130 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6131 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6132 Op.getNumOperands() == 4)));
6134 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6135 // from two other 128-bit ones.
6137 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6138 return LowerAVXCONCAT_VECTORS(Op, DAG);
6142 //===----------------------------------------------------------------------===//
6143 // Vector shuffle lowering
6145 // This is an experimental code path for lowering vector shuffles on x86. It is
6146 // designed to handle arbitrary vector shuffles and blends, gracefully
6147 // degrading performance as necessary. It works hard to recognize idiomatic
6148 // shuffles and lower them to optimal instruction patterns without leaving
6149 // a framework that allows reasonably efficient handling of all vector shuffle
6151 //===----------------------------------------------------------------------===//
6153 /// \brief Tiny helper function to identify a no-op mask.
6155 /// This is a somewhat boring predicate function. It checks whether the mask
6156 /// array input, which is assumed to be a single-input shuffle mask of the kind
6157 /// used by the X86 shuffle instructions (not a fully general
6158 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6159 /// in-place shuffle are 'no-op's.
6160 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6161 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6162 if (Mask[i] != -1 && Mask[i] != i)
6167 /// \brief Helper function to classify a mask as a single-input mask.
6169 /// This isn't a generic single-input test because in the vector shuffle
6170 /// lowering we canonicalize single inputs to be the first input operand. This
6171 /// means we can more quickly test for a single input by only checking whether
6172 /// an input from the second operand exists. We also assume that the size of
6173 /// mask corresponds to the size of the input vectors which isn't true in the
6174 /// fully general case.
6175 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6177 if (M >= (int)Mask.size())
6182 /// \brief Test whether there are elements crossing 128-bit lanes in this
6185 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6186 /// and we routinely test for these.
6187 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6188 int LaneSize = 128 / VT.getScalarSizeInBits();
6189 int Size = Mask.size();
6190 for (int i = 0; i < Size; ++i)
6191 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6196 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6198 /// This checks a shuffle mask to see if it is performing the same
6199 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6200 /// that it is also not lane-crossing. It may however involve a blend from the
6201 /// same lane of a second vector.
6203 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6204 /// non-trivial to compute in the face of undef lanes. The representation is
6205 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6206 /// entries from both V1 and V2 inputs to the wider mask.
6208 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6209 SmallVectorImpl<int> &RepeatedMask) {
6210 int LaneSize = 128 / VT.getScalarSizeInBits();
6211 RepeatedMask.resize(LaneSize, -1);
6212 int Size = Mask.size();
6213 for (int i = 0; i < Size; ++i) {
6216 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6217 // This entry crosses lanes, so there is no way to model this shuffle.
6220 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6221 if (RepeatedMask[i % LaneSize] == -1)
6222 // This is the first non-undef entry in this slot of a 128-bit lane.
6223 RepeatedMask[i % LaneSize] =
6224 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6225 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6226 // Found a mismatch with the repeated mask.
6232 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6235 /// This is a fast way to test a shuffle mask against a fixed pattern:
6237 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6239 /// It returns true if the mask is exactly as wide as the argument list, and
6240 /// each element of the mask is either -1 (signifying undef) or the value given
6241 /// in the argument.
6242 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6243 ArrayRef<int> ExpectedMask) {
6244 if (Mask.size() != ExpectedMask.size())
6247 int Size = Mask.size();
6249 // If the values are build vectors, we can look through them to find
6250 // equivalent inputs that make the shuffles equivalent.
6251 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6252 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6254 for (int i = 0; i < Size; ++i)
6255 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6256 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6257 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6258 if (!MaskBV || !ExpectedBV ||
6259 MaskBV->getOperand(Mask[i] % Size) !=
6260 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6267 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6269 /// This helper function produces an 8-bit shuffle immediate corresponding to
6270 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6271 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6274 /// NB: We rely heavily on "undef" masks preserving the input lane.
6275 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6276 SelectionDAG &DAG) {
6277 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6278 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6279 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6280 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6281 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6284 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6285 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6286 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6287 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6288 return DAG.getConstant(Imm, DL, MVT::i8);
6291 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6293 /// This is used as a fallback approach when first class blend instructions are
6294 /// unavailable. Currently it is only suitable for integer vectors, but could
6295 /// be generalized for floating point vectors if desirable.
6296 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6297 SDValue V2, ArrayRef<int> Mask,
6298 SelectionDAG &DAG) {
6299 assert(VT.isInteger() && "Only supports integer vector types!");
6300 MVT EltVT = VT.getScalarType();
6301 int NumEltBits = EltVT.getSizeInBits();
6302 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6303 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6305 SmallVector<SDValue, 16> MaskOps;
6306 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6307 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6308 return SDValue(); // Shuffled input!
6309 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6312 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6313 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6314 // We have to cast V2 around.
6315 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6316 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6317 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6318 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6319 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6320 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6323 /// \brief Try to emit a blend instruction for a shuffle.
6325 /// This doesn't do any checks for the availability of instructions for blending
6326 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6327 /// be matched in the backend with the type given. What it does check for is
6328 /// that the shuffle mask is in fact a blend.
6329 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6330 SDValue V2, ArrayRef<int> Mask,
6331 const X86Subtarget *Subtarget,
6332 SelectionDAG &DAG) {
6333 unsigned BlendMask = 0;
6334 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6335 if (Mask[i] >= Size) {
6336 if (Mask[i] != i + Size)
6337 return SDValue(); // Shuffled V2 input!
6338 BlendMask |= 1u << i;
6341 if (Mask[i] >= 0 && Mask[i] != i)
6342 return SDValue(); // Shuffled V1 input!
6344 switch (VT.SimpleTy) {
6349 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6350 DAG.getConstant(BlendMask, DL, MVT::i8));
6354 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6358 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6359 // that instruction.
6360 if (Subtarget->hasAVX2()) {
6361 // Scale the blend by the number of 32-bit dwords per element.
6362 int Scale = VT.getScalarSizeInBits() / 32;
6364 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6365 if (Mask[i] >= Size)
6366 for (int j = 0; j < Scale; ++j)
6367 BlendMask |= 1u << (i * Scale + j);
6369 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6370 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6371 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6372 return DAG.getNode(ISD::BITCAST, DL, VT,
6373 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6374 DAG.getConstant(BlendMask, DL, MVT::i8)));
6378 // For integer shuffles we need to expand the mask and cast the inputs to
6379 // v8i16s prior to blending.
6380 int Scale = 8 / VT.getVectorNumElements();
6382 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6383 if (Mask[i] >= Size)
6384 for (int j = 0; j < Scale; ++j)
6385 BlendMask |= 1u << (i * Scale + j);
6387 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6388 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6389 return DAG.getNode(ISD::BITCAST, DL, VT,
6390 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6391 DAG.getConstant(BlendMask, DL, MVT::i8)));
6395 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6396 SmallVector<int, 8> RepeatedMask;
6397 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6398 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6399 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6401 for (int i = 0; i < 8; ++i)
6402 if (RepeatedMask[i] >= 16)
6403 BlendMask |= 1u << i;
6404 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6405 DAG.getConstant(BlendMask, DL, MVT::i8));
6411 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6412 "256-bit byte-blends require AVX2 support!");
6414 // Scale the blend by the number of bytes per element.
6415 int Scale = VT.getScalarSizeInBits() / 8;
6417 // This form of blend is always done on bytes. Compute the byte vector
6419 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6421 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6422 // mix of LLVM's code generator and the x86 backend. We tell the code
6423 // generator that boolean values in the elements of an x86 vector register
6424 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6425 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6426 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6427 // of the element (the remaining are ignored) and 0 in that high bit would
6428 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6429 // the LLVM model for boolean values in vector elements gets the relevant
6430 // bit set, it is set backwards and over constrained relative to x86's
6432 SmallVector<SDValue, 32> VSELECTMask;
6433 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6434 for (int j = 0; j < Scale; ++j)
6435 VSELECTMask.push_back(
6436 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6437 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6440 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6441 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6443 ISD::BITCAST, DL, VT,
6444 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6445 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6450 llvm_unreachable("Not a supported integer vector type!");
6454 /// \brief Try to lower as a blend of elements from two inputs followed by
6455 /// a single-input permutation.
6457 /// This matches the pattern where we can blend elements from two inputs and
6458 /// then reduce the shuffle to a single-input permutation.
6459 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6462 SelectionDAG &DAG) {
6463 // We build up the blend mask while checking whether a blend is a viable way
6464 // to reduce the shuffle.
6465 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6466 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6468 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6472 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6474 if (BlendMask[Mask[i] % Size] == -1)
6475 BlendMask[Mask[i] % Size] = Mask[i];
6476 else if (BlendMask[Mask[i] % Size] != Mask[i])
6477 return SDValue(); // Can't blend in the needed input!
6479 PermuteMask[i] = Mask[i] % Size;
6482 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6483 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6486 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6487 /// blends and permutes.
6489 /// This matches the extremely common pattern for handling combined
6490 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6491 /// operations. It will try to pick the best arrangement of shuffles and
6493 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6497 SelectionDAG &DAG) {
6498 // Shuffle the input elements into the desired positions in V1 and V2 and
6499 // blend them together.
6500 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6501 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6502 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6503 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6504 if (Mask[i] >= 0 && Mask[i] < Size) {
6505 V1Mask[i] = Mask[i];
6507 } else if (Mask[i] >= Size) {
6508 V2Mask[i] = Mask[i] - Size;
6509 BlendMask[i] = i + Size;
6512 // Try to lower with the simpler initial blend strategy unless one of the
6513 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6514 // shuffle may be able to fold with a load or other benefit. However, when
6515 // we'll have to do 2x as many shuffles in order to achieve this, blending
6516 // first is a better strategy.
6517 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6518 if (SDValue BlendPerm =
6519 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6522 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6523 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6524 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6527 /// \brief Try to lower a vector shuffle as a byte rotation.
6529 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6530 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6531 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6532 /// try to generically lower a vector shuffle through such an pattern. It
6533 /// does not check for the profitability of lowering either as PALIGNR or
6534 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6535 /// This matches shuffle vectors that look like:
6537 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6539 /// Essentially it concatenates V1 and V2, shifts right by some number of
6540 /// elements, and takes the low elements as the result. Note that while this is
6541 /// specified as a *right shift* because x86 is little-endian, it is a *left
6542 /// rotate* of the vector lanes.
6543 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6546 const X86Subtarget *Subtarget,
6547 SelectionDAG &DAG) {
6548 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6550 int NumElts = Mask.size();
6551 int NumLanes = VT.getSizeInBits() / 128;
6552 int NumLaneElts = NumElts / NumLanes;
6554 // We need to detect various ways of spelling a rotation:
6555 // [11, 12, 13, 14, 15, 0, 1, 2]
6556 // [-1, 12, 13, 14, -1, -1, 1, -1]
6557 // [-1, -1, -1, -1, -1, -1, 1, 2]
6558 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6559 // [-1, 4, 5, 6, -1, -1, 9, -1]
6560 // [-1, 4, 5, 6, -1, -1, -1, -1]
6563 for (int l = 0; l < NumElts; l += NumLaneElts) {
6564 for (int i = 0; i < NumLaneElts; ++i) {
6565 if (Mask[l + i] == -1)
6567 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6569 // Get the mod-Size index and lane correct it.
6570 int LaneIdx = (Mask[l + i] % NumElts) - l;
6571 // Make sure it was in this lane.
6572 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6575 // Determine where a rotated vector would have started.
6576 int StartIdx = i - LaneIdx;
6578 // The identity rotation isn't interesting, stop.
6581 // If we found the tail of a vector the rotation must be the missing
6582 // front. If we found the head of a vector, it must be how much of the
6584 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6587 Rotation = CandidateRotation;
6588 else if (Rotation != CandidateRotation)
6589 // The rotations don't match, so we can't match this mask.
6592 // Compute which value this mask is pointing at.
6593 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6595 // Compute which of the two target values this index should be assigned
6596 // to. This reflects whether the high elements are remaining or the low
6597 // elements are remaining.
6598 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6600 // Either set up this value if we've not encountered it before, or check
6601 // that it remains consistent.
6604 else if (TargetV != MaskV)
6605 // This may be a rotation, but it pulls from the inputs in some
6606 // unsupported interleaving.
6611 // Check that we successfully analyzed the mask, and normalize the results.
6612 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6613 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6619 // The actual rotate instruction rotates bytes, so we need to scale the
6620 // rotation based on how many bytes are in the vector lane.
6621 int Scale = 16 / NumLaneElts;
6623 // SSSE3 targets can use the palignr instruction.
6624 if (Subtarget->hasSSSE3()) {
6625 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6626 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6627 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6628 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6630 return DAG.getNode(ISD::BITCAST, DL, VT,
6631 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6632 DAG.getConstant(Rotation * Scale, DL,
6636 assert(VT.getSizeInBits() == 128 &&
6637 "Rotate-based lowering only supports 128-bit lowering!");
6638 assert(Mask.size() <= 16 &&
6639 "Can shuffle at most 16 bytes in a 128-bit vector!");
6641 // Default SSE2 implementation
6642 int LoByteShift = 16 - Rotation * Scale;
6643 int HiByteShift = Rotation * Scale;
6645 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6646 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6647 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6649 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6650 DAG.getConstant(LoByteShift, DL, MVT::i8));
6651 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6652 DAG.getConstant(HiByteShift, DL, MVT::i8));
6653 return DAG.getNode(ISD::BITCAST, DL, VT,
6654 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6657 /// \brief Compute whether each element of a shuffle is zeroable.
6659 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6660 /// Either it is an undef element in the shuffle mask, the element of the input
6661 /// referenced is undef, or the element of the input referenced is known to be
6662 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6663 /// as many lanes with this technique as possible to simplify the remaining
6665 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6666 SDValue V1, SDValue V2) {
6667 SmallBitVector Zeroable(Mask.size(), false);
6669 while (V1.getOpcode() == ISD::BITCAST)
6670 V1 = V1->getOperand(0);
6671 while (V2.getOpcode() == ISD::BITCAST)
6672 V2 = V2->getOperand(0);
6674 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6675 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6677 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6679 // Handle the easy cases.
6680 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6685 // If this is an index into a build_vector node (which has the same number
6686 // of elements), dig out the input value and use it.
6687 SDValue V = M < Size ? V1 : V2;
6688 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6691 SDValue Input = V.getOperand(M % Size);
6692 // The UNDEF opcode check really should be dead code here, but not quite
6693 // worth asserting on (it isn't invalid, just unexpected).
6694 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6701 /// \brief Try to emit a bitmask instruction for a shuffle.
6703 /// This handles cases where we can model a blend exactly as a bitmask due to
6704 /// one of the inputs being zeroable.
6705 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6706 SDValue V2, ArrayRef<int> Mask,
6707 SelectionDAG &DAG) {
6708 MVT EltVT = VT.getScalarType();
6709 int NumEltBits = EltVT.getSizeInBits();
6710 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6711 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6712 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6714 if (EltVT.isFloatingPoint()) {
6715 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6716 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6718 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6719 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6721 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6724 if (Mask[i] % Size != i)
6725 return SDValue(); // Not a blend.
6727 V = Mask[i] < Size ? V1 : V2;
6728 else if (V != (Mask[i] < Size ? V1 : V2))
6729 return SDValue(); // Can only let one input through the mask.
6731 VMaskOps[i] = AllOnes;
6734 return SDValue(); // No non-zeroable elements!
6736 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6737 V = DAG.getNode(VT.isFloatingPoint()
6738 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6743 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6745 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6746 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6747 /// matches elements from one of the input vectors shuffled to the left or
6748 /// right with zeroable elements 'shifted in'. It handles both the strictly
6749 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6752 /// PSHL : (little-endian) left bit shift.
6753 /// [ zz, 0, zz, 2 ]
6754 /// [ -1, 4, zz, -1 ]
6755 /// PSRL : (little-endian) right bit shift.
6757 /// [ -1, -1, 7, zz]
6758 /// PSLLDQ : (little-endian) left byte shift
6759 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6760 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6761 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6762 /// PSRLDQ : (little-endian) right byte shift
6763 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6764 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6765 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6766 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6767 SDValue V2, ArrayRef<int> Mask,
6768 SelectionDAG &DAG) {
6769 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6771 int Size = Mask.size();
6772 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6774 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6775 for (int i = 0; i < Size; i += Scale)
6776 for (int j = 0; j < Shift; ++j)
6777 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6783 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6784 for (int i = 0; i != Size; i += Scale) {
6785 unsigned Pos = Left ? i + Shift : i;
6786 unsigned Low = Left ? i : i + Shift;
6787 unsigned Len = Scale - Shift;
6788 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6789 Low + (V == V1 ? 0 : Size)))
6793 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6794 bool ByteShift = ShiftEltBits > 64;
6795 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6796 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6797 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6799 // Normalize the scale for byte shifts to still produce an i64 element
6801 Scale = ByteShift ? Scale / 2 : Scale;
6803 // We need to round trip through the appropriate type for the shift.
6804 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6805 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6806 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6807 "Illegal integer vector type");
6808 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6810 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6811 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6812 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6815 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6816 // keep doubling the size of the integer elements up to that. We can
6817 // then shift the elements of the integer vector by whole multiples of
6818 // their width within the elements of the larger integer vector. Test each
6819 // multiple to see if we can find a match with the moved element indices
6820 // and that the shifted in elements are all zeroable.
6821 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6822 for (int Shift = 1; Shift != Scale; ++Shift)
6823 for (bool Left : {true, false})
6824 if (CheckZeros(Shift, Scale, Left))
6825 for (SDValue V : {V1, V2})
6826 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6833 /// \brief Lower a vector shuffle as a zero or any extension.
6835 /// Given a specific number of elements, element bit width, and extension
6836 /// stride, produce either a zero or any extension based on the available
6837 /// features of the subtarget.
6838 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6839 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6840 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6841 assert(Scale > 1 && "Need a scale to extend.");
6842 int NumElements = VT.getVectorNumElements();
6843 int EltBits = VT.getScalarSizeInBits();
6844 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6845 "Only 8, 16, and 32 bit elements can be extended.");
6846 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6848 // Found a valid zext mask! Try various lowering strategies based on the
6849 // input type and available ISA extensions.
6850 if (Subtarget->hasSSE41()) {
6851 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6852 NumElements / Scale);
6853 return DAG.getNode(ISD::BITCAST, DL, VT,
6854 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6857 // For any extends we can cheat for larger element sizes and use shuffle
6858 // instructions that can fold with a load and/or copy.
6859 if (AnyExt && EltBits == 32) {
6860 int PSHUFDMask[4] = {0, -1, 1, -1};
6862 ISD::BITCAST, DL, VT,
6863 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6864 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6865 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
6867 if (AnyExt && EltBits == 16 && Scale > 2) {
6868 int PSHUFDMask[4] = {0, -1, 0, -1};
6869 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6870 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6871 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
6872 int PSHUFHWMask[4] = {1, -1, -1, -1};
6874 ISD::BITCAST, DL, VT,
6875 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6876 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6877 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
6880 // If this would require more than 2 unpack instructions to expand, use
6881 // pshufb when available. We can only use more than 2 unpack instructions
6882 // when zero extending i8 elements which also makes it easier to use pshufb.
6883 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6884 assert(NumElements == 16 && "Unexpected byte vector width!");
6885 SDValue PSHUFBMask[16];
6886 for (int i = 0; i < 16; ++i)
6888 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
6889 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6890 return DAG.getNode(ISD::BITCAST, DL, VT,
6891 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6892 DAG.getNode(ISD::BUILD_VECTOR, DL,
6893 MVT::v16i8, PSHUFBMask)));
6896 // Otherwise emit a sequence of unpacks.
6898 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6899 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6900 : getZeroVector(InputVT, Subtarget, DAG, DL);
6901 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6902 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6906 } while (Scale > 1);
6907 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6910 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6912 /// This routine will try to do everything in its power to cleverly lower
6913 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6914 /// check for the profitability of this lowering, it tries to aggressively
6915 /// match this pattern. It will use all of the micro-architectural details it
6916 /// can to emit an efficient lowering. It handles both blends with all-zero
6917 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6918 /// masking out later).
6920 /// The reason we have dedicated lowering for zext-style shuffles is that they
6921 /// are both incredibly common and often quite performance sensitive.
6922 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6923 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6924 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6925 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6927 int Bits = VT.getSizeInBits();
6928 int NumElements = VT.getVectorNumElements();
6929 assert(VT.getScalarSizeInBits() <= 32 &&
6930 "Exceeds 32-bit integer zero extension limit");
6931 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6933 // Define a helper function to check a particular ext-scale and lower to it if
6935 auto Lower = [&](int Scale) -> SDValue {
6938 for (int i = 0; i < NumElements; ++i) {
6940 continue; // Valid anywhere but doesn't tell us anything.
6941 if (i % Scale != 0) {
6942 // Each of the extended elements need to be zeroable.
6946 // We no longer are in the anyext case.
6951 // Each of the base elements needs to be consecutive indices into the
6952 // same input vector.
6953 SDValue V = Mask[i] < NumElements ? V1 : V2;
6956 else if (InputV != V)
6957 return SDValue(); // Flip-flopping inputs.
6959 if (Mask[i] % NumElements != i / Scale)
6960 return SDValue(); // Non-consecutive strided elements.
6963 // If we fail to find an input, we have a zero-shuffle which should always
6964 // have already been handled.
6965 // FIXME: Maybe handle this here in case during blending we end up with one?
6969 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6970 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6973 // The widest scale possible for extending is to a 64-bit integer.
6974 assert(Bits % 64 == 0 &&
6975 "The number of bits in a vector must be divisible by 64 on x86!");
6976 int NumExtElements = Bits / 64;
6978 // Each iteration, try extending the elements half as much, but into twice as
6980 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6981 assert(NumElements % NumExtElements == 0 &&
6982 "The input vector size must be divisible by the extended size.");
6983 if (SDValue V = Lower(NumElements / NumExtElements))
6987 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6991 // Returns one of the source operands if the shuffle can be reduced to a
6992 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6993 auto CanZExtLowHalf = [&]() {
6994 for (int i = NumElements / 2; i != NumElements; ++i)
6997 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6999 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7004 if (SDValue V = CanZExtLowHalf()) {
7005 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
7006 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7007 return DAG.getNode(ISD::BITCAST, DL, VT, V);
7010 // No viable ext lowering found.
7014 /// \brief Try to get a scalar value for a specific element of a vector.
7016 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7017 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7018 SelectionDAG &DAG) {
7019 MVT VT = V.getSimpleValueType();
7020 MVT EltVT = VT.getVectorElementType();
7021 while (V.getOpcode() == ISD::BITCAST)
7022 V = V.getOperand(0);
7023 // If the bitcasts shift the element size, we can't extract an equivalent
7025 MVT NewVT = V.getSimpleValueType();
7026 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7029 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7030 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7031 // Ensure the scalar operand is the same size as the destination.
7032 // FIXME: Add support for scalar truncation where possible.
7033 SDValue S = V.getOperand(Idx);
7034 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7035 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7041 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7043 /// This is particularly important because the set of instructions varies
7044 /// significantly based on whether the operand is a load or not.
7045 static bool isShuffleFoldableLoad(SDValue V) {
7046 while (V.getOpcode() == ISD::BITCAST)
7047 V = V.getOperand(0);
7049 return ISD::isNON_EXTLoad(V.getNode());
7052 /// \brief Try to lower insertion of a single element into a zero vector.
7054 /// This is a common pattern that we have especially efficient patterns to lower
7055 /// across all subtarget feature sets.
7056 static SDValue lowerVectorShuffleAsElementInsertion(
7057 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7058 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7059 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7061 MVT EltVT = VT.getVectorElementType();
7063 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7064 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7066 bool IsV1Zeroable = true;
7067 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7068 if (i != V2Index && !Zeroable[i]) {
7069 IsV1Zeroable = false;
7073 // Check for a single input from a SCALAR_TO_VECTOR node.
7074 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7075 // all the smarts here sunk into that routine. However, the current
7076 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7077 // vector shuffle lowering is dead.
7078 if (SDValue V2S = getScalarValueForVectorElement(
7079 V2, Mask[V2Index] - Mask.size(), DAG)) {
7080 // We need to zext the scalar if it is smaller than an i32.
7081 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
7082 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7083 // Using zext to expand a narrow element won't work for non-zero
7088 // Zero-extend directly to i32.
7090 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7092 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7093 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7094 EltVT == MVT::i16) {
7095 // Either not inserting from the low element of the input or the input
7096 // element size is too small to use VZEXT_MOVL to clear the high bits.
7100 if (!IsV1Zeroable) {
7101 // If V1 can't be treated as a zero vector we have fewer options to lower
7102 // this. We can't support integer vectors or non-zero targets cheaply, and
7103 // the V1 elements can't be permuted in any way.
7104 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7105 if (!VT.isFloatingPoint() || V2Index != 0)
7107 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7108 V1Mask[V2Index] = -1;
7109 if (!isNoopShuffleMask(V1Mask))
7111 // This is essentially a special case blend operation, but if we have
7112 // general purpose blend operations, they are always faster. Bail and let
7113 // the rest of the lowering handle these as blends.
7114 if (Subtarget->hasSSE41())
7117 // Otherwise, use MOVSD or MOVSS.
7118 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7119 "Only two types of floating point element types to handle!");
7120 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7124 // This lowering only works for the low element with floating point vectors.
7125 if (VT.isFloatingPoint() && V2Index != 0)
7128 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7130 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7133 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7134 // the desired position. Otherwise it is more efficient to do a vector
7135 // shift left. We know that we can do a vector shift left because all
7136 // the inputs are zero.
7137 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7138 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7139 V2Shuffle[V2Index] = 0;
7140 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7142 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
7144 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7146 V2Index * EltVT.getSizeInBits()/8, DL,
7147 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7148 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7154 /// \brief Try to lower broadcast of a single element.
7156 /// For convenience, this code also bundles all of the subtarget feature set
7157 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7158 /// a convenient way to factor it out.
7159 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7161 const X86Subtarget *Subtarget,
7162 SelectionDAG &DAG) {
7163 if (!Subtarget->hasAVX())
7165 if (VT.isInteger() && !Subtarget->hasAVX2())
7168 // Check that the mask is a broadcast.
7169 int BroadcastIdx = -1;
7171 if (M >= 0 && BroadcastIdx == -1)
7173 else if (M >= 0 && M != BroadcastIdx)
7176 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7177 "a sorted mask where the broadcast "
7180 // Go up the chain of (vector) values to find a scalar load that we can
7181 // combine with the broadcast.
7183 switch (V.getOpcode()) {
7184 case ISD::CONCAT_VECTORS: {
7185 int OperandSize = Mask.size() / V.getNumOperands();
7186 V = V.getOperand(BroadcastIdx / OperandSize);
7187 BroadcastIdx %= OperandSize;
7191 case ISD::INSERT_SUBVECTOR: {
7192 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7193 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7197 int BeginIdx = (int)ConstantIdx->getZExtValue();
7199 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7200 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7201 BroadcastIdx -= BeginIdx;
7212 // Check if this is a broadcast of a scalar. We special case lowering
7213 // for scalars so that we can more effectively fold with loads.
7214 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7215 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7216 V = V.getOperand(BroadcastIdx);
7218 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7219 // Only AVX2 has register broadcasts.
7220 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7222 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7223 // We can't broadcast from a vector register without AVX2, and we can only
7224 // broadcast from the zero-element of a vector register.
7228 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7231 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7232 // INSERTPS when the V1 elements are already in the correct locations
7233 // because otherwise we can just always use two SHUFPS instructions which
7234 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7235 // perform INSERTPS if a single V1 element is out of place and all V2
7236 // elements are zeroable.
7237 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7239 SelectionDAG &DAG) {
7240 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7241 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7242 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7243 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7245 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7248 int V1DstIndex = -1;
7249 int V2DstIndex = -1;
7250 bool V1UsedInPlace = false;
7252 for (int i = 0; i < 4; ++i) {
7253 // Synthesize a zero mask from the zeroable elements (includes undefs).
7259 // Flag if we use any V1 inputs in place.
7261 V1UsedInPlace = true;
7265 // We can only insert a single non-zeroable element.
7266 if (V1DstIndex != -1 || V2DstIndex != -1)
7270 // V1 input out of place for insertion.
7273 // V2 input for insertion.
7278 // Don't bother if we have no (non-zeroable) element for insertion.
7279 if (V1DstIndex == -1 && V2DstIndex == -1)
7282 // Determine element insertion src/dst indices. The src index is from the
7283 // start of the inserted vector, not the start of the concatenated vector.
7284 unsigned V2SrcIndex = 0;
7285 if (V1DstIndex != -1) {
7286 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7287 // and don't use the original V2 at all.
7288 V2SrcIndex = Mask[V1DstIndex];
7289 V2DstIndex = V1DstIndex;
7292 V2SrcIndex = Mask[V2DstIndex] - 4;
7295 // If no V1 inputs are used in place, then the result is created only from
7296 // the zero mask and the V2 insertion - so remove V1 dependency.
7298 V1 = DAG.getUNDEF(MVT::v4f32);
7300 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7301 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7303 // Insert the V2 element into the desired position.
7305 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7306 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7309 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7310 /// UNPCK instruction.
7312 /// This specifically targets cases where we end up with alternating between
7313 /// the two inputs, and so can permute them into something that feeds a single
7314 /// UNPCK instruction. Note that this routine only targets integer vectors
7315 /// because for floating point vectors we have a generalized SHUFPS lowering
7316 /// strategy that handles everything that doesn't *exactly* match an unpack,
7317 /// making this clever lowering unnecessary.
7318 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7319 SDValue V2, ArrayRef<int> Mask,
7320 SelectionDAG &DAG) {
7321 assert(!VT.isFloatingPoint() &&
7322 "This routine only supports integer vectors.");
7323 assert(!isSingleInputShuffleMask(Mask) &&
7324 "This routine should only be used when blending two inputs.");
7325 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7327 int Size = Mask.size();
7329 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7330 return M >= 0 && M % Size < Size / 2;
7332 int NumHiInputs = std::count_if(
7333 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7335 bool UnpackLo = NumLoInputs >= NumHiInputs;
7337 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7338 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7339 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7341 for (int i = 0; i < Size; ++i) {
7345 // Each element of the unpack contains Scale elements from this mask.
7346 int UnpackIdx = i / Scale;
7348 // We only handle the case where V1 feeds the first slots of the unpack.
7349 // We rely on canonicalization to ensure this is the case.
7350 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7353 // Setup the mask for this input. The indexing is tricky as we have to
7354 // handle the unpack stride.
7355 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7356 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7360 // If we will have to shuffle both inputs to use the unpack, check whether
7361 // we can just unpack first and shuffle the result. If so, skip this unpack.
7362 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7363 !isNoopShuffleMask(V2Mask))
7366 // Shuffle the inputs into place.
7367 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7368 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7370 // Cast the inputs to the type we will use to unpack them.
7371 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7372 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7374 // Unpack the inputs and cast the result back to the desired type.
7375 return DAG.getNode(ISD::BITCAST, DL, VT,
7376 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7377 DL, UnpackVT, V1, V2));
7380 // We try each unpack from the largest to the smallest to try and find one
7381 // that fits this mask.
7382 int OrigNumElements = VT.getVectorNumElements();
7383 int OrigScalarSize = VT.getScalarSizeInBits();
7384 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7385 int Scale = ScalarSize / OrigScalarSize;
7386 int NumElements = OrigNumElements / Scale;
7387 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7388 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7392 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7394 if (NumLoInputs == 0 || NumHiInputs == 0) {
7395 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7396 "We have to have *some* inputs!");
7397 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7399 // FIXME: We could consider the total complexity of the permute of each
7400 // possible unpacking. Or at the least we should consider how many
7401 // half-crossings are created.
7402 // FIXME: We could consider commuting the unpacks.
7404 SmallVector<int, 32> PermMask;
7405 PermMask.assign(Size, -1);
7406 for (int i = 0; i < Size; ++i) {
7410 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7413 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7415 return DAG.getVectorShuffle(
7416 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7418 DAG.getUNDEF(VT), PermMask);
7424 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7426 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7427 /// support for floating point shuffles but not integer shuffles. These
7428 /// instructions will incur a domain crossing penalty on some chips though so
7429 /// it is better to avoid lowering through this for integer vectors where
7431 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7432 const X86Subtarget *Subtarget,
7433 SelectionDAG &DAG) {
7435 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7436 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7437 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7438 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7439 ArrayRef<int> Mask = SVOp->getMask();
7440 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7442 if (isSingleInputShuffleMask(Mask)) {
7443 // Use low duplicate instructions for masks that match their pattern.
7444 if (Subtarget->hasSSE3())
7445 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7446 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7448 // Straight shuffle of a single input vector. Simulate this by using the
7449 // single input as both of the "inputs" to this instruction..
7450 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7452 if (Subtarget->hasAVX()) {
7453 // If we have AVX, we can use VPERMILPS which will allow folding a load
7454 // into the shuffle.
7455 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7456 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7459 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7460 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7462 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7463 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7465 // If we have a single input, insert that into V1 if we can do so cheaply.
7466 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7467 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7468 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7470 // Try inverting the insertion since for v2 masks it is easy to do and we
7471 // can't reliably sort the mask one way or the other.
7472 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7473 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7474 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7475 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7479 // Try to use one of the special instruction patterns to handle two common
7480 // blend patterns if a zero-blend above didn't work.
7481 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7482 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7483 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7484 // We can either use a special instruction to load over the low double or
7485 // to move just the low double.
7487 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7489 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7491 if (Subtarget->hasSSE41())
7492 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7496 // Use dedicated unpack instructions for masks that match their pattern.
7497 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7498 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7499 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7500 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7502 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7503 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7504 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7507 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7509 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7510 /// the integer unit to minimize domain crossing penalties. However, for blends
7511 /// it falls back to the floating point shuffle operation with appropriate bit
7513 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7514 const X86Subtarget *Subtarget,
7515 SelectionDAG &DAG) {
7517 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7518 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7519 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7520 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7521 ArrayRef<int> Mask = SVOp->getMask();
7522 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7524 if (isSingleInputShuffleMask(Mask)) {
7525 // Check for being able to broadcast a single element.
7526 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7527 Mask, Subtarget, DAG))
7530 // Straight shuffle of a single input vector. For everything from SSE2
7531 // onward this has a single fast instruction with no scary immediates.
7532 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7533 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7534 int WidenedMask[4] = {
7535 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7536 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7538 ISD::BITCAST, DL, MVT::v2i64,
7539 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7540 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7542 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7543 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7544 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7545 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7547 // If we have a blend of two PACKUS operations an the blend aligns with the
7548 // low and half halves, we can just merge the PACKUS operations. This is
7549 // particularly important as it lets us merge shuffles that this routine itself
7551 auto GetPackNode = [](SDValue V) {
7552 while (V.getOpcode() == ISD::BITCAST)
7553 V = V.getOperand(0);
7555 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7557 if (SDValue V1Pack = GetPackNode(V1))
7558 if (SDValue V2Pack = GetPackNode(V2))
7559 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7560 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7561 Mask[0] == 0 ? V1Pack.getOperand(0)
7562 : V1Pack.getOperand(1),
7563 Mask[1] == 2 ? V2Pack.getOperand(0)
7564 : V2Pack.getOperand(1)));
7566 // Try to use shift instructions.
7568 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7571 // When loading a scalar and then shuffling it into a vector we can often do
7572 // the insertion cheaply.
7573 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7574 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7576 // Try inverting the insertion since for v2 masks it is easy to do and we
7577 // can't reliably sort the mask one way or the other.
7578 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7579 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7580 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7583 // We have different paths for blend lowering, but they all must use the
7584 // *exact* same predicate.
7585 bool IsBlendSupported = Subtarget->hasSSE41();
7586 if (IsBlendSupported)
7587 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7591 // Use dedicated unpack instructions for masks that match their pattern.
7592 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7593 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7594 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7595 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7597 // Try to use byte rotation instructions.
7598 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7599 if (Subtarget->hasSSSE3())
7600 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7601 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7604 // If we have direct support for blends, we should lower by decomposing into
7605 // a permute. That will be faster than the domain cross.
7606 if (IsBlendSupported)
7607 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7610 // We implement this with SHUFPD which is pretty lame because it will likely
7611 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7612 // However, all the alternatives are still more cycles and newer chips don't
7613 // have this problem. It would be really nice if x86 had better shuffles here.
7614 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7615 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7616 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7617 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7620 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7622 /// This is used to disable more specialized lowerings when the shufps lowering
7623 /// will happen to be efficient.
7624 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7625 // This routine only handles 128-bit shufps.
7626 assert(Mask.size() == 4 && "Unsupported mask size!");
7628 // To lower with a single SHUFPS we need to have the low half and high half
7629 // each requiring a single input.
7630 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7632 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7638 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7640 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7641 /// It makes no assumptions about whether this is the *best* lowering, it simply
7643 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7644 ArrayRef<int> Mask, SDValue V1,
7645 SDValue V2, SelectionDAG &DAG) {
7646 SDValue LowV = V1, HighV = V2;
7647 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7650 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7652 if (NumV2Elements == 1) {
7654 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7657 // Compute the index adjacent to V2Index and in the same half by toggling
7659 int V2AdjIndex = V2Index ^ 1;
7661 if (Mask[V2AdjIndex] == -1) {
7662 // Handles all the cases where we have a single V2 element and an undef.
7663 // This will only ever happen in the high lanes because we commute the
7664 // vector otherwise.
7666 std::swap(LowV, HighV);
7667 NewMask[V2Index] -= 4;
7669 // Handle the case where the V2 element ends up adjacent to a V1 element.
7670 // To make this work, blend them together as the first step.
7671 int V1Index = V2AdjIndex;
7672 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7673 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7674 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7676 // Now proceed to reconstruct the final blend as we have the necessary
7677 // high or low half formed.
7684 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7685 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7687 } else if (NumV2Elements == 2) {
7688 if (Mask[0] < 4 && Mask[1] < 4) {
7689 // Handle the easy case where we have V1 in the low lanes and V2 in the
7693 } else if (Mask[2] < 4 && Mask[3] < 4) {
7694 // We also handle the reversed case because this utility may get called
7695 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7696 // arrange things in the right direction.
7702 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7703 // trying to place elements directly, just blend them and set up the final
7704 // shuffle to place them.
7706 // The first two blend mask elements are for V1, the second two are for
7708 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7709 Mask[2] < 4 ? Mask[2] : Mask[3],
7710 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7711 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7712 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7713 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7715 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7718 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7719 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7720 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7721 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7724 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7725 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
7728 /// \brief Lower 4-lane 32-bit floating point shuffles.
7730 /// Uses instructions exclusively from the floating point unit to minimize
7731 /// domain crossing penalties, as these are sufficient to implement all v4f32
7733 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7734 const X86Subtarget *Subtarget,
7735 SelectionDAG &DAG) {
7737 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7738 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7739 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7740 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7741 ArrayRef<int> Mask = SVOp->getMask();
7742 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7745 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7747 if (NumV2Elements == 0) {
7748 // Check for being able to broadcast a single element.
7749 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7750 Mask, Subtarget, DAG))
7753 // Use even/odd duplicate instructions for masks that match their pattern.
7754 if (Subtarget->hasSSE3()) {
7755 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7756 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7757 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7758 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7761 if (Subtarget->hasAVX()) {
7762 // If we have AVX, we can use VPERMILPS which will allow folding a load
7763 // into the shuffle.
7764 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7765 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7768 // Otherwise, use a straight shuffle of a single input vector. We pass the
7769 // input vector to both operands to simulate this with a SHUFPS.
7770 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7771 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7774 // There are special ways we can lower some single-element blends. However, we
7775 // have custom ways we can lower more complex single-element blends below that
7776 // we defer to if both this and BLENDPS fail to match, so restrict this to
7777 // when the V2 input is targeting element 0 of the mask -- that is the fast
7779 if (NumV2Elements == 1 && Mask[0] >= 4)
7780 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7781 Mask, Subtarget, DAG))
7784 if (Subtarget->hasSSE41()) {
7785 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7789 // Use INSERTPS if we can complete the shuffle efficiently.
7790 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7793 if (!isSingleSHUFPSMask(Mask))
7794 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7795 DL, MVT::v4f32, V1, V2, Mask, DAG))
7799 // Use dedicated unpack instructions for masks that match their pattern.
7800 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7801 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7802 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7803 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7804 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7805 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7806 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7807 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7809 // Otherwise fall back to a SHUFPS lowering strategy.
7810 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7813 /// \brief Lower 4-lane i32 vector shuffles.
7815 /// We try to handle these with integer-domain shuffles where we can, but for
7816 /// blends we use the floating point domain blend instructions.
7817 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7818 const X86Subtarget *Subtarget,
7819 SelectionDAG &DAG) {
7821 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7822 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7823 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7824 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7825 ArrayRef<int> Mask = SVOp->getMask();
7826 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7828 // Whenever we can lower this as a zext, that instruction is strictly faster
7829 // than any alternative. It also allows us to fold memory operands into the
7830 // shuffle in many cases.
7831 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7832 Mask, Subtarget, DAG))
7836 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7838 if (NumV2Elements == 0) {
7839 // Check for being able to broadcast a single element.
7840 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7841 Mask, Subtarget, DAG))
7844 // Straight shuffle of a single input vector. For everything from SSE2
7845 // onward this has a single fast instruction with no scary immediates.
7846 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7847 // but we aren't actually going to use the UNPCK instruction because doing
7848 // so prevents folding a load into this instruction or making a copy.
7849 const int UnpackLoMask[] = {0, 0, 1, 1};
7850 const int UnpackHiMask[] = {2, 2, 3, 3};
7851 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7852 Mask = UnpackLoMask;
7853 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7854 Mask = UnpackHiMask;
7856 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7857 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7860 // Try to use shift instructions.
7862 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7865 // There are special ways we can lower some single-element blends.
7866 if (NumV2Elements == 1)
7867 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7868 Mask, Subtarget, DAG))
7871 // We have different paths for blend lowering, but they all must use the
7872 // *exact* same predicate.
7873 bool IsBlendSupported = Subtarget->hasSSE41();
7874 if (IsBlendSupported)
7875 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7879 if (SDValue Masked =
7880 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7883 // Use dedicated unpack instructions for masks that match their pattern.
7884 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7885 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7886 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7887 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7888 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7889 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7890 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7891 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7893 // Try to use byte rotation instructions.
7894 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7895 if (Subtarget->hasSSSE3())
7896 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7897 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7900 // If we have direct support for blends, we should lower by decomposing into
7901 // a permute. That will be faster than the domain cross.
7902 if (IsBlendSupported)
7903 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7906 // Try to lower by permuting the inputs into an unpack instruction.
7907 if (SDValue Unpack =
7908 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7911 // We implement this with SHUFPS because it can blend from two vectors.
7912 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7913 // up the inputs, bypassing domain shift penalties that we would encur if we
7914 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7916 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7917 DAG.getVectorShuffle(
7919 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7920 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7923 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7924 /// shuffle lowering, and the most complex part.
7926 /// The lowering strategy is to try to form pairs of input lanes which are
7927 /// targeted at the same half of the final vector, and then use a dword shuffle
7928 /// to place them onto the right half, and finally unpack the paired lanes into
7929 /// their final position.
7931 /// The exact breakdown of how to form these dword pairs and align them on the
7932 /// correct sides is really tricky. See the comments within the function for
7933 /// more of the details.
7935 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
7936 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
7937 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
7938 /// vector, form the analogous 128-bit 8-element Mask.
7939 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7940 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
7941 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7942 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
7943 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
7945 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
7946 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7947 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7949 SmallVector<int, 4> LoInputs;
7950 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7951 [](int M) { return M >= 0; });
7952 std::sort(LoInputs.begin(), LoInputs.end());
7953 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7954 SmallVector<int, 4> HiInputs;
7955 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7956 [](int M) { return M >= 0; });
7957 std::sort(HiInputs.begin(), HiInputs.end());
7958 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7960 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7961 int NumHToL = LoInputs.size() - NumLToL;
7963 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7964 int NumHToH = HiInputs.size() - NumLToH;
7965 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7966 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7967 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7968 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7970 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7971 // such inputs we can swap two of the dwords across the half mark and end up
7972 // with <=2 inputs to each half in each half. Once there, we can fall through
7973 // to the generic code below. For example:
7975 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7976 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7978 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7979 // and an existing 2-into-2 on the other half. In this case we may have to
7980 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7981 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7982 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7983 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7984 // half than the one we target for fixing) will be fixed when we re-enter this
7985 // path. We will also combine away any sequence of PSHUFD instructions that
7986 // result into a single instruction. Here is an example of the tricky case:
7988 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7989 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7991 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7993 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7994 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7996 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7997 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7999 // The result is fine to be handled by the generic logic.
8000 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8001 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8002 int AOffset, int BOffset) {
8003 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8004 "Must call this with A having 3 or 1 inputs from the A half.");
8005 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8006 "Must call this with B having 1 or 3 inputs from the B half.");
8007 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8008 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8010 // Compute the index of dword with only one word among the three inputs in
8011 // a half by taking the sum of the half with three inputs and subtracting
8012 // the sum of the actual three inputs. The difference is the remaining
8015 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8016 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8017 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8018 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8019 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8020 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8021 int TripleNonInputIdx =
8022 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8023 TripleDWord = TripleNonInputIdx / 2;
8025 // We use xor with one to compute the adjacent DWord to whichever one the
8027 OneInputDWord = (OneInput / 2) ^ 1;
8029 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8030 // and BToA inputs. If there is also such a problem with the BToB and AToB
8031 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8032 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8033 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8034 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8035 // Compute how many inputs will be flipped by swapping these DWords. We
8037 // to balance this to ensure we don't form a 3-1 shuffle in the other
8039 int NumFlippedAToBInputs =
8040 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8041 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8042 int NumFlippedBToBInputs =
8043 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8044 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8045 if ((NumFlippedAToBInputs == 1 &&
8046 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8047 (NumFlippedBToBInputs == 1 &&
8048 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8049 // We choose whether to fix the A half or B half based on whether that
8050 // half has zero flipped inputs. At zero, we may not be able to fix it
8051 // with that half. We also bias towards fixing the B half because that
8052 // will more commonly be the high half, and we have to bias one way.
8053 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8054 ArrayRef<int> Inputs) {
8055 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8056 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8057 PinnedIdx ^ 1) != Inputs.end();
8058 // Determine whether the free index is in the flipped dword or the
8059 // unflipped dword based on where the pinned index is. We use this bit
8060 // in an xor to conditionally select the adjacent dword.
8061 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8062 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8063 FixFreeIdx) != Inputs.end();
8064 if (IsFixIdxInput == IsFixFreeIdxInput)
8066 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8067 FixFreeIdx) != Inputs.end();
8068 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8069 "We need to be changing the number of flipped inputs!");
8070 int PSHUFHalfMask[] = {0, 1, 2, 3};
8071 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8072 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8074 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8077 if (M != -1 && M == FixIdx)
8079 else if (M != -1 && M == FixFreeIdx)
8082 if (NumFlippedBToBInputs != 0) {
8084 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8085 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8087 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8089 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8090 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8095 int PSHUFDMask[] = {0, 1, 2, 3};
8096 PSHUFDMask[ADWord] = BDWord;
8097 PSHUFDMask[BDWord] = ADWord;
8098 V = DAG.getNode(ISD::BITCAST, DL, VT,
8099 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8100 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8101 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8104 // Adjust the mask to match the new locations of A and B.
8106 if (M != -1 && M/2 == ADWord)
8107 M = 2 * BDWord + M % 2;
8108 else if (M != -1 && M/2 == BDWord)
8109 M = 2 * ADWord + M % 2;
8111 // Recurse back into this routine to re-compute state now that this isn't
8112 // a 3 and 1 problem.
8113 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8116 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8117 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8118 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8119 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8121 // At this point there are at most two inputs to the low and high halves from
8122 // each half. That means the inputs can always be grouped into dwords and
8123 // those dwords can then be moved to the correct half with a dword shuffle.
8124 // We use at most one low and one high word shuffle to collect these paired
8125 // inputs into dwords, and finally a dword shuffle to place them.
8126 int PSHUFLMask[4] = {-1, -1, -1, -1};
8127 int PSHUFHMask[4] = {-1, -1, -1, -1};
8128 int PSHUFDMask[4] = {-1, -1, -1, -1};
8130 // First fix the masks for all the inputs that are staying in their
8131 // original halves. This will then dictate the targets of the cross-half
8133 auto fixInPlaceInputs =
8134 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8135 MutableArrayRef<int> SourceHalfMask,
8136 MutableArrayRef<int> HalfMask, int HalfOffset) {
8137 if (InPlaceInputs.empty())
8139 if (InPlaceInputs.size() == 1) {
8140 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8141 InPlaceInputs[0] - HalfOffset;
8142 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8145 if (IncomingInputs.empty()) {
8146 // Just fix all of the in place inputs.
8147 for (int Input : InPlaceInputs) {
8148 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8149 PSHUFDMask[Input / 2] = Input / 2;
8154 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8155 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8156 InPlaceInputs[0] - HalfOffset;
8157 // Put the second input next to the first so that they are packed into
8158 // a dword. We find the adjacent index by toggling the low bit.
8159 int AdjIndex = InPlaceInputs[0] ^ 1;
8160 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8161 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8162 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8164 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8165 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8167 // Now gather the cross-half inputs and place them into a free dword of
8168 // their target half.
8169 // FIXME: This operation could almost certainly be simplified dramatically to
8170 // look more like the 3-1 fixing operation.
8171 auto moveInputsToRightHalf = [&PSHUFDMask](
8172 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8173 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8174 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8176 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8177 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8179 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8181 int LowWord = Word & ~1;
8182 int HighWord = Word | 1;
8183 return isWordClobbered(SourceHalfMask, LowWord) ||
8184 isWordClobbered(SourceHalfMask, HighWord);
8187 if (IncomingInputs.empty())
8190 if (ExistingInputs.empty()) {
8191 // Map any dwords with inputs from them into the right half.
8192 for (int Input : IncomingInputs) {
8193 // If the source half mask maps over the inputs, turn those into
8194 // swaps and use the swapped lane.
8195 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8196 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8197 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8198 Input - SourceOffset;
8199 // We have to swap the uses in our half mask in one sweep.
8200 for (int &M : HalfMask)
8201 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8203 else if (M == Input)
8204 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8206 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8207 Input - SourceOffset &&
8208 "Previous placement doesn't match!");
8210 // Note that this correctly re-maps both when we do a swap and when
8211 // we observe the other side of the swap above. We rely on that to
8212 // avoid swapping the members of the input list directly.
8213 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8216 // Map the input's dword into the correct half.
8217 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8218 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8220 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8222 "Previous placement doesn't match!");
8225 // And just directly shift any other-half mask elements to be same-half
8226 // as we will have mirrored the dword containing the element into the
8227 // same position within that half.
8228 for (int &M : HalfMask)
8229 if (M >= SourceOffset && M < SourceOffset + 4) {
8230 M = M - SourceOffset + DestOffset;
8231 assert(M >= 0 && "This should never wrap below zero!");
8236 // Ensure we have the input in a viable dword of its current half. This
8237 // is particularly tricky because the original position may be clobbered
8238 // by inputs being moved and *staying* in that half.
8239 if (IncomingInputs.size() == 1) {
8240 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8241 int InputFixed = std::find(std::begin(SourceHalfMask),
8242 std::end(SourceHalfMask), -1) -
8243 std::begin(SourceHalfMask) + SourceOffset;
8244 SourceHalfMask[InputFixed - SourceOffset] =
8245 IncomingInputs[0] - SourceOffset;
8246 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8248 IncomingInputs[0] = InputFixed;
8250 } else if (IncomingInputs.size() == 2) {
8251 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8252 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8253 // We have two non-adjacent or clobbered inputs we need to extract from
8254 // the source half. To do this, we need to map them into some adjacent
8255 // dword slot in the source mask.
8256 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8257 IncomingInputs[1] - SourceOffset};
8259 // If there is a free slot in the source half mask adjacent to one of
8260 // the inputs, place the other input in it. We use (Index XOR 1) to
8261 // compute an adjacent index.
8262 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8263 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8264 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8265 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8266 InputsFixed[1] = InputsFixed[0] ^ 1;
8267 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8268 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8269 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8270 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8271 InputsFixed[0] = InputsFixed[1] ^ 1;
8272 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8273 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8274 // The two inputs are in the same DWord but it is clobbered and the
8275 // adjacent DWord isn't used at all. Move both inputs to the free
8277 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8278 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8279 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8280 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8282 // The only way we hit this point is if there is no clobbering
8283 // (because there are no off-half inputs to this half) and there is no
8284 // free slot adjacent to one of the inputs. In this case, we have to
8285 // swap an input with a non-input.
8286 for (int i = 0; i < 4; ++i)
8287 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8288 "We can't handle any clobbers here!");
8289 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8290 "Cannot have adjacent inputs here!");
8292 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8293 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8295 // We also have to update the final source mask in this case because
8296 // it may need to undo the above swap.
8297 for (int &M : FinalSourceHalfMask)
8298 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8299 M = InputsFixed[1] + SourceOffset;
8300 else if (M == InputsFixed[1] + SourceOffset)
8301 M = (InputsFixed[0] ^ 1) + SourceOffset;
8303 InputsFixed[1] = InputsFixed[0] ^ 1;
8306 // Point everything at the fixed inputs.
8307 for (int &M : HalfMask)
8308 if (M == IncomingInputs[0])
8309 M = InputsFixed[0] + SourceOffset;
8310 else if (M == IncomingInputs[1])
8311 M = InputsFixed[1] + SourceOffset;
8313 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8314 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8317 llvm_unreachable("Unhandled input size!");
8320 // Now hoist the DWord down to the right half.
8321 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8322 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8323 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8324 for (int &M : HalfMask)
8325 for (int Input : IncomingInputs)
8327 M = FreeDWord * 2 + Input % 2;
8329 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8330 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8331 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8332 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8334 // Now enact all the shuffles we've computed to move the inputs into their
8336 if (!isNoopShuffleMask(PSHUFLMask))
8337 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8338 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8339 if (!isNoopShuffleMask(PSHUFHMask))
8340 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8341 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8342 if (!isNoopShuffleMask(PSHUFDMask))
8343 V = DAG.getNode(ISD::BITCAST, DL, VT,
8344 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8345 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8346 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8349 // At this point, each half should contain all its inputs, and we can then
8350 // just shuffle them into their final position.
8351 assert(std::count_if(LoMask.begin(), LoMask.end(),
8352 [](int M) { return M >= 4; }) == 0 &&
8353 "Failed to lift all the high half inputs to the low mask!");
8354 assert(std::count_if(HiMask.begin(), HiMask.end(),
8355 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8356 "Failed to lift all the low half inputs to the high mask!");
8358 // Do a half shuffle for the low mask.
8359 if (!isNoopShuffleMask(LoMask))
8360 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8361 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8363 // Do a half shuffle with the high mask after shifting its values down.
8364 for (int &M : HiMask)
8367 if (!isNoopShuffleMask(HiMask))
8368 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8369 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8374 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8375 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8376 SDValue V2, ArrayRef<int> Mask,
8377 SelectionDAG &DAG, bool &V1InUse,
8379 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8385 int Size = Mask.size();
8386 int Scale = 16 / Size;
8387 for (int i = 0; i < 16; ++i) {
8388 if (Mask[i / Scale] == -1) {
8389 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8391 const int ZeroMask = 0x80;
8392 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8394 int V2Idx = Mask[i / Scale] < Size
8396 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8397 if (Zeroable[i / Scale])
8398 V1Idx = V2Idx = ZeroMask;
8399 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8400 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8401 V1InUse |= (ZeroMask != V1Idx);
8402 V2InUse |= (ZeroMask != V2Idx);
8407 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8408 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8409 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8411 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8412 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8413 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8415 // If we need shuffled inputs from both, blend the two.
8417 if (V1InUse && V2InUse)
8418 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8420 V = V1InUse ? V1 : V2;
8422 // Cast the result back to the correct type.
8423 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8426 /// \brief Generic lowering of 8-lane i16 shuffles.
8428 /// This handles both single-input shuffles and combined shuffle/blends with
8429 /// two inputs. The single input shuffles are immediately delegated to
8430 /// a dedicated lowering routine.
8432 /// The blends are lowered in one of three fundamental ways. If there are few
8433 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8434 /// of the input is significantly cheaper when lowered as an interleaving of
8435 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8436 /// halves of the inputs separately (making them have relatively few inputs)
8437 /// and then concatenate them.
8438 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8439 const X86Subtarget *Subtarget,
8440 SelectionDAG &DAG) {
8442 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8443 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8444 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8445 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8446 ArrayRef<int> OrigMask = SVOp->getMask();
8447 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8448 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8449 MutableArrayRef<int> Mask(MaskStorage);
8451 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8453 // Whenever we can lower this as a zext, that instruction is strictly faster
8454 // than any alternative.
8455 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8456 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8459 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8461 auto isV2 = [](int M) { return M >= 8; };
8463 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8465 if (NumV2Inputs == 0) {
8466 // Check for being able to broadcast a single element.
8467 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8468 Mask, Subtarget, DAG))
8471 // Try to use shift instructions.
8473 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8476 // Use dedicated unpack instructions for masks that match their pattern.
8477 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8478 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8479 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8480 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8482 // Try to use byte rotation instructions.
8483 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8484 Mask, Subtarget, DAG))
8487 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8491 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8492 "All single-input shuffles should be canonicalized to be V1-input "
8495 // Try to use shift instructions.
8497 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8500 // There are special ways we can lower some single-element blends.
8501 if (NumV2Inputs == 1)
8502 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8503 Mask, Subtarget, DAG))
8506 // We have different paths for blend lowering, but they all must use the
8507 // *exact* same predicate.
8508 bool IsBlendSupported = Subtarget->hasSSE41();
8509 if (IsBlendSupported)
8510 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8514 if (SDValue Masked =
8515 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8518 // Use dedicated unpack instructions for masks that match their pattern.
8519 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8520 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8521 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8522 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8524 // Try to use byte rotation instructions.
8525 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8526 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8529 if (SDValue BitBlend =
8530 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8533 if (SDValue Unpack =
8534 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8537 // If we can't directly blend but can use PSHUFB, that will be better as it
8538 // can both shuffle and set up the inefficient blend.
8539 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8540 bool V1InUse, V2InUse;
8541 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8545 // We can always bit-blend if we have to so the fallback strategy is to
8546 // decompose into single-input permutes and blends.
8547 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8551 /// \brief Check whether a compaction lowering can be done by dropping even
8552 /// elements and compute how many times even elements must be dropped.
8554 /// This handles shuffles which take every Nth element where N is a power of
8555 /// two. Example shuffle masks:
8557 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8558 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8559 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8560 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8561 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8562 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8564 /// Any of these lanes can of course be undef.
8566 /// This routine only supports N <= 3.
8567 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8570 /// \returns N above, or the number of times even elements must be dropped if
8571 /// there is such a number. Otherwise returns zero.
8572 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8573 // Figure out whether we're looping over two inputs or just one.
8574 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8576 // The modulus for the shuffle vector entries is based on whether this is
8577 // a single input or not.
8578 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8579 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8580 "We should only be called with masks with a power-of-2 size!");
8582 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8584 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8585 // and 2^3 simultaneously. This is because we may have ambiguity with
8586 // partially undef inputs.
8587 bool ViableForN[3] = {true, true, true};
8589 for (int i = 0, e = Mask.size(); i < e; ++i) {
8590 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8595 bool IsAnyViable = false;
8596 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8597 if (ViableForN[j]) {
8600 // The shuffle mask must be equal to (i * 2^N) % M.
8601 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8604 ViableForN[j] = false;
8606 // Early exit if we exhaust the possible powers of two.
8611 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8615 // Return 0 as there is no viable power of two.
8619 /// \brief Generic lowering of v16i8 shuffles.
8621 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8622 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8623 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8624 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8626 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8627 const X86Subtarget *Subtarget,
8628 SelectionDAG &DAG) {
8630 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8631 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8632 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8633 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8634 ArrayRef<int> Mask = SVOp->getMask();
8635 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8637 // Try to use shift instructions.
8639 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8642 // Try to use byte rotation instructions.
8643 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8644 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8647 // Try to use a zext lowering.
8648 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8649 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8653 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8655 // For single-input shuffles, there are some nicer lowering tricks we can use.
8656 if (NumV2Elements == 0) {
8657 // Check for being able to broadcast a single element.
8658 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8659 Mask, Subtarget, DAG))
8662 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8663 // Notably, this handles splat and partial-splat shuffles more efficiently.
8664 // However, it only makes sense if the pre-duplication shuffle simplifies
8665 // things significantly. Currently, this means we need to be able to
8666 // express the pre-duplication shuffle as an i16 shuffle.
8668 // FIXME: We should check for other patterns which can be widened into an
8669 // i16 shuffle as well.
8670 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8671 for (int i = 0; i < 16; i += 2)
8672 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8677 auto tryToWidenViaDuplication = [&]() -> SDValue {
8678 if (!canWidenViaDuplication(Mask))
8680 SmallVector<int, 4> LoInputs;
8681 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8682 [](int M) { return M >= 0 && M < 8; });
8683 std::sort(LoInputs.begin(), LoInputs.end());
8684 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8686 SmallVector<int, 4> HiInputs;
8687 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8688 [](int M) { return M >= 8; });
8689 std::sort(HiInputs.begin(), HiInputs.end());
8690 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8693 bool TargetLo = LoInputs.size() >= HiInputs.size();
8694 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8695 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8697 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8698 SmallDenseMap<int, int, 8> LaneMap;
8699 for (int I : InPlaceInputs) {
8700 PreDupI16Shuffle[I/2] = I/2;
8703 int j = TargetLo ? 0 : 4, je = j + 4;
8704 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8705 // Check if j is already a shuffle of this input. This happens when
8706 // there are two adjacent bytes after we move the low one.
8707 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8708 // If we haven't yet mapped the input, search for a slot into which
8710 while (j < je && PreDupI16Shuffle[j] != -1)
8714 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8717 // Map this input with the i16 shuffle.
8718 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8721 // Update the lane map based on the mapping we ended up with.
8722 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8725 ISD::BITCAST, DL, MVT::v16i8,
8726 DAG.getVectorShuffle(MVT::v8i16, DL,
8727 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8728 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8730 // Unpack the bytes to form the i16s that will be shuffled into place.
8731 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8732 MVT::v16i8, V1, V1);
8734 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8735 for (int i = 0; i < 16; ++i)
8736 if (Mask[i] != -1) {
8737 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8738 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8739 if (PostDupI16Shuffle[i / 2] == -1)
8740 PostDupI16Shuffle[i / 2] = MappedMask;
8742 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8743 "Conflicting entrties in the original shuffle!");
8746 ISD::BITCAST, DL, MVT::v16i8,
8747 DAG.getVectorShuffle(MVT::v8i16, DL,
8748 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8749 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8751 if (SDValue V = tryToWidenViaDuplication())
8755 // Use dedicated unpack instructions for masks that match their pattern.
8756 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8757 0, 16, 1, 17, 2, 18, 3, 19,
8759 4, 20, 5, 21, 6, 22, 7, 23}))
8760 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8761 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8762 8, 24, 9, 25, 10, 26, 11, 27,
8764 12, 28, 13, 29, 14, 30, 15, 31}))
8765 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8767 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8768 // with PSHUFB. It is important to do this before we attempt to generate any
8769 // blends but after all of the single-input lowerings. If the single input
8770 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8771 // want to preserve that and we can DAG combine any longer sequences into
8772 // a PSHUFB in the end. But once we start blending from multiple inputs,
8773 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8774 // and there are *very* few patterns that would actually be faster than the
8775 // PSHUFB approach because of its ability to zero lanes.
8777 // FIXME: The only exceptions to the above are blends which are exact
8778 // interleavings with direct instructions supporting them. We currently don't
8779 // handle those well here.
8780 if (Subtarget->hasSSSE3()) {
8781 bool V1InUse = false;
8782 bool V2InUse = false;
8784 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8785 DAG, V1InUse, V2InUse);
8787 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8788 // do so. This avoids using them to handle blends-with-zero which is
8789 // important as a single pshufb is significantly faster for that.
8790 if (V1InUse && V2InUse) {
8791 if (Subtarget->hasSSE41())
8792 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8793 Mask, Subtarget, DAG))
8796 // We can use an unpack to do the blending rather than an or in some
8797 // cases. Even though the or may be (very minorly) more efficient, we
8798 // preference this lowering because there are common cases where part of
8799 // the complexity of the shuffles goes away when we do the final blend as
8801 // FIXME: It might be worth trying to detect if the unpack-feeding
8802 // shuffles will both be pshufb, in which case we shouldn't bother with
8804 if (SDValue Unpack =
8805 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8812 // There are special ways we can lower some single-element blends.
8813 if (NumV2Elements == 1)
8814 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8815 Mask, Subtarget, DAG))
8818 if (SDValue BitBlend =
8819 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8822 // Check whether a compaction lowering can be done. This handles shuffles
8823 // which take every Nth element for some even N. See the helper function for
8826 // We special case these as they can be particularly efficiently handled with
8827 // the PACKUSB instruction on x86 and they show up in common patterns of
8828 // rearranging bytes to truncate wide elements.
8829 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8830 // NumEvenDrops is the power of two stride of the elements. Another way of
8831 // thinking about it is that we need to drop the even elements this many
8832 // times to get the original input.
8833 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8835 // First we need to zero all the dropped bytes.
8836 assert(NumEvenDrops <= 3 &&
8837 "No support for dropping even elements more than 3 times.");
8838 // We use the mask type to pick which bytes are preserved based on how many
8839 // elements are dropped.
8840 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8841 SDValue ByteClearMask =
8842 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8843 DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
8844 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8846 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8848 // Now pack things back together.
8849 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8850 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8851 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8852 for (int i = 1; i < NumEvenDrops; ++i) {
8853 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8854 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8860 // Handle multi-input cases by blending single-input shuffles.
8861 if (NumV2Elements > 0)
8862 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8865 // The fallback path for single-input shuffles widens this into two v8i16
8866 // vectors with unpacks, shuffles those, and then pulls them back together
8870 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8871 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8872 for (int i = 0; i < 16; ++i)
8874 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8876 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8878 SDValue VLoHalf, VHiHalf;
8879 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8880 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8882 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8883 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8884 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8885 [](int M) { return M >= 0 && M % 2 == 1; })) {
8886 // Use a mask to drop the high bytes.
8887 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8888 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8889 DAG.getConstant(0x00FF, DL, MVT::v8i16));
8891 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8892 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8894 // Squash the masks to point directly into VLoHalf.
8895 for (int &M : LoBlendMask)
8898 for (int &M : HiBlendMask)
8902 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8903 // VHiHalf so that we can blend them as i16s.
8904 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8905 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8906 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8907 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8910 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8911 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8913 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8916 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8918 /// This routine breaks down the specific type of 128-bit shuffle and
8919 /// dispatches to the lowering routines accordingly.
8920 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8921 MVT VT, const X86Subtarget *Subtarget,
8922 SelectionDAG &DAG) {
8923 switch (VT.SimpleTy) {
8925 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8927 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8929 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8931 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8933 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8935 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8938 llvm_unreachable("Unimplemented!");
8942 /// \brief Helper function to test whether a shuffle mask could be
8943 /// simplified by widening the elements being shuffled.
8945 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8946 /// leaves it in an unspecified state.
8948 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8949 /// shuffle masks. The latter have the special property of a '-2' representing
8950 /// a zero-ed lane of a vector.
8951 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8952 SmallVectorImpl<int> &WidenedMask) {
8953 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8954 // If both elements are undef, its trivial.
8955 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8956 WidenedMask.push_back(SM_SentinelUndef);
8960 // Check for an undef mask and a mask value properly aligned to fit with
8961 // a pair of values. If we find such a case, use the non-undef mask's value.
8962 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8963 WidenedMask.push_back(Mask[i + 1] / 2);
8966 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8967 WidenedMask.push_back(Mask[i] / 2);
8971 // When zeroing, we need to spread the zeroing across both lanes to widen.
8972 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8973 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8974 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8975 WidenedMask.push_back(SM_SentinelZero);
8981 // Finally check if the two mask values are adjacent and aligned with
8983 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8984 WidenedMask.push_back(Mask[i] / 2);
8988 // Otherwise we can't safely widen the elements used in this shuffle.
8991 assert(WidenedMask.size() == Mask.size() / 2 &&
8992 "Incorrect size of mask after widening the elements!");
8997 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8999 /// This routine just extracts two subvectors, shuffles them independently, and
9000 /// then concatenates them back together. This should work effectively with all
9001 /// AVX vector shuffle types.
9002 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9003 SDValue V2, ArrayRef<int> Mask,
9004 SelectionDAG &DAG) {
9005 assert(VT.getSizeInBits() >= 256 &&
9006 "Only for 256-bit or wider vector shuffles!");
9007 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9008 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9010 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9011 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9013 int NumElements = VT.getVectorNumElements();
9014 int SplitNumElements = NumElements / 2;
9015 MVT ScalarVT = VT.getScalarType();
9016 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9018 // Rather than splitting build-vectors, just build two narrower build
9019 // vectors. This helps shuffling with splats and zeros.
9020 auto SplitVector = [&](SDValue V) {
9021 while (V.getOpcode() == ISD::BITCAST)
9022 V = V->getOperand(0);
9024 MVT OrigVT = V.getSimpleValueType();
9025 int OrigNumElements = OrigVT.getVectorNumElements();
9026 int OrigSplitNumElements = OrigNumElements / 2;
9027 MVT OrigScalarVT = OrigVT.getScalarType();
9028 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9032 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9034 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9035 DAG.getIntPtrConstant(0, DL));
9036 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9037 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9040 SmallVector<SDValue, 16> LoOps, HiOps;
9041 for (int i = 0; i < OrigSplitNumElements; ++i) {
9042 LoOps.push_back(BV->getOperand(i));
9043 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9045 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9046 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9048 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
9049 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
9052 SDValue LoV1, HiV1, LoV2, HiV2;
9053 std::tie(LoV1, HiV1) = SplitVector(V1);
9054 std::tie(LoV2, HiV2) = SplitVector(V2);
9056 // Now create two 4-way blends of these half-width vectors.
9057 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9058 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9059 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9060 for (int i = 0; i < SplitNumElements; ++i) {
9061 int M = HalfMask[i];
9062 if (M >= NumElements) {
9063 if (M >= NumElements + SplitNumElements)
9067 V2BlendMask.push_back(M - NumElements);
9068 V1BlendMask.push_back(-1);
9069 BlendMask.push_back(SplitNumElements + i);
9070 } else if (M >= 0) {
9071 if (M >= SplitNumElements)
9075 V2BlendMask.push_back(-1);
9076 V1BlendMask.push_back(M);
9077 BlendMask.push_back(i);
9079 V2BlendMask.push_back(-1);
9080 V1BlendMask.push_back(-1);
9081 BlendMask.push_back(-1);
9085 // Because the lowering happens after all combining takes place, we need to
9086 // manually combine these blend masks as much as possible so that we create
9087 // a minimal number of high-level vector shuffle nodes.
9089 // First try just blending the halves of V1 or V2.
9090 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9091 return DAG.getUNDEF(SplitVT);
9092 if (!UseLoV2 && !UseHiV2)
9093 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9094 if (!UseLoV1 && !UseHiV1)
9095 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9097 SDValue V1Blend, V2Blend;
9098 if (UseLoV1 && UseHiV1) {
9100 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9102 // We only use half of V1 so map the usage down into the final blend mask.
9103 V1Blend = UseLoV1 ? LoV1 : HiV1;
9104 for (int i = 0; i < SplitNumElements; ++i)
9105 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9106 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9108 if (UseLoV2 && UseHiV2) {
9110 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9112 // We only use half of V2 so map the usage down into the final blend mask.
9113 V2Blend = UseLoV2 ? LoV2 : HiV2;
9114 for (int i = 0; i < SplitNumElements; ++i)
9115 if (BlendMask[i] >= SplitNumElements)
9116 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9118 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9120 SDValue Lo = HalfBlend(LoMask);
9121 SDValue Hi = HalfBlend(HiMask);
9122 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9125 /// \brief Either split a vector in halves or decompose the shuffles and the
9128 /// This is provided as a good fallback for many lowerings of non-single-input
9129 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9130 /// between splitting the shuffle into 128-bit components and stitching those
9131 /// back together vs. extracting the single-input shuffles and blending those
9133 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9134 SDValue V2, ArrayRef<int> Mask,
9135 SelectionDAG &DAG) {
9136 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9137 "lower single-input shuffles as it "
9138 "could then recurse on itself.");
9139 int Size = Mask.size();
9141 // If this can be modeled as a broadcast of two elements followed by a blend,
9142 // prefer that lowering. This is especially important because broadcasts can
9143 // often fold with memory operands.
9144 auto DoBothBroadcast = [&] {
9145 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9148 if (V2BroadcastIdx == -1)
9149 V2BroadcastIdx = M - Size;
9150 else if (M - Size != V2BroadcastIdx)
9152 } else if (M >= 0) {
9153 if (V1BroadcastIdx == -1)
9155 else if (M != V1BroadcastIdx)
9160 if (DoBothBroadcast())
9161 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9164 // If the inputs all stem from a single 128-bit lane of each input, then we
9165 // split them rather than blending because the split will decompose to
9166 // unusually few instructions.
9167 int LaneCount = VT.getSizeInBits() / 128;
9168 int LaneSize = Size / LaneCount;
9169 SmallBitVector LaneInputs[2];
9170 LaneInputs[0].resize(LaneCount, false);
9171 LaneInputs[1].resize(LaneCount, false);
9172 for (int i = 0; i < Size; ++i)
9174 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9175 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9176 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9178 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9179 // that the decomposed single-input shuffles don't end up here.
9180 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9183 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9184 /// a permutation and blend of those lanes.
9186 /// This essentially blends the out-of-lane inputs to each lane into the lane
9187 /// from a permuted copy of the vector. This lowering strategy results in four
9188 /// instructions in the worst case for a single-input cross lane shuffle which
9189 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9190 /// of. Special cases for each particular shuffle pattern should be handled
9191 /// prior to trying this lowering.
9192 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9193 SDValue V1, SDValue V2,
9195 SelectionDAG &DAG) {
9196 // FIXME: This should probably be generalized for 512-bit vectors as well.
9197 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9198 int LaneSize = Mask.size() / 2;
9200 // If there are only inputs from one 128-bit lane, splitting will in fact be
9201 // less expensive. The flags track whether the given lane contains an element
9202 // that crosses to another lane.
9203 bool LaneCrossing[2] = {false, false};
9204 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9205 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9206 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9207 if (!LaneCrossing[0] || !LaneCrossing[1])
9208 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9210 if (isSingleInputShuffleMask(Mask)) {
9211 SmallVector<int, 32> FlippedBlendMask;
9212 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9213 FlippedBlendMask.push_back(
9214 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9216 : Mask[i] % LaneSize +
9217 (i / LaneSize) * LaneSize + Size));
9219 // Flip the vector, and blend the results which should now be in-lane. The
9220 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9221 // 5 for the high source. The value 3 selects the high half of source 2 and
9222 // the value 2 selects the low half of source 2. We only use source 2 to
9223 // allow folding it into a memory operand.
9224 unsigned PERMMask = 3 | 2 << 4;
9225 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9226 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9227 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9230 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9231 // will be handled by the above logic and a blend of the results, much like
9232 // other patterns in AVX.
9233 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9236 /// \brief Handle lowering 2-lane 128-bit shuffles.
9237 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9238 SDValue V2, ArrayRef<int> Mask,
9239 const X86Subtarget *Subtarget,
9240 SelectionDAG &DAG) {
9241 // TODO: If minimizing size and one of the inputs is a zero vector and the
9242 // the zero vector has only one use, we could use a VPERM2X128 to save the
9243 // instruction bytes needed to explicitly generate the zero vector.
9245 // Blends are faster and handle all the non-lane-crossing cases.
9246 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9250 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9251 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9253 // If either input operand is a zero vector, use VPERM2X128 because its mask
9254 // allows us to replace the zero input with an implicit zero.
9255 if (!IsV1Zero && !IsV2Zero) {
9256 // Check for patterns which can be matched with a single insert of a 128-bit
9258 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9259 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9260 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9261 VT.getVectorNumElements() / 2);
9262 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9263 DAG.getIntPtrConstant(0, DL));
9264 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9265 OnlyUsesV1 ? V1 : V2,
9266 DAG.getIntPtrConstant(0, DL));
9267 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9271 // Otherwise form a 128-bit permutation. After accounting for undefs,
9272 // convert the 64-bit shuffle mask selection values into 128-bit
9273 // selection bits by dividing the indexes by 2 and shifting into positions
9274 // defined by a vperm2*128 instruction's immediate control byte.
9276 // The immediate permute control byte looks like this:
9277 // [1:0] - select 128 bits from sources for low half of destination
9279 // [3] - zero low half of destination
9280 // [5:4] - select 128 bits from sources for high half of destination
9282 // [7] - zero high half of destination
9284 int MaskLO = Mask[0];
9285 if (MaskLO == SM_SentinelUndef)
9286 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9288 int MaskHI = Mask[2];
9289 if (MaskHI == SM_SentinelUndef)
9290 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9292 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9294 // If either input is a zero vector, replace it with an undef input.
9295 // Shuffle mask values < 4 are selecting elements of V1.
9296 // Shuffle mask values >= 4 are selecting elements of V2.
9297 // Adjust each half of the permute mask by clearing the half that was
9298 // selecting the zero vector and setting the zero mask bit.
9300 V1 = DAG.getUNDEF(VT);
9302 PermMask = (PermMask & 0xf0) | 0x08;
9304 PermMask = (PermMask & 0x0f) | 0x80;
9307 V2 = DAG.getUNDEF(VT);
9309 PermMask = (PermMask & 0xf0) | 0x08;
9311 PermMask = (PermMask & 0x0f) | 0x80;
9314 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9315 DAG.getConstant(PermMask, DL, MVT::i8));
9318 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9319 /// shuffling each lane.
9321 /// This will only succeed when the result of fixing the 128-bit lanes results
9322 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9323 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9324 /// the lane crosses early and then use simpler shuffles within each lane.
9326 /// FIXME: It might be worthwhile at some point to support this without
9327 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9328 /// in x86 only floating point has interesting non-repeating shuffles, and even
9329 /// those are still *marginally* more expensive.
9330 static SDValue lowerVectorShuffleByMerging128BitLanes(
9331 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9332 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9333 assert(!isSingleInputShuffleMask(Mask) &&
9334 "This is only useful with multiple inputs.");
9336 int Size = Mask.size();
9337 int LaneSize = 128 / VT.getScalarSizeInBits();
9338 int NumLanes = Size / LaneSize;
9339 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9341 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9342 // check whether the in-128-bit lane shuffles share a repeating pattern.
9343 SmallVector<int, 4> Lanes;
9344 Lanes.resize(NumLanes, -1);
9345 SmallVector<int, 4> InLaneMask;
9346 InLaneMask.resize(LaneSize, -1);
9347 for (int i = 0; i < Size; ++i) {
9351 int j = i / LaneSize;
9354 // First entry we've seen for this lane.
9355 Lanes[j] = Mask[i] / LaneSize;
9356 } else if (Lanes[j] != Mask[i] / LaneSize) {
9357 // This doesn't match the lane selected previously!
9361 // Check that within each lane we have a consistent shuffle mask.
9362 int k = i % LaneSize;
9363 if (InLaneMask[k] < 0) {
9364 InLaneMask[k] = Mask[i] % LaneSize;
9365 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9366 // This doesn't fit a repeating in-lane mask.
9371 // First shuffle the lanes into place.
9372 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9373 VT.getSizeInBits() / 64);
9374 SmallVector<int, 8> LaneMask;
9375 LaneMask.resize(NumLanes * 2, -1);
9376 for (int i = 0; i < NumLanes; ++i)
9377 if (Lanes[i] >= 0) {
9378 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9379 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9382 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9383 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9384 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9386 // Cast it back to the type we actually want.
9387 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9389 // Now do a simple shuffle that isn't lane crossing.
9390 SmallVector<int, 8> NewMask;
9391 NewMask.resize(Size, -1);
9392 for (int i = 0; i < Size; ++i)
9394 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9395 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9396 "Must not introduce lane crosses at this point!");
9398 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9401 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9404 /// This returns true if the elements from a particular input are already in the
9405 /// slot required by the given mask and require no permutation.
9406 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9407 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9408 int Size = Mask.size();
9409 for (int i = 0; i < Size; ++i)
9410 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9416 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9418 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9419 /// isn't available.
9420 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9421 const X86Subtarget *Subtarget,
9422 SelectionDAG &DAG) {
9424 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9425 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9426 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9427 ArrayRef<int> Mask = SVOp->getMask();
9428 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9430 SmallVector<int, 4> WidenedMask;
9431 if (canWidenShuffleElements(Mask, WidenedMask))
9432 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9435 if (isSingleInputShuffleMask(Mask)) {
9436 // Check for being able to broadcast a single element.
9437 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9438 Mask, Subtarget, DAG))
9441 // Use low duplicate instructions for masks that match their pattern.
9442 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9443 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9445 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9446 // Non-half-crossing single input shuffles can be lowerid with an
9447 // interleaved permutation.
9448 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9449 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9450 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9451 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9454 // With AVX2 we have direct support for this permutation.
9455 if (Subtarget->hasAVX2())
9456 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9457 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9459 // Otherwise, fall back.
9460 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9464 // X86 has dedicated unpack instructions that can handle specific blend
9465 // operations: UNPCKH and UNPCKL.
9466 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9467 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9468 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9469 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9470 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9471 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9472 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9473 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9475 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9479 // Check if the blend happens to exactly fit that of SHUFPD.
9480 if ((Mask[0] == -1 || Mask[0] < 2) &&
9481 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9482 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9483 (Mask[3] == -1 || Mask[3] >= 6)) {
9484 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9485 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9486 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9487 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9489 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9490 (Mask[1] == -1 || Mask[1] < 2) &&
9491 (Mask[2] == -1 || Mask[2] >= 6) &&
9492 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9493 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9494 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9495 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9496 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9499 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9500 // shuffle. However, if we have AVX2 and either inputs are already in place,
9501 // we will be able to shuffle even across lanes the other input in a single
9502 // instruction so skip this pattern.
9503 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9504 isShuffleMaskInputInPlace(1, Mask))))
9505 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9506 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9509 // If we have AVX2 then we always want to lower with a blend because an v4 we
9510 // can fully permute the elements.
9511 if (Subtarget->hasAVX2())
9512 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9515 // Otherwise fall back on generic lowering.
9516 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9519 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9521 /// This routine is only called when we have AVX2 and thus a reasonable
9522 /// instruction set for v4i64 shuffling..
9523 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9524 const X86Subtarget *Subtarget,
9525 SelectionDAG &DAG) {
9527 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9528 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9529 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9530 ArrayRef<int> Mask = SVOp->getMask();
9531 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9532 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9534 SmallVector<int, 4> WidenedMask;
9535 if (canWidenShuffleElements(Mask, WidenedMask))
9536 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9539 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9543 // Check for being able to broadcast a single element.
9544 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9545 Mask, Subtarget, DAG))
9548 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9549 // use lower latency instructions that will operate on both 128-bit lanes.
9550 SmallVector<int, 2> RepeatedMask;
9551 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9552 if (isSingleInputShuffleMask(Mask)) {
9553 int PSHUFDMask[] = {-1, -1, -1, -1};
9554 for (int i = 0; i < 2; ++i)
9555 if (RepeatedMask[i] >= 0) {
9556 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9557 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9560 ISD::BITCAST, DL, MVT::v4i64,
9561 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9562 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9563 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9567 // AVX2 provides a direct instruction for permuting a single input across
9569 if (isSingleInputShuffleMask(Mask))
9570 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9571 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9573 // Try to use shift instructions.
9575 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9578 // Use dedicated unpack instructions for masks that match their pattern.
9579 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9580 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9581 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9582 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9583 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9584 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9585 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9586 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9588 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9589 // shuffle. However, if we have AVX2 and either inputs are already in place,
9590 // we will be able to shuffle even across lanes the other input in a single
9591 // instruction so skip this pattern.
9592 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9593 isShuffleMaskInputInPlace(1, Mask))))
9594 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9595 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9598 // Otherwise fall back on generic blend lowering.
9599 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9603 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9605 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9606 /// isn't available.
9607 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9608 const X86Subtarget *Subtarget,
9609 SelectionDAG &DAG) {
9611 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9612 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9613 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9614 ArrayRef<int> Mask = SVOp->getMask();
9615 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9617 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9621 // Check for being able to broadcast a single element.
9622 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9623 Mask, Subtarget, DAG))
9626 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9627 // options to efficiently lower the shuffle.
9628 SmallVector<int, 4> RepeatedMask;
9629 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9630 assert(RepeatedMask.size() == 4 &&
9631 "Repeated masks must be half the mask width!");
9633 // Use even/odd duplicate instructions for masks that match their pattern.
9634 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9635 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9636 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9637 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9639 if (isSingleInputShuffleMask(Mask))
9640 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9641 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9643 // Use dedicated unpack instructions for masks that match their pattern.
9644 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9645 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9646 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9647 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9648 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9649 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9650 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9651 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9653 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9654 // have already handled any direct blends. We also need to squash the
9655 // repeated mask into a simulated v4f32 mask.
9656 for (int i = 0; i < 4; ++i)
9657 if (RepeatedMask[i] >= 8)
9658 RepeatedMask[i] -= 4;
9659 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9662 // If we have a single input shuffle with different shuffle patterns in the
9663 // two 128-bit lanes use the variable mask to VPERMILPS.
9664 if (isSingleInputShuffleMask(Mask)) {
9665 SDValue VPermMask[8];
9666 for (int i = 0; i < 8; ++i)
9667 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9668 : DAG.getConstant(Mask[i], DL, MVT::i32);
9669 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9671 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9672 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9674 if (Subtarget->hasAVX2())
9675 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9676 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9677 DAG.getNode(ISD::BUILD_VECTOR, DL,
9678 MVT::v8i32, VPermMask)),
9681 // Otherwise, fall back.
9682 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9686 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9688 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9689 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9692 // If we have AVX2 then we always want to lower with a blend because at v8 we
9693 // can fully permute the elements.
9694 if (Subtarget->hasAVX2())
9695 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9698 // Otherwise fall back on generic lowering.
9699 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9702 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9704 /// This routine is only called when we have AVX2 and thus a reasonable
9705 /// instruction set for v8i32 shuffling..
9706 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9707 const X86Subtarget *Subtarget,
9708 SelectionDAG &DAG) {
9710 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9711 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9713 ArrayRef<int> Mask = SVOp->getMask();
9714 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9715 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9717 // Whenever we can lower this as a zext, that instruction is strictly faster
9718 // than any alternative. It also allows us to fold memory operands into the
9719 // shuffle in many cases.
9720 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9721 Mask, Subtarget, DAG))
9724 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9728 // Check for being able to broadcast a single element.
9729 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9730 Mask, Subtarget, DAG))
9733 // If the shuffle mask is repeated in each 128-bit lane we can use more
9734 // efficient instructions that mirror the shuffles across the two 128-bit
9736 SmallVector<int, 4> RepeatedMask;
9737 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9738 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9739 if (isSingleInputShuffleMask(Mask))
9740 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9741 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9743 // Use dedicated unpack instructions for masks that match their pattern.
9744 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9745 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9746 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9747 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9748 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9749 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9750 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9751 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9754 // Try to use shift instructions.
9756 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9759 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9760 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9763 // If the shuffle patterns aren't repeated but it is a single input, directly
9764 // generate a cross-lane VPERMD instruction.
9765 if (isSingleInputShuffleMask(Mask)) {
9766 SDValue VPermMask[8];
9767 for (int i = 0; i < 8; ++i)
9768 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9769 : DAG.getConstant(Mask[i], DL, MVT::i32);
9771 X86ISD::VPERMV, DL, MVT::v8i32,
9772 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9775 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9777 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9778 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9781 // Otherwise fall back on generic blend lowering.
9782 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9786 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9788 /// This routine is only called when we have AVX2 and thus a reasonable
9789 /// instruction set for v16i16 shuffling..
9790 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9791 const X86Subtarget *Subtarget,
9792 SelectionDAG &DAG) {
9794 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9795 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9796 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9797 ArrayRef<int> Mask = SVOp->getMask();
9798 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9799 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9801 // Whenever we can lower this as a zext, that instruction is strictly faster
9802 // than any alternative. It also allows us to fold memory operands into the
9803 // shuffle in many cases.
9804 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9805 Mask, Subtarget, DAG))
9808 // Check for being able to broadcast a single element.
9809 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9810 Mask, Subtarget, DAG))
9813 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9817 // Use dedicated unpack instructions for masks that match their pattern.
9818 if (isShuffleEquivalent(V1, V2, Mask,
9819 {// First 128-bit lane:
9820 0, 16, 1, 17, 2, 18, 3, 19,
9821 // Second 128-bit lane:
9822 8, 24, 9, 25, 10, 26, 11, 27}))
9823 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9824 if (isShuffleEquivalent(V1, V2, Mask,
9825 {// First 128-bit lane:
9826 4, 20, 5, 21, 6, 22, 7, 23,
9827 // Second 128-bit lane:
9828 12, 28, 13, 29, 14, 30, 15, 31}))
9829 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9831 // Try to use shift instructions.
9833 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9836 // Try to use byte rotation instructions.
9837 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9838 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9841 if (isSingleInputShuffleMask(Mask)) {
9842 // There are no generalized cross-lane shuffle operations available on i16
9844 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9845 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9848 SmallVector<int, 8> RepeatedMask;
9849 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
9850 // As this is a single-input shuffle, the repeated mask should be
9851 // a strictly valid v8i16 mask that we can pass through to the v8i16
9852 // lowering to handle even the v16 case.
9853 return lowerV8I16GeneralSingleInputVectorShuffle(
9854 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
9857 SDValue PSHUFBMask[32];
9858 for (int i = 0; i < 16; ++i) {
9859 if (Mask[i] == -1) {
9860 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9864 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9865 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9866 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
9867 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
9870 ISD::BITCAST, DL, MVT::v16i16,
9872 X86ISD::PSHUFB, DL, MVT::v32i8,
9873 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9874 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9877 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9879 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9880 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9883 // Otherwise fall back on generic lowering.
9884 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9887 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9889 /// This routine is only called when we have AVX2 and thus a reasonable
9890 /// instruction set for v32i8 shuffling..
9891 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9892 const X86Subtarget *Subtarget,
9893 SelectionDAG &DAG) {
9895 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9896 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9897 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9898 ArrayRef<int> Mask = SVOp->getMask();
9899 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9900 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9902 // Whenever we can lower this as a zext, that instruction is strictly faster
9903 // than any alternative. It also allows us to fold memory operands into the
9904 // shuffle in many cases.
9905 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9906 Mask, Subtarget, DAG))
9909 // Check for being able to broadcast a single element.
9910 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9911 Mask, Subtarget, DAG))
9914 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9918 // Use dedicated unpack instructions for masks that match their pattern.
9919 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9921 if (isShuffleEquivalent(
9923 {// First 128-bit lane:
9924 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9925 // Second 128-bit lane:
9926 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9927 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9928 if (isShuffleEquivalent(
9930 {// First 128-bit lane:
9931 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9932 // Second 128-bit lane:
9933 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9934 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9936 // Try to use shift instructions.
9938 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9941 // Try to use byte rotation instructions.
9942 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9943 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9946 if (isSingleInputShuffleMask(Mask)) {
9947 // There are no generalized cross-lane shuffle operations available on i8
9949 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9950 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9953 SDValue PSHUFBMask[32];
9954 for (int i = 0; i < 32; ++i)
9957 ? DAG.getUNDEF(MVT::i8)
9958 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
9962 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9963 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9966 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9968 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9969 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9972 // Otherwise fall back on generic lowering.
9973 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9976 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9978 /// This routine either breaks down the specific type of a 256-bit x86 vector
9979 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9980 /// together based on the available instructions.
9981 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9982 MVT VT, const X86Subtarget *Subtarget,
9983 SelectionDAG &DAG) {
9985 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9986 ArrayRef<int> Mask = SVOp->getMask();
9988 // If we have a single input to the zero element, insert that into V1 if we
9989 // can do so cheaply.
9990 int NumElts = VT.getVectorNumElements();
9991 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
9992 return M >= NumElts;
9995 if (NumV2Elements == 1 && Mask[0] >= NumElts)
9996 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9997 DL, VT, V1, V2, Mask, Subtarget, DAG))
10000 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10001 // check for those subtargets here and avoid much of the subtarget querying in
10002 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10003 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10004 // floating point types there eventually, just immediately cast everything to
10005 // a float and operate entirely in that domain.
10006 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10007 int ElementBits = VT.getScalarSizeInBits();
10008 if (ElementBits < 32)
10009 // No floating point type available, decompose into 128-bit vectors.
10010 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10012 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10013 VT.getVectorNumElements());
10014 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
10015 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
10016 return DAG.getNode(ISD::BITCAST, DL, VT,
10017 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10020 switch (VT.SimpleTy) {
10022 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10024 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10026 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10028 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10030 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10032 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10035 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10039 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10040 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10041 const X86Subtarget *Subtarget,
10042 SelectionDAG &DAG) {
10044 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10045 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10046 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10047 ArrayRef<int> Mask = SVOp->getMask();
10048 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10050 // X86 has dedicated unpack instructions that can handle specific blend
10051 // operations: UNPCKH and UNPCKL.
10052 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10053 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10054 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10055 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10057 // FIXME: Implement direct support for this type!
10058 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10061 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10062 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10063 const X86Subtarget *Subtarget,
10064 SelectionDAG &DAG) {
10066 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10067 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10068 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10069 ArrayRef<int> Mask = SVOp->getMask();
10070 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10072 // Use dedicated unpack instructions for masks that match their pattern.
10073 if (isShuffleEquivalent(V1, V2, Mask,
10074 {// First 128-bit lane.
10075 0, 16, 1, 17, 4, 20, 5, 21,
10076 // Second 128-bit lane.
10077 8, 24, 9, 25, 12, 28, 13, 29}))
10078 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10079 if (isShuffleEquivalent(V1, V2, Mask,
10080 {// First 128-bit lane.
10081 2, 18, 3, 19, 6, 22, 7, 23,
10082 // Second 128-bit lane.
10083 10, 26, 11, 27, 14, 30, 15, 31}))
10084 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10086 // FIXME: Implement direct support for this type!
10087 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10090 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10091 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10092 const X86Subtarget *Subtarget,
10093 SelectionDAG &DAG) {
10095 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10096 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10097 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10098 ArrayRef<int> Mask = SVOp->getMask();
10099 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10101 // X86 has dedicated unpack instructions that can handle specific blend
10102 // operations: UNPCKH and UNPCKL.
10103 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10104 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10105 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10106 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10108 // FIXME: Implement direct support for this type!
10109 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10112 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10113 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10114 const X86Subtarget *Subtarget,
10115 SelectionDAG &DAG) {
10117 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10118 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10119 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10120 ArrayRef<int> Mask = SVOp->getMask();
10121 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10123 // Use dedicated unpack instructions for masks that match their pattern.
10124 if (isShuffleEquivalent(V1, V2, Mask,
10125 {// First 128-bit lane.
10126 0, 16, 1, 17, 4, 20, 5, 21,
10127 // Second 128-bit lane.
10128 8, 24, 9, 25, 12, 28, 13, 29}))
10129 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10130 if (isShuffleEquivalent(V1, V2, Mask,
10131 {// First 128-bit lane.
10132 2, 18, 3, 19, 6, 22, 7, 23,
10133 // Second 128-bit lane.
10134 10, 26, 11, 27, 14, 30, 15, 31}))
10135 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10137 // FIXME: Implement direct support for this type!
10138 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10141 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10142 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10143 const X86Subtarget *Subtarget,
10144 SelectionDAG &DAG) {
10146 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10147 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10148 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10149 ArrayRef<int> Mask = SVOp->getMask();
10150 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10151 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10153 // FIXME: Implement direct support for this type!
10154 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10157 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10158 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10159 const X86Subtarget *Subtarget,
10160 SelectionDAG &DAG) {
10162 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10163 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10164 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10165 ArrayRef<int> Mask = SVOp->getMask();
10166 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10167 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10169 // FIXME: Implement direct support for this type!
10170 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10173 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10175 /// This routine either breaks down the specific type of a 512-bit x86 vector
10176 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10177 /// together based on the available instructions.
10178 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10179 MVT VT, const X86Subtarget *Subtarget,
10180 SelectionDAG &DAG) {
10182 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10183 ArrayRef<int> Mask = SVOp->getMask();
10184 assert(Subtarget->hasAVX512() &&
10185 "Cannot lower 512-bit vectors w/ basic ISA!");
10187 // Check for being able to broadcast a single element.
10188 if (SDValue Broadcast =
10189 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10192 // Dispatch to each element type for lowering. If we don't have supprot for
10193 // specific element type shuffles at 512 bits, immediately split them and
10194 // lower them. Each lowering routine of a given type is allowed to assume that
10195 // the requisite ISA extensions for that element type are available.
10196 switch (VT.SimpleTy) {
10198 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10200 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10202 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10204 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10206 if (Subtarget->hasBWI())
10207 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10210 if (Subtarget->hasBWI())
10211 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10215 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10218 // Otherwise fall back on splitting.
10219 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10222 /// \brief Top-level lowering for x86 vector shuffles.
10224 /// This handles decomposition, canonicalization, and lowering of all x86
10225 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10226 /// above in helper routines. The canonicalization attempts to widen shuffles
10227 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10228 /// s.t. only one of the two inputs needs to be tested, etc.
10229 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10230 SelectionDAG &DAG) {
10231 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10232 ArrayRef<int> Mask = SVOp->getMask();
10233 SDValue V1 = Op.getOperand(0);
10234 SDValue V2 = Op.getOperand(1);
10235 MVT VT = Op.getSimpleValueType();
10236 int NumElements = VT.getVectorNumElements();
10239 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10241 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10242 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10243 if (V1IsUndef && V2IsUndef)
10244 return DAG.getUNDEF(VT);
10246 // When we create a shuffle node we put the UNDEF node to second operand,
10247 // but in some cases the first operand may be transformed to UNDEF.
10248 // In this case we should just commute the node.
10250 return DAG.getCommutedVectorShuffle(*SVOp);
10252 // Check for non-undef masks pointing at an undef vector and make the masks
10253 // undef as well. This makes it easier to match the shuffle based solely on
10257 if (M >= NumElements) {
10258 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10259 for (int &M : NewMask)
10260 if (M >= NumElements)
10262 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10265 // We actually see shuffles that are entirely re-arrangements of a set of
10266 // zero inputs. This mostly happens while decomposing complex shuffles into
10267 // simple ones. Directly lower these as a buildvector of zeros.
10268 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10269 if (Zeroable.all())
10270 return getZeroVector(VT, Subtarget, DAG, dl);
10272 // Try to collapse shuffles into using a vector type with fewer elements but
10273 // wider element types. We cap this to not form integers or floating point
10274 // elements wider than 64 bits, but it might be interesting to form i128
10275 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10276 SmallVector<int, 16> WidenedMask;
10277 if (VT.getScalarSizeInBits() < 64 &&
10278 canWidenShuffleElements(Mask, WidenedMask)) {
10279 MVT NewEltVT = VT.isFloatingPoint()
10280 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10281 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10282 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10283 // Make sure that the new vector type is legal. For example, v2f64 isn't
10285 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10286 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
10287 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
10288 return DAG.getNode(ISD::BITCAST, dl, VT,
10289 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10293 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10294 for (int M : SVOp->getMask())
10296 ++NumUndefElements;
10297 else if (M < NumElements)
10302 // Commute the shuffle as needed such that more elements come from V1 than
10303 // V2. This allows us to match the shuffle pattern strictly on how many
10304 // elements come from V1 without handling the symmetric cases.
10305 if (NumV2Elements > NumV1Elements)
10306 return DAG.getCommutedVectorShuffle(*SVOp);
10308 // When the number of V1 and V2 elements are the same, try to minimize the
10309 // number of uses of V2 in the low half of the vector. When that is tied,
10310 // ensure that the sum of indices for V1 is equal to or lower than the sum
10311 // indices for V2. When those are equal, try to ensure that the number of odd
10312 // indices for V1 is lower than the number of odd indices for V2.
10313 if (NumV1Elements == NumV2Elements) {
10314 int LowV1Elements = 0, LowV2Elements = 0;
10315 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10316 if (M >= NumElements)
10320 if (LowV2Elements > LowV1Elements) {
10321 return DAG.getCommutedVectorShuffle(*SVOp);
10322 } else if (LowV2Elements == LowV1Elements) {
10323 int SumV1Indices = 0, SumV2Indices = 0;
10324 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10325 if (SVOp->getMask()[i] >= NumElements)
10327 else if (SVOp->getMask()[i] >= 0)
10329 if (SumV2Indices < SumV1Indices) {
10330 return DAG.getCommutedVectorShuffle(*SVOp);
10331 } else if (SumV2Indices == SumV1Indices) {
10332 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10333 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10334 if (SVOp->getMask()[i] >= NumElements)
10335 NumV2OddIndices += i % 2;
10336 else if (SVOp->getMask()[i] >= 0)
10337 NumV1OddIndices += i % 2;
10338 if (NumV2OddIndices < NumV1OddIndices)
10339 return DAG.getCommutedVectorShuffle(*SVOp);
10344 // For each vector width, delegate to a specialized lowering routine.
10345 if (VT.getSizeInBits() == 128)
10346 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10348 if (VT.getSizeInBits() == 256)
10349 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10351 // Force AVX-512 vectors to be scalarized for now.
10352 // FIXME: Implement AVX-512 support!
10353 if (VT.getSizeInBits() == 512)
10354 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10356 llvm_unreachable("Unimplemented!");
10359 // This function assumes its argument is a BUILD_VECTOR of constants or
10360 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10362 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10363 unsigned &MaskValue) {
10365 unsigned NumElems = BuildVector->getNumOperands();
10366 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10367 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10368 unsigned NumElemsInLane = NumElems / NumLanes;
10370 // Blend for v16i16 should be symetric for the both lanes.
10371 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10372 SDValue EltCond = BuildVector->getOperand(i);
10373 SDValue SndLaneEltCond =
10374 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10376 int Lane1Cond = -1, Lane2Cond = -1;
10377 if (isa<ConstantSDNode>(EltCond))
10378 Lane1Cond = !isZero(EltCond);
10379 if (isa<ConstantSDNode>(SndLaneEltCond))
10380 Lane2Cond = !isZero(SndLaneEltCond);
10382 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10383 // Lane1Cond != 0, means we want the first argument.
10384 // Lane1Cond == 0, means we want the second argument.
10385 // The encoding of this argument is 0 for the first argument, 1
10386 // for the second. Therefore, invert the condition.
10387 MaskValue |= !Lane1Cond << i;
10388 else if (Lane1Cond < 0)
10389 MaskValue |= !Lane2Cond << i;
10396 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10397 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10398 const X86Subtarget *Subtarget,
10399 SelectionDAG &DAG) {
10400 SDValue Cond = Op.getOperand(0);
10401 SDValue LHS = Op.getOperand(1);
10402 SDValue RHS = Op.getOperand(2);
10404 MVT VT = Op.getSimpleValueType();
10406 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10408 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10410 // Only non-legal VSELECTs reach this lowering, convert those into generic
10411 // shuffles and re-use the shuffle lowering path for blends.
10412 SmallVector<int, 32> Mask;
10413 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10414 SDValue CondElt = CondBV->getOperand(i);
10416 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10418 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10421 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10422 // A vselect where all conditions and data are constants can be optimized into
10423 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10424 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10425 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10426 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10429 // Try to lower this to a blend-style vector shuffle. This can handle all
10430 // constant condition cases.
10431 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10434 // Variable blends are only legal from SSE4.1 onward.
10435 if (!Subtarget->hasSSE41())
10438 // Only some types will be legal on some subtargets. If we can emit a legal
10439 // VSELECT-matching blend, return Op, and but if we need to expand, return
10441 switch (Op.getSimpleValueType().SimpleTy) {
10443 // Most of the vector types have blends past SSE4.1.
10447 // The byte blends for AVX vectors were introduced only in AVX2.
10448 if (Subtarget->hasAVX2())
10455 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10456 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10459 // FIXME: We should custom lower this by fixing the condition and using i8
10465 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10466 MVT VT = Op.getSimpleValueType();
10469 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10472 if (VT.getSizeInBits() == 8) {
10473 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10474 Op.getOperand(0), Op.getOperand(1));
10475 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10476 DAG.getValueType(VT));
10477 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10480 if (VT.getSizeInBits() == 16) {
10481 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10482 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10484 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10485 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10486 DAG.getNode(ISD::BITCAST, dl,
10489 Op.getOperand(1)));
10490 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10491 Op.getOperand(0), Op.getOperand(1));
10492 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10493 DAG.getValueType(VT));
10494 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10497 if (VT == MVT::f32) {
10498 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10499 // the result back to FR32 register. It's only worth matching if the
10500 // result has a single use which is a store or a bitcast to i32. And in
10501 // the case of a store, it's not worth it if the index is a constant 0,
10502 // because a MOVSSmr can be used instead, which is smaller and faster.
10503 if (!Op.hasOneUse())
10505 SDNode *User = *Op.getNode()->use_begin();
10506 if ((User->getOpcode() != ISD::STORE ||
10507 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10508 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10509 (User->getOpcode() != ISD::BITCAST ||
10510 User->getValueType(0) != MVT::i32))
10512 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10513 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10516 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10519 if (VT == MVT::i32 || VT == MVT::i64) {
10520 // ExtractPS/pextrq works with constant index.
10521 if (isa<ConstantSDNode>(Op.getOperand(1)))
10527 /// Extract one bit from mask vector, like v16i1 or v8i1.
10528 /// AVX-512 feature.
10530 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10531 SDValue Vec = Op.getOperand(0);
10533 MVT VecVT = Vec.getSimpleValueType();
10534 SDValue Idx = Op.getOperand(1);
10535 MVT EltVT = Op.getSimpleValueType();
10537 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10538 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10539 "Unexpected vector type in ExtractBitFromMaskVector");
10541 // variable index can't be handled in mask registers,
10542 // extend vector to VR512
10543 if (!isa<ConstantSDNode>(Idx)) {
10544 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10545 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10546 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10547 ExtVT.getVectorElementType(), Ext, Idx);
10548 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10551 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10552 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10553 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10554 rc = getRegClassFor(MVT::v16i1);
10555 unsigned MaxSift = rc->getSize()*8 - 1;
10556 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10557 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10558 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10559 DAG.getConstant(MaxSift, dl, MVT::i8));
10560 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10561 DAG.getIntPtrConstant(0, dl));
10565 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10566 SelectionDAG &DAG) const {
10568 SDValue Vec = Op.getOperand(0);
10569 MVT VecVT = Vec.getSimpleValueType();
10570 SDValue Idx = Op.getOperand(1);
10572 if (Op.getSimpleValueType() == MVT::i1)
10573 return ExtractBitFromMaskVector(Op, DAG);
10575 if (!isa<ConstantSDNode>(Idx)) {
10576 if (VecVT.is512BitVector() ||
10577 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10578 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10581 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10582 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10583 MaskEltVT.getSizeInBits());
10585 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10586 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10587 getZeroVector(MaskVT, Subtarget, DAG, dl),
10588 Idx, DAG.getConstant(0, dl, getPointerTy()));
10589 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10590 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10591 Perm, DAG.getConstant(0, dl, getPointerTy()));
10596 // If this is a 256-bit vector result, first extract the 128-bit vector and
10597 // then extract the element from the 128-bit vector.
10598 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10600 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10601 // Get the 128-bit vector.
10602 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10603 MVT EltVT = VecVT.getVectorElementType();
10605 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10607 //if (IdxVal >= NumElems/2)
10608 // IdxVal -= NumElems/2;
10609 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10610 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10611 DAG.getConstant(IdxVal, dl, MVT::i32));
10614 assert(VecVT.is128BitVector() && "Unexpected vector length");
10616 if (Subtarget->hasSSE41()) {
10617 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10622 MVT VT = Op.getSimpleValueType();
10623 // TODO: handle v16i8.
10624 if (VT.getSizeInBits() == 16) {
10625 SDValue Vec = Op.getOperand(0);
10626 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10628 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10629 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10630 DAG.getNode(ISD::BITCAST, dl,
10632 Op.getOperand(1)));
10633 // Transform it so it match pextrw which produces a 32-bit result.
10634 MVT EltVT = MVT::i32;
10635 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10636 Op.getOperand(0), Op.getOperand(1));
10637 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10638 DAG.getValueType(VT));
10639 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10642 if (VT.getSizeInBits() == 32) {
10643 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10647 // SHUFPS the element to the lowest double word, then movss.
10648 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10649 MVT VVT = Op.getOperand(0).getSimpleValueType();
10650 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10651 DAG.getUNDEF(VVT), Mask);
10652 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10653 DAG.getIntPtrConstant(0, dl));
10656 if (VT.getSizeInBits() == 64) {
10657 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10658 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10659 // to match extract_elt for f64.
10660 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10664 // UNPCKHPD the element to the lowest double word, then movsd.
10665 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10666 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10667 int Mask[2] = { 1, -1 };
10668 MVT VVT = Op.getOperand(0).getSimpleValueType();
10669 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10670 DAG.getUNDEF(VVT), Mask);
10671 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10672 DAG.getIntPtrConstant(0, dl));
10678 /// Insert one bit to mask vector, like v16i1 or v8i1.
10679 /// AVX-512 feature.
10681 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10683 SDValue Vec = Op.getOperand(0);
10684 SDValue Elt = Op.getOperand(1);
10685 SDValue Idx = Op.getOperand(2);
10686 MVT VecVT = Vec.getSimpleValueType();
10688 if (!isa<ConstantSDNode>(Idx)) {
10689 // Non constant index. Extend source and destination,
10690 // insert element and then truncate the result.
10691 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10692 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10693 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10694 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10695 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10696 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10699 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10700 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10702 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10703 DAG.getConstant(IdxVal, dl, MVT::i8));
10704 if (Vec.getOpcode() == ISD::UNDEF)
10706 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10709 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10710 SelectionDAG &DAG) const {
10711 MVT VT = Op.getSimpleValueType();
10712 MVT EltVT = VT.getVectorElementType();
10714 if (EltVT == MVT::i1)
10715 return InsertBitToMaskVector(Op, DAG);
10718 SDValue N0 = Op.getOperand(0);
10719 SDValue N1 = Op.getOperand(1);
10720 SDValue N2 = Op.getOperand(2);
10721 if (!isa<ConstantSDNode>(N2))
10723 auto *N2C = cast<ConstantSDNode>(N2);
10724 unsigned IdxVal = N2C->getZExtValue();
10726 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10727 // into that, and then insert the subvector back into the result.
10728 if (VT.is256BitVector() || VT.is512BitVector()) {
10729 // With a 256-bit vector, we can insert into the zero element efficiently
10730 // using a blend if we have AVX or AVX2 and the right data type.
10731 if (VT.is256BitVector() && IdxVal == 0) {
10732 // TODO: It is worthwhile to cast integer to floating point and back
10733 // and incur a domain crossing penalty if that's what we'll end up
10734 // doing anyway after extracting to a 128-bit vector.
10735 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
10736 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
10737 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
10738 N2 = DAG.getIntPtrConstant(1, dl);
10739 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
10743 // Get the desired 128-bit vector chunk.
10744 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10746 // Insert the element into the desired chunk.
10747 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10748 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10750 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10751 DAG.getConstant(IdxIn128, dl, MVT::i32));
10753 // Insert the changed part back into the bigger vector
10754 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10756 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10758 if (Subtarget->hasSSE41()) {
10759 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10761 if (VT == MVT::v8i16) {
10762 Opc = X86ISD::PINSRW;
10764 assert(VT == MVT::v16i8);
10765 Opc = X86ISD::PINSRB;
10768 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10770 if (N1.getValueType() != MVT::i32)
10771 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10772 if (N2.getValueType() != MVT::i32)
10773 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10774 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10777 if (EltVT == MVT::f32) {
10778 // Bits [7:6] of the constant are the source select. This will always be
10779 // zero here. The DAG Combiner may combine an extract_elt index into
10780 // these bits. For example (insert (extract, 3), 2) could be matched by
10781 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
10782 // Bits [5:4] of the constant are the destination select. This is the
10783 // value of the incoming immediate.
10784 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10785 // combine either bitwise AND or insert of float 0.0 to set these bits.
10787 const Function *F = DAG.getMachineFunction().getFunction();
10788 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
10789 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
10790 // If this is an insertion of 32-bits into the low 32-bits of
10791 // a vector, we prefer to generate a blend with immediate rather
10792 // than an insertps. Blends are simpler operations in hardware and so
10793 // will always have equal or better performance than insertps.
10794 // But if optimizing for size and there's a load folding opportunity,
10795 // generate insertps because blendps does not have a 32-bit memory
10797 N2 = DAG.getIntPtrConstant(1, dl);
10798 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10799 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
10801 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
10802 // Create this as a scalar to vector..
10803 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10804 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10807 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10808 // PINSR* works with constant index.
10813 if (EltVT == MVT::i8)
10816 if (EltVT.getSizeInBits() == 16) {
10817 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10818 // as its second argument.
10819 if (N1.getValueType() != MVT::i32)
10820 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10821 if (N2.getValueType() != MVT::i32)
10822 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10823 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10828 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10830 MVT OpVT = Op.getSimpleValueType();
10832 // If this is a 256-bit vector result, first insert into a 128-bit
10833 // vector and then insert into the 256-bit vector.
10834 if (!OpVT.is128BitVector()) {
10835 // Insert into a 128-bit vector.
10836 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10837 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10838 OpVT.getVectorNumElements() / SizeFactor);
10840 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10842 // Insert the 128-bit vector.
10843 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10846 if (OpVT == MVT::v1i64 &&
10847 Op.getOperand(0).getValueType() == MVT::i64)
10848 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10850 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10851 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10852 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10853 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10856 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10857 // a simple subregister reference or explicit instructions to grab
10858 // upper bits of a vector.
10859 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10860 SelectionDAG &DAG) {
10862 SDValue In = Op.getOperand(0);
10863 SDValue Idx = Op.getOperand(1);
10864 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10865 MVT ResVT = Op.getSimpleValueType();
10866 MVT InVT = In.getSimpleValueType();
10868 if (Subtarget->hasFp256()) {
10869 if (ResVT.is128BitVector() &&
10870 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10871 isa<ConstantSDNode>(Idx)) {
10872 return Extract128BitVector(In, IdxVal, DAG, dl);
10874 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10875 isa<ConstantSDNode>(Idx)) {
10876 return Extract256BitVector(In, IdxVal, DAG, dl);
10882 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10883 // simple superregister reference or explicit instructions to insert
10884 // the upper bits of a vector.
10885 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10886 SelectionDAG &DAG) {
10887 if (!Subtarget->hasAVX())
10891 SDValue Vec = Op.getOperand(0);
10892 SDValue SubVec = Op.getOperand(1);
10893 SDValue Idx = Op.getOperand(2);
10895 if (!isa<ConstantSDNode>(Idx))
10898 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10899 MVT OpVT = Op.getSimpleValueType();
10900 MVT SubVecVT = SubVec.getSimpleValueType();
10902 // Fold two 16-byte subvector loads into one 32-byte load:
10903 // (insert_subvector (insert_subvector undef, (load addr), 0),
10904 // (load addr + 16), Elts/2)
10906 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10907 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10908 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10909 !Subtarget->isUnalignedMem32Slow()) {
10910 SDValue SubVec2 = Vec.getOperand(1);
10911 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10912 if (Idx2->getZExtValue() == 0) {
10913 SDValue Ops[] = { SubVec2, SubVec };
10914 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10921 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10922 SubVecVT.is128BitVector())
10923 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10925 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10926 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10928 if (OpVT.getVectorElementType() == MVT::i1) {
10929 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
10931 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
10932 SDValue Undef = DAG.getUNDEF(OpVT);
10933 unsigned NumElems = OpVT.getVectorNumElements();
10934 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
10936 if (IdxVal == OpVT.getVectorNumElements() / 2) {
10937 // Zero upper bits of the Vec
10938 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10939 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10941 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10943 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10944 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10947 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10949 // Zero upper bits of the Vec2
10950 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10951 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
10952 // Zero lower bits of the Vec
10953 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10954 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10955 // Merge them together
10956 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10962 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10963 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10964 // one of the above mentioned nodes. It has to be wrapped because otherwise
10965 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10966 // be used to form addressing mode. These wrapped nodes will be selected
10969 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10970 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10972 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10973 // global base reg.
10974 unsigned char OpFlag = 0;
10975 unsigned WrapperKind = X86ISD::Wrapper;
10976 CodeModel::Model M = DAG.getTarget().getCodeModel();
10978 if (Subtarget->isPICStyleRIPRel() &&
10979 (M == CodeModel::Small || M == CodeModel::Kernel))
10980 WrapperKind = X86ISD::WrapperRIP;
10981 else if (Subtarget->isPICStyleGOT())
10982 OpFlag = X86II::MO_GOTOFF;
10983 else if (Subtarget->isPICStyleStubPIC())
10984 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10986 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10987 CP->getAlignment(),
10988 CP->getOffset(), OpFlag);
10990 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10991 // With PIC, the address is actually $g + Offset.
10993 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10994 DAG.getNode(X86ISD::GlobalBaseReg,
10995 SDLoc(), getPointerTy()),
11002 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11003 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11005 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11006 // global base reg.
11007 unsigned char OpFlag = 0;
11008 unsigned WrapperKind = X86ISD::Wrapper;
11009 CodeModel::Model M = DAG.getTarget().getCodeModel();
11011 if (Subtarget->isPICStyleRIPRel() &&
11012 (M == CodeModel::Small || M == CodeModel::Kernel))
11013 WrapperKind = X86ISD::WrapperRIP;
11014 else if (Subtarget->isPICStyleGOT())
11015 OpFlag = X86II::MO_GOTOFF;
11016 else if (Subtarget->isPICStyleStubPIC())
11017 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11019 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
11022 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11024 // With PIC, the address is actually $g + Offset.
11026 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11027 DAG.getNode(X86ISD::GlobalBaseReg,
11028 SDLoc(), getPointerTy()),
11035 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11036 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11038 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11039 // global base reg.
11040 unsigned char OpFlag = 0;
11041 unsigned WrapperKind = X86ISD::Wrapper;
11042 CodeModel::Model M = DAG.getTarget().getCodeModel();
11044 if (Subtarget->isPICStyleRIPRel() &&
11045 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11046 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11047 OpFlag = X86II::MO_GOTPCREL;
11048 WrapperKind = X86ISD::WrapperRIP;
11049 } else if (Subtarget->isPICStyleGOT()) {
11050 OpFlag = X86II::MO_GOT;
11051 } else if (Subtarget->isPICStyleStubPIC()) {
11052 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11053 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11054 OpFlag = X86II::MO_DARWIN_NONLAZY;
11057 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
11060 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11062 // With PIC, the address is actually $g + Offset.
11063 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11064 !Subtarget->is64Bit()) {
11065 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11066 DAG.getNode(X86ISD::GlobalBaseReg,
11067 SDLoc(), getPointerTy()),
11071 // For symbols that require a load from a stub to get the address, emit the
11073 if (isGlobalStubReference(OpFlag))
11074 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
11075 MachinePointerInfo::getGOT(), false, false, false, 0);
11081 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11082 // Create the TargetBlockAddressAddress node.
11083 unsigned char OpFlags =
11084 Subtarget->ClassifyBlockAddressReference();
11085 CodeModel::Model M = DAG.getTarget().getCodeModel();
11086 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11087 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11089 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
11092 if (Subtarget->isPICStyleRIPRel() &&
11093 (M == CodeModel::Small || M == CodeModel::Kernel))
11094 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11096 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11098 // With PIC, the address is actually $g + Offset.
11099 if (isGlobalRelativeToPICBase(OpFlags)) {
11100 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11101 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11109 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11110 int64_t Offset, SelectionDAG &DAG) const {
11111 // Create the TargetGlobalAddress node, folding in the constant
11112 // offset if it is legal.
11113 unsigned char OpFlags =
11114 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11115 CodeModel::Model M = DAG.getTarget().getCodeModel();
11117 if (OpFlags == X86II::MO_NO_FLAG &&
11118 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11119 // A direct static reference to a global.
11120 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
11123 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
11126 if (Subtarget->isPICStyleRIPRel() &&
11127 (M == CodeModel::Small || M == CodeModel::Kernel))
11128 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11130 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11132 // With PIC, the address is actually $g + Offset.
11133 if (isGlobalRelativeToPICBase(OpFlags)) {
11134 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11135 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11139 // For globals that require a load from a stub to get the address, emit the
11141 if (isGlobalStubReference(OpFlags))
11142 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
11143 MachinePointerInfo::getGOT(), false, false, false, 0);
11145 // If there was a non-zero offset that we didn't fold, create an explicit
11146 // addition for it.
11148 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
11149 DAG.getConstant(Offset, dl, getPointerTy()));
11155 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11156 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11157 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11158 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11162 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11163 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11164 unsigned char OperandFlags, bool LocalDynamic = false) {
11165 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11166 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11168 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11169 GA->getValueType(0),
11173 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11177 SDValue Ops[] = { Chain, TGA, *InFlag };
11178 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11180 SDValue Ops[] = { Chain, TGA };
11181 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11184 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11185 MFI->setAdjustsStack(true);
11186 MFI->setHasCalls(true);
11188 SDValue Flag = Chain.getValue(1);
11189 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11192 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11194 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11197 SDLoc dl(GA); // ? function entry point might be better
11198 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11199 DAG.getNode(X86ISD::GlobalBaseReg,
11200 SDLoc(), PtrVT), InFlag);
11201 InFlag = Chain.getValue(1);
11203 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11206 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11208 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11210 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11211 X86::RAX, X86II::MO_TLSGD);
11214 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11220 // Get the start address of the TLS block for this module.
11221 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11222 .getInfo<X86MachineFunctionInfo>();
11223 MFI->incNumLocalDynamicTLSAccesses();
11227 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11228 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11231 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11232 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11233 InFlag = Chain.getValue(1);
11234 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11235 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11238 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11242 unsigned char OperandFlags = X86II::MO_DTPOFF;
11243 unsigned WrapperKind = X86ISD::Wrapper;
11244 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11245 GA->getValueType(0),
11246 GA->getOffset(), OperandFlags);
11247 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11249 // Add x@dtpoff with the base.
11250 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11253 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11254 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11255 const EVT PtrVT, TLSModel::Model model,
11256 bool is64Bit, bool isPIC) {
11259 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11260 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11261 is64Bit ? 257 : 256));
11263 SDValue ThreadPointer =
11264 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11265 MachinePointerInfo(Ptr), false, false, false, 0);
11267 unsigned char OperandFlags = 0;
11268 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11270 unsigned WrapperKind = X86ISD::Wrapper;
11271 if (model == TLSModel::LocalExec) {
11272 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11273 } else if (model == TLSModel::InitialExec) {
11275 OperandFlags = X86II::MO_GOTTPOFF;
11276 WrapperKind = X86ISD::WrapperRIP;
11278 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11281 llvm_unreachable("Unexpected model");
11284 // emit "addl x@ntpoff,%eax" (local exec)
11285 // or "addl x@indntpoff,%eax" (initial exec)
11286 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11288 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11289 GA->getOffset(), OperandFlags);
11290 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11292 if (model == TLSModel::InitialExec) {
11293 if (isPIC && !is64Bit) {
11294 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11295 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11299 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11300 MachinePointerInfo::getGOT(), false, false, false, 0);
11303 // The address of the thread local variable is the add of the thread
11304 // pointer with the offset of the variable.
11305 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11309 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11311 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11312 const GlobalValue *GV = GA->getGlobal();
11314 if (Subtarget->isTargetELF()) {
11315 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11317 case TLSModel::GeneralDynamic:
11318 if (Subtarget->is64Bit())
11319 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
11320 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
11321 case TLSModel::LocalDynamic:
11322 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
11323 Subtarget->is64Bit());
11324 case TLSModel::InitialExec:
11325 case TLSModel::LocalExec:
11326 return LowerToTLSExecModel(
11327 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
11328 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
11330 llvm_unreachable("Unknown TLS model.");
11333 if (Subtarget->isTargetDarwin()) {
11334 // Darwin only has one model of TLS. Lower to that.
11335 unsigned char OpFlag = 0;
11336 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11337 X86ISD::WrapperRIP : X86ISD::Wrapper;
11339 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11340 // global base reg.
11341 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11342 !Subtarget->is64Bit();
11344 OpFlag = X86II::MO_TLVP_PIC_BASE;
11346 OpFlag = X86II::MO_TLVP;
11348 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11349 GA->getValueType(0),
11350 GA->getOffset(), OpFlag);
11351 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11353 // With PIC32, the address is actually $g + Offset.
11355 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11356 DAG.getNode(X86ISD::GlobalBaseReg,
11357 SDLoc(), getPointerTy()),
11360 // Lowering the machine isd will make sure everything is in the right
11362 SDValue Chain = DAG.getEntryNode();
11363 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11364 SDValue Args[] = { Chain, Offset };
11365 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11367 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11368 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11369 MFI->setAdjustsStack(true);
11371 // And our return value (tls address) is in the standard call return value
11373 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11374 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11375 Chain.getValue(1));
11378 if (Subtarget->isTargetKnownWindowsMSVC() ||
11379 Subtarget->isTargetWindowsGNU()) {
11380 // Just use the implicit TLS architecture
11381 // Need to generate someting similar to:
11382 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11384 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11385 // mov rcx, qword [rdx+rcx*8]
11386 // mov eax, .tls$:tlsvar
11387 // [rax+rcx] contains the address
11388 // Windows 64bit: gs:0x58
11389 // Windows 32bit: fs:__tls_array
11392 SDValue Chain = DAG.getEntryNode();
11394 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11395 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11396 // use its literal value of 0x2C.
11397 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11398 ? Type::getInt8PtrTy(*DAG.getContext(),
11400 : Type::getInt32PtrTy(*DAG.getContext(),
11404 Subtarget->is64Bit()
11405 ? DAG.getIntPtrConstant(0x58, dl)
11406 : (Subtarget->isTargetWindowsGNU()
11407 ? DAG.getIntPtrConstant(0x2C, dl)
11408 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11410 SDValue ThreadPointer =
11411 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11412 MachinePointerInfo(Ptr), false, false, false, 0);
11415 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11416 res = ThreadPointer;
11418 // Load the _tls_index variable
11419 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11420 if (Subtarget->is64Bit())
11421 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain, IDX,
11422 MachinePointerInfo(), MVT::i32, false, false,
11425 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11426 false, false, false, 0);
11428 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()), dl,
11430 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11432 res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11435 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11436 false, false, false, 0);
11438 // Get the offset of start of .tls section
11439 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11440 GA->getValueType(0),
11441 GA->getOffset(), X86II::MO_SECREL);
11442 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11444 // The address of the thread local variable is the add of the thread
11445 // pointer with the offset of the variable.
11446 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11449 llvm_unreachable("TLS not implemented for this target.");
11452 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11453 /// and take a 2 x i32 value to shift plus a shift amount.
11454 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11455 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11456 MVT VT = Op.getSimpleValueType();
11457 unsigned VTBits = VT.getSizeInBits();
11459 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11460 SDValue ShOpLo = Op.getOperand(0);
11461 SDValue ShOpHi = Op.getOperand(1);
11462 SDValue ShAmt = Op.getOperand(2);
11463 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11464 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11466 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11467 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11468 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11469 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11470 : DAG.getConstant(0, dl, VT);
11472 SDValue Tmp2, Tmp3;
11473 if (Op.getOpcode() == ISD::SHL_PARTS) {
11474 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11475 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11477 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11478 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11481 // If the shift amount is larger or equal than the width of a part we can't
11482 // rely on the results of shld/shrd. Insert a test and select the appropriate
11483 // values for large shift amounts.
11484 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11485 DAG.getConstant(VTBits, dl, MVT::i8));
11486 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11487 AndNode, DAG.getConstant(0, dl, MVT::i8));
11490 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11491 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11492 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11494 if (Op.getOpcode() == ISD::SHL_PARTS) {
11495 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11496 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11498 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11499 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11502 SDValue Ops[2] = { Lo, Hi };
11503 return DAG.getMergeValues(Ops, dl);
11506 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11507 SelectionDAG &DAG) const {
11508 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11511 if (SrcVT.isVector()) {
11512 if (SrcVT.getVectorElementType() == MVT::i1) {
11513 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11514 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11515 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11516 Op.getOperand(0)));
11521 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11522 "Unknown SINT_TO_FP to lower!");
11524 // These are really Legal; return the operand so the caller accepts it as
11526 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11528 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11529 Subtarget->is64Bit()) {
11533 unsigned Size = SrcVT.getSizeInBits()/8;
11534 MachineFunction &MF = DAG.getMachineFunction();
11535 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11536 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11537 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11539 MachinePointerInfo::getFixedStack(SSFI),
11541 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11544 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11546 SelectionDAG &DAG) const {
11550 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11552 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11554 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11556 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11558 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11559 MachineMemOperand *MMO;
11561 int SSFI = FI->getIndex();
11563 DAG.getMachineFunction()
11564 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11565 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11567 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11568 StackSlot = StackSlot.getOperand(1);
11570 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11571 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11573 Tys, Ops, SrcVT, MMO);
11576 Chain = Result.getValue(1);
11577 SDValue InFlag = Result.getValue(2);
11579 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11580 // shouldn't be necessary except that RFP cannot be live across
11581 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11582 MachineFunction &MF = DAG.getMachineFunction();
11583 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11584 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11585 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11586 Tys = DAG.getVTList(MVT::Other);
11588 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11590 MachineMemOperand *MMO =
11591 DAG.getMachineFunction()
11592 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11593 MachineMemOperand::MOStore, SSFISize, SSFISize);
11595 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11596 Ops, Op.getValueType(), MMO);
11597 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11598 MachinePointerInfo::getFixedStack(SSFI),
11599 false, false, false, 0);
11605 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11606 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11607 SelectionDAG &DAG) const {
11608 // This algorithm is not obvious. Here it is what we're trying to output:
11611 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11612 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11614 haddpd %xmm0, %xmm0
11616 pshufd $0x4e, %xmm0, %xmm1
11622 LLVMContext *Context = DAG.getContext();
11624 // Build some magic constants.
11625 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11626 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11627 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11629 SmallVector<Constant*,2> CV1;
11631 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11632 APInt(64, 0x4330000000000000ULL))));
11634 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11635 APInt(64, 0x4530000000000000ULL))));
11636 Constant *C1 = ConstantVector::get(CV1);
11637 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11639 // Load the 64-bit value into an XMM register.
11640 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11642 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11643 MachinePointerInfo::getConstantPool(),
11644 false, false, false, 16);
11645 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11646 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11649 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11650 MachinePointerInfo::getConstantPool(),
11651 false, false, false, 16);
11652 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11653 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11656 if (Subtarget->hasSSE3()) {
11657 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11658 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11660 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11661 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11663 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11664 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11668 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11669 DAG.getIntPtrConstant(0, dl));
11672 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11673 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11674 SelectionDAG &DAG) const {
11676 // FP constant to bias correct the final result.
11677 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
11680 // Load the 32-bit value into an XMM register.
11681 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11684 // Zero out the upper parts of the register.
11685 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11687 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11688 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11689 DAG.getIntPtrConstant(0, dl));
11691 // Or the load with the bias.
11692 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11693 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11694 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11695 MVT::v2f64, Load)),
11696 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11697 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11698 MVT::v2f64, Bias)));
11699 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11700 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11701 DAG.getIntPtrConstant(0, dl));
11703 // Subtract the bias.
11704 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11706 // Handle final rounding.
11707 EVT DestVT = Op.getValueType();
11709 if (DestVT.bitsLT(MVT::f64))
11710 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11711 DAG.getIntPtrConstant(0, dl));
11712 if (DestVT.bitsGT(MVT::f64))
11713 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11715 // Handle final rounding.
11719 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11720 const X86Subtarget &Subtarget) {
11721 // The algorithm is the following:
11722 // #ifdef __SSE4_1__
11723 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11724 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11725 // (uint4) 0x53000000, 0xaa);
11727 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11728 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11730 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11731 // return (float4) lo + fhi;
11734 SDValue V = Op->getOperand(0);
11735 EVT VecIntVT = V.getValueType();
11736 bool Is128 = VecIntVT == MVT::v4i32;
11737 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11738 // If we convert to something else than the supported type, e.g., to v4f64,
11740 if (VecFloatVT != Op->getValueType(0))
11743 unsigned NumElts = VecIntVT.getVectorNumElements();
11744 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11745 "Unsupported custom type");
11746 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11748 // In the #idef/#else code, we have in common:
11749 // - The vector of constants:
11755 // Create the splat vector for 0x4b000000.
11756 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
11757 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11758 CstLow, CstLow, CstLow, CstLow};
11759 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11760 makeArrayRef(&CstLowArray[0], NumElts));
11761 // Create the splat vector for 0x53000000.
11762 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
11763 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11764 CstHigh, CstHigh, CstHigh, CstHigh};
11765 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11766 makeArrayRef(&CstHighArray[0], NumElts));
11768 // Create the right shift.
11769 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
11770 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11771 CstShift, CstShift, CstShift, CstShift};
11772 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11773 makeArrayRef(&CstShiftArray[0], NumElts));
11774 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11777 if (Subtarget.hasSSE41()) {
11778 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11779 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11780 SDValue VecCstLowBitcast =
11781 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11782 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11783 // Low will be bitcasted right away, so do not bother bitcasting back to its
11785 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11786 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11787 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11788 // (uint4) 0x53000000, 0xaa);
11789 SDValue VecCstHighBitcast =
11790 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11791 SDValue VecShiftBitcast =
11792 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11793 // High will be bitcasted right away, so do not bother bitcasting back to
11794 // its original type.
11795 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11796 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11798 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
11799 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11800 CstMask, CstMask, CstMask);
11801 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11802 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11803 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11805 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11806 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11809 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11810 SDValue CstFAdd = DAG.getConstantFP(
11811 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
11812 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11813 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11814 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11815 makeArrayRef(&CstFAddArray[0], NumElts));
11817 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11818 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11820 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11821 // return (float4) lo + fhi;
11822 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11823 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11826 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11827 SelectionDAG &DAG) const {
11828 SDValue N0 = Op.getOperand(0);
11829 MVT SVT = N0.getSimpleValueType();
11832 switch (SVT.SimpleTy) {
11834 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11839 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11840 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11841 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11845 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11848 if (Subtarget->hasAVX512())
11849 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
11850 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
11852 llvm_unreachable(nullptr);
11855 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11856 SelectionDAG &DAG) const {
11857 SDValue N0 = Op.getOperand(0);
11860 if (Op.getValueType().isVector())
11861 return lowerUINT_TO_FP_vec(Op, DAG);
11863 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11864 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11865 // the optimization here.
11866 if (DAG.SignBitIsZero(N0))
11867 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11869 MVT SrcVT = N0.getSimpleValueType();
11870 MVT DstVT = Op.getSimpleValueType();
11871 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11872 return LowerUINT_TO_FP_i64(Op, DAG);
11873 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11874 return LowerUINT_TO_FP_i32(Op, DAG);
11875 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11878 // Make a 64-bit buffer, and use it to build an FILD.
11879 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11880 if (SrcVT == MVT::i32) {
11881 SDValue WordOff = DAG.getConstant(4, dl, getPointerTy());
11882 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11883 getPointerTy(), StackSlot, WordOff);
11884 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11885 StackSlot, MachinePointerInfo(),
11887 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
11888 OffsetSlot, MachinePointerInfo(),
11890 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11894 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11895 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11896 StackSlot, MachinePointerInfo(),
11898 // For i64 source, we need to add the appropriate power of 2 if the input
11899 // was negative. This is the same as the optimization in
11900 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11901 // we must be careful to do the computation in x87 extended precision, not
11902 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11903 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11904 MachineMemOperand *MMO =
11905 DAG.getMachineFunction()
11906 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11907 MachineMemOperand::MOLoad, 8, 8);
11909 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11910 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11911 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11914 APInt FF(32, 0x5F800000ULL);
11916 // Check whether the sign bit is set.
11917 SDValue SignSet = DAG.getSetCC(dl,
11918 getSetCCResultType(*DAG.getContext(), MVT::i64),
11920 DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
11922 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11923 SDValue FudgePtr = DAG.getConstantPool(
11924 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11927 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11928 SDValue Zero = DAG.getIntPtrConstant(0, dl);
11929 SDValue Four = DAG.getIntPtrConstant(4, dl);
11930 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11932 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11934 // Load the value out, extending it from f32 to f80.
11935 // FIXME: Avoid the extend by constructing the right constant pool?
11936 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11937 FudgePtr, MachinePointerInfo::getConstantPool(),
11938 MVT::f32, false, false, false, 4);
11939 // Extend everything to 80 bits to force it to be done on x87.
11940 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11941 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
11942 DAG.getIntPtrConstant(0, dl));
11945 std::pair<SDValue,SDValue>
11946 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11947 bool IsSigned, bool IsReplace) const {
11950 EVT DstTy = Op.getValueType();
11952 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11953 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11957 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11958 DstTy.getSimpleVT() >= MVT::i16 &&
11959 "Unknown FP_TO_INT to lower!");
11961 // These are really Legal.
11962 if (DstTy == MVT::i32 &&
11963 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11964 return std::make_pair(SDValue(), SDValue());
11965 if (Subtarget->is64Bit() &&
11966 DstTy == MVT::i64 &&
11967 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11968 return std::make_pair(SDValue(), SDValue());
11970 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11971 // stack slot, or into the FTOL runtime function.
11972 MachineFunction &MF = DAG.getMachineFunction();
11973 unsigned MemSize = DstTy.getSizeInBits()/8;
11974 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11975 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11978 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11979 Opc = X86ISD::WIN_FTOL;
11981 switch (DstTy.getSimpleVT().SimpleTy) {
11982 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11983 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11984 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11985 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11988 SDValue Chain = DAG.getEntryNode();
11989 SDValue Value = Op.getOperand(0);
11990 EVT TheVT = Op.getOperand(0).getValueType();
11991 // FIXME This causes a redundant load/store if the SSE-class value is already
11992 // in memory, such as if it is on the callstack.
11993 if (isScalarFPTypeInSSEReg(TheVT)) {
11994 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11995 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11996 MachinePointerInfo::getFixedStack(SSFI),
11998 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12000 Chain, StackSlot, DAG.getValueType(TheVT)
12003 MachineMemOperand *MMO =
12004 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12005 MachineMemOperand::MOLoad, MemSize, MemSize);
12006 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12007 Chain = Value.getValue(1);
12008 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12009 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12012 MachineMemOperand *MMO =
12013 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12014 MachineMemOperand::MOStore, MemSize, MemSize);
12016 if (Opc != X86ISD::WIN_FTOL) {
12017 // Build the FP_TO_INT*_IN_MEM
12018 SDValue Ops[] = { Chain, Value, StackSlot };
12019 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12021 return std::make_pair(FIST, StackSlot);
12023 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12024 DAG.getVTList(MVT::Other, MVT::Glue),
12026 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12027 MVT::i32, ftol.getValue(1));
12028 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12029 MVT::i32, eax.getValue(2));
12030 SDValue Ops[] = { eax, edx };
12031 SDValue pair = IsReplace
12032 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12033 : DAG.getMergeValues(Ops, DL);
12034 return std::make_pair(pair, SDValue());
12038 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12039 const X86Subtarget *Subtarget) {
12040 MVT VT = Op->getSimpleValueType(0);
12041 SDValue In = Op->getOperand(0);
12042 MVT InVT = In.getSimpleValueType();
12045 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12046 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12048 // Optimize vectors in AVX mode:
12051 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12052 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12053 // Concat upper and lower parts.
12056 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12057 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12058 // Concat upper and lower parts.
12061 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12062 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12063 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12066 if (Subtarget->hasInt256())
12067 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12069 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12070 SDValue Undef = DAG.getUNDEF(InVT);
12071 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12072 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12073 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12075 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12076 VT.getVectorNumElements()/2);
12078 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
12079 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
12081 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12084 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12085 SelectionDAG &DAG) {
12086 MVT VT = Op->getSimpleValueType(0);
12087 SDValue In = Op->getOperand(0);
12088 MVT InVT = In.getSimpleValueType();
12090 unsigned int NumElts = VT.getVectorNumElements();
12091 if (NumElts != 8 && NumElts != 16)
12094 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12095 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12097 assert(InVT.getVectorElementType() == MVT::i1);
12098 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12100 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12102 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12104 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12105 if (VT.is512BitVector())
12107 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12110 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12111 SelectionDAG &DAG) {
12112 if (Subtarget->hasFp256()) {
12113 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12121 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12122 SelectionDAG &DAG) {
12124 MVT VT = Op.getSimpleValueType();
12125 SDValue In = Op.getOperand(0);
12126 MVT SVT = In.getSimpleValueType();
12128 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12129 return LowerZERO_EXTEND_AVX512(Op, DAG);
12131 if (Subtarget->hasFp256()) {
12132 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12137 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12138 VT.getVectorNumElements() != SVT.getVectorNumElements());
12142 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12144 MVT VT = Op.getSimpleValueType();
12145 SDValue In = Op.getOperand(0);
12146 MVT InVT = In.getSimpleValueType();
12148 if (VT == MVT::i1) {
12149 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12150 "Invalid scalar TRUNCATE operation");
12151 if (InVT.getSizeInBits() >= 32)
12153 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12154 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12156 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12157 "Invalid TRUNCATE operation");
12159 // move vector to mask - truncate solution for SKX
12160 if (VT.getVectorElementType() == MVT::i1) {
12161 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12162 Subtarget->hasBWI())
12163 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12164 if ((InVT.is256BitVector() || InVT.is128BitVector())
12165 && InVT.getScalarSizeInBits() <= 16 &&
12166 Subtarget->hasBWI() && Subtarget->hasVLX())
12167 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12168 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12169 Subtarget->hasDQI())
12170 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12171 if ((InVT.is256BitVector() || InVT.is128BitVector())
12172 && InVT.getScalarSizeInBits() >= 32 &&
12173 Subtarget->hasDQI() && Subtarget->hasVLX())
12174 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12176 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
12177 if (VT.getVectorElementType().getSizeInBits() >=8)
12178 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12180 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12181 unsigned NumElts = InVT.getVectorNumElements();
12182 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12183 if (InVT.getSizeInBits() < 512) {
12184 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12185 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12190 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12191 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12192 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12195 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12196 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12197 if (Subtarget->hasInt256()) {
12198 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12199 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
12200 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12202 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12203 DAG.getIntPtrConstant(0, DL));
12206 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12207 DAG.getIntPtrConstant(0, DL));
12208 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12209 DAG.getIntPtrConstant(2, DL));
12210 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12211 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12212 static const int ShufMask[] = {0, 2, 4, 6};
12213 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12216 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12217 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12218 if (Subtarget->hasInt256()) {
12219 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
12221 SmallVector<SDValue,32> pshufbMask;
12222 for (unsigned i = 0; i < 2; ++i) {
12223 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12224 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12225 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12226 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12227 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12228 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12229 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12230 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12231 for (unsigned j = 0; j < 8; ++j)
12232 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12234 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12235 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12236 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
12238 static const int ShufMask[] = {0, 2, -1, -1};
12239 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12241 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12242 DAG.getIntPtrConstant(0, DL));
12243 return DAG.getNode(ISD::BITCAST, DL, VT, In);
12246 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12247 DAG.getIntPtrConstant(0, DL));
12249 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12250 DAG.getIntPtrConstant(4, DL));
12252 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
12253 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
12255 // The PSHUFB mask:
12256 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12257 -1, -1, -1, -1, -1, -1, -1, -1};
12259 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12260 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12261 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12263 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12264 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12266 // The MOVLHPS Mask:
12267 static const int ShufMask2[] = {0, 1, 4, 5};
12268 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12269 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
12272 // Handle truncation of V256 to V128 using shuffles.
12273 if (!VT.is128BitVector() || !InVT.is256BitVector())
12276 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12278 unsigned NumElems = VT.getVectorNumElements();
12279 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12281 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12282 // Prepare truncation shuffle mask
12283 for (unsigned i = 0; i != NumElems; ++i)
12284 MaskVec[i] = i * 2;
12285 SDValue V = DAG.getVectorShuffle(NVT, DL,
12286 DAG.getNode(ISD::BITCAST, DL, NVT, In),
12287 DAG.getUNDEF(NVT), &MaskVec[0]);
12288 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12289 DAG.getIntPtrConstant(0, DL));
12292 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12293 SelectionDAG &DAG) const {
12294 assert(!Op.getSimpleValueType().isVector());
12296 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12297 /*IsSigned=*/ true, /*IsReplace=*/ false);
12298 SDValue FIST = Vals.first, StackSlot = Vals.second;
12299 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12300 if (!FIST.getNode()) return Op;
12302 if (StackSlot.getNode())
12303 // Load the result.
12304 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12305 FIST, StackSlot, MachinePointerInfo(),
12306 false, false, false, 0);
12308 // The node is the result.
12312 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12313 SelectionDAG &DAG) const {
12314 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12315 /*IsSigned=*/ false, /*IsReplace=*/ false);
12316 SDValue FIST = Vals.first, StackSlot = Vals.second;
12317 assert(FIST.getNode() && "Unexpected failure");
12319 if (StackSlot.getNode())
12320 // Load the result.
12321 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12322 FIST, StackSlot, MachinePointerInfo(),
12323 false, false, false, 0);
12325 // The node is the result.
12329 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12331 MVT VT = Op.getSimpleValueType();
12332 SDValue In = Op.getOperand(0);
12333 MVT SVT = In.getSimpleValueType();
12335 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12337 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12338 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12339 In, DAG.getUNDEF(SVT)));
12342 /// The only differences between FABS and FNEG are the mask and the logic op.
12343 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12344 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12345 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12346 "Wrong opcode for lowering FABS or FNEG.");
12348 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12350 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12351 // into an FNABS. We'll lower the FABS after that if it is still in use.
12353 for (SDNode *User : Op->uses())
12354 if (User->getOpcode() == ISD::FNEG)
12357 SDValue Op0 = Op.getOperand(0);
12358 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12361 MVT VT = Op.getSimpleValueType();
12362 // Assume scalar op for initialization; update for vector if needed.
12363 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
12364 // generate a 16-byte vector constant and logic op even for the scalar case.
12365 // Using a 16-byte mask allows folding the load of the mask with
12366 // the logic op, so it can save (~4 bytes) on code size.
12368 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
12369 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12370 // decide if we should generate a 16-byte constant mask when we only need 4 or
12371 // 8 bytes for the scalar case.
12372 if (VT.isVector()) {
12373 EltVT = VT.getVectorElementType();
12374 NumElts = VT.getVectorNumElements();
12377 unsigned EltBits = EltVT.getSizeInBits();
12378 LLVMContext *Context = DAG.getContext();
12379 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12381 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12382 Constant *C = ConstantInt::get(*Context, MaskElt);
12383 C = ConstantVector::getSplat(NumElts, C);
12384 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12385 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
12386 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12387 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12388 MachinePointerInfo::getConstantPool(),
12389 false, false, false, Alignment);
12391 if (VT.isVector()) {
12392 // For a vector, cast operands to a vector type, perform the logic op,
12393 // and cast the result back to the original value type.
12394 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12395 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12396 SDValue Operand = IsFNABS ?
12397 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12398 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12399 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12400 return DAG.getNode(ISD::BITCAST, dl, VT,
12401 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12404 // If not vector, then scalar.
12405 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12406 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12407 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12410 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12411 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12412 LLVMContext *Context = DAG.getContext();
12413 SDValue Op0 = Op.getOperand(0);
12414 SDValue Op1 = Op.getOperand(1);
12416 MVT VT = Op.getSimpleValueType();
12417 MVT SrcVT = Op1.getSimpleValueType();
12419 // If second operand is smaller, extend it first.
12420 if (SrcVT.bitsLT(VT)) {
12421 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12424 // And if it is bigger, shrink it first.
12425 if (SrcVT.bitsGT(VT)) {
12426 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12430 // At this point the operands and the result should have the same
12431 // type, and that won't be f80 since that is not custom lowered.
12433 const fltSemantics &Sem =
12434 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12435 const unsigned SizeInBits = VT.getSizeInBits();
12437 SmallVector<Constant *, 4> CV(
12438 VT == MVT::f64 ? 2 : 4,
12439 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12441 // First, clear all bits but the sign bit from the second operand (sign).
12442 CV[0] = ConstantFP::get(*Context,
12443 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12444 Constant *C = ConstantVector::get(CV);
12445 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12446 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12447 MachinePointerInfo::getConstantPool(),
12448 false, false, false, 16);
12449 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12451 // Next, clear the sign bit from the first operand (magnitude).
12452 // If it's a constant, we can clear it here.
12453 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12454 APFloat APF = Op0CN->getValueAPF();
12455 // If the magnitude is a positive zero, the sign bit alone is enough.
12456 if (APF.isPosZero())
12459 CV[0] = ConstantFP::get(*Context, APF);
12461 CV[0] = ConstantFP::get(
12463 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12465 C = ConstantVector::get(CV);
12466 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12467 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12468 MachinePointerInfo::getConstantPool(),
12469 false, false, false, 16);
12470 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12471 if (!isa<ConstantFPSDNode>(Op0))
12472 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12474 // OR the magnitude value with the sign bit.
12475 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12478 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12479 SDValue N0 = Op.getOperand(0);
12481 MVT VT = Op.getSimpleValueType();
12483 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12484 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12485 DAG.getConstant(1, dl, VT));
12486 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12489 // Check whether an OR'd tree is PTEST-able.
12490 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12491 SelectionDAG &DAG) {
12492 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12494 if (!Subtarget->hasSSE41())
12497 if (!Op->hasOneUse())
12500 SDNode *N = Op.getNode();
12503 SmallVector<SDValue, 8> Opnds;
12504 DenseMap<SDValue, unsigned> VecInMap;
12505 SmallVector<SDValue, 8> VecIns;
12506 EVT VT = MVT::Other;
12508 // Recognize a special case where a vector is casted into wide integer to
12510 Opnds.push_back(N->getOperand(0));
12511 Opnds.push_back(N->getOperand(1));
12513 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12514 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12515 // BFS traverse all OR'd operands.
12516 if (I->getOpcode() == ISD::OR) {
12517 Opnds.push_back(I->getOperand(0));
12518 Opnds.push_back(I->getOperand(1));
12519 // Re-evaluate the number of nodes to be traversed.
12520 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12524 // Quit if a non-EXTRACT_VECTOR_ELT
12525 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12528 // Quit if without a constant index.
12529 SDValue Idx = I->getOperand(1);
12530 if (!isa<ConstantSDNode>(Idx))
12533 SDValue ExtractedFromVec = I->getOperand(0);
12534 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12535 if (M == VecInMap.end()) {
12536 VT = ExtractedFromVec.getValueType();
12537 // Quit if not 128/256-bit vector.
12538 if (!VT.is128BitVector() && !VT.is256BitVector())
12540 // Quit if not the same type.
12541 if (VecInMap.begin() != VecInMap.end() &&
12542 VT != VecInMap.begin()->first.getValueType())
12544 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12545 VecIns.push_back(ExtractedFromVec);
12547 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12550 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12551 "Not extracted from 128-/256-bit vector.");
12553 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12555 for (DenseMap<SDValue, unsigned>::const_iterator
12556 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12557 // Quit if not all elements are used.
12558 if (I->second != FullMask)
12562 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12564 // Cast all vectors into TestVT for PTEST.
12565 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12566 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12568 // If more than one full vectors are evaluated, OR them first before PTEST.
12569 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12570 // Each iteration will OR 2 nodes and append the result until there is only
12571 // 1 node left, i.e. the final OR'd value of all vectors.
12572 SDValue LHS = VecIns[Slot];
12573 SDValue RHS = VecIns[Slot + 1];
12574 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12577 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12578 VecIns.back(), VecIns.back());
12581 /// \brief return true if \c Op has a use that doesn't just read flags.
12582 static bool hasNonFlagsUse(SDValue Op) {
12583 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12585 SDNode *User = *UI;
12586 unsigned UOpNo = UI.getOperandNo();
12587 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12588 // Look pass truncate.
12589 UOpNo = User->use_begin().getOperandNo();
12590 User = *User->use_begin();
12593 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12594 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12600 /// Emit nodes that will be selected as "test Op0,Op0", or something
12602 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12603 SelectionDAG &DAG) const {
12604 if (Op.getValueType() == MVT::i1) {
12605 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12606 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12607 DAG.getConstant(0, dl, MVT::i8));
12609 // CF and OF aren't always set the way we want. Determine which
12610 // of these we need.
12611 bool NeedCF = false;
12612 bool NeedOF = false;
12615 case X86::COND_A: case X86::COND_AE:
12616 case X86::COND_B: case X86::COND_BE:
12619 case X86::COND_G: case X86::COND_GE:
12620 case X86::COND_L: case X86::COND_LE:
12621 case X86::COND_O: case X86::COND_NO: {
12622 // Check if we really need to set the
12623 // Overflow flag. If NoSignedWrap is present
12624 // that is not actually needed.
12625 switch (Op->getOpcode()) {
12630 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
12631 if (BinNode->Flags.hasNoSignedWrap())
12641 // See if we can use the EFLAGS value from the operand instead of
12642 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12643 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12644 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12645 // Emit a CMP with 0, which is the TEST pattern.
12646 //if (Op.getValueType() == MVT::i1)
12647 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12648 // DAG.getConstant(0, MVT::i1));
12649 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12650 DAG.getConstant(0, dl, Op.getValueType()));
12652 unsigned Opcode = 0;
12653 unsigned NumOperands = 0;
12655 // Truncate operations may prevent the merge of the SETCC instruction
12656 // and the arithmetic instruction before it. Attempt to truncate the operands
12657 // of the arithmetic instruction and use a reduced bit-width instruction.
12658 bool NeedTruncation = false;
12659 SDValue ArithOp = Op;
12660 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12661 SDValue Arith = Op->getOperand(0);
12662 // Both the trunc and the arithmetic op need to have one user each.
12663 if (Arith->hasOneUse())
12664 switch (Arith.getOpcode()) {
12671 NeedTruncation = true;
12677 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12678 // which may be the result of a CAST. We use the variable 'Op', which is the
12679 // non-casted variable when we check for possible users.
12680 switch (ArithOp.getOpcode()) {
12682 // Due to an isel shortcoming, be conservative if this add is likely to be
12683 // selected as part of a load-modify-store instruction. When the root node
12684 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12685 // uses of other nodes in the match, such as the ADD in this case. This
12686 // leads to the ADD being left around and reselected, with the result being
12687 // two adds in the output. Alas, even if none our users are stores, that
12688 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12689 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12690 // climbing the DAG back to the root, and it doesn't seem to be worth the
12692 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12693 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12694 if (UI->getOpcode() != ISD::CopyToReg &&
12695 UI->getOpcode() != ISD::SETCC &&
12696 UI->getOpcode() != ISD::STORE)
12699 if (ConstantSDNode *C =
12700 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12701 // An add of one will be selected as an INC.
12702 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12703 Opcode = X86ISD::INC;
12708 // An add of negative one (subtract of one) will be selected as a DEC.
12709 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12710 Opcode = X86ISD::DEC;
12716 // Otherwise use a regular EFLAGS-setting add.
12717 Opcode = X86ISD::ADD;
12722 // If we have a constant logical shift that's only used in a comparison
12723 // against zero turn it into an equivalent AND. This allows turning it into
12724 // a TEST instruction later.
12725 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12726 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12727 EVT VT = Op.getValueType();
12728 unsigned BitWidth = VT.getSizeInBits();
12729 unsigned ShAmt = Op->getConstantOperandVal(1);
12730 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12732 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12733 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12734 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12735 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12737 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12738 DAG.getConstant(Mask, dl, VT));
12739 DAG.ReplaceAllUsesWith(Op, New);
12745 // If the primary and result isn't used, don't bother using X86ISD::AND,
12746 // because a TEST instruction will be better.
12747 if (!hasNonFlagsUse(Op))
12753 // Due to the ISEL shortcoming noted above, be conservative if this op is
12754 // likely to be selected as part of a load-modify-store instruction.
12755 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12756 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12757 if (UI->getOpcode() == ISD::STORE)
12760 // Otherwise use a regular EFLAGS-setting instruction.
12761 switch (ArithOp.getOpcode()) {
12762 default: llvm_unreachable("unexpected operator!");
12763 case ISD::SUB: Opcode = X86ISD::SUB; break;
12764 case ISD::XOR: Opcode = X86ISD::XOR; break;
12765 case ISD::AND: Opcode = X86ISD::AND; break;
12767 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12768 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12769 if (EFLAGS.getNode())
12772 Opcode = X86ISD::OR;
12786 return SDValue(Op.getNode(), 1);
12792 // If we found that truncation is beneficial, perform the truncation and
12794 if (NeedTruncation) {
12795 EVT VT = Op.getValueType();
12796 SDValue WideVal = Op->getOperand(0);
12797 EVT WideVT = WideVal.getValueType();
12798 unsigned ConvertedOp = 0;
12799 // Use a target machine opcode to prevent further DAGCombine
12800 // optimizations that may separate the arithmetic operations
12801 // from the setcc node.
12802 switch (WideVal.getOpcode()) {
12804 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12805 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12806 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12807 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12808 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12813 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12814 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12815 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12816 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12822 // Emit a CMP with 0, which is the TEST pattern.
12823 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12824 DAG.getConstant(0, dl, Op.getValueType()));
12826 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12827 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12829 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12830 DAG.ReplaceAllUsesWith(Op, New);
12831 return SDValue(New.getNode(), 1);
12834 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12836 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12837 SDLoc dl, SelectionDAG &DAG) const {
12838 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12839 if (C->getAPIntValue() == 0)
12840 return EmitTest(Op0, X86CC, dl, DAG);
12842 if (Op0.getValueType() == MVT::i1)
12843 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12846 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12847 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12848 // Do the comparison at i32 if it's smaller, besides the Atom case.
12849 // This avoids subregister aliasing issues. Keep the smaller reference
12850 // if we're optimizing for size, however, as that'll allow better folding
12851 // of memory operations.
12852 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12853 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12854 Attribute::MinSize) &&
12855 !Subtarget->isAtom()) {
12856 unsigned ExtendOp =
12857 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12858 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12859 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12861 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12862 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12863 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12865 return SDValue(Sub.getNode(), 1);
12867 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12870 /// Convert a comparison if required by the subtarget.
12871 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12872 SelectionDAG &DAG) const {
12873 // If the subtarget does not support the FUCOMI instruction, floating-point
12874 // comparisons have to be converted.
12875 if (Subtarget->hasCMov() ||
12876 Cmp.getOpcode() != X86ISD::CMP ||
12877 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12878 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12881 // The instruction selector will select an FUCOM instruction instead of
12882 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12883 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12884 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12886 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12887 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12888 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12889 DAG.getConstant(8, dl, MVT::i8));
12890 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12891 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12894 /// The minimum architected relative accuracy is 2^-12. We need one
12895 /// Newton-Raphson step to have a good float result (24 bits of precision).
12896 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12897 DAGCombinerInfo &DCI,
12898 unsigned &RefinementSteps,
12899 bool &UseOneConstNR) const {
12900 // FIXME: We should use instruction latency models to calculate the cost of
12901 // each potential sequence, but this is very hard to do reliably because
12902 // at least Intel's Core* chips have variable timing based on the number of
12903 // significant digits in the divisor and/or sqrt operand.
12904 if (!Subtarget->useSqrtEst())
12907 EVT VT = Op.getValueType();
12909 // SSE1 has rsqrtss and rsqrtps.
12910 // TODO: Add support for AVX512 (v16f32).
12911 // It is likely not profitable to do this for f64 because a double-precision
12912 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12913 // instructions: convert to single, rsqrtss, convert back to double, refine
12914 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12915 // along with FMA, this could be a throughput win.
12916 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12917 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12918 RefinementSteps = 1;
12919 UseOneConstNR = false;
12920 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12925 /// The minimum architected relative accuracy is 2^-12. We need one
12926 /// Newton-Raphson step to have a good float result (24 bits of precision).
12927 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12928 DAGCombinerInfo &DCI,
12929 unsigned &RefinementSteps) const {
12930 // FIXME: We should use instruction latency models to calculate the cost of
12931 // each potential sequence, but this is very hard to do reliably because
12932 // at least Intel's Core* chips have variable timing based on the number of
12933 // significant digits in the divisor.
12934 if (!Subtarget->useReciprocalEst())
12937 EVT VT = Op.getValueType();
12939 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12940 // TODO: Add support for AVX512 (v16f32).
12941 // It is likely not profitable to do this for f64 because a double-precision
12942 // reciprocal estimate with refinement on x86 prior to FMA requires
12943 // 15 instructions: convert to single, rcpss, convert back to double, refine
12944 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12945 // along with FMA, this could be a throughput win.
12946 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12947 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12948 RefinementSteps = ReciprocalEstimateRefinementSteps;
12949 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12954 /// If we have at least two divisions that use the same divisor, convert to
12955 /// multplication by a reciprocal. This may need to be adjusted for a given
12956 /// CPU if a division's cost is not at least twice the cost of a multiplication.
12957 /// This is because we still need one division to calculate the reciprocal and
12958 /// then we need two multiplies by that reciprocal as replacements for the
12959 /// original divisions.
12960 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
12961 return NumUsers > 1;
12964 static bool isAllOnes(SDValue V) {
12965 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12966 return C && C->isAllOnesValue();
12969 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12970 /// if it's possible.
12971 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12972 SDLoc dl, SelectionDAG &DAG) const {
12973 SDValue Op0 = And.getOperand(0);
12974 SDValue Op1 = And.getOperand(1);
12975 if (Op0.getOpcode() == ISD::TRUNCATE)
12976 Op0 = Op0.getOperand(0);
12977 if (Op1.getOpcode() == ISD::TRUNCATE)
12978 Op1 = Op1.getOperand(0);
12981 if (Op1.getOpcode() == ISD::SHL)
12982 std::swap(Op0, Op1);
12983 if (Op0.getOpcode() == ISD::SHL) {
12984 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12985 if (And00C->getZExtValue() == 1) {
12986 // If we looked past a truncate, check that it's only truncating away
12988 unsigned BitWidth = Op0.getValueSizeInBits();
12989 unsigned AndBitWidth = And.getValueSizeInBits();
12990 if (BitWidth > AndBitWidth) {
12992 DAG.computeKnownBits(Op0, Zeros, Ones);
12993 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12997 RHS = Op0.getOperand(1);
12999 } else if (Op1.getOpcode() == ISD::Constant) {
13000 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13001 uint64_t AndRHSVal = AndRHS->getZExtValue();
13002 SDValue AndLHS = Op0;
13004 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13005 LHS = AndLHS.getOperand(0);
13006 RHS = AndLHS.getOperand(1);
13009 // Use BT if the immediate can't be encoded in a TEST instruction.
13010 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13012 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13016 if (LHS.getNode()) {
13017 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13018 // instruction. Since the shift amount is in-range-or-undefined, we know
13019 // that doing a bittest on the i32 value is ok. We extend to i32 because
13020 // the encoding for the i16 version is larger than the i32 version.
13021 // Also promote i16 to i32 for performance / code size reason.
13022 if (LHS.getValueType() == MVT::i8 ||
13023 LHS.getValueType() == MVT::i16)
13024 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13026 // If the operand types disagree, extend the shift amount to match. Since
13027 // BT ignores high bits (like shifts) we can use anyextend.
13028 if (LHS.getValueType() != RHS.getValueType())
13029 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13031 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13032 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13033 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13034 DAG.getConstant(Cond, dl, MVT::i8), BT);
13040 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13042 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13047 // SSE Condition code mapping:
13056 switch (SetCCOpcode) {
13057 default: llvm_unreachable("Unexpected SETCC condition");
13059 case ISD::SETEQ: SSECC = 0; break;
13061 case ISD::SETGT: Swap = true; // Fallthrough
13063 case ISD::SETOLT: SSECC = 1; break;
13065 case ISD::SETGE: Swap = true; // Fallthrough
13067 case ISD::SETOLE: SSECC = 2; break;
13068 case ISD::SETUO: SSECC = 3; break;
13070 case ISD::SETNE: SSECC = 4; break;
13071 case ISD::SETULE: Swap = true; // Fallthrough
13072 case ISD::SETUGE: SSECC = 5; break;
13073 case ISD::SETULT: Swap = true; // Fallthrough
13074 case ISD::SETUGT: SSECC = 6; break;
13075 case ISD::SETO: SSECC = 7; break;
13077 case ISD::SETONE: SSECC = 8; break;
13080 std::swap(Op0, Op1);
13085 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13086 // ones, and then concatenate the result back.
13087 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13088 MVT VT = Op.getSimpleValueType();
13090 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13091 "Unsupported value type for operation");
13093 unsigned NumElems = VT.getVectorNumElements();
13095 SDValue CC = Op.getOperand(2);
13097 // Extract the LHS vectors
13098 SDValue LHS = Op.getOperand(0);
13099 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13100 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13102 // Extract the RHS vectors
13103 SDValue RHS = Op.getOperand(1);
13104 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13105 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13107 // Issue the operation on the smaller types and concatenate the result back
13108 MVT EltVT = VT.getVectorElementType();
13109 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13110 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13111 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13112 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13115 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13116 SDValue Op0 = Op.getOperand(0);
13117 SDValue Op1 = Op.getOperand(1);
13118 SDValue CC = Op.getOperand(2);
13119 MVT VT = Op.getSimpleValueType();
13122 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13123 "Unexpected type for boolean compare operation");
13124 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13125 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13126 DAG.getConstant(-1, dl, VT));
13127 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13128 DAG.getConstant(-1, dl, VT));
13129 switch (SetCCOpcode) {
13130 default: llvm_unreachable("Unexpected SETCC condition");
13132 // (x != y) -> ~(x ^ y)
13133 return DAG.getNode(ISD::XOR, dl, VT,
13134 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13135 DAG.getConstant(-1, dl, VT));
13137 // (x == y) -> (x ^ y)
13138 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13141 // (x > y) -> (x & ~y)
13142 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13145 // (x < y) -> (~x & y)
13146 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13149 // (x <= y) -> (~x | y)
13150 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13153 // (x >=y) -> (x | ~y)
13154 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13158 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13159 const X86Subtarget *Subtarget) {
13160 SDValue Op0 = Op.getOperand(0);
13161 SDValue Op1 = Op.getOperand(1);
13162 SDValue CC = Op.getOperand(2);
13163 MVT VT = Op.getSimpleValueType();
13166 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13167 Op.getValueType().getScalarType() == MVT::i1 &&
13168 "Cannot set masked compare for this operation");
13170 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13172 bool Unsigned = false;
13175 switch (SetCCOpcode) {
13176 default: llvm_unreachable("Unexpected SETCC condition");
13177 case ISD::SETNE: SSECC = 4; break;
13178 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13179 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13180 case ISD::SETLT: Swap = true; //fall-through
13181 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13182 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13183 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13184 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13185 case ISD::SETULE: Unsigned = true; //fall-through
13186 case ISD::SETLE: SSECC = 2; break;
13190 std::swap(Op0, Op1);
13192 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13193 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13194 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13195 DAG.getConstant(SSECC, dl, MVT::i8));
13198 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13199 /// operand \p Op1. If non-trivial (for example because it's not constant)
13200 /// return an empty value.
13201 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13203 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13207 MVT VT = Op1.getSimpleValueType();
13208 MVT EVT = VT.getVectorElementType();
13209 unsigned n = VT.getVectorNumElements();
13210 SmallVector<SDValue, 8> ULTOp1;
13212 for (unsigned i = 0; i < n; ++i) {
13213 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13214 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13217 // Avoid underflow.
13218 APInt Val = Elt->getAPIntValue();
13222 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13225 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13228 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13229 SelectionDAG &DAG) {
13230 SDValue Op0 = Op.getOperand(0);
13231 SDValue Op1 = Op.getOperand(1);
13232 SDValue CC = Op.getOperand(2);
13233 MVT VT = Op.getSimpleValueType();
13234 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13235 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13240 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13241 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13244 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13245 unsigned Opc = X86ISD::CMPP;
13246 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13247 assert(VT.getVectorNumElements() <= 16);
13248 Opc = X86ISD::CMPM;
13250 // In the two special cases we can't handle, emit two comparisons.
13253 unsigned CombineOpc;
13254 if (SetCCOpcode == ISD::SETUEQ) {
13255 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13257 assert(SetCCOpcode == ISD::SETONE);
13258 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13261 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13262 DAG.getConstant(CC0, dl, MVT::i8));
13263 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13264 DAG.getConstant(CC1, dl, MVT::i8));
13265 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13267 // Handle all other FP comparisons here.
13268 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13269 DAG.getConstant(SSECC, dl, MVT::i8));
13272 // Break 256-bit integer vector compare into smaller ones.
13273 if (VT.is256BitVector() && !Subtarget->hasInt256())
13274 return Lower256IntVSETCC(Op, DAG);
13276 EVT OpVT = Op1.getValueType();
13277 if (OpVT.getVectorElementType() == MVT::i1)
13278 return LowerBoolVSETCC_AVX512(Op, DAG);
13280 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13281 if (Subtarget->hasAVX512()) {
13282 if (Op1.getValueType().is512BitVector() ||
13283 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13284 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13285 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13287 // In AVX-512 architecture setcc returns mask with i1 elements,
13288 // But there is no compare instruction for i8 and i16 elements in KNL.
13289 // We are not talking about 512-bit operands in this case, these
13290 // types are illegal.
13292 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13293 OpVT.getVectorElementType().getSizeInBits() >= 8))
13294 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13295 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13298 // We are handling one of the integer comparisons here. Since SSE only has
13299 // GT and EQ comparisons for integer, swapping operands and multiple
13300 // operations may be required for some comparisons.
13302 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13303 bool Subus = false;
13305 switch (SetCCOpcode) {
13306 default: llvm_unreachable("Unexpected SETCC condition");
13307 case ISD::SETNE: Invert = true;
13308 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13309 case ISD::SETLT: Swap = true;
13310 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13311 case ISD::SETGE: Swap = true;
13312 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13313 Invert = true; break;
13314 case ISD::SETULT: Swap = true;
13315 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13316 FlipSigns = true; break;
13317 case ISD::SETUGE: Swap = true;
13318 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13319 FlipSigns = true; Invert = true; break;
13322 // Special case: Use min/max operations for SETULE/SETUGE
13323 MVT VET = VT.getVectorElementType();
13325 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13326 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13329 switch (SetCCOpcode) {
13331 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
13332 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
13335 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13338 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13339 if (!MinMax && hasSubus) {
13340 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13342 // t = psubus Op0, Op1
13343 // pcmpeq t, <0..0>
13344 switch (SetCCOpcode) {
13346 case ISD::SETULT: {
13347 // If the comparison is against a constant we can turn this into a
13348 // setule. With psubus, setule does not require a swap. This is
13349 // beneficial because the constant in the register is no longer
13350 // destructed as the destination so it can be hoisted out of a loop.
13351 // Only do this pre-AVX since vpcmp* is no longer destructive.
13352 if (Subtarget->hasAVX())
13354 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13355 if (ULEOp1.getNode()) {
13357 Subus = true; Invert = false; Swap = false;
13361 // Psubus is better than flip-sign because it requires no inversion.
13362 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13363 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13367 Opc = X86ISD::SUBUS;
13373 std::swap(Op0, Op1);
13375 // Check that the operation in question is available (most are plain SSE2,
13376 // but PCMPGTQ and PCMPEQQ have different requirements).
13377 if (VT == MVT::v2i64) {
13378 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13379 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13381 // First cast everything to the right type.
13382 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13383 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13385 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13386 // bits of the inputs before performing those operations. The lower
13387 // compare is always unsigned.
13390 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13392 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13393 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13394 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13395 Sign, Zero, Sign, Zero);
13397 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13398 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13400 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13401 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13402 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13404 // Create masks for only the low parts/high parts of the 64 bit integers.
13405 static const int MaskHi[] = { 1, 1, 3, 3 };
13406 static const int MaskLo[] = { 0, 0, 2, 2 };
13407 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13408 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13409 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13411 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13412 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13415 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13417 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13420 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13421 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13422 // pcmpeqd + pshufd + pand.
13423 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13425 // First cast everything to the right type.
13426 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13427 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13430 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13432 // Make sure the lower and upper halves are both all-ones.
13433 static const int Mask[] = { 1, 0, 3, 2 };
13434 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13435 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13438 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13440 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13444 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13445 // bits of the inputs before performing those operations.
13447 EVT EltVT = VT.getVectorElementType();
13448 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13450 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13451 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13454 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13456 // If the logical-not of the result is required, perform that now.
13458 Result = DAG.getNOT(dl, Result, VT);
13461 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13464 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13465 getZeroVector(VT, Subtarget, DAG, dl));
13470 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13472 MVT VT = Op.getSimpleValueType();
13474 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13476 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13477 && "SetCC type must be 8-bit or 1-bit integer");
13478 SDValue Op0 = Op.getOperand(0);
13479 SDValue Op1 = Op.getOperand(1);
13481 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13483 // Optimize to BT if possible.
13484 // Lower (X & (1 << N)) == 0 to BT(X, N).
13485 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13486 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13487 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13488 Op1.getOpcode() == ISD::Constant &&
13489 cast<ConstantSDNode>(Op1)->isNullValue() &&
13490 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13491 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13492 if (NewSetCC.getNode()) {
13494 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13499 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13501 if (Op1.getOpcode() == ISD::Constant &&
13502 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13503 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13504 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13506 // If the input is a setcc, then reuse the input setcc or use a new one with
13507 // the inverted condition.
13508 if (Op0.getOpcode() == X86ISD::SETCC) {
13509 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13510 bool Invert = (CC == ISD::SETNE) ^
13511 cast<ConstantSDNode>(Op1)->isNullValue();
13515 CCode = X86::GetOppositeBranchCondition(CCode);
13516 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13517 DAG.getConstant(CCode, dl, MVT::i8),
13518 Op0.getOperand(1));
13520 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13524 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13525 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13526 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13528 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13529 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13532 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13533 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13534 if (X86CC == X86::COND_INVALID)
13537 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13538 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13539 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13540 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13542 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13546 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13547 static bool isX86LogicalCmp(SDValue Op) {
13548 unsigned Opc = Op.getNode()->getOpcode();
13549 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13550 Opc == X86ISD::SAHF)
13552 if (Op.getResNo() == 1 &&
13553 (Opc == X86ISD::ADD ||
13554 Opc == X86ISD::SUB ||
13555 Opc == X86ISD::ADC ||
13556 Opc == X86ISD::SBB ||
13557 Opc == X86ISD::SMUL ||
13558 Opc == X86ISD::UMUL ||
13559 Opc == X86ISD::INC ||
13560 Opc == X86ISD::DEC ||
13561 Opc == X86ISD::OR ||
13562 Opc == X86ISD::XOR ||
13563 Opc == X86ISD::AND))
13566 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13572 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13573 if (V.getOpcode() != ISD::TRUNCATE)
13576 SDValue VOp0 = V.getOperand(0);
13577 unsigned InBits = VOp0.getValueSizeInBits();
13578 unsigned Bits = V.getValueSizeInBits();
13579 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13582 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13583 bool addTest = true;
13584 SDValue Cond = Op.getOperand(0);
13585 SDValue Op1 = Op.getOperand(1);
13586 SDValue Op2 = Op.getOperand(2);
13588 EVT VT = Op1.getValueType();
13591 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13592 // are available or VBLENDV if AVX is available.
13593 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13594 if (Cond.getOpcode() == ISD::SETCC &&
13595 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13596 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13597 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13598 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13599 int SSECC = translateX86FSETCC(
13600 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13603 if (Subtarget->hasAVX512()) {
13604 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13605 DAG.getConstant(SSECC, DL, MVT::i8));
13606 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13609 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13610 DAG.getConstant(SSECC, DL, MVT::i8));
13612 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13613 // of 3 logic instructions for size savings and potentially speed.
13614 // Unfortunately, there is no scalar form of VBLENDV.
13616 // If either operand is a constant, don't try this. We can expect to
13617 // optimize away at least one of the logic instructions later in that
13618 // case, so that sequence would be faster than a variable blend.
13620 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13621 // uses XMM0 as the selection register. That may need just as many
13622 // instructions as the AND/ANDN/OR sequence due to register moves, so
13625 if (Subtarget->hasAVX() &&
13626 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13628 // Convert to vectors, do a VSELECT, and convert back to scalar.
13629 // All of the conversions should be optimized away.
13631 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13632 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13633 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13634 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13636 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13637 VCmp = DAG.getNode(ISD::BITCAST, DL, VCmpVT, VCmp);
13639 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13641 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13642 VSel, DAG.getIntPtrConstant(0, DL));
13644 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13645 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13646 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13650 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
13652 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
13653 Op1Scalar = ConvertI1VectorToInterger(Op1, DAG);
13654 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
13655 Op1Scalar = Op1.getOperand(0);
13657 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
13658 Op2Scalar = ConvertI1VectorToInterger(Op2, DAG);
13659 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
13660 Op2Scalar = Op2.getOperand(0);
13661 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
13662 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
13663 Op1Scalar.getValueType(),
13664 Cond, Op1Scalar, Op2Scalar);
13665 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
13666 return DAG.getNode(ISD::BITCAST, DL, VT, newSelect);
13667 SDValue ExtVec = DAG.getNode(ISD::BITCAST, DL, MVT::v8i1, newSelect);
13668 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
13669 DAG.getIntPtrConstant(0, DL));
13673 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
13674 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
13675 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13676 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
13677 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13678 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
13679 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
13681 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
13684 if (Cond.getOpcode() == ISD::SETCC) {
13685 SDValue NewCond = LowerSETCC(Cond, DAG);
13686 if (NewCond.getNode())
13690 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13691 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13692 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13693 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13694 if (Cond.getOpcode() == X86ISD::SETCC &&
13695 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13696 isZero(Cond.getOperand(1).getOperand(1))) {
13697 SDValue Cmp = Cond.getOperand(1);
13699 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13701 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13702 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13703 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13705 SDValue CmpOp0 = Cmp.getOperand(0);
13706 // Apply further optimizations for special cases
13707 // (select (x != 0), -1, 0) -> neg & sbb
13708 // (select (x == 0), 0, -1) -> neg & sbb
13709 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13710 if (YC->isNullValue() &&
13711 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13712 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13713 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13714 DAG.getConstant(0, DL,
13715 CmpOp0.getValueType()),
13717 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13718 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13719 SDValue(Neg.getNode(), 1));
13723 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13724 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
13725 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13727 SDValue Res = // Res = 0 or -1.
13728 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13729 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
13731 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13732 Res = DAG.getNOT(DL, Res, Res.getValueType());
13734 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13735 if (!N2C || !N2C->isNullValue())
13736 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13741 // Look past (and (setcc_carry (cmp ...)), 1).
13742 if (Cond.getOpcode() == ISD::AND &&
13743 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13744 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13745 if (C && C->getAPIntValue() == 1)
13746 Cond = Cond.getOperand(0);
13749 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13750 // setting operand in place of the X86ISD::SETCC.
13751 unsigned CondOpcode = Cond.getOpcode();
13752 if (CondOpcode == X86ISD::SETCC ||
13753 CondOpcode == X86ISD::SETCC_CARRY) {
13754 CC = Cond.getOperand(0);
13756 SDValue Cmp = Cond.getOperand(1);
13757 unsigned Opc = Cmp.getOpcode();
13758 MVT VT = Op.getSimpleValueType();
13760 bool IllegalFPCMov = false;
13761 if (VT.isFloatingPoint() && !VT.isVector() &&
13762 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13763 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13765 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13766 Opc == X86ISD::BT) { // FIXME
13770 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13771 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13772 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13773 Cond.getOperand(0).getValueType() != MVT::i8)) {
13774 SDValue LHS = Cond.getOperand(0);
13775 SDValue RHS = Cond.getOperand(1);
13776 unsigned X86Opcode;
13779 switch (CondOpcode) {
13780 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13781 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13782 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13783 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13784 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13785 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13786 default: llvm_unreachable("unexpected overflowing operator");
13788 if (CondOpcode == ISD::UMULO)
13789 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13792 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13794 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13796 if (CondOpcode == ISD::UMULO)
13797 Cond = X86Op.getValue(2);
13799 Cond = X86Op.getValue(1);
13801 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
13806 // Look pass the truncate if the high bits are known zero.
13807 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13808 Cond = Cond.getOperand(0);
13810 // We know the result of AND is compared against zero. Try to match
13812 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13813 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13814 if (NewSetCC.getNode()) {
13815 CC = NewSetCC.getOperand(0);
13816 Cond = NewSetCC.getOperand(1);
13823 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
13824 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13827 // a < b ? -1 : 0 -> RES = ~setcc_carry
13828 // a < b ? 0 : -1 -> RES = setcc_carry
13829 // a >= b ? -1 : 0 -> RES = setcc_carry
13830 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13831 if (Cond.getOpcode() == X86ISD::SUB) {
13832 Cond = ConvertCmpIfNecessary(Cond, DAG);
13833 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13835 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13836 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13837 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13838 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13840 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13841 return DAG.getNOT(DL, Res, Res.getValueType());
13846 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13847 // widen the cmov and push the truncate through. This avoids introducing a new
13848 // branch during isel and doesn't add any extensions.
13849 if (Op.getValueType() == MVT::i8 &&
13850 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13851 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13852 if (T1.getValueType() == T2.getValueType() &&
13853 // Blacklist CopyFromReg to avoid partial register stalls.
13854 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13855 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13856 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13857 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13861 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13862 // condition is true.
13863 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13864 SDValue Ops[] = { Op2, Op1, CC, Cond };
13865 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13868 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13869 SelectionDAG &DAG) {
13870 MVT VT = Op->getSimpleValueType(0);
13871 SDValue In = Op->getOperand(0);
13872 MVT InVT = In.getSimpleValueType();
13873 MVT VTElt = VT.getVectorElementType();
13874 MVT InVTElt = InVT.getVectorElementType();
13878 if ((InVTElt == MVT::i1) &&
13879 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13880 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13882 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13883 VTElt.getSizeInBits() <= 16)) ||
13885 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13886 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13888 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13889 VTElt.getSizeInBits() >= 32))))
13890 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13892 unsigned int NumElts = VT.getVectorNumElements();
13894 if (NumElts != 8 && NumElts != 16)
13897 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13898 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13899 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13900 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13903 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13904 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13906 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
13909 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
13911 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
13912 if (VT.is512BitVector())
13914 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
13917 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13918 SelectionDAG &DAG) {
13919 MVT VT = Op->getSimpleValueType(0);
13920 SDValue In = Op->getOperand(0);
13921 MVT InVT = In.getSimpleValueType();
13924 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13925 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13927 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13928 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13929 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13932 if (Subtarget->hasInt256())
13933 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13935 // Optimize vectors in AVX mode
13936 // Sign extend v8i16 to v8i32 and
13939 // Divide input vector into two parts
13940 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13941 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13942 // concat the vectors to original VT
13944 unsigned NumElems = InVT.getVectorNumElements();
13945 SDValue Undef = DAG.getUNDEF(InVT);
13947 SmallVector<int,8> ShufMask1(NumElems, -1);
13948 for (unsigned i = 0; i != NumElems/2; ++i)
13951 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13953 SmallVector<int,8> ShufMask2(NumElems, -1);
13954 for (unsigned i = 0; i != NumElems/2; ++i)
13955 ShufMask2[i] = i + NumElems/2;
13957 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13959 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13960 VT.getVectorNumElements()/2);
13962 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13963 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13965 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13968 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13969 // may emit an illegal shuffle but the expansion is still better than scalar
13970 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13971 // we'll emit a shuffle and a arithmetic shift.
13972 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13973 // TODO: It is possible to support ZExt by zeroing the undef values during
13974 // the shuffle phase or after the shuffle.
13975 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13976 SelectionDAG &DAG) {
13977 MVT RegVT = Op.getSimpleValueType();
13978 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13979 assert(RegVT.isInteger() &&
13980 "We only custom lower integer vector sext loads.");
13982 // Nothing useful we can do without SSE2 shuffles.
13983 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13985 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13987 EVT MemVT = Ld->getMemoryVT();
13988 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13989 unsigned RegSz = RegVT.getSizeInBits();
13991 ISD::LoadExtType Ext = Ld->getExtensionType();
13993 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13994 && "Only anyext and sext are currently implemented.");
13995 assert(MemVT != RegVT && "Cannot extend to the same type");
13996 assert(MemVT.isVector() && "Must load a vector from memory");
13998 unsigned NumElems = RegVT.getVectorNumElements();
13999 unsigned MemSz = MemVT.getSizeInBits();
14000 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14002 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14003 // The only way in which we have a legal 256-bit vector result but not the
14004 // integer 256-bit operations needed to directly lower a sextload is if we
14005 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14006 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14007 // correctly legalized. We do this late to allow the canonical form of
14008 // sextload to persist throughout the rest of the DAG combiner -- it wants
14009 // to fold together any extensions it can, and so will fuse a sign_extend
14010 // of an sextload into a sextload targeting a wider value.
14012 if (MemSz == 128) {
14013 // Just switch this to a normal load.
14014 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14015 "it must be a legal 128-bit vector "
14017 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14018 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14019 Ld->isInvariant(), Ld->getAlignment());
14021 assert(MemSz < 128 &&
14022 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14023 // Do an sext load to a 128-bit vector type. We want to use the same
14024 // number of elements, but elements half as wide. This will end up being
14025 // recursively lowered by this routine, but will succeed as we definitely
14026 // have all the necessary features if we're using AVX1.
14028 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14029 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14031 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14032 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14033 Ld->isNonTemporal(), Ld->isInvariant(),
14034 Ld->getAlignment());
14037 // Replace chain users with the new chain.
14038 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14039 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14041 // Finally, do a normal sign-extend to the desired register.
14042 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14045 // All sizes must be a power of two.
14046 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14047 "Non-power-of-two elements are not custom lowered!");
14049 // Attempt to load the original value using scalar loads.
14050 // Find the largest scalar type that divides the total loaded size.
14051 MVT SclrLoadTy = MVT::i8;
14052 for (MVT Tp : MVT::integer_valuetypes()) {
14053 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14058 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14059 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14061 SclrLoadTy = MVT::f64;
14063 // Calculate the number of scalar loads that we need to perform
14064 // in order to load our vector from memory.
14065 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14067 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14068 "Can only lower sext loads with a single scalar load!");
14070 unsigned loadRegZize = RegSz;
14071 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14074 // Represent our vector as a sequence of elements which are the
14075 // largest scalar that we can load.
14076 EVT LoadUnitVecVT = EVT::getVectorVT(
14077 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14079 // Represent the data using the same element type that is stored in
14080 // memory. In practice, we ''widen'' MemVT.
14082 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14083 loadRegZize / MemVT.getScalarType().getSizeInBits());
14085 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14086 "Invalid vector type");
14088 // We can't shuffle using an illegal type.
14089 assert(TLI.isTypeLegal(WideVecVT) &&
14090 "We only lower types that form legal widened vector types");
14092 SmallVector<SDValue, 8> Chains;
14093 SDValue Ptr = Ld->getBasePtr();
14094 SDValue Increment =
14095 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl, TLI.getPointerTy());
14096 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14098 for (unsigned i = 0; i < NumLoads; ++i) {
14099 // Perform a single load.
14100 SDValue ScalarLoad =
14101 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14102 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14103 Ld->getAlignment());
14104 Chains.push_back(ScalarLoad.getValue(1));
14105 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14106 // another round of DAGCombining.
14108 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14110 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14111 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14113 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14116 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14118 // Bitcast the loaded value to a vector of the original element type, in
14119 // the size of the target vector type.
14120 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
14121 unsigned SizeRatio = RegSz / MemSz;
14123 if (Ext == ISD::SEXTLOAD) {
14124 // If we have SSE4.1, we can directly emit a VSEXT node.
14125 if (Subtarget->hasSSE41()) {
14126 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14127 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14131 // Otherwise we'll shuffle the small elements in the high bits of the
14132 // larger type and perform an arithmetic shift. If the shift is not legal
14133 // it's better to scalarize.
14134 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14135 "We can't implement a sext load without an arithmetic right shift!");
14137 // Redistribute the loaded elements into the different locations.
14138 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14139 for (unsigned i = 0; i != NumElems; ++i)
14140 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14142 SDValue Shuff = DAG.getVectorShuffle(
14143 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14145 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14147 // Build the arithmetic shift.
14148 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14149 MemVT.getVectorElementType().getSizeInBits();
14151 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14152 DAG.getConstant(Amt, dl, RegVT));
14154 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14158 // Redistribute the loaded elements into the different locations.
14159 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14160 for (unsigned i = 0; i != NumElems; ++i)
14161 ShuffleVec[i * SizeRatio] = i;
14163 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14164 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14166 // Bitcast to the requested type.
14167 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14168 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14172 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14173 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14174 // from the AND / OR.
14175 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14176 Opc = Op.getOpcode();
14177 if (Opc != ISD::OR && Opc != ISD::AND)
14179 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14180 Op.getOperand(0).hasOneUse() &&
14181 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14182 Op.getOperand(1).hasOneUse());
14185 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14186 // 1 and that the SETCC node has a single use.
14187 static bool isXor1OfSetCC(SDValue Op) {
14188 if (Op.getOpcode() != ISD::XOR)
14190 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14191 if (N1C && N1C->getAPIntValue() == 1) {
14192 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14193 Op.getOperand(0).hasOneUse();
14198 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14199 bool addTest = true;
14200 SDValue Chain = Op.getOperand(0);
14201 SDValue Cond = Op.getOperand(1);
14202 SDValue Dest = Op.getOperand(2);
14205 bool Inverted = false;
14207 if (Cond.getOpcode() == ISD::SETCC) {
14208 // Check for setcc([su]{add,sub,mul}o == 0).
14209 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14210 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14211 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14212 Cond.getOperand(0).getResNo() == 1 &&
14213 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14214 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14215 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14216 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14217 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14218 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14220 Cond = Cond.getOperand(0);
14222 SDValue NewCond = LowerSETCC(Cond, DAG);
14223 if (NewCond.getNode())
14228 // FIXME: LowerXALUO doesn't handle these!!
14229 else if (Cond.getOpcode() == X86ISD::ADD ||
14230 Cond.getOpcode() == X86ISD::SUB ||
14231 Cond.getOpcode() == X86ISD::SMUL ||
14232 Cond.getOpcode() == X86ISD::UMUL)
14233 Cond = LowerXALUO(Cond, DAG);
14236 // Look pass (and (setcc_carry (cmp ...)), 1).
14237 if (Cond.getOpcode() == ISD::AND &&
14238 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14239 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14240 if (C && C->getAPIntValue() == 1)
14241 Cond = Cond.getOperand(0);
14244 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14245 // setting operand in place of the X86ISD::SETCC.
14246 unsigned CondOpcode = Cond.getOpcode();
14247 if (CondOpcode == X86ISD::SETCC ||
14248 CondOpcode == X86ISD::SETCC_CARRY) {
14249 CC = Cond.getOperand(0);
14251 SDValue Cmp = Cond.getOperand(1);
14252 unsigned Opc = Cmp.getOpcode();
14253 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14254 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14258 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14262 // These can only come from an arithmetic instruction with overflow,
14263 // e.g. SADDO, UADDO.
14264 Cond = Cond.getNode()->getOperand(1);
14270 CondOpcode = Cond.getOpcode();
14271 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14272 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14273 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14274 Cond.getOperand(0).getValueType() != MVT::i8)) {
14275 SDValue LHS = Cond.getOperand(0);
14276 SDValue RHS = Cond.getOperand(1);
14277 unsigned X86Opcode;
14280 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14281 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14283 switch (CondOpcode) {
14284 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14286 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14288 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14291 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14292 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14294 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14296 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14299 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14300 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14301 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14302 default: llvm_unreachable("unexpected overflowing operator");
14305 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14306 if (CondOpcode == ISD::UMULO)
14307 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14310 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14312 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14314 if (CondOpcode == ISD::UMULO)
14315 Cond = X86Op.getValue(2);
14317 Cond = X86Op.getValue(1);
14319 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14323 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14324 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14325 if (CondOpc == ISD::OR) {
14326 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14327 // two branches instead of an explicit OR instruction with a
14329 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14330 isX86LogicalCmp(Cmp)) {
14331 CC = Cond.getOperand(0).getOperand(0);
14332 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14333 Chain, Dest, CC, Cmp);
14334 CC = Cond.getOperand(1).getOperand(0);
14338 } else { // ISD::AND
14339 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14340 // two branches instead of an explicit AND instruction with a
14341 // separate test. However, we only do this if this block doesn't
14342 // have a fall-through edge, because this requires an explicit
14343 // jmp when the condition is false.
14344 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14345 isX86LogicalCmp(Cmp) &&
14346 Op.getNode()->hasOneUse()) {
14347 X86::CondCode CCode =
14348 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14349 CCode = X86::GetOppositeBranchCondition(CCode);
14350 CC = DAG.getConstant(CCode, dl, MVT::i8);
14351 SDNode *User = *Op.getNode()->use_begin();
14352 // Look for an unconditional branch following this conditional branch.
14353 // We need this because we need to reverse the successors in order
14354 // to implement FCMP_OEQ.
14355 if (User->getOpcode() == ISD::BR) {
14356 SDValue FalseBB = User->getOperand(1);
14358 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14359 assert(NewBR == User);
14363 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14364 Chain, Dest, CC, Cmp);
14365 X86::CondCode CCode =
14366 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14367 CCode = X86::GetOppositeBranchCondition(CCode);
14368 CC = DAG.getConstant(CCode, dl, MVT::i8);
14374 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14375 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14376 // It should be transformed during dag combiner except when the condition
14377 // is set by a arithmetics with overflow node.
14378 X86::CondCode CCode =
14379 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14380 CCode = X86::GetOppositeBranchCondition(CCode);
14381 CC = DAG.getConstant(CCode, dl, MVT::i8);
14382 Cond = Cond.getOperand(0).getOperand(1);
14384 } else if (Cond.getOpcode() == ISD::SETCC &&
14385 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14386 // For FCMP_OEQ, we can emit
14387 // two branches instead of an explicit AND instruction with a
14388 // separate test. However, we only do this if this block doesn't
14389 // have a fall-through edge, because this requires an explicit
14390 // jmp when the condition is false.
14391 if (Op.getNode()->hasOneUse()) {
14392 SDNode *User = *Op.getNode()->use_begin();
14393 // Look for an unconditional branch following this conditional branch.
14394 // We need this because we need to reverse the successors in order
14395 // to implement FCMP_OEQ.
14396 if (User->getOpcode() == ISD::BR) {
14397 SDValue FalseBB = User->getOperand(1);
14399 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14400 assert(NewBR == User);
14404 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14405 Cond.getOperand(0), Cond.getOperand(1));
14406 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14407 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14408 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14409 Chain, Dest, CC, Cmp);
14410 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14415 } else if (Cond.getOpcode() == ISD::SETCC &&
14416 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14417 // For FCMP_UNE, we can emit
14418 // two branches instead of an explicit AND instruction with a
14419 // separate test. However, we only do this if this block doesn't
14420 // have a fall-through edge, because this requires an explicit
14421 // jmp when the condition is false.
14422 if (Op.getNode()->hasOneUse()) {
14423 SDNode *User = *Op.getNode()->use_begin();
14424 // Look for an unconditional branch following this conditional branch.
14425 // We need this because we need to reverse the successors in order
14426 // to implement FCMP_UNE.
14427 if (User->getOpcode() == ISD::BR) {
14428 SDValue FalseBB = User->getOperand(1);
14430 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14431 assert(NewBR == User);
14434 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14435 Cond.getOperand(0), Cond.getOperand(1));
14436 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14437 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14438 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14439 Chain, Dest, CC, Cmp);
14440 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14450 // Look pass the truncate if the high bits are known zero.
14451 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14452 Cond = Cond.getOperand(0);
14454 // We know the result of AND is compared against zero. Try to match
14456 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14457 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14458 if (NewSetCC.getNode()) {
14459 CC = NewSetCC.getOperand(0);
14460 Cond = NewSetCC.getOperand(1);
14467 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14468 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14469 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14471 Cond = ConvertCmpIfNecessary(Cond, DAG);
14472 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14473 Chain, Dest, CC, Cond);
14476 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14477 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14478 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14479 // that the guard pages used by the OS virtual memory manager are allocated in
14480 // correct sequence.
14482 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14483 SelectionDAG &DAG) const {
14484 MachineFunction &MF = DAG.getMachineFunction();
14485 bool SplitStack = MF.shouldSplitStack();
14486 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14491 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14492 SDNode* Node = Op.getNode();
14494 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14495 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14496 " not tell us which reg is the stack pointer!");
14497 EVT VT = Node->getValueType(0);
14498 SDValue Tmp1 = SDValue(Node, 0);
14499 SDValue Tmp2 = SDValue(Node, 1);
14500 SDValue Tmp3 = Node->getOperand(2);
14501 SDValue Chain = Tmp1.getOperand(0);
14503 // Chain the dynamic stack allocation so that it doesn't modify the stack
14504 // pointer when other instructions are using the stack.
14505 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14508 SDValue Size = Tmp2.getOperand(1);
14509 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14510 Chain = SP.getValue(1);
14511 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14512 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14513 unsigned StackAlign = TFI.getStackAlignment();
14514 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14515 if (Align > StackAlign)
14516 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14517 DAG.getConstant(-(uint64_t)Align, dl, VT));
14518 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14520 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14521 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14524 SDValue Ops[2] = { Tmp1, Tmp2 };
14525 return DAG.getMergeValues(Ops, dl);
14529 SDValue Chain = Op.getOperand(0);
14530 SDValue Size = Op.getOperand(1);
14531 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14532 EVT VT = Op.getNode()->getValueType(0);
14534 bool Is64Bit = Subtarget->is64Bit();
14535 EVT SPTy = getPointerTy();
14538 MachineRegisterInfo &MRI = MF.getRegInfo();
14541 // The 64 bit implementation of segmented stacks needs to clobber both r10
14542 // r11. This makes it impossible to use it along with nested parameters.
14543 const Function *F = MF.getFunction();
14545 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14547 if (I->hasNestAttr())
14548 report_fatal_error("Cannot use segmented stacks with functions that "
14549 "have nested arguments.");
14552 const TargetRegisterClass *AddrRegClass =
14553 getRegClassFor(getPointerTy());
14554 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14555 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14556 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14557 DAG.getRegister(Vreg, SPTy));
14558 SDValue Ops1[2] = { Value, Chain };
14559 return DAG.getMergeValues(Ops1, dl);
14562 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14564 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14565 Flag = Chain.getValue(1);
14566 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14568 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14570 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14571 unsigned SPReg = RegInfo->getStackRegister();
14572 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14573 Chain = SP.getValue(1);
14576 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14577 DAG.getConstant(-(uint64_t)Align, dl, VT));
14578 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14581 SDValue Ops1[2] = { SP, Chain };
14582 return DAG.getMergeValues(Ops1, dl);
14586 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14587 MachineFunction &MF = DAG.getMachineFunction();
14588 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14590 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14593 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14594 // vastart just stores the address of the VarArgsFrameIndex slot into the
14595 // memory location argument.
14596 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14598 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14599 MachinePointerInfo(SV), false, false, 0);
14603 // gp_offset (0 - 6 * 8)
14604 // fp_offset (48 - 48 + 8 * 16)
14605 // overflow_arg_area (point to parameters coming in memory).
14607 SmallVector<SDValue, 8> MemOps;
14608 SDValue FIN = Op.getOperand(1);
14610 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14611 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14613 FIN, MachinePointerInfo(SV), false, false, 0);
14614 MemOps.push_back(Store);
14617 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14618 FIN, DAG.getIntPtrConstant(4, DL));
14619 Store = DAG.getStore(Op.getOperand(0), DL,
14620 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
14622 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14623 MemOps.push_back(Store);
14625 // Store ptr to overflow_arg_area
14626 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14627 FIN, DAG.getIntPtrConstant(4, DL));
14628 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14630 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14631 MachinePointerInfo(SV, 8),
14633 MemOps.push_back(Store);
14635 // Store ptr to reg_save_area.
14636 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14637 FIN, DAG.getIntPtrConstant(8, DL));
14638 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14640 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14641 MachinePointerInfo(SV, 16), false, false, 0);
14642 MemOps.push_back(Store);
14643 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14646 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14647 assert(Subtarget->is64Bit() &&
14648 "LowerVAARG only handles 64-bit va_arg!");
14649 assert((Subtarget->isTargetLinux() ||
14650 Subtarget->isTargetDarwin()) &&
14651 "Unhandled target in LowerVAARG");
14652 assert(Op.getNode()->getNumOperands() == 4);
14653 SDValue Chain = Op.getOperand(0);
14654 SDValue SrcPtr = Op.getOperand(1);
14655 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14656 unsigned Align = Op.getConstantOperandVal(3);
14659 EVT ArgVT = Op.getNode()->getValueType(0);
14660 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14661 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14664 // Decide which area this value should be read from.
14665 // TODO: Implement the AMD64 ABI in its entirety. This simple
14666 // selection mechanism works only for the basic types.
14667 if (ArgVT == MVT::f80) {
14668 llvm_unreachable("va_arg for f80 not yet implemented");
14669 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14670 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14671 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14672 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14674 llvm_unreachable("Unhandled argument type in LowerVAARG");
14677 if (ArgMode == 2) {
14678 // Sanity Check: Make sure using fp_offset makes sense.
14679 assert(!Subtarget->useSoftFloat() &&
14680 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14681 Attribute::NoImplicitFloat)) &&
14682 Subtarget->hasSSE1());
14685 // Insert VAARG_64 node into the DAG
14686 // VAARG_64 returns two values: Variable Argument Address, Chain
14687 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
14688 DAG.getConstant(ArgMode, dl, MVT::i8),
14689 DAG.getConstant(Align, dl, MVT::i32)};
14690 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14691 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14692 VTs, InstOps, MVT::i64,
14693 MachinePointerInfo(SV),
14695 /*Volatile=*/false,
14697 /*WriteMem=*/true);
14698 Chain = VAARG.getValue(1);
14700 // Load the next argument and return it
14701 return DAG.getLoad(ArgVT, dl,
14704 MachinePointerInfo(),
14705 false, false, false, 0);
14708 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14709 SelectionDAG &DAG) {
14710 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14711 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14712 SDValue Chain = Op.getOperand(0);
14713 SDValue DstPtr = Op.getOperand(1);
14714 SDValue SrcPtr = Op.getOperand(2);
14715 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14716 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14719 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14720 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
14722 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14725 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14726 // amount is a constant. Takes immediate version of shift as input.
14727 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14728 SDValue SrcOp, uint64_t ShiftAmt,
14729 SelectionDAG &DAG) {
14730 MVT ElementType = VT.getVectorElementType();
14732 // Fold this packed shift into its first operand if ShiftAmt is 0.
14736 // Check for ShiftAmt >= element width
14737 if (ShiftAmt >= ElementType.getSizeInBits()) {
14738 if (Opc == X86ISD::VSRAI)
14739 ShiftAmt = ElementType.getSizeInBits() - 1;
14741 return DAG.getConstant(0, dl, VT);
14744 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14745 && "Unknown target vector shift-by-constant node");
14747 // Fold this packed vector shift into a build vector if SrcOp is a
14748 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14749 if (VT == SrcOp.getSimpleValueType() &&
14750 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14751 SmallVector<SDValue, 8> Elts;
14752 unsigned NumElts = SrcOp->getNumOperands();
14753 ConstantSDNode *ND;
14756 default: llvm_unreachable(nullptr);
14757 case X86ISD::VSHLI:
14758 for (unsigned i=0; i!=NumElts; ++i) {
14759 SDValue CurrentOp = SrcOp->getOperand(i);
14760 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14761 Elts.push_back(CurrentOp);
14764 ND = cast<ConstantSDNode>(CurrentOp);
14765 const APInt &C = ND->getAPIntValue();
14766 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
14769 case X86ISD::VSRLI:
14770 for (unsigned i=0; i!=NumElts; ++i) {
14771 SDValue CurrentOp = SrcOp->getOperand(i);
14772 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14773 Elts.push_back(CurrentOp);
14776 ND = cast<ConstantSDNode>(CurrentOp);
14777 const APInt &C = ND->getAPIntValue();
14778 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
14781 case X86ISD::VSRAI:
14782 for (unsigned i=0; i!=NumElts; ++i) {
14783 SDValue CurrentOp = SrcOp->getOperand(i);
14784 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14785 Elts.push_back(CurrentOp);
14788 ND = cast<ConstantSDNode>(CurrentOp);
14789 const APInt &C = ND->getAPIntValue();
14790 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
14795 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14798 return DAG.getNode(Opc, dl, VT, SrcOp,
14799 DAG.getConstant(ShiftAmt, dl, MVT::i8));
14802 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14803 // may or may not be a constant. Takes immediate version of shift as input.
14804 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14805 SDValue SrcOp, SDValue ShAmt,
14806 SelectionDAG &DAG) {
14807 MVT SVT = ShAmt.getSimpleValueType();
14808 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14810 // Catch shift-by-constant.
14811 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14812 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14813 CShAmt->getZExtValue(), DAG);
14815 // Change opcode to non-immediate version
14817 default: llvm_unreachable("Unknown target vector shift node");
14818 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14819 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14820 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14823 const X86Subtarget &Subtarget =
14824 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14825 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14826 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14827 // Let the shuffle legalizer expand this shift amount node.
14828 SDValue Op0 = ShAmt.getOperand(0);
14829 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14830 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14832 // Need to build a vector containing shift amount.
14833 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14834 SmallVector<SDValue, 4> ShOps;
14835 ShOps.push_back(ShAmt);
14836 if (SVT == MVT::i32) {
14837 ShOps.push_back(DAG.getConstant(0, dl, SVT));
14838 ShOps.push_back(DAG.getUNDEF(SVT));
14840 ShOps.push_back(DAG.getUNDEF(SVT));
14842 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14843 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14846 // The return type has to be a 128-bit type with the same element
14847 // type as the input type.
14848 MVT EltVT = VT.getVectorElementType();
14849 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14851 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14852 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14855 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14856 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14857 /// necessary casting for \p Mask when lowering masking intrinsics.
14858 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14859 SDValue PreservedSrc,
14860 const X86Subtarget *Subtarget,
14861 SelectionDAG &DAG) {
14862 EVT VT = Op.getValueType();
14863 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14864 MVT::i1, VT.getVectorNumElements());
14865 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14866 Mask.getValueType().getSizeInBits());
14869 assert(MaskVT.isSimple() && "invalid mask type");
14871 if (isAllOnes(Mask))
14874 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14875 // are extracted by EXTRACT_SUBVECTOR.
14876 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14877 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14878 DAG.getIntPtrConstant(0, dl));
14880 switch (Op.getOpcode()) {
14882 case X86ISD::PCMPEQM:
14883 case X86ISD::PCMPGTM:
14885 case X86ISD::CMPMU:
14886 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14888 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14889 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14890 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14893 /// \brief Creates an SDNode for a predicated scalar operation.
14894 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14895 /// The mask is comming as MVT::i8 and it should be truncated
14896 /// to MVT::i1 while lowering masking intrinsics.
14897 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14898 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14899 /// a scalar instruction.
14900 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14901 SDValue PreservedSrc,
14902 const X86Subtarget *Subtarget,
14903 SelectionDAG &DAG) {
14904 if (isAllOnes(Mask))
14907 EVT VT = Op.getValueType();
14909 // The mask should be of type MVT::i1
14910 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14912 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14913 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14914 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14917 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14918 SelectionDAG &DAG) {
14920 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14921 EVT VT = Op.getValueType();
14922 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14924 switch(IntrData->Type) {
14925 case INTR_TYPE_1OP:
14926 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14927 case INTR_TYPE_2OP:
14928 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14930 case INTR_TYPE_3OP:
14931 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14932 Op.getOperand(2), Op.getOperand(3));
14933 case INTR_TYPE_1OP_MASK_RM: {
14934 SDValue Src = Op.getOperand(1);
14935 SDValue Src0 = Op.getOperand(2);
14936 SDValue Mask = Op.getOperand(3);
14937 SDValue RoundingMode = Op.getOperand(4);
14938 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14940 Mask, Src0, Subtarget, DAG);
14942 case INTR_TYPE_SCALAR_MASK_RM: {
14943 SDValue Src1 = Op.getOperand(1);
14944 SDValue Src2 = Op.getOperand(2);
14945 SDValue Src0 = Op.getOperand(3);
14946 SDValue Mask = Op.getOperand(4);
14947 // There are 2 kinds of intrinsics in this group:
14948 // (1) With supress-all-exceptions (sae) or rounding mode- 6 operands
14949 // (2) With rounding mode and sae - 7 operands.
14950 if (Op.getNumOperands() == 6) {
14951 SDValue Sae = Op.getOperand(5);
14952 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
14953 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
14955 Mask, Src0, Subtarget, DAG);
14957 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
14958 SDValue RoundingMode = Op.getOperand(5);
14959 SDValue Sae = Op.getOperand(6);
14960 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14961 RoundingMode, Sae),
14962 Mask, Src0, Subtarget, DAG);
14964 case INTR_TYPE_2OP_MASK: {
14965 SDValue Src1 = Op.getOperand(1);
14966 SDValue Src2 = Op.getOperand(2);
14967 SDValue PassThru = Op.getOperand(3);
14968 SDValue Mask = Op.getOperand(4);
14969 // We specify 2 possible opcodes for intrinsics with rounding modes.
14970 // First, we check if the intrinsic may have non-default rounding mode,
14971 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14972 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14973 if (IntrWithRoundingModeOpcode != 0) {
14974 SDValue Rnd = Op.getOperand(5);
14975 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14976 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14977 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14978 dl, Op.getValueType(),
14980 Mask, PassThru, Subtarget, DAG);
14983 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14985 Mask, PassThru, Subtarget, DAG);
14987 case FMA_OP_MASK: {
14988 SDValue Src1 = Op.getOperand(1);
14989 SDValue Src2 = Op.getOperand(2);
14990 SDValue Src3 = Op.getOperand(3);
14991 SDValue Mask = Op.getOperand(4);
14992 // We specify 2 possible opcodes for intrinsics with rounding modes.
14993 // First, we check if the intrinsic may have non-default rounding mode,
14994 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14995 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14996 if (IntrWithRoundingModeOpcode != 0) {
14997 SDValue Rnd = Op.getOperand(5);
14998 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14999 X86::STATIC_ROUNDING::CUR_DIRECTION)
15000 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15001 dl, Op.getValueType(),
15002 Src1, Src2, Src3, Rnd),
15003 Mask, Src1, Subtarget, DAG);
15005 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15006 dl, Op.getValueType(),
15008 Mask, Src1, Subtarget, DAG);
15011 case CMP_MASK_CC: {
15012 // Comparison intrinsics with masks.
15013 // Example of transformation:
15014 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15015 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15017 // (v8i1 (insert_subvector undef,
15018 // (v2i1 (and (PCMPEQM %a, %b),
15019 // (extract_subvector
15020 // (v8i1 (bitcast %mask)), 0))), 0))))
15021 EVT VT = Op.getOperand(1).getValueType();
15022 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15023 VT.getVectorNumElements());
15024 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15025 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15026 Mask.getValueType().getSizeInBits());
15028 if (IntrData->Type == CMP_MASK_CC) {
15029 SDValue CC = Op.getOperand(3);
15030 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15031 // We specify 2 possible opcodes for intrinsics with rounding modes.
15032 // First, we check if the intrinsic may have non-default rounding mode,
15033 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15034 if (IntrData->Opc1 != 0) {
15035 SDValue Rnd = Op.getOperand(5);
15036 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15037 X86::STATIC_ROUNDING::CUR_DIRECTION)
15038 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15039 Op.getOperand(2), CC, Rnd);
15041 //default rounding mode
15043 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15044 Op.getOperand(2), CC);
15047 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15048 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15051 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15052 DAG.getTargetConstant(0, dl,
15055 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15056 DAG.getUNDEF(BitcastVT), CmpMask,
15057 DAG.getIntPtrConstant(0, dl));
15058 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
15060 case COMI: { // Comparison intrinsics
15061 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15062 SDValue LHS = Op.getOperand(1);
15063 SDValue RHS = Op.getOperand(2);
15064 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15065 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15066 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15067 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15068 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15069 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15072 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15073 Op.getOperand(1), Op.getOperand(2), DAG);
15075 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15076 Op.getSimpleValueType(),
15078 Op.getOperand(2), DAG),
15079 Op.getOperand(4), Op.getOperand(3), Subtarget,
15081 case COMPRESS_EXPAND_IN_REG: {
15082 SDValue Mask = Op.getOperand(3);
15083 SDValue DataToCompress = Op.getOperand(1);
15084 SDValue PassThru = Op.getOperand(2);
15085 if (isAllOnes(Mask)) // return data as is
15086 return Op.getOperand(1);
15087 EVT VT = Op.getValueType();
15088 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15089 VT.getVectorNumElements());
15090 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15091 Mask.getValueType().getSizeInBits());
15093 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15094 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15095 DAG.getIntPtrConstant(0, dl));
15097 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
15101 SDValue Mask = Op.getOperand(3);
15102 EVT VT = Op.getValueType();
15103 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15104 VT.getVectorNumElements());
15105 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15106 Mask.getValueType().getSizeInBits());
15108 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15109 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15110 DAG.getIntPtrConstant(0, dl));
15111 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15120 default: return SDValue(); // Don't custom lower most intrinsics.
15122 case Intrinsic::x86_avx2_permd:
15123 case Intrinsic::x86_avx2_permps:
15124 // Operands intentionally swapped. Mask is last operand to intrinsic,
15125 // but second operand for node/instruction.
15126 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15127 Op.getOperand(2), Op.getOperand(1));
15129 case Intrinsic::x86_avx512_mask_valign_q_512:
15130 case Intrinsic::x86_avx512_mask_valign_d_512:
15131 // Vector source operands are swapped.
15132 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
15133 Op.getValueType(), Op.getOperand(2),
15136 Op.getOperand(5), Op.getOperand(4),
15139 // ptest and testp intrinsics. The intrinsic these come from are designed to
15140 // return an integer value, not just an instruction so lower it to the ptest
15141 // or testp pattern and a setcc for the result.
15142 case Intrinsic::x86_sse41_ptestz:
15143 case Intrinsic::x86_sse41_ptestc:
15144 case Intrinsic::x86_sse41_ptestnzc:
15145 case Intrinsic::x86_avx_ptestz_256:
15146 case Intrinsic::x86_avx_ptestc_256:
15147 case Intrinsic::x86_avx_ptestnzc_256:
15148 case Intrinsic::x86_avx_vtestz_ps:
15149 case Intrinsic::x86_avx_vtestc_ps:
15150 case Intrinsic::x86_avx_vtestnzc_ps:
15151 case Intrinsic::x86_avx_vtestz_pd:
15152 case Intrinsic::x86_avx_vtestc_pd:
15153 case Intrinsic::x86_avx_vtestnzc_pd:
15154 case Intrinsic::x86_avx_vtestz_ps_256:
15155 case Intrinsic::x86_avx_vtestc_ps_256:
15156 case Intrinsic::x86_avx_vtestnzc_ps_256:
15157 case Intrinsic::x86_avx_vtestz_pd_256:
15158 case Intrinsic::x86_avx_vtestc_pd_256:
15159 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15160 bool IsTestPacked = false;
15163 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15164 case Intrinsic::x86_avx_vtestz_ps:
15165 case Intrinsic::x86_avx_vtestz_pd:
15166 case Intrinsic::x86_avx_vtestz_ps_256:
15167 case Intrinsic::x86_avx_vtestz_pd_256:
15168 IsTestPacked = true; // Fallthrough
15169 case Intrinsic::x86_sse41_ptestz:
15170 case Intrinsic::x86_avx_ptestz_256:
15172 X86CC = X86::COND_E;
15174 case Intrinsic::x86_avx_vtestc_ps:
15175 case Intrinsic::x86_avx_vtestc_pd:
15176 case Intrinsic::x86_avx_vtestc_ps_256:
15177 case Intrinsic::x86_avx_vtestc_pd_256:
15178 IsTestPacked = true; // Fallthrough
15179 case Intrinsic::x86_sse41_ptestc:
15180 case Intrinsic::x86_avx_ptestc_256:
15182 X86CC = X86::COND_B;
15184 case Intrinsic::x86_avx_vtestnzc_ps:
15185 case Intrinsic::x86_avx_vtestnzc_pd:
15186 case Intrinsic::x86_avx_vtestnzc_ps_256:
15187 case Intrinsic::x86_avx_vtestnzc_pd_256:
15188 IsTestPacked = true; // Fallthrough
15189 case Intrinsic::x86_sse41_ptestnzc:
15190 case Intrinsic::x86_avx_ptestnzc_256:
15192 X86CC = X86::COND_A;
15196 SDValue LHS = Op.getOperand(1);
15197 SDValue RHS = Op.getOperand(2);
15198 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15199 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15200 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15201 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15202 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15204 case Intrinsic::x86_avx512_kortestz_w:
15205 case Intrinsic::x86_avx512_kortestc_w: {
15206 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15207 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
15208 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
15209 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15210 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15211 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15212 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15215 case Intrinsic::x86_sse42_pcmpistria128:
15216 case Intrinsic::x86_sse42_pcmpestria128:
15217 case Intrinsic::x86_sse42_pcmpistric128:
15218 case Intrinsic::x86_sse42_pcmpestric128:
15219 case Intrinsic::x86_sse42_pcmpistrio128:
15220 case Intrinsic::x86_sse42_pcmpestrio128:
15221 case Intrinsic::x86_sse42_pcmpistris128:
15222 case Intrinsic::x86_sse42_pcmpestris128:
15223 case Intrinsic::x86_sse42_pcmpistriz128:
15224 case Intrinsic::x86_sse42_pcmpestriz128: {
15228 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15229 case Intrinsic::x86_sse42_pcmpistria128:
15230 Opcode = X86ISD::PCMPISTRI;
15231 X86CC = X86::COND_A;
15233 case Intrinsic::x86_sse42_pcmpestria128:
15234 Opcode = X86ISD::PCMPESTRI;
15235 X86CC = X86::COND_A;
15237 case Intrinsic::x86_sse42_pcmpistric128:
15238 Opcode = X86ISD::PCMPISTRI;
15239 X86CC = X86::COND_B;
15241 case Intrinsic::x86_sse42_pcmpestric128:
15242 Opcode = X86ISD::PCMPESTRI;
15243 X86CC = X86::COND_B;
15245 case Intrinsic::x86_sse42_pcmpistrio128:
15246 Opcode = X86ISD::PCMPISTRI;
15247 X86CC = X86::COND_O;
15249 case Intrinsic::x86_sse42_pcmpestrio128:
15250 Opcode = X86ISD::PCMPESTRI;
15251 X86CC = X86::COND_O;
15253 case Intrinsic::x86_sse42_pcmpistris128:
15254 Opcode = X86ISD::PCMPISTRI;
15255 X86CC = X86::COND_S;
15257 case Intrinsic::x86_sse42_pcmpestris128:
15258 Opcode = X86ISD::PCMPESTRI;
15259 X86CC = X86::COND_S;
15261 case Intrinsic::x86_sse42_pcmpistriz128:
15262 Opcode = X86ISD::PCMPISTRI;
15263 X86CC = X86::COND_E;
15265 case Intrinsic::x86_sse42_pcmpestriz128:
15266 Opcode = X86ISD::PCMPESTRI;
15267 X86CC = X86::COND_E;
15270 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15271 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15272 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15273 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15274 DAG.getConstant(X86CC, dl, MVT::i8),
15275 SDValue(PCMP.getNode(), 1));
15276 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15279 case Intrinsic::x86_sse42_pcmpistri128:
15280 case Intrinsic::x86_sse42_pcmpestri128: {
15282 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15283 Opcode = X86ISD::PCMPISTRI;
15285 Opcode = X86ISD::PCMPESTRI;
15287 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15288 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15289 return DAG.getNode(Opcode, dl, VTs, NewOps);
15292 case Intrinsic::x86_seh_lsda: {
15293 // Compute the symbol for the LSDA. We know it'll get emitted later.
15294 MachineFunction &MF = DAG.getMachineFunction();
15295 SDValue Op1 = Op.getOperand(1);
15297 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15298 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15299 GlobalValue::getRealLinkageName(Fn->getName()));
15300 StringRef Name = LSDASym->getName();
15301 assert(Name.data()[Name.size()] == '\0' && "not null terminated");
15303 // Generate a simple absolute symbol reference. This intrinsic is only
15304 // supported on 32-bit Windows, which isn't PIC.
15306 DAG.getTargetExternalSymbol(Name.data(), VT, X86II::MO_NOPREFIX);
15307 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15312 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15313 SDValue Src, SDValue Mask, SDValue Base,
15314 SDValue Index, SDValue ScaleOp, SDValue Chain,
15315 const X86Subtarget * Subtarget) {
15317 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15318 assert(C && "Invalid scale type");
15319 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15320 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15321 Index.getSimpleValueType().getVectorNumElements());
15323 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15325 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15327 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15328 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15329 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15330 SDValue Segment = DAG.getRegister(0, MVT::i32);
15331 if (Src.getOpcode() == ISD::UNDEF)
15332 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15333 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15334 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15335 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15336 return DAG.getMergeValues(RetOps, dl);
15339 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15340 SDValue Src, SDValue Mask, SDValue Base,
15341 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15343 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15344 assert(C && "Invalid scale type");
15345 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15346 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15347 SDValue Segment = DAG.getRegister(0, MVT::i32);
15348 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15349 Index.getSimpleValueType().getVectorNumElements());
15351 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15353 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15355 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15356 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15357 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15358 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15359 return SDValue(Res, 1);
15362 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15363 SDValue Mask, SDValue Base, SDValue Index,
15364 SDValue ScaleOp, SDValue Chain) {
15366 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15367 assert(C && "Invalid scale type");
15368 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15369 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15370 SDValue Segment = DAG.getRegister(0, MVT::i32);
15372 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15374 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15376 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15378 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15379 //SDVTList VTs = DAG.getVTList(MVT::Other);
15380 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15381 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15382 return SDValue(Res, 0);
15385 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15386 // read performance monitor counters (x86_rdpmc).
15387 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15388 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15389 SmallVectorImpl<SDValue> &Results) {
15390 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15391 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15394 // The ECX register is used to select the index of the performance counter
15396 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
15398 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
15400 // Reads the content of a 64-bit performance counter and returns it in the
15401 // registers EDX:EAX.
15402 if (Subtarget->is64Bit()) {
15403 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15404 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15407 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15408 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15411 Chain = HI.getValue(1);
15413 if (Subtarget->is64Bit()) {
15414 // The EAX register is loaded with the low-order 32 bits. The EDX register
15415 // is loaded with the supported high-order bits of the counter.
15416 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15417 DAG.getConstant(32, DL, MVT::i8));
15418 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15419 Results.push_back(Chain);
15423 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15424 SDValue Ops[] = { LO, HI };
15425 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15426 Results.push_back(Pair);
15427 Results.push_back(Chain);
15430 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
15431 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
15432 // also used to custom lower READCYCLECOUNTER nodes.
15433 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
15434 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15435 SmallVectorImpl<SDValue> &Results) {
15436 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15437 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
15440 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
15441 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
15442 // and the EAX register is loaded with the low-order 32 bits.
15443 if (Subtarget->is64Bit()) {
15444 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15445 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15448 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15449 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15452 SDValue Chain = HI.getValue(1);
15454 if (Opcode == X86ISD::RDTSCP_DAG) {
15455 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15457 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
15458 // the ECX register. Add 'ecx' explicitly to the chain.
15459 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
15461 // Explicitly store the content of ECX at the location passed in input
15462 // to the 'rdtscp' intrinsic.
15463 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
15464 MachinePointerInfo(), false, false, 0);
15467 if (Subtarget->is64Bit()) {
15468 // The EDX register is loaded with the high-order 32 bits of the MSR, and
15469 // the EAX register is loaded with the low-order 32 bits.
15470 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15471 DAG.getConstant(32, DL, MVT::i8));
15472 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15473 Results.push_back(Chain);
15477 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15478 SDValue Ops[] = { LO, HI };
15479 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15480 Results.push_back(Pair);
15481 Results.push_back(Chain);
15484 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
15485 SelectionDAG &DAG) {
15486 SmallVector<SDValue, 2> Results;
15488 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
15490 return DAG.getMergeValues(Results, DL);
15494 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15495 SelectionDAG &DAG) {
15496 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
15498 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
15503 switch(IntrData->Type) {
15505 llvm_unreachable("Unknown Intrinsic Type");
15509 // Emit the node with the right value type.
15510 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
15511 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15513 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
15514 // Otherwise return the value from Rand, which is always 0, casted to i32.
15515 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
15516 DAG.getConstant(1, dl, Op->getValueType(1)),
15517 DAG.getConstant(X86::COND_B, dl, MVT::i32),
15518 SDValue(Result.getNode(), 1) };
15519 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
15520 DAG.getVTList(Op->getValueType(1), MVT::Glue),
15523 // Return { result, isValid, chain }.
15524 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
15525 SDValue(Result.getNode(), 2));
15528 //gather(v1, mask, index, base, scale);
15529 SDValue Chain = Op.getOperand(0);
15530 SDValue Src = Op.getOperand(2);
15531 SDValue Base = Op.getOperand(3);
15532 SDValue Index = Op.getOperand(4);
15533 SDValue Mask = Op.getOperand(5);
15534 SDValue Scale = Op.getOperand(6);
15535 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
15539 //scatter(base, mask, index, v1, scale);
15540 SDValue Chain = Op.getOperand(0);
15541 SDValue Base = Op.getOperand(2);
15542 SDValue Mask = Op.getOperand(3);
15543 SDValue Index = Op.getOperand(4);
15544 SDValue Src = Op.getOperand(5);
15545 SDValue Scale = Op.getOperand(6);
15546 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
15550 SDValue Hint = Op.getOperand(6);
15551 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
15552 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
15553 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
15554 SDValue Chain = Op.getOperand(0);
15555 SDValue Mask = Op.getOperand(2);
15556 SDValue Index = Op.getOperand(3);
15557 SDValue Base = Op.getOperand(4);
15558 SDValue Scale = Op.getOperand(5);
15559 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15561 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15563 SmallVector<SDValue, 2> Results;
15564 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
15566 return DAG.getMergeValues(Results, dl);
15568 // Read Performance Monitoring Counters.
15570 SmallVector<SDValue, 2> Results;
15571 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15572 return DAG.getMergeValues(Results, dl);
15574 // XTEST intrinsics.
15576 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15577 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15578 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15579 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
15581 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15582 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15583 Ret, SDValue(InTrans.getNode(), 1));
15587 SmallVector<SDValue, 2> Results;
15588 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15589 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15590 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15591 DAG.getConstant(-1, dl, MVT::i8));
15592 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15593 Op.getOperand(4), GenCF.getValue(1));
15594 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15595 Op.getOperand(5), MachinePointerInfo(),
15597 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15598 DAG.getConstant(X86::COND_B, dl, MVT::i8),
15600 Results.push_back(SetCC);
15601 Results.push_back(Store);
15602 return DAG.getMergeValues(Results, dl);
15604 case COMPRESS_TO_MEM: {
15606 SDValue Mask = Op.getOperand(4);
15607 SDValue DataToCompress = Op.getOperand(3);
15608 SDValue Addr = Op.getOperand(2);
15609 SDValue Chain = Op.getOperand(0);
15611 if (isAllOnes(Mask)) // return just a store
15612 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15613 MachinePointerInfo(), false, false, 0);
15615 EVT VT = DataToCompress.getValueType();
15616 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15617 VT.getVectorNumElements());
15618 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15619 Mask.getValueType().getSizeInBits());
15620 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15621 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15622 DAG.getIntPtrConstant(0, dl));
15624 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15625 DataToCompress, DAG.getUNDEF(VT));
15626 return DAG.getStore(Chain, dl, Compressed, Addr,
15627 MachinePointerInfo(), false, false, 0);
15629 case EXPAND_FROM_MEM: {
15631 SDValue Mask = Op.getOperand(4);
15632 SDValue PathThru = Op.getOperand(3);
15633 SDValue Addr = Op.getOperand(2);
15634 SDValue Chain = Op.getOperand(0);
15635 EVT VT = Op.getValueType();
15637 if (isAllOnes(Mask)) // return just a load
15638 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15640 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15641 VT.getVectorNumElements());
15642 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15643 Mask.getValueType().getSizeInBits());
15644 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15645 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15646 DAG.getIntPtrConstant(0, dl));
15648 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15649 false, false, false, 0);
15651 SDValue Results[] = {
15652 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15654 return DAG.getMergeValues(Results, dl);
15659 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15660 SelectionDAG &DAG) const {
15661 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15662 MFI->setReturnAddressIsTaken(true);
15664 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15667 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15669 EVT PtrVT = getPointerTy();
15672 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15673 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15674 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
15675 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15676 DAG.getNode(ISD::ADD, dl, PtrVT,
15677 FrameAddr, Offset),
15678 MachinePointerInfo(), false, false, false, 0);
15681 // Just load the return address.
15682 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15683 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15684 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15687 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15688 MachineFunction &MF = DAG.getMachineFunction();
15689 MachineFrameInfo *MFI = MF.getFrameInfo();
15690 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15691 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15692 EVT VT = Op.getValueType();
15694 MFI->setFrameAddressIsTaken(true);
15696 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15697 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15698 // is not possible to crawl up the stack without looking at the unwind codes
15700 int FrameAddrIndex = FuncInfo->getFAIndex();
15701 if (!FrameAddrIndex) {
15702 // Set up a frame object for the return address.
15703 unsigned SlotSize = RegInfo->getSlotSize();
15704 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15705 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
15706 FuncInfo->setFAIndex(FrameAddrIndex);
15708 return DAG.getFrameIndex(FrameAddrIndex, VT);
15711 unsigned FrameReg =
15712 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15713 SDLoc dl(Op); // FIXME probably not meaningful
15714 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15715 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15716 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15717 "Invalid Frame Register!");
15718 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15720 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15721 MachinePointerInfo(),
15722 false, false, false, 0);
15726 // FIXME? Maybe this could be a TableGen attribute on some registers and
15727 // this table could be generated automatically from RegInfo.
15728 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15730 unsigned Reg = StringSwitch<unsigned>(RegName)
15731 .Case("esp", X86::ESP)
15732 .Case("rsp", X86::RSP)
15736 report_fatal_error("Invalid register name global variable");
15739 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15740 SelectionDAG &DAG) const {
15741 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15742 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
15745 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15746 SDValue Chain = Op.getOperand(0);
15747 SDValue Offset = Op.getOperand(1);
15748 SDValue Handler = Op.getOperand(2);
15751 EVT PtrVT = getPointerTy();
15752 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15753 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15754 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15755 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15756 "Invalid Frame Register!");
15757 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15758 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15760 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15761 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
15763 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15764 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15766 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15768 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15769 DAG.getRegister(StoreAddrReg, PtrVT));
15772 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15773 SelectionDAG &DAG) const {
15775 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15776 DAG.getVTList(MVT::i32, MVT::Other),
15777 Op.getOperand(0), Op.getOperand(1));
15780 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15781 SelectionDAG &DAG) const {
15783 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15784 Op.getOperand(0), Op.getOperand(1));
15787 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15788 return Op.getOperand(0);
15791 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15792 SelectionDAG &DAG) const {
15793 SDValue Root = Op.getOperand(0);
15794 SDValue Trmp = Op.getOperand(1); // trampoline
15795 SDValue FPtr = Op.getOperand(2); // nested function
15796 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15799 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15800 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15802 if (Subtarget->is64Bit()) {
15803 SDValue OutChains[6];
15805 // Large code-model.
15806 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15807 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15809 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15810 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15812 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15814 // Load the pointer to the nested function into R11.
15815 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15816 SDValue Addr = Trmp;
15817 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15818 Addr, MachinePointerInfo(TrmpAddr),
15821 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15822 DAG.getConstant(2, dl, MVT::i64));
15823 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15824 MachinePointerInfo(TrmpAddr, 2),
15827 // Load the 'nest' parameter value into R10.
15828 // R10 is specified in X86CallingConv.td
15829 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15830 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15831 DAG.getConstant(10, dl, MVT::i64));
15832 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15833 Addr, MachinePointerInfo(TrmpAddr, 10),
15836 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15837 DAG.getConstant(12, dl, MVT::i64));
15838 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15839 MachinePointerInfo(TrmpAddr, 12),
15842 // Jump to the nested function.
15843 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15844 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15845 DAG.getConstant(20, dl, MVT::i64));
15846 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15847 Addr, MachinePointerInfo(TrmpAddr, 20),
15850 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15851 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15852 DAG.getConstant(22, dl, MVT::i64));
15853 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
15854 Addr, MachinePointerInfo(TrmpAddr, 22),
15857 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15859 const Function *Func =
15860 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15861 CallingConv::ID CC = Func->getCallingConv();
15866 llvm_unreachable("Unsupported calling convention");
15867 case CallingConv::C:
15868 case CallingConv::X86_StdCall: {
15869 // Pass 'nest' parameter in ECX.
15870 // Must be kept in sync with X86CallingConv.td
15871 NestReg = X86::ECX;
15873 // Check that ECX wasn't needed by an 'inreg' parameter.
15874 FunctionType *FTy = Func->getFunctionType();
15875 const AttributeSet &Attrs = Func->getAttributes();
15877 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15878 unsigned InRegCount = 0;
15881 for (FunctionType::param_iterator I = FTy->param_begin(),
15882 E = FTy->param_end(); I != E; ++I, ++Idx)
15883 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15884 // FIXME: should only count parameters that are lowered to integers.
15885 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15887 if (InRegCount > 2) {
15888 report_fatal_error("Nest register in use - reduce number of inreg"
15894 case CallingConv::X86_FastCall:
15895 case CallingConv::X86_ThisCall:
15896 case CallingConv::Fast:
15897 // Pass 'nest' parameter in EAX.
15898 // Must be kept in sync with X86CallingConv.td
15899 NestReg = X86::EAX;
15903 SDValue OutChains[4];
15904 SDValue Addr, Disp;
15906 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15907 DAG.getConstant(10, dl, MVT::i32));
15908 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15910 // This is storing the opcode for MOV32ri.
15911 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15912 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15913 OutChains[0] = DAG.getStore(Root, dl,
15914 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
15915 Trmp, MachinePointerInfo(TrmpAddr),
15918 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15919 DAG.getConstant(1, dl, MVT::i32));
15920 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15921 MachinePointerInfo(TrmpAddr, 1),
15924 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15925 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15926 DAG.getConstant(5, dl, MVT::i32));
15927 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
15928 Addr, MachinePointerInfo(TrmpAddr, 5),
15931 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15932 DAG.getConstant(6, dl, MVT::i32));
15933 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15934 MachinePointerInfo(TrmpAddr, 6),
15937 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15941 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15942 SelectionDAG &DAG) const {
15944 The rounding mode is in bits 11:10 of FPSR, and has the following
15946 00 Round to nearest
15951 FLT_ROUNDS, on the other hand, expects the following:
15958 To perform the conversion, we do:
15959 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15962 MachineFunction &MF = DAG.getMachineFunction();
15963 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15964 unsigned StackAlignment = TFI.getStackAlignment();
15965 MVT VT = Op.getSimpleValueType();
15968 // Save FP Control Word to stack slot
15969 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15970 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15972 MachineMemOperand *MMO =
15973 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15974 MachineMemOperand::MOStore, 2, 2);
15976 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15977 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15978 DAG.getVTList(MVT::Other),
15979 Ops, MVT::i16, MMO);
15981 // Load FP Control Word from stack slot
15982 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15983 MachinePointerInfo(), false, false, false, 0);
15985 // Transform as necessary
15987 DAG.getNode(ISD::SRL, DL, MVT::i16,
15988 DAG.getNode(ISD::AND, DL, MVT::i16,
15989 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
15990 DAG.getConstant(11, DL, MVT::i8));
15992 DAG.getNode(ISD::SRL, DL, MVT::i16,
15993 DAG.getNode(ISD::AND, DL, MVT::i16,
15994 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
15995 DAG.getConstant(9, DL, MVT::i8));
15998 DAG.getNode(ISD::AND, DL, MVT::i16,
15999 DAG.getNode(ISD::ADD, DL, MVT::i16,
16000 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16001 DAG.getConstant(1, DL, MVT::i16)),
16002 DAG.getConstant(3, DL, MVT::i16));
16004 return DAG.getNode((VT.getSizeInBits() < 16 ?
16005 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16008 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16009 MVT VT = Op.getSimpleValueType();
16011 unsigned NumBits = VT.getSizeInBits();
16014 Op = Op.getOperand(0);
16015 if (VT == MVT::i8) {
16016 // Zero extend to i32 since there is not an i8 bsr.
16018 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16021 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16022 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16023 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16025 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16028 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16029 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16032 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16034 // Finally xor with NumBits-1.
16035 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16036 DAG.getConstant(NumBits - 1, dl, OpVT));
16039 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16043 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16044 MVT VT = Op.getSimpleValueType();
16046 unsigned NumBits = VT.getSizeInBits();
16049 Op = Op.getOperand(0);
16050 if (VT == MVT::i8) {
16051 // Zero extend to i32 since there is not an i8 bsr.
16053 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16056 // Issue a bsr (scan bits in reverse).
16057 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16058 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16060 // And xor with NumBits-1.
16061 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16062 DAG.getConstant(NumBits - 1, dl, OpVT));
16065 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16069 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16070 MVT VT = Op.getSimpleValueType();
16071 unsigned NumBits = VT.getSizeInBits();
16073 Op = Op.getOperand(0);
16075 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16076 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16077 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16079 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16082 DAG.getConstant(NumBits, dl, VT),
16083 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16086 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16089 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16090 // ones, and then concatenate the result back.
16091 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16092 MVT VT = Op.getSimpleValueType();
16094 assert(VT.is256BitVector() && VT.isInteger() &&
16095 "Unsupported value type for operation");
16097 unsigned NumElems = VT.getVectorNumElements();
16100 // Extract the LHS vectors
16101 SDValue LHS = Op.getOperand(0);
16102 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16103 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16105 // Extract the RHS vectors
16106 SDValue RHS = Op.getOperand(1);
16107 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16108 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16110 MVT EltVT = VT.getVectorElementType();
16111 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16113 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16114 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16115 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16118 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16119 assert(Op.getSimpleValueType().is256BitVector() &&
16120 Op.getSimpleValueType().isInteger() &&
16121 "Only handle AVX 256-bit vector integer operation");
16122 return Lower256IntArith(Op, DAG);
16125 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16126 assert(Op.getSimpleValueType().is256BitVector() &&
16127 Op.getSimpleValueType().isInteger() &&
16128 "Only handle AVX 256-bit vector integer operation");
16129 return Lower256IntArith(Op, DAG);
16132 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16133 SelectionDAG &DAG) {
16135 MVT VT = Op.getSimpleValueType();
16137 // Decompose 256-bit ops into smaller 128-bit ops.
16138 if (VT.is256BitVector() && !Subtarget->hasInt256())
16139 return Lower256IntArith(Op, DAG);
16141 SDValue A = Op.getOperand(0);
16142 SDValue B = Op.getOperand(1);
16144 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16145 // pairs, multiply and truncate.
16146 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16147 if (Subtarget->hasInt256()) {
16148 if (VT == MVT::v32i8) {
16149 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16150 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16151 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16152 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16153 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16154 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16155 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16156 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16157 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16158 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16161 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16162 return DAG.getNode(
16163 ISD::TRUNCATE, dl, VT,
16164 DAG.getNode(ISD::MUL, dl, ExVT,
16165 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16166 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16169 assert(VT == MVT::v16i8 &&
16170 "Pre-AVX2 support only supports v16i8 multiplication");
16171 MVT ExVT = MVT::v8i16;
16173 // Extract the lo parts and sign extend to i16
16175 if (Subtarget->hasSSE41()) {
16176 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16177 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16179 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16180 -1, 4, -1, 5, -1, 6, -1, 7};
16181 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16182 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16183 ALo = DAG.getNode(ISD::BITCAST, dl, ExVT, ALo);
16184 BLo = DAG.getNode(ISD::BITCAST, dl, ExVT, BLo);
16185 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16186 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16189 // Extract the hi parts and sign extend to i16
16191 if (Subtarget->hasSSE41()) {
16192 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16193 -1, -1, -1, -1, -1, -1, -1, -1};
16194 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16195 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16196 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16197 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16199 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16200 -1, 12, -1, 13, -1, 14, -1, 15};
16201 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16202 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16203 AHi = DAG.getNode(ISD::BITCAST, dl, ExVT, AHi);
16204 BHi = DAG.getNode(ISD::BITCAST, dl, ExVT, BHi);
16205 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16206 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16209 // Multiply, mask the lower 8bits of the lo/hi results and pack
16210 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16211 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16212 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16213 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16214 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16217 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16218 if (VT == MVT::v4i32) {
16219 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16220 "Should not custom lower when pmuldq is available!");
16222 // Extract the odd parts.
16223 static const int UnpackMask[] = { 1, -1, 3, -1 };
16224 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16225 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16227 // Multiply the even parts.
16228 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16229 // Now multiply odd parts.
16230 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16232 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
16233 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
16235 // Merge the two vectors back together with a shuffle. This expands into 2
16237 static const int ShufMask[] = { 0, 4, 2, 6 };
16238 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16241 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16242 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16244 // Ahi = psrlqi(a, 32);
16245 // Bhi = psrlqi(b, 32);
16247 // AloBlo = pmuludq(a, b);
16248 // AloBhi = pmuludq(a, Bhi);
16249 // AhiBlo = pmuludq(Ahi, b);
16251 // AloBhi = psllqi(AloBhi, 32);
16252 // AhiBlo = psllqi(AhiBlo, 32);
16253 // return AloBlo + AloBhi + AhiBlo;
16255 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16256 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16258 // Bit cast to 32-bit vectors for MULUDQ
16259 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16260 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16261 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
16262 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
16263 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
16264 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
16266 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16267 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16268 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16270 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16271 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16273 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16274 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16277 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16278 assert(Subtarget->isTargetWin64() && "Unexpected target");
16279 EVT VT = Op.getValueType();
16280 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16281 "Unexpected return type for lowering");
16285 switch (Op->getOpcode()) {
16286 default: llvm_unreachable("Unexpected request for libcall!");
16287 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16288 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16289 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16290 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16291 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16292 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16296 SDValue InChain = DAG.getEntryNode();
16298 TargetLowering::ArgListTy Args;
16299 TargetLowering::ArgListEntry Entry;
16300 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16301 EVT ArgVT = Op->getOperand(i).getValueType();
16302 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16303 "Unexpected argument type for lowering");
16304 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16305 Entry.Node = StackPtr;
16306 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16308 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16309 Entry.Ty = PointerType::get(ArgTy,0);
16310 Entry.isSExt = false;
16311 Entry.isZExt = false;
16312 Args.push_back(Entry);
16315 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16318 TargetLowering::CallLoweringInfo CLI(DAG);
16319 CLI.setDebugLoc(dl).setChain(InChain)
16320 .setCallee(getLibcallCallingConv(LC),
16321 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16322 Callee, std::move(Args), 0)
16323 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16325 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16326 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
16329 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16330 SelectionDAG &DAG) {
16331 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16332 EVT VT = Op0.getValueType();
16335 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16336 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16338 // PMULxD operations multiply each even value (starting at 0) of LHS with
16339 // the related value of RHS and produce a widen result.
16340 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16341 // => <2 x i64> <ae|cg>
16343 // In other word, to have all the results, we need to perform two PMULxD:
16344 // 1. one with the even values.
16345 // 2. one with the odd values.
16346 // To achieve #2, with need to place the odd values at an even position.
16348 // Place the odd value at an even position (basically, shift all values 1
16349 // step to the left):
16350 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16351 // <a|b|c|d> => <b|undef|d|undef>
16352 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16353 // <e|f|g|h> => <f|undef|h|undef>
16354 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16356 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16358 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16359 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16361 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16362 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16363 // => <2 x i64> <ae|cg>
16364 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
16365 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
16366 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
16367 // => <2 x i64> <bf|dh>
16368 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
16369 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
16371 // Shuffle it back into the right order.
16372 SDValue Highs, Lows;
16373 if (VT == MVT::v8i32) {
16374 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
16375 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16376 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
16377 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16379 const int HighMask[] = {1, 5, 3, 7};
16380 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16381 const int LowMask[] = {0, 4, 2, 6};
16382 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16385 // If we have a signed multiply but no PMULDQ fix up the high parts of a
16386 // unsigned multiply.
16387 if (IsSigned && !Subtarget->hasSSE41()) {
16389 DAG.getConstant(31, dl,
16390 DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
16391 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
16392 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
16393 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
16394 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
16396 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
16397 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
16400 // The first result of MUL_LOHI is actually the low value, followed by the
16402 SDValue Ops[] = {Lows, Highs};
16403 return DAG.getMergeValues(Ops, dl);
16406 // Return true if the requred (according to Opcode) shift-imm form is natively
16407 // supported by the Subtarget
16408 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
16410 if (VT.getScalarSizeInBits() < 16)
16413 if (VT.is512BitVector() &&
16414 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
16417 bool LShift = VT.is128BitVector() ||
16418 (VT.is256BitVector() && Subtarget->hasInt256());
16420 bool AShift = LShift && (Subtarget->hasVLX() ||
16421 (VT != MVT::v2i64 && VT != MVT::v4i64));
16422 return (Opcode == ISD::SRA) ? AShift : LShift;
16425 // The shift amount is a variable, but it is the same for all vector lanes.
16426 // These instrcutions are defined together with shift-immediate.
16428 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
16430 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
16433 // Return true if the requred (according to Opcode) variable-shift form is
16434 // natively supported by the Subtarget
16435 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
16438 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
16441 // vXi16 supported only on AVX-512, BWI
16442 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
16445 if (VT.is512BitVector() || Subtarget->hasVLX())
16448 bool LShift = VT.is128BitVector() || VT.is256BitVector();
16449 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
16450 return (Opcode == ISD::SRA) ? AShift : LShift;
16453 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
16454 const X86Subtarget *Subtarget) {
16455 MVT VT = Op.getSimpleValueType();
16457 SDValue R = Op.getOperand(0);
16458 SDValue Amt = Op.getOperand(1);
16460 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16461 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16463 // Optimize shl/srl/sra with constant shift amount.
16464 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
16465 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
16466 uint64_t ShiftAmt = ShiftConst->getZExtValue();
16468 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
16469 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16471 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
16472 unsigned NumElts = VT.getVectorNumElements();
16473 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
16475 if (Op.getOpcode() == ISD::SHL) {
16476 // Make a large shift.
16477 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
16479 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
16480 // Zero out the rightmost bits.
16481 SmallVector<SDValue, 32> V(
16482 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
16483 return DAG.getNode(ISD::AND, dl, VT, SHL,
16484 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16486 if (Op.getOpcode() == ISD::SRL) {
16487 // Make a large shift.
16488 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
16490 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
16491 // Zero out the leftmost bits.
16492 SmallVector<SDValue, 32> V(
16493 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
16494 return DAG.getNode(ISD::AND, dl, VT, SRL,
16495 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16497 if (Op.getOpcode() == ISD::SRA) {
16498 if (ShiftAmt == 7) {
16499 // R s>> 7 === R s< 0
16500 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
16501 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
16504 // R s>> a === ((R u>> a) ^ m) - m
16505 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
16506 SmallVector<SDValue, 32> V(NumElts,
16507 DAG.getConstant(128 >> ShiftAmt, dl,
16509 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
16510 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
16511 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
16514 llvm_unreachable("Unknown shift opcode.");
16519 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16520 if (!Subtarget->is64Bit() &&
16521 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
16522 Amt.getOpcode() == ISD::BITCAST &&
16523 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16524 Amt = Amt.getOperand(0);
16525 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16526 VT.getVectorNumElements();
16527 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
16528 uint64_t ShiftAmt = 0;
16529 for (unsigned i = 0; i != Ratio; ++i) {
16530 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
16534 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
16536 // Check remaining shift amounts.
16537 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16538 uint64_t ShAmt = 0;
16539 for (unsigned j = 0; j != Ratio; ++j) {
16540 ConstantSDNode *C =
16541 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
16545 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
16547 if (ShAmt != ShiftAmt)
16550 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16556 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
16557 const X86Subtarget* Subtarget) {
16558 MVT VT = Op.getSimpleValueType();
16560 SDValue R = Op.getOperand(0);
16561 SDValue Amt = Op.getOperand(1);
16563 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16564 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16566 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
16567 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
16569 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
16571 EVT EltVT = VT.getVectorElementType();
16573 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
16574 // Check if this build_vector node is doing a splat.
16575 // If so, then set BaseShAmt equal to the splat value.
16576 BaseShAmt = BV->getSplatValue();
16577 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
16578 BaseShAmt = SDValue();
16580 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
16581 Amt = Amt.getOperand(0);
16583 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
16584 if (SVN && SVN->isSplat()) {
16585 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
16586 SDValue InVec = Amt.getOperand(0);
16587 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
16588 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
16589 "Unexpected shuffle index found!");
16590 BaseShAmt = InVec.getOperand(SplatIdx);
16591 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
16592 if (ConstantSDNode *C =
16593 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
16594 if (C->getZExtValue() == SplatIdx)
16595 BaseShAmt = InVec.getOperand(1);
16600 // Avoid introducing an extract element from a shuffle.
16601 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
16602 DAG.getIntPtrConstant(SplatIdx, dl));
16606 if (BaseShAmt.getNode()) {
16607 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
16608 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
16609 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
16610 else if (EltVT.bitsLT(MVT::i32))
16611 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
16613 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
16617 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16618 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
16619 Amt.getOpcode() == ISD::BITCAST &&
16620 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16621 Amt = Amt.getOperand(0);
16622 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16623 VT.getVectorNumElements();
16624 std::vector<SDValue> Vals(Ratio);
16625 for (unsigned i = 0; i != Ratio; ++i)
16626 Vals[i] = Amt.getOperand(i);
16627 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16628 for (unsigned j = 0; j != Ratio; ++j)
16629 if (Vals[j] != Amt.getOperand(i + j))
16632 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
16637 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16638 SelectionDAG &DAG) {
16639 MVT VT = Op.getSimpleValueType();
16641 SDValue R = Op.getOperand(0);
16642 SDValue Amt = Op.getOperand(1);
16644 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16645 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16647 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
16650 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
16653 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
16656 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
16657 // shifts per-lane and then shuffle the partial results back together.
16658 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
16659 // Splat the shift amounts so the scalar shifts above will catch it.
16660 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
16661 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
16662 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
16663 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
16664 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
16667 // If possible, lower this packed shift into a vector multiply instead of
16668 // expanding it into a sequence of scalar shifts.
16669 // Do this only if the vector shift count is a constant build_vector.
16670 if (Op.getOpcode() == ISD::SHL &&
16671 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16672 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16673 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16674 SmallVector<SDValue, 8> Elts;
16675 EVT SVT = VT.getScalarType();
16676 unsigned SVTBits = SVT.getSizeInBits();
16677 const APInt &One = APInt(SVTBits, 1);
16678 unsigned NumElems = VT.getVectorNumElements();
16680 for (unsigned i=0; i !=NumElems; ++i) {
16681 SDValue Op = Amt->getOperand(i);
16682 if (Op->getOpcode() == ISD::UNDEF) {
16683 Elts.push_back(Op);
16687 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16688 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16689 uint64_t ShAmt = C.getZExtValue();
16690 if (ShAmt >= SVTBits) {
16691 Elts.push_back(DAG.getUNDEF(SVT));
16694 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
16696 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16697 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16700 // Lower SHL with variable shift amount.
16701 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16702 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
16704 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
16705 DAG.getConstant(0x3f800000U, dl, VT));
16706 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16707 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16708 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16711 // If possible, lower this shift as a sequence of two shifts by
16712 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16714 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16716 // Could be rewritten as:
16717 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16719 // The advantage is that the two shifts from the example would be
16720 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16721 // the vector shift into four scalar shifts plus four pairs of vector
16723 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16724 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16725 unsigned TargetOpcode = X86ISD::MOVSS;
16726 bool CanBeSimplified;
16727 // The splat value for the first packed shift (the 'X' from the example).
16728 SDValue Amt1 = Amt->getOperand(0);
16729 // The splat value for the second packed shift (the 'Y' from the example).
16730 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16731 Amt->getOperand(2);
16733 // See if it is possible to replace this node with a sequence of
16734 // two shifts followed by a MOVSS/MOVSD
16735 if (VT == MVT::v4i32) {
16736 // Check if it is legal to use a MOVSS.
16737 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16738 Amt2 == Amt->getOperand(3);
16739 if (!CanBeSimplified) {
16740 // Otherwise, check if we can still simplify this node using a MOVSD.
16741 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16742 Amt->getOperand(2) == Amt->getOperand(3);
16743 TargetOpcode = X86ISD::MOVSD;
16744 Amt2 = Amt->getOperand(2);
16747 // Do similar checks for the case where the machine value type
16749 CanBeSimplified = Amt1 == Amt->getOperand(1);
16750 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16751 CanBeSimplified = Amt2 == Amt->getOperand(i);
16753 if (!CanBeSimplified) {
16754 TargetOpcode = X86ISD::MOVSD;
16755 CanBeSimplified = true;
16756 Amt2 = Amt->getOperand(4);
16757 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16758 CanBeSimplified = Amt1 == Amt->getOperand(i);
16759 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16760 CanBeSimplified = Amt2 == Amt->getOperand(j);
16764 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16765 isa<ConstantSDNode>(Amt2)) {
16766 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16767 EVT CastVT = MVT::v4i32;
16769 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
16770 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16772 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
16773 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16774 if (TargetOpcode == X86ISD::MOVSD)
16775 CastVT = MVT::v2i64;
16776 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16777 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16778 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16780 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16784 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16785 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
16786 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, dl, VT));
16788 SDValue VSelM = DAG.getConstant(0x80, dl, VT);
16789 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16790 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16792 // r = VSELECT(r, shl(r, 4), a);
16793 SDValue M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(4, dl, VT));
16794 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16797 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16798 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16799 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16801 // r = VSELECT(r, shl(r, 2), a);
16802 M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(2, dl, VT));
16803 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16806 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16807 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16808 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16810 // return VSELECT(r, r+r, a);
16811 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16812 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16816 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16817 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16818 // solution better.
16819 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16820 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16822 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16823 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16824 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16825 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16826 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16829 // Decompose 256-bit shifts into smaller 128-bit shifts.
16830 if (VT.is256BitVector()) {
16831 unsigned NumElems = VT.getVectorNumElements();
16832 MVT EltVT = VT.getVectorElementType();
16833 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16835 // Extract the two vectors
16836 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16837 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16839 // Recreate the shift amount vectors
16840 SDValue Amt1, Amt2;
16841 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16842 // Constant shift amount
16843 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
16844 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
16845 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
16847 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16848 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16850 // Variable shift amount
16851 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16852 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16855 // Issue new vector shifts for the smaller types
16856 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16857 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16859 // Concatenate the result back
16860 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16866 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16867 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16868 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16869 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16870 // has only one use.
16871 SDNode *N = Op.getNode();
16872 SDValue LHS = N->getOperand(0);
16873 SDValue RHS = N->getOperand(1);
16874 unsigned BaseOp = 0;
16877 switch (Op.getOpcode()) {
16878 default: llvm_unreachable("Unknown ovf instruction!");
16880 // A subtract of one will be selected as a INC. Note that INC doesn't
16881 // set CF, so we can't do this for UADDO.
16882 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16884 BaseOp = X86ISD::INC;
16885 Cond = X86::COND_O;
16888 BaseOp = X86ISD::ADD;
16889 Cond = X86::COND_O;
16892 BaseOp = X86ISD::ADD;
16893 Cond = X86::COND_B;
16896 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16897 // set CF, so we can't do this for USUBO.
16898 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16900 BaseOp = X86ISD::DEC;
16901 Cond = X86::COND_O;
16904 BaseOp = X86ISD::SUB;
16905 Cond = X86::COND_O;
16908 BaseOp = X86ISD::SUB;
16909 Cond = X86::COND_B;
16912 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16913 Cond = X86::COND_O;
16915 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16916 if (N->getValueType(0) == MVT::i8) {
16917 BaseOp = X86ISD::UMUL8;
16918 Cond = X86::COND_O;
16921 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16923 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16926 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16927 DAG.getConstant(X86::COND_O, DL, MVT::i32),
16928 SDValue(Sum.getNode(), 2));
16930 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16934 // Also sets EFLAGS.
16935 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16936 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16939 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16940 DAG.getConstant(Cond, DL, MVT::i32),
16941 SDValue(Sum.getNode(), 1));
16943 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16946 /// Returns true if the operand type is exactly twice the native width, and
16947 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16948 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16949 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16950 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16951 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16954 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16955 else if (OpWidth == 128)
16956 return Subtarget->hasCmpxchg16b();
16961 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16962 return needsCmpXchgNb(SI->getValueOperand()->getType());
16965 // Note: this turns large loads into lock cmpxchg8b/16b.
16966 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16967 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16968 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16969 return needsCmpXchgNb(PTy->getElementType());
16972 TargetLoweringBase::AtomicRMWExpansionKind
16973 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16974 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16975 const Type *MemType = AI->getType();
16977 // If the operand is too big, we must see if cmpxchg8/16b is available
16978 // and default to library calls otherwise.
16979 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
16980 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
16981 : AtomicRMWExpansionKind::None;
16984 AtomicRMWInst::BinOp Op = AI->getOperation();
16987 llvm_unreachable("Unknown atomic operation");
16988 case AtomicRMWInst::Xchg:
16989 case AtomicRMWInst::Add:
16990 case AtomicRMWInst::Sub:
16991 // It's better to use xadd, xsub or xchg for these in all cases.
16992 return AtomicRMWExpansionKind::None;
16993 case AtomicRMWInst::Or:
16994 case AtomicRMWInst::And:
16995 case AtomicRMWInst::Xor:
16996 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16997 // prefix to a normal instruction for these operations.
16998 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
16999 : AtomicRMWExpansionKind::None;
17000 case AtomicRMWInst::Nand:
17001 case AtomicRMWInst::Max:
17002 case AtomicRMWInst::Min:
17003 case AtomicRMWInst::UMax:
17004 case AtomicRMWInst::UMin:
17005 // These always require a non-trivial set of data operations on x86. We must
17006 // use a cmpxchg loop.
17007 return AtomicRMWExpansionKind::CmpXChg;
17011 static bool hasMFENCE(const X86Subtarget& Subtarget) {
17012 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
17013 // no-sse2). There isn't any reason to disable it if the target processor
17015 return Subtarget.hasSSE2() || Subtarget.is64Bit();
17019 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
17020 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17021 const Type *MemType = AI->getType();
17022 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
17023 // there is no benefit in turning such RMWs into loads, and it is actually
17024 // harmful as it introduces a mfence.
17025 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
17028 auto Builder = IRBuilder<>(AI);
17029 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
17030 auto SynchScope = AI->getSynchScope();
17031 // We must restrict the ordering to avoid generating loads with Release or
17032 // ReleaseAcquire orderings.
17033 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
17034 auto Ptr = AI->getPointerOperand();
17036 // Before the load we need a fence. Here is an example lifted from
17037 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
17040 // x.store(1, relaxed);
17041 // r1 = y.fetch_add(0, release);
17043 // y.fetch_add(42, acquire);
17044 // r2 = x.load(relaxed);
17045 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
17046 // lowered to just a load without a fence. A mfence flushes the store buffer,
17047 // making the optimization clearly correct.
17048 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
17049 // otherwise, we might be able to be more agressive on relaxed idempotent
17050 // rmw. In practice, they do not look useful, so we don't try to be
17051 // especially clever.
17052 if (SynchScope == SingleThread)
17053 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
17054 // the IR level, so we must wrap it in an intrinsic.
17057 if (!hasMFENCE(*Subtarget))
17058 // FIXME: it might make sense to use a locked operation here but on a
17059 // different cache-line to prevent cache-line bouncing. In practice it
17060 // is probably a small win, and x86 processors without mfence are rare
17061 // enough that we do not bother.
17065 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
17066 Builder.CreateCall(MFence, {});
17068 // Finally we can emit the atomic load.
17069 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
17070 AI->getType()->getPrimitiveSizeInBits());
17071 Loaded->setAtomic(Order, SynchScope);
17072 AI->replaceAllUsesWith(Loaded);
17073 AI->eraseFromParent();
17077 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
17078 SelectionDAG &DAG) {
17080 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
17081 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
17082 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
17083 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
17085 // The only fence that needs an instruction is a sequentially-consistent
17086 // cross-thread fence.
17087 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
17088 if (hasMFENCE(*Subtarget))
17089 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
17091 SDValue Chain = Op.getOperand(0);
17092 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
17094 DAG.getRegister(X86::ESP, MVT::i32), // Base
17095 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
17096 DAG.getRegister(0, MVT::i32), // Index
17097 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
17098 DAG.getRegister(0, MVT::i32), // Segment.
17102 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
17103 return SDValue(Res, 0);
17106 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
17107 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17110 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17111 SelectionDAG &DAG) {
17112 MVT T = Op.getSimpleValueType();
17116 switch(T.SimpleTy) {
17117 default: llvm_unreachable("Invalid value type!");
17118 case MVT::i8: Reg = X86::AL; size = 1; break;
17119 case MVT::i16: Reg = X86::AX; size = 2; break;
17120 case MVT::i32: Reg = X86::EAX; size = 4; break;
17122 assert(Subtarget->is64Bit() && "Node not type legal!");
17123 Reg = X86::RAX; size = 8;
17126 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17127 Op.getOperand(2), SDValue());
17128 SDValue Ops[] = { cpIn.getValue(0),
17131 DAG.getTargetConstant(size, DL, MVT::i8),
17132 cpIn.getValue(1) };
17133 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17134 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17135 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17139 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17140 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17141 MVT::i32, cpOut.getValue(2));
17142 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17143 DAG.getConstant(X86::COND_E, DL, MVT::i8),
17146 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17147 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17148 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17152 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17153 SelectionDAG &DAG) {
17154 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17155 MVT DstVT = Op.getSimpleValueType();
17157 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17158 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17159 if (DstVT != MVT::f64)
17160 // This conversion needs to be expanded.
17163 SDValue InVec = Op->getOperand(0);
17165 unsigned NumElts = SrcVT.getVectorNumElements();
17166 EVT SVT = SrcVT.getVectorElementType();
17168 // Widen the vector in input in the case of MVT::v2i32.
17169 // Example: from MVT::v2i32 to MVT::v4i32.
17170 SmallVector<SDValue, 16> Elts;
17171 for (unsigned i = 0, e = NumElts; i != e; ++i)
17172 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17173 DAG.getIntPtrConstant(i, dl)));
17175 // Explicitly mark the extra elements as Undef.
17176 Elts.append(NumElts, DAG.getUNDEF(SVT));
17178 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17179 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17180 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
17181 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17182 DAG.getIntPtrConstant(0, dl));
17185 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17186 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17187 assert((DstVT == MVT::i64 ||
17188 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17189 "Unexpected custom BITCAST");
17190 // i64 <=> MMX conversions are Legal.
17191 if (SrcVT==MVT::i64 && DstVT.isVector())
17193 if (DstVT==MVT::i64 && SrcVT.isVector())
17195 // MMX <=> MMX conversions are Legal.
17196 if (SrcVT.isVector() && DstVT.isVector())
17198 // All other conversions need to be expanded.
17202 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17203 SelectionDAG &DAG) {
17204 SDNode *Node = Op.getNode();
17207 Op = Op.getOperand(0);
17208 EVT VT = Op.getValueType();
17209 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17210 "CTPOP lowering only implemented for 128/256-bit wide vector types");
17212 unsigned NumElts = VT.getVectorNumElements();
17213 EVT EltVT = VT.getVectorElementType();
17214 unsigned Len = EltVT.getSizeInBits();
17216 // This is the vectorized version of the "best" algorithm from
17217 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
17218 // with a minor tweak to use a series of adds + shifts instead of vector
17219 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
17221 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
17222 // v8i32 => Always profitable
17224 // FIXME: There a couple of possible improvements:
17226 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
17227 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
17229 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
17230 "CTPOP not implemented for this vector element type.");
17232 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
17233 // extra legalization.
17234 bool NeedsBitcast = EltVT == MVT::i32;
17235 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
17237 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl,
17239 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl,
17241 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl,
17244 // v = v - ((v >> 1) & 0x55555555...)
17245 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, dl, EltVT));
17246 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
17247 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
17249 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17251 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
17252 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
17254 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
17256 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
17257 if (VT != And.getValueType())
17258 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17259 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
17261 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
17262 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
17263 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
17264 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, dl, EltVT));
17265 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
17267 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
17268 if (NeedsBitcast) {
17269 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17270 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
17271 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
17274 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
17275 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
17276 if (VT != AndRHS.getValueType()) {
17277 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
17278 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
17280 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
17282 // v = (v + (v >> 4)) & 0x0F0F0F0F...
17283 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, dl, EltVT));
17284 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
17285 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
17286 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17288 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
17289 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
17290 if (NeedsBitcast) {
17291 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17292 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
17294 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
17295 if (VT != And.getValueType())
17296 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17298 // The algorithm mentioned above uses:
17299 // v = (v * 0x01010101...) >> (Len - 8)
17301 // Change it to use vector adds + vector shifts which yield faster results on
17302 // Haswell than using vector integer multiplication.
17304 // For i32 elements:
17305 // v = v + (v >> 8)
17306 // v = v + (v >> 16)
17308 // For i64 elements:
17309 // v = v + (v >> 8)
17310 // v = v + (v >> 16)
17311 // v = v + (v >> 32)
17314 SmallVector<SDValue, 8> Csts;
17315 for (unsigned i = 8; i <= Len/2; i *= 2) {
17316 Csts.assign(NumElts, DAG.getConstant(i, dl, EltVT));
17317 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
17318 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
17319 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17323 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
17324 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), dl,
17326 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
17327 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
17328 if (NeedsBitcast) {
17329 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17330 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
17332 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
17333 if (VT != And.getValueType())
17334 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17339 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
17340 SDNode *Node = Op.getNode();
17342 EVT T = Node->getValueType(0);
17343 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
17344 DAG.getConstant(0, dl, T), Node->getOperand(2));
17345 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
17346 cast<AtomicSDNode>(Node)->getMemoryVT(),
17347 Node->getOperand(0),
17348 Node->getOperand(1), negOp,
17349 cast<AtomicSDNode>(Node)->getMemOperand(),
17350 cast<AtomicSDNode>(Node)->getOrdering(),
17351 cast<AtomicSDNode>(Node)->getSynchScope());
17354 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
17355 SDNode *Node = Op.getNode();
17357 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
17359 // Convert seq_cst store -> xchg
17360 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
17361 // FIXME: On 32-bit, store -> fist or movq would be more efficient
17362 // (The only way to get a 16-byte store is cmpxchg16b)
17363 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
17364 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
17365 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17366 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
17367 cast<AtomicSDNode>(Node)->getMemoryVT(),
17368 Node->getOperand(0),
17369 Node->getOperand(1), Node->getOperand(2),
17370 cast<AtomicSDNode>(Node)->getMemOperand(),
17371 cast<AtomicSDNode>(Node)->getOrdering(),
17372 cast<AtomicSDNode>(Node)->getSynchScope());
17373 return Swap.getValue(1);
17375 // Other atomic stores have a simple pattern.
17379 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
17380 EVT VT = Op.getNode()->getSimpleValueType(0);
17382 // Let legalize expand this if it isn't a legal type yet.
17383 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17386 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17389 bool ExtraOp = false;
17390 switch (Op.getOpcode()) {
17391 default: llvm_unreachable("Invalid code");
17392 case ISD::ADDC: Opc = X86ISD::ADD; break;
17393 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
17394 case ISD::SUBC: Opc = X86ISD::SUB; break;
17395 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
17399 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17401 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17402 Op.getOperand(1), Op.getOperand(2));
17405 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
17406 SelectionDAG &DAG) {
17407 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
17409 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
17410 // which returns the values as { float, float } (in XMM0) or
17411 // { double, double } (which is returned in XMM0, XMM1).
17413 SDValue Arg = Op.getOperand(0);
17414 EVT ArgVT = Arg.getValueType();
17415 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17417 TargetLowering::ArgListTy Args;
17418 TargetLowering::ArgListEntry Entry;
17422 Entry.isSExt = false;
17423 Entry.isZExt = false;
17424 Args.push_back(Entry);
17426 bool isF64 = ArgVT == MVT::f64;
17427 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
17428 // the small struct {f32, f32} is returned in (eax, edx). For f64,
17429 // the results are returned via SRet in memory.
17430 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
17431 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17432 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
17434 Type *RetTy = isF64
17435 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
17436 : (Type*)VectorType::get(ArgTy, 4);
17438 TargetLowering::CallLoweringInfo CLI(DAG);
17439 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
17440 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
17442 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
17445 // Returned in xmm0 and xmm1.
17446 return CallResult.first;
17448 // Returned in bits 0:31 and 32:64 xmm0.
17449 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17450 CallResult.first, DAG.getIntPtrConstant(0, dl));
17451 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17452 CallResult.first, DAG.getIntPtrConstant(1, dl));
17453 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
17454 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
17457 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
17458 SelectionDAG &DAG) {
17459 assert(Subtarget->hasAVX512() &&
17460 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17462 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
17463 EVT VT = N->getValue().getValueType();
17464 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
17467 // X86 scatter kills mask register, so its type should be added to
17468 // the list of return values
17469 if (N->getNumValues() == 1) {
17470 SDValue Index = N->getIndex();
17471 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17472 !Index.getValueType().is512BitVector())
17473 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17475 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
17476 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17477 N->getOperand(3), Index };
17479 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
17480 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
17481 return SDValue(NewScatter.getNode(), 0);
17486 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
17487 SelectionDAG &DAG) {
17488 assert(Subtarget->hasAVX512() &&
17489 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17491 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
17492 EVT VT = Op.getValueType();
17493 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
17496 SDValue Index = N->getIndex();
17497 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17498 !Index.getValueType().is512BitVector()) {
17499 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17500 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17501 N->getOperand(3), Index };
17502 DAG.UpdateNodeOperands(N, Ops);
17507 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
17508 SelectionDAG &DAG) const {
17509 // TODO: Eventually, the lowering of these nodes should be informed by or
17510 // deferred to the GC strategy for the function in which they appear. For
17511 // now, however, they must be lowered to something. Since they are logically
17512 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17513 // require special handling for these nodes), lower them as literal NOOPs for
17515 SmallVector<SDValue, 2> Ops;
17517 Ops.push_back(Op.getOperand(0));
17518 if (Op->getGluedNode())
17519 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17522 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17523 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17528 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
17529 SelectionDAG &DAG) const {
17530 // TODO: Eventually, the lowering of these nodes should be informed by or
17531 // deferred to the GC strategy for the function in which they appear. For
17532 // now, however, they must be lowered to something. Since they are logically
17533 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17534 // require special handling for these nodes), lower them as literal NOOPs for
17536 SmallVector<SDValue, 2> Ops;
17538 Ops.push_back(Op.getOperand(0));
17539 if (Op->getGluedNode())
17540 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17543 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17544 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17549 /// LowerOperation - Provide custom lowering hooks for some operations.
17551 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
17552 switch (Op.getOpcode()) {
17553 default: llvm_unreachable("Should not custom lower this!");
17554 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
17555 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
17556 return LowerCMP_SWAP(Op, Subtarget, DAG);
17557 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
17558 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
17559 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
17560 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
17561 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
17562 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
17563 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
17564 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
17565 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
17566 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
17567 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
17568 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
17569 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
17570 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
17571 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
17572 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
17573 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
17574 case ISD::SHL_PARTS:
17575 case ISD::SRA_PARTS:
17576 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
17577 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
17578 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
17579 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
17580 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
17581 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
17582 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
17583 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
17584 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
17585 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
17586 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
17588 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
17589 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
17590 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
17591 case ISD::SETCC: return LowerSETCC(Op, DAG);
17592 case ISD::SELECT: return LowerSELECT(Op, DAG);
17593 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
17594 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
17595 case ISD::VASTART: return LowerVASTART(Op, DAG);
17596 case ISD::VAARG: return LowerVAARG(Op, DAG);
17597 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
17598 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
17599 case ISD::INTRINSIC_VOID:
17600 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
17601 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
17602 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
17603 case ISD::FRAME_TO_ARGS_OFFSET:
17604 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
17605 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
17606 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
17607 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
17608 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
17609 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
17610 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
17611 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
17612 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
17613 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
17614 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
17615 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
17616 case ISD::UMUL_LOHI:
17617 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
17620 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
17626 case ISD::UMULO: return LowerXALUO(Op, DAG);
17627 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
17628 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
17632 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
17633 case ISD::ADD: return LowerADD(Op, DAG);
17634 case ISD::SUB: return LowerSUB(Op, DAG);
17635 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
17636 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
17637 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
17638 case ISD::GC_TRANSITION_START:
17639 return LowerGC_TRANSITION_START(Op, DAG);
17640 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
17644 /// ReplaceNodeResults - Replace a node with an illegal result type
17645 /// with a new node built out of custom code.
17646 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
17647 SmallVectorImpl<SDValue>&Results,
17648 SelectionDAG &DAG) const {
17650 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17651 switch (N->getOpcode()) {
17653 llvm_unreachable("Do not know how to custom type legalize this operation!");
17654 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17655 case X86ISD::FMINC:
17657 case X86ISD::FMAXC:
17658 case X86ISD::FMAX: {
17659 EVT VT = N->getValueType(0);
17660 if (VT != MVT::v2f32)
17661 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17662 SDValue UNDEF = DAG.getUNDEF(VT);
17663 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17664 N->getOperand(0), UNDEF);
17665 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17666 N->getOperand(1), UNDEF);
17667 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17670 case ISD::SIGN_EXTEND_INREG:
17675 // We don't want to expand or promote these.
17682 case ISD::UDIVREM: {
17683 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17684 Results.push_back(V);
17687 case ISD::FP_TO_SINT:
17688 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
17689 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
17690 if (N->getOperand(0).getValueType() == MVT::f16)
17693 case ISD::FP_TO_UINT: {
17694 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17696 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17699 std::pair<SDValue,SDValue> Vals =
17700 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17701 SDValue FIST = Vals.first, StackSlot = Vals.second;
17702 if (FIST.getNode()) {
17703 EVT VT = N->getValueType(0);
17704 // Return a load from the stack slot.
17705 if (StackSlot.getNode())
17706 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17707 MachinePointerInfo(),
17708 false, false, false, 0));
17710 Results.push_back(FIST);
17714 case ISD::UINT_TO_FP: {
17715 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17716 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17717 N->getValueType(0) != MVT::v2f32)
17719 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17721 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
17723 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17724 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17725 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17726 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17727 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17728 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17731 case ISD::FP_ROUND: {
17732 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17734 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17735 Results.push_back(V);
17738 case ISD::FP_EXTEND: {
17739 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
17740 // No other ValueType for FP_EXTEND should reach this point.
17741 assert(N->getValueType(0) == MVT::v2f32 &&
17742 "Do not know how to legalize this Node");
17745 case ISD::INTRINSIC_W_CHAIN: {
17746 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17748 default : llvm_unreachable("Do not know how to custom type "
17749 "legalize this intrinsic operation!");
17750 case Intrinsic::x86_rdtsc:
17751 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17753 case Intrinsic::x86_rdtscp:
17754 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17756 case Intrinsic::x86_rdpmc:
17757 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17760 case ISD::READCYCLECOUNTER: {
17761 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17764 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17765 EVT T = N->getValueType(0);
17766 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17767 bool Regs64bit = T == MVT::i128;
17768 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17769 SDValue cpInL, cpInH;
17770 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17771 DAG.getConstant(0, dl, HalfT));
17772 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17773 DAG.getConstant(1, dl, HalfT));
17774 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17775 Regs64bit ? X86::RAX : X86::EAX,
17777 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17778 Regs64bit ? X86::RDX : X86::EDX,
17779 cpInH, cpInL.getValue(1));
17780 SDValue swapInL, swapInH;
17781 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17782 DAG.getConstant(0, dl, HalfT));
17783 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17784 DAG.getConstant(1, dl, HalfT));
17785 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17786 Regs64bit ? X86::RBX : X86::EBX,
17787 swapInL, cpInH.getValue(1));
17788 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17789 Regs64bit ? X86::RCX : X86::ECX,
17790 swapInH, swapInL.getValue(1));
17791 SDValue Ops[] = { swapInH.getValue(0),
17793 swapInH.getValue(1) };
17794 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17795 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17796 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17797 X86ISD::LCMPXCHG8_DAG;
17798 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17799 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17800 Regs64bit ? X86::RAX : X86::EAX,
17801 HalfT, Result.getValue(1));
17802 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17803 Regs64bit ? X86::RDX : X86::EDX,
17804 HalfT, cpOutL.getValue(2));
17805 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17807 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17808 MVT::i32, cpOutH.getValue(2));
17810 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17811 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
17812 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17814 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17815 Results.push_back(Success);
17816 Results.push_back(EFLAGS.getValue(1));
17819 case ISD::ATOMIC_SWAP:
17820 case ISD::ATOMIC_LOAD_ADD:
17821 case ISD::ATOMIC_LOAD_SUB:
17822 case ISD::ATOMIC_LOAD_AND:
17823 case ISD::ATOMIC_LOAD_OR:
17824 case ISD::ATOMIC_LOAD_XOR:
17825 case ISD::ATOMIC_LOAD_NAND:
17826 case ISD::ATOMIC_LOAD_MIN:
17827 case ISD::ATOMIC_LOAD_MAX:
17828 case ISD::ATOMIC_LOAD_UMIN:
17829 case ISD::ATOMIC_LOAD_UMAX:
17830 case ISD::ATOMIC_LOAD: {
17831 // Delegate to generic TypeLegalization. Situations we can really handle
17832 // should have already been dealt with by AtomicExpandPass.cpp.
17835 case ISD::BITCAST: {
17836 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17837 EVT DstVT = N->getValueType(0);
17838 EVT SrcVT = N->getOperand(0)->getValueType(0);
17840 if (SrcVT != MVT::f64 ||
17841 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17844 unsigned NumElts = DstVT.getVectorNumElements();
17845 EVT SVT = DstVT.getVectorElementType();
17846 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17847 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17848 MVT::v2f64, N->getOperand(0));
17849 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17851 if (ExperimentalVectorWideningLegalization) {
17852 // If we are legalizing vectors by widening, we already have the desired
17853 // legal vector type, just return it.
17854 Results.push_back(ToVecInt);
17858 SmallVector<SDValue, 8> Elts;
17859 for (unsigned i = 0, e = NumElts; i != e; ++i)
17860 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17861 ToVecInt, DAG.getIntPtrConstant(i, dl)));
17863 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17868 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17869 switch ((X86ISD::NodeType)Opcode) {
17870 case X86ISD::FIRST_NUMBER: break;
17871 case X86ISD::BSF: return "X86ISD::BSF";
17872 case X86ISD::BSR: return "X86ISD::BSR";
17873 case X86ISD::SHLD: return "X86ISD::SHLD";
17874 case X86ISD::SHRD: return "X86ISD::SHRD";
17875 case X86ISD::FAND: return "X86ISD::FAND";
17876 case X86ISD::FANDN: return "X86ISD::FANDN";
17877 case X86ISD::FOR: return "X86ISD::FOR";
17878 case X86ISD::FXOR: return "X86ISD::FXOR";
17879 case X86ISD::FSRL: return "X86ISD::FSRL";
17880 case X86ISD::FILD: return "X86ISD::FILD";
17881 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17882 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17883 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17884 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17885 case X86ISD::FLD: return "X86ISD::FLD";
17886 case X86ISD::FST: return "X86ISD::FST";
17887 case X86ISD::CALL: return "X86ISD::CALL";
17888 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17889 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17890 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17891 case X86ISD::BT: return "X86ISD::BT";
17892 case X86ISD::CMP: return "X86ISD::CMP";
17893 case X86ISD::COMI: return "X86ISD::COMI";
17894 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17895 case X86ISD::CMPM: return "X86ISD::CMPM";
17896 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17897 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
17898 case X86ISD::SETCC: return "X86ISD::SETCC";
17899 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17900 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17901 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
17902 case X86ISD::CMOV: return "X86ISD::CMOV";
17903 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17904 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17905 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17906 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17907 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17908 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17909 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17910 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
17911 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
17912 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
17913 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17914 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17915 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17916 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17917 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17918 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
17919 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17920 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17921 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17922 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17923 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17924 case X86ISD::ADDUS: return "X86ISD::ADDUS";
17925 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17926 case X86ISD::HADD: return "X86ISD::HADD";
17927 case X86ISD::HSUB: return "X86ISD::HSUB";
17928 case X86ISD::FHADD: return "X86ISD::FHADD";
17929 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17930 case X86ISD::UMAX: return "X86ISD::UMAX";
17931 case X86ISD::UMIN: return "X86ISD::UMIN";
17932 case X86ISD::SMAX: return "X86ISD::SMAX";
17933 case X86ISD::SMIN: return "X86ISD::SMIN";
17934 case X86ISD::FMAX: return "X86ISD::FMAX";
17935 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
17936 case X86ISD::FMIN: return "X86ISD::FMIN";
17937 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
17938 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17939 case X86ISD::FMINC: return "X86ISD::FMINC";
17940 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17941 case X86ISD::FRCP: return "X86ISD::FRCP";
17942 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17943 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17944 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17945 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17946 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17947 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17948 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17949 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17950 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17951 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17952 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17953 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17954 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17955 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17956 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17957 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17958 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17959 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17960 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17961 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17962 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17963 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17964 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17965 case X86ISD::VSHL: return "X86ISD::VSHL";
17966 case X86ISD::VSRL: return "X86ISD::VSRL";
17967 case X86ISD::VSRA: return "X86ISD::VSRA";
17968 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17969 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17970 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17971 case X86ISD::CMPP: return "X86ISD::CMPP";
17972 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17973 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17974 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17975 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17976 case X86ISD::ADD: return "X86ISD::ADD";
17977 case X86ISD::SUB: return "X86ISD::SUB";
17978 case X86ISD::ADC: return "X86ISD::ADC";
17979 case X86ISD::SBB: return "X86ISD::SBB";
17980 case X86ISD::SMUL: return "X86ISD::SMUL";
17981 case X86ISD::UMUL: return "X86ISD::UMUL";
17982 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17983 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17984 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17985 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17986 case X86ISD::INC: return "X86ISD::INC";
17987 case X86ISD::DEC: return "X86ISD::DEC";
17988 case X86ISD::OR: return "X86ISD::OR";
17989 case X86ISD::XOR: return "X86ISD::XOR";
17990 case X86ISD::AND: return "X86ISD::AND";
17991 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17992 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17993 case X86ISD::PTEST: return "X86ISD::PTEST";
17994 case X86ISD::TESTP: return "X86ISD::TESTP";
17995 case X86ISD::TESTM: return "X86ISD::TESTM";
17996 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17997 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17998 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17999 case X86ISD::PACKUS: return "X86ISD::PACKUS";
18000 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
18001 case X86ISD::VALIGN: return "X86ISD::VALIGN";
18002 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
18003 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
18004 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
18005 case X86ISD::SHUFP: return "X86ISD::SHUFP";
18006 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
18007 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
18008 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
18009 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
18010 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
18011 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
18012 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
18013 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
18014 case X86ISD::MOVSD: return "X86ISD::MOVSD";
18015 case X86ISD::MOVSS: return "X86ISD::MOVSS";
18016 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
18017 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
18018 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
18019 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
18020 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
18021 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
18022 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
18023 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
18024 case X86ISD::VPERMV: return "X86ISD::VPERMV";
18025 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
18026 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
18027 case X86ISD::VPERMI: return "X86ISD::VPERMI";
18028 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
18029 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
18030 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
18031 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
18032 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
18033 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
18034 case X86ISD::MFENCE: return "X86ISD::MFENCE";
18035 case X86ISD::SFENCE: return "X86ISD::SFENCE";
18036 case X86ISD::LFENCE: return "X86ISD::LFENCE";
18037 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
18038 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
18039 case X86ISD::SAHF: return "X86ISD::SAHF";
18040 case X86ISD::RDRAND: return "X86ISD::RDRAND";
18041 case X86ISD::RDSEED: return "X86ISD::RDSEED";
18042 case X86ISD::FMADD: return "X86ISD::FMADD";
18043 case X86ISD::FMSUB: return "X86ISD::FMSUB";
18044 case X86ISD::FNMADD: return "X86ISD::FNMADD";
18045 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
18046 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
18047 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
18048 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
18049 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
18050 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
18051 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
18052 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
18053 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
18054 case X86ISD::RNDSCALE: return "X86ISD::RNDSCALE";
18055 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
18056 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
18057 case X86ISD::XTEST: return "X86ISD::XTEST";
18058 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
18059 case X86ISD::EXPAND: return "X86ISD::EXPAND";
18060 case X86ISD::SELECT: return "X86ISD::SELECT";
18061 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
18062 case X86ISD::RCP28: return "X86ISD::RCP28";
18063 case X86ISD::EXP2: return "X86ISD::EXP2";
18064 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
18065 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
18066 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
18067 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
18068 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
18069 case X86ISD::ADDS: return "X86ISD::ADDS";
18070 case X86ISD::SUBS: return "X86ISD::SUBS";
18075 // isLegalAddressingMode - Return true if the addressing mode represented
18076 // by AM is legal for this target, for a load/store of the specified type.
18077 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
18079 // X86 supports extremely general addressing modes.
18080 CodeModel::Model M = getTargetMachine().getCodeModel();
18081 Reloc::Model R = getTargetMachine().getRelocationModel();
18083 // X86 allows a sign-extended 32-bit immediate field as a displacement.
18084 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
18089 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
18091 // If a reference to this global requires an extra load, we can't fold it.
18092 if (isGlobalStubReference(GVFlags))
18095 // If BaseGV requires a register for the PIC base, we cannot also have a
18096 // BaseReg specified.
18097 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
18100 // If lower 4G is not available, then we must use rip-relative addressing.
18101 if ((M != CodeModel::Small || R != Reloc::Static) &&
18102 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
18106 switch (AM.Scale) {
18112 // These scales always work.
18117 // These scales are formed with basereg+scalereg. Only accept if there is
18122 default: // Other stuff never works.
18129 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
18130 unsigned Bits = Ty->getScalarSizeInBits();
18132 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
18133 // particularly cheaper than those without.
18137 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
18138 // variable shifts just as cheap as scalar ones.
18139 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
18142 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
18143 // fully general vector.
18147 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
18148 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18150 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
18151 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
18152 return NumBits1 > NumBits2;
18155 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
18156 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18159 if (!isTypeLegal(EVT::getEVT(Ty1)))
18162 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
18164 // Assuming the caller doesn't have a zeroext or signext return parameter,
18165 // truncation all the way down to i1 is valid.
18169 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
18170 return isInt<32>(Imm);
18173 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
18174 // Can also use sub to handle negated immediates.
18175 return isInt<32>(Imm);
18178 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
18179 if (!VT1.isInteger() || !VT2.isInteger())
18181 unsigned NumBits1 = VT1.getSizeInBits();
18182 unsigned NumBits2 = VT2.getSizeInBits();
18183 return NumBits1 > NumBits2;
18186 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
18187 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18188 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
18191 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
18192 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18193 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
18196 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
18197 EVT VT1 = Val.getValueType();
18198 if (isZExtFree(VT1, VT2))
18201 if (Val.getOpcode() != ISD::LOAD)
18204 if (!VT1.isSimple() || !VT1.isInteger() ||
18205 !VT2.isSimple() || !VT2.isInteger())
18208 switch (VT1.getSimpleVT().SimpleTy) {
18213 // X86 has 8, 16, and 32-bit zero-extending loads.
18220 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
18223 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
18224 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
18227 VT = VT.getScalarType();
18229 if (!VT.isSimple())
18232 switch (VT.getSimpleVT().SimpleTy) {
18243 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
18244 // i16 instructions are longer (0x66 prefix) and potentially slower.
18245 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18248 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18249 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18250 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18251 /// are assumed to be legal.
18253 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18255 if (!VT.isSimple())
18258 // Not for i1 vectors
18259 if (VT.getScalarType() == MVT::i1)
18262 // Very little shuffling can be done for 64-bit vectors right now.
18263 if (VT.getSizeInBits() == 64)
18266 // We only care that the types being shuffled are legal. The lowering can
18267 // handle any possible shuffle mask that results.
18268 return isTypeLegal(VT.getSimpleVT());
18272 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18274 // Just delegate to the generic legality, clear masks aren't special.
18275 return isShuffleMaskLegal(Mask, VT);
18278 //===----------------------------------------------------------------------===//
18279 // X86 Scheduler Hooks
18280 //===----------------------------------------------------------------------===//
18282 /// Utility function to emit xbegin specifying the start of an RTM region.
18283 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18284 const TargetInstrInfo *TII) {
18285 DebugLoc DL = MI->getDebugLoc();
18287 const BasicBlock *BB = MBB->getBasicBlock();
18288 MachineFunction::iterator I = MBB;
18291 // For the v = xbegin(), we generate
18302 MachineBasicBlock *thisMBB = MBB;
18303 MachineFunction *MF = MBB->getParent();
18304 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18305 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18306 MF->insert(I, mainMBB);
18307 MF->insert(I, sinkMBB);
18309 // Transfer the remainder of BB and its successor edges to sinkMBB.
18310 sinkMBB->splice(sinkMBB->begin(), MBB,
18311 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18312 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18316 // # fallthrough to mainMBB
18317 // # abortion to sinkMBB
18318 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18319 thisMBB->addSuccessor(mainMBB);
18320 thisMBB->addSuccessor(sinkMBB);
18324 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18325 mainMBB->addSuccessor(sinkMBB);
18328 // EAX is live into the sinkMBB
18329 sinkMBB->addLiveIn(X86::EAX);
18330 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18331 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18334 MI->eraseFromParent();
18338 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18339 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18340 // in the .td file.
18341 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18342 const TargetInstrInfo *TII) {
18344 switch (MI->getOpcode()) {
18345 default: llvm_unreachable("illegal opcode!");
18346 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18347 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18348 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18349 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18350 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18351 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18352 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18353 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18356 DebugLoc dl = MI->getDebugLoc();
18357 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18359 unsigned NumArgs = MI->getNumOperands();
18360 for (unsigned i = 1; i < NumArgs; ++i) {
18361 MachineOperand &Op = MI->getOperand(i);
18362 if (!(Op.isReg() && Op.isImplicit()))
18363 MIB.addOperand(Op);
18365 if (MI->hasOneMemOperand())
18366 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18368 BuildMI(*BB, MI, dl,
18369 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18370 .addReg(X86::XMM0);
18372 MI->eraseFromParent();
18376 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18377 // defs in an instruction pattern
18378 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18379 const TargetInstrInfo *TII) {
18381 switch (MI->getOpcode()) {
18382 default: llvm_unreachable("illegal opcode!");
18383 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18384 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18385 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18386 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18387 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18388 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18389 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18390 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18393 DebugLoc dl = MI->getDebugLoc();
18394 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18396 unsigned NumArgs = MI->getNumOperands(); // remove the results
18397 for (unsigned i = 1; i < NumArgs; ++i) {
18398 MachineOperand &Op = MI->getOperand(i);
18399 if (!(Op.isReg() && Op.isImplicit()))
18400 MIB.addOperand(Op);
18402 if (MI->hasOneMemOperand())
18403 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18405 BuildMI(*BB, MI, dl,
18406 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18409 MI->eraseFromParent();
18413 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18414 const X86Subtarget *Subtarget) {
18415 DebugLoc dl = MI->getDebugLoc();
18416 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18417 // Address into RAX/EAX, other two args into ECX, EDX.
18418 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18419 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18420 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18421 for (int i = 0; i < X86::AddrNumOperands; ++i)
18422 MIB.addOperand(MI->getOperand(i));
18424 unsigned ValOps = X86::AddrNumOperands;
18425 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18426 .addReg(MI->getOperand(ValOps).getReg());
18427 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18428 .addReg(MI->getOperand(ValOps+1).getReg());
18430 // The instruction doesn't actually take any operands though.
18431 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18433 MI->eraseFromParent(); // The pseudo is gone now.
18437 MachineBasicBlock *
18438 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
18439 MachineBasicBlock *MBB) const {
18440 // Emit va_arg instruction on X86-64.
18442 // Operands to this pseudo-instruction:
18443 // 0 ) Output : destination address (reg)
18444 // 1-5) Input : va_list address (addr, i64mem)
18445 // 6 ) ArgSize : Size (in bytes) of vararg type
18446 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18447 // 8 ) Align : Alignment of type
18448 // 9 ) EFLAGS (implicit-def)
18450 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18451 static_assert(X86::AddrNumOperands == 5,
18452 "VAARG_64 assumes 5 address operands");
18454 unsigned DestReg = MI->getOperand(0).getReg();
18455 MachineOperand &Base = MI->getOperand(1);
18456 MachineOperand &Scale = MI->getOperand(2);
18457 MachineOperand &Index = MI->getOperand(3);
18458 MachineOperand &Disp = MI->getOperand(4);
18459 MachineOperand &Segment = MI->getOperand(5);
18460 unsigned ArgSize = MI->getOperand(6).getImm();
18461 unsigned ArgMode = MI->getOperand(7).getImm();
18462 unsigned Align = MI->getOperand(8).getImm();
18464 // Memory Reference
18465 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
18466 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18467 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18469 // Machine Information
18470 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18471 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
18472 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
18473 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
18474 DebugLoc DL = MI->getDebugLoc();
18476 // struct va_list {
18479 // i64 overflow_area (address)
18480 // i64 reg_save_area (address)
18482 // sizeof(va_list) = 24
18483 // alignment(va_list) = 8
18485 unsigned TotalNumIntRegs = 6;
18486 unsigned TotalNumXMMRegs = 8;
18487 bool UseGPOffset = (ArgMode == 1);
18488 bool UseFPOffset = (ArgMode == 2);
18489 unsigned MaxOffset = TotalNumIntRegs * 8 +
18490 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
18492 /* Align ArgSize to a multiple of 8 */
18493 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
18494 bool NeedsAlign = (Align > 8);
18496 MachineBasicBlock *thisMBB = MBB;
18497 MachineBasicBlock *overflowMBB;
18498 MachineBasicBlock *offsetMBB;
18499 MachineBasicBlock *endMBB;
18501 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
18502 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
18503 unsigned OffsetReg = 0;
18505 if (!UseGPOffset && !UseFPOffset) {
18506 // If we only pull from the overflow region, we don't create a branch.
18507 // We don't need to alter control flow.
18508 OffsetDestReg = 0; // unused
18509 OverflowDestReg = DestReg;
18511 offsetMBB = nullptr;
18512 overflowMBB = thisMBB;
18515 // First emit code to check if gp_offset (or fp_offset) is below the bound.
18516 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
18517 // If not, pull from overflow_area. (branch to overflowMBB)
18522 // offsetMBB overflowMBB
18527 // Registers for the PHI in endMBB
18528 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
18529 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
18531 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18532 MachineFunction *MF = MBB->getParent();
18533 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18534 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18535 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18537 MachineFunction::iterator MBBIter = MBB;
18540 // Insert the new basic blocks
18541 MF->insert(MBBIter, offsetMBB);
18542 MF->insert(MBBIter, overflowMBB);
18543 MF->insert(MBBIter, endMBB);
18545 // Transfer the remainder of MBB and its successor edges to endMBB.
18546 endMBB->splice(endMBB->begin(), thisMBB,
18547 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
18548 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
18550 // Make offsetMBB and overflowMBB successors of thisMBB
18551 thisMBB->addSuccessor(offsetMBB);
18552 thisMBB->addSuccessor(overflowMBB);
18554 // endMBB is a successor of both offsetMBB and overflowMBB
18555 offsetMBB->addSuccessor(endMBB);
18556 overflowMBB->addSuccessor(endMBB);
18558 // Load the offset value into a register
18559 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18560 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
18564 .addDisp(Disp, UseFPOffset ? 4 : 0)
18565 .addOperand(Segment)
18566 .setMemRefs(MMOBegin, MMOEnd);
18568 // Check if there is enough room left to pull this argument.
18569 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
18571 .addImm(MaxOffset + 8 - ArgSizeA8);
18573 // Branch to "overflowMBB" if offset >= max
18574 // Fall through to "offsetMBB" otherwise
18575 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
18576 .addMBB(overflowMBB);
18579 // In offsetMBB, emit code to use the reg_save_area.
18581 assert(OffsetReg != 0);
18583 // Read the reg_save_area address.
18584 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
18585 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
18590 .addOperand(Segment)
18591 .setMemRefs(MMOBegin, MMOEnd);
18593 // Zero-extend the offset
18594 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
18595 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
18598 .addImm(X86::sub_32bit);
18600 // Add the offset to the reg_save_area to get the final address.
18601 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
18602 .addReg(OffsetReg64)
18603 .addReg(RegSaveReg);
18605 // Compute the offset for the next argument
18606 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18607 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
18609 .addImm(UseFPOffset ? 16 : 8);
18611 // Store it back into the va_list.
18612 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
18616 .addDisp(Disp, UseFPOffset ? 4 : 0)
18617 .addOperand(Segment)
18618 .addReg(NextOffsetReg)
18619 .setMemRefs(MMOBegin, MMOEnd);
18622 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
18627 // Emit code to use overflow area
18630 // Load the overflow_area address into a register.
18631 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
18632 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
18637 .addOperand(Segment)
18638 .setMemRefs(MMOBegin, MMOEnd);
18640 // If we need to align it, do so. Otherwise, just copy the address
18641 // to OverflowDestReg.
18643 // Align the overflow address
18644 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
18645 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
18647 // aligned_addr = (addr + (align-1)) & ~(align-1)
18648 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
18649 .addReg(OverflowAddrReg)
18652 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
18654 .addImm(~(uint64_t)(Align-1));
18656 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
18657 .addReg(OverflowAddrReg);
18660 // Compute the next overflow address after this argument.
18661 // (the overflow address should be kept 8-byte aligned)
18662 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
18663 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
18664 .addReg(OverflowDestReg)
18665 .addImm(ArgSizeA8);
18667 // Store the new overflow address.
18668 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
18673 .addOperand(Segment)
18674 .addReg(NextAddrReg)
18675 .setMemRefs(MMOBegin, MMOEnd);
18677 // If we branched, emit the PHI to the front of endMBB.
18679 BuildMI(*endMBB, endMBB->begin(), DL,
18680 TII->get(X86::PHI), DestReg)
18681 .addReg(OffsetDestReg).addMBB(offsetMBB)
18682 .addReg(OverflowDestReg).addMBB(overflowMBB);
18685 // Erase the pseudo instruction
18686 MI->eraseFromParent();
18691 MachineBasicBlock *
18692 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18694 MachineBasicBlock *MBB) const {
18695 // Emit code to save XMM registers to the stack. The ABI says that the
18696 // number of registers to save is given in %al, so it's theoretically
18697 // possible to do an indirect jump trick to avoid saving all of them,
18698 // however this code takes a simpler approach and just executes all
18699 // of the stores if %al is non-zero. It's less code, and it's probably
18700 // easier on the hardware branch predictor, and stores aren't all that
18701 // expensive anyway.
18703 // Create the new basic blocks. One block contains all the XMM stores,
18704 // and one block is the final destination regardless of whether any
18705 // stores were performed.
18706 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18707 MachineFunction *F = MBB->getParent();
18708 MachineFunction::iterator MBBIter = MBB;
18710 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
18711 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
18712 F->insert(MBBIter, XMMSaveMBB);
18713 F->insert(MBBIter, EndMBB);
18715 // Transfer the remainder of MBB and its successor edges to EndMBB.
18716 EndMBB->splice(EndMBB->begin(), MBB,
18717 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18718 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
18720 // The original block will now fall through to the XMM save block.
18721 MBB->addSuccessor(XMMSaveMBB);
18722 // The XMMSaveMBB will fall through to the end block.
18723 XMMSaveMBB->addSuccessor(EndMBB);
18725 // Now add the instructions.
18726 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18727 DebugLoc DL = MI->getDebugLoc();
18729 unsigned CountReg = MI->getOperand(0).getReg();
18730 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
18731 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
18733 if (!Subtarget->isTargetWin64()) {
18734 // If %al is 0, branch around the XMM save block.
18735 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
18736 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18737 MBB->addSuccessor(EndMBB);
18740 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18741 // that was just emitted, but clearly shouldn't be "saved".
18742 assert((MI->getNumOperands() <= 3 ||
18743 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18744 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18745 && "Expected last argument to be EFLAGS");
18746 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18747 // In the XMM save block, save all the XMM argument registers.
18748 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18749 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18750 MachineMemOperand *MMO =
18751 F->getMachineMemOperand(
18752 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18753 MachineMemOperand::MOStore,
18754 /*Size=*/16, /*Align=*/16);
18755 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18756 .addFrameIndex(RegSaveFrameIndex)
18757 .addImm(/*Scale=*/1)
18758 .addReg(/*IndexReg=*/0)
18759 .addImm(/*Disp=*/Offset)
18760 .addReg(/*Segment=*/0)
18761 .addReg(MI->getOperand(i).getReg())
18762 .addMemOperand(MMO);
18765 MI->eraseFromParent(); // The pseudo instruction is gone now.
18770 // The EFLAGS operand of SelectItr might be missing a kill marker
18771 // because there were multiple uses of EFLAGS, and ISel didn't know
18772 // which to mark. Figure out whether SelectItr should have had a
18773 // kill marker, and set it if it should. Returns the correct kill
18775 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18776 MachineBasicBlock* BB,
18777 const TargetRegisterInfo* TRI) {
18778 // Scan forward through BB for a use/def of EFLAGS.
18779 MachineBasicBlock::iterator miI(std::next(SelectItr));
18780 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18781 const MachineInstr& mi = *miI;
18782 if (mi.readsRegister(X86::EFLAGS))
18784 if (mi.definesRegister(X86::EFLAGS))
18785 break; // Should have kill-flag - update below.
18788 // If we hit the end of the block, check whether EFLAGS is live into a
18790 if (miI == BB->end()) {
18791 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18792 sEnd = BB->succ_end();
18793 sItr != sEnd; ++sItr) {
18794 MachineBasicBlock* succ = *sItr;
18795 if (succ->isLiveIn(X86::EFLAGS))
18800 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18801 // out. SelectMI should have a kill flag on EFLAGS.
18802 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18806 MachineBasicBlock *
18807 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18808 MachineBasicBlock *BB) const {
18809 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18810 DebugLoc DL = MI->getDebugLoc();
18812 // To "insert" a SELECT_CC instruction, we actually have to insert the
18813 // diamond control-flow pattern. The incoming instruction knows the
18814 // destination vreg to set, the condition code register to branch on, the
18815 // true/false values to select between, and a branch opcode to use.
18816 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18817 MachineFunction::iterator It = BB;
18823 // cmpTY ccX, r1, r2
18825 // fallthrough --> copy0MBB
18826 MachineBasicBlock *thisMBB = BB;
18827 MachineFunction *F = BB->getParent();
18829 // We also lower double CMOVs:
18830 // (CMOV (CMOV F, T, cc1), T, cc2)
18831 // to two successives branches. For that, we look for another CMOV as the
18832 // following instruction.
18834 // Without this, we would add a PHI between the two jumps, which ends up
18835 // creating a few copies all around. For instance, for
18837 // (sitofp (zext (fcmp une)))
18839 // we would generate:
18841 // ucomiss %xmm1, %xmm0
18842 // movss <1.0f>, %xmm0
18843 // movaps %xmm0, %xmm1
18845 // xorps %xmm1, %xmm1
18848 // movaps %xmm1, %xmm0
18852 // because this custom-inserter would have generated:
18864 // A: X = ...; Y = ...
18866 // C: Z = PHI [X, A], [Y, B]
18868 // E: PHI [X, C], [Z, D]
18870 // If we lower both CMOVs in a single step, we can instead generate:
18882 // A: X = ...; Y = ...
18884 // E: PHI [X, A], [X, C], [Y, D]
18886 // Which, in our sitofp/fcmp example, gives us something like:
18888 // ucomiss %xmm1, %xmm0
18889 // movss <1.0f>, %xmm0
18892 // xorps %xmm0, %xmm0
18896 MachineInstr *NextCMOV = nullptr;
18897 MachineBasicBlock::iterator NextMIIt =
18898 std::next(MachineBasicBlock::iterator(MI));
18899 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
18900 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
18901 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
18902 NextCMOV = &*NextMIIt;
18904 MachineBasicBlock *jcc1MBB = nullptr;
18906 // If we have a double CMOV, we lower it to two successive branches to
18907 // the same block. EFLAGS is used by both, so mark it as live in the second.
18909 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
18910 F->insert(It, jcc1MBB);
18911 jcc1MBB->addLiveIn(X86::EFLAGS);
18914 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18915 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18916 F->insert(It, copy0MBB);
18917 F->insert(It, sinkMBB);
18919 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18920 // live into the sink and copy blocks.
18921 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18923 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
18924 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
18925 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
18926 copy0MBB->addLiveIn(X86::EFLAGS);
18927 sinkMBB->addLiveIn(X86::EFLAGS);
18930 // Transfer the remainder of BB and its successor edges to sinkMBB.
18931 sinkMBB->splice(sinkMBB->begin(), BB,
18932 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18933 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18935 // Add the true and fallthrough blocks as its successors.
18937 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
18938 BB->addSuccessor(jcc1MBB);
18940 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
18941 // jump to the sinkMBB.
18942 jcc1MBB->addSuccessor(copy0MBB);
18943 jcc1MBB->addSuccessor(sinkMBB);
18945 BB->addSuccessor(copy0MBB);
18948 // The true block target of the first (or only) branch is always sinkMBB.
18949 BB->addSuccessor(sinkMBB);
18951 // Create the conditional branch instruction.
18953 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18954 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18957 unsigned Opc2 = X86::GetCondBranchFromCond(
18958 (X86::CondCode)NextCMOV->getOperand(3).getImm());
18959 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
18963 // %FalseValue = ...
18964 // # fallthrough to sinkMBB
18965 copy0MBB->addSuccessor(sinkMBB);
18968 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18970 MachineInstrBuilder MIB =
18971 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
18972 MI->getOperand(0).getReg())
18973 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18974 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18976 // If we have a double CMOV, the second Jcc provides the same incoming
18977 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
18979 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
18980 // Copy the PHI result to the register defined by the second CMOV.
18981 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
18982 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
18983 .addReg(MI->getOperand(0).getReg());
18984 NextCMOV->eraseFromParent();
18987 MI->eraseFromParent(); // The pseudo instruction is gone now.
18991 MachineBasicBlock *
18992 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18993 MachineBasicBlock *BB) const {
18994 MachineFunction *MF = BB->getParent();
18995 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18996 DebugLoc DL = MI->getDebugLoc();
18997 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18999 assert(MF->shouldSplitStack());
19001 const bool Is64Bit = Subtarget->is64Bit();
19002 const bool IsLP64 = Subtarget->isTarget64BitLP64();
19004 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
19005 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
19008 // ... [Till the alloca]
19009 // If stacklet is not large enough, jump to mallocMBB
19012 // Allocate by subtracting from RSP
19013 // Jump to continueMBB
19016 // Allocate by call to runtime
19020 // [rest of original BB]
19023 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19024 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19025 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19027 MachineRegisterInfo &MRI = MF->getRegInfo();
19028 const TargetRegisterClass *AddrRegClass =
19029 getRegClassFor(getPointerTy());
19031 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19032 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19033 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
19034 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
19035 sizeVReg = MI->getOperand(1).getReg(),
19036 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
19038 MachineFunction::iterator MBBIter = BB;
19041 MF->insert(MBBIter, bumpMBB);
19042 MF->insert(MBBIter, mallocMBB);
19043 MF->insert(MBBIter, continueMBB);
19045 continueMBB->splice(continueMBB->begin(), BB,
19046 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19047 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
19049 // Add code to the main basic block to check if the stack limit has been hit,
19050 // and if so, jump to mallocMBB otherwise to bumpMBB.
19051 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
19052 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
19053 .addReg(tmpSPVReg).addReg(sizeVReg);
19054 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
19055 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
19056 .addReg(SPLimitVReg);
19057 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
19059 // bumpMBB simply decreases the stack pointer, since we know the current
19060 // stacklet has enough space.
19061 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
19062 .addReg(SPLimitVReg);
19063 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
19064 .addReg(SPLimitVReg);
19065 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19067 // Calls into a routine in libgcc to allocate more space from the heap.
19068 const uint32_t *RegMask =
19069 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
19071 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
19073 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19074 .addExternalSymbol("__morestack_allocate_stack_space")
19075 .addRegMask(RegMask)
19076 .addReg(X86::RDI, RegState::Implicit)
19077 .addReg(X86::RAX, RegState::ImplicitDefine);
19078 } else if (Is64Bit) {
19079 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
19081 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19082 .addExternalSymbol("__morestack_allocate_stack_space")
19083 .addRegMask(RegMask)
19084 .addReg(X86::EDI, RegState::Implicit)
19085 .addReg(X86::EAX, RegState::ImplicitDefine);
19087 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
19089 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
19090 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
19091 .addExternalSymbol("__morestack_allocate_stack_space")
19092 .addRegMask(RegMask)
19093 .addReg(X86::EAX, RegState::ImplicitDefine);
19097 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
19100 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
19101 .addReg(IsLP64 ? X86::RAX : X86::EAX);
19102 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19104 // Set up the CFG correctly.
19105 BB->addSuccessor(bumpMBB);
19106 BB->addSuccessor(mallocMBB);
19107 mallocMBB->addSuccessor(continueMBB);
19108 bumpMBB->addSuccessor(continueMBB);
19110 // Take care of the PHI nodes.
19111 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
19112 MI->getOperand(0).getReg())
19113 .addReg(mallocPtrVReg).addMBB(mallocMBB)
19114 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
19116 // Delete the original pseudo instruction.
19117 MI->eraseFromParent();
19120 return continueMBB;
19123 MachineBasicBlock *
19124 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
19125 MachineBasicBlock *BB) const {
19126 DebugLoc DL = MI->getDebugLoc();
19128 assert(!Subtarget->isTargetMachO());
19130 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
19132 MI->eraseFromParent(); // The pseudo instruction is gone now.
19136 MachineBasicBlock *
19137 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
19138 MachineBasicBlock *BB) const {
19139 // This is pretty easy. We're taking the value that we received from
19140 // our load from the relocation, sticking it in either RDI (x86-64)
19141 // or EAX and doing an indirect call. The return value will then
19142 // be in the normal return register.
19143 MachineFunction *F = BB->getParent();
19144 const X86InstrInfo *TII = Subtarget->getInstrInfo();
19145 DebugLoc DL = MI->getDebugLoc();
19147 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
19148 assert(MI->getOperand(3).isGlobal() && "This should be a global");
19150 // Get a register mask for the lowered call.
19151 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
19152 // proper register mask.
19153 const uint32_t *RegMask =
19154 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
19155 if (Subtarget->is64Bit()) {
19156 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19157 TII->get(X86::MOV64rm), X86::RDI)
19159 .addImm(0).addReg(0)
19160 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19161 MI->getOperand(3).getTargetFlags())
19163 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
19164 addDirectMem(MIB, X86::RDI);
19165 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
19166 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
19167 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19168 TII->get(X86::MOV32rm), X86::EAX)
19170 .addImm(0).addReg(0)
19171 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19172 MI->getOperand(3).getTargetFlags())
19174 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19175 addDirectMem(MIB, X86::EAX);
19176 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19178 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19179 TII->get(X86::MOV32rm), X86::EAX)
19180 .addReg(TII->getGlobalBaseReg(F))
19181 .addImm(0).addReg(0)
19182 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19183 MI->getOperand(3).getTargetFlags())
19185 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19186 addDirectMem(MIB, X86::EAX);
19187 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19190 MI->eraseFromParent(); // The pseudo instruction is gone now.
19194 MachineBasicBlock *
19195 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
19196 MachineBasicBlock *MBB) const {
19197 DebugLoc DL = MI->getDebugLoc();
19198 MachineFunction *MF = MBB->getParent();
19199 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19200 MachineRegisterInfo &MRI = MF->getRegInfo();
19202 const BasicBlock *BB = MBB->getBasicBlock();
19203 MachineFunction::iterator I = MBB;
19206 // Memory Reference
19207 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19208 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19211 unsigned MemOpndSlot = 0;
19213 unsigned CurOp = 0;
19215 DstReg = MI->getOperand(CurOp++).getReg();
19216 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
19217 assert(RC->hasType(MVT::i32) && "Invalid destination!");
19218 unsigned mainDstReg = MRI.createVirtualRegister(RC);
19219 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
19221 MemOpndSlot = CurOp;
19223 MVT PVT = getPointerTy();
19224 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19225 "Invalid Pointer Size!");
19227 // For v = setjmp(buf), we generate
19230 // buf[LabelOffset] = restoreMBB
19231 // SjLjSetup restoreMBB
19237 // v = phi(main, restore)
19240 // if base pointer being used, load it from frame
19243 MachineBasicBlock *thisMBB = MBB;
19244 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19245 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19246 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
19247 MF->insert(I, mainMBB);
19248 MF->insert(I, sinkMBB);
19249 MF->push_back(restoreMBB);
19251 MachineInstrBuilder MIB;
19253 // Transfer the remainder of BB and its successor edges to sinkMBB.
19254 sinkMBB->splice(sinkMBB->begin(), MBB,
19255 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19256 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19259 unsigned PtrStoreOpc = 0;
19260 unsigned LabelReg = 0;
19261 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19262 Reloc::Model RM = MF->getTarget().getRelocationModel();
19263 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19264 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19266 // Prepare IP either in reg or imm.
19267 if (!UseImmLabel) {
19268 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19269 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19270 LabelReg = MRI.createVirtualRegister(PtrRC);
19271 if (Subtarget->is64Bit()) {
19272 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19276 .addMBB(restoreMBB)
19279 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19280 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19281 .addReg(XII->getGlobalBaseReg(MF))
19284 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19288 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19290 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19291 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19292 if (i == X86::AddrDisp)
19293 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19295 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19298 MIB.addReg(LabelReg);
19300 MIB.addMBB(restoreMBB);
19301 MIB.setMemRefs(MMOBegin, MMOEnd);
19303 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19304 .addMBB(restoreMBB);
19306 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19307 MIB.addRegMask(RegInfo->getNoPreservedMask());
19308 thisMBB->addSuccessor(mainMBB);
19309 thisMBB->addSuccessor(restoreMBB);
19313 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19314 mainMBB->addSuccessor(sinkMBB);
19317 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19318 TII->get(X86::PHI), DstReg)
19319 .addReg(mainDstReg).addMBB(mainMBB)
19320 .addReg(restoreDstReg).addMBB(restoreMBB);
19323 if (RegInfo->hasBasePointer(*MF)) {
19324 const bool Uses64BitFramePtr =
19325 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
19326 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
19327 X86FI->setRestoreBasePointer(MF);
19328 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
19329 unsigned BasePtr = RegInfo->getBaseRegister();
19330 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
19331 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
19332 FramePtr, true, X86FI->getRestoreBasePointerOffset())
19333 .setMIFlag(MachineInstr::FrameSetup);
19335 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19336 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
19337 restoreMBB->addSuccessor(sinkMBB);
19339 MI->eraseFromParent();
19343 MachineBasicBlock *
19344 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19345 MachineBasicBlock *MBB) const {
19346 DebugLoc DL = MI->getDebugLoc();
19347 MachineFunction *MF = MBB->getParent();
19348 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19349 MachineRegisterInfo &MRI = MF->getRegInfo();
19351 // Memory Reference
19352 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19353 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19355 MVT PVT = getPointerTy();
19356 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19357 "Invalid Pointer Size!");
19359 const TargetRegisterClass *RC =
19360 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19361 unsigned Tmp = MRI.createVirtualRegister(RC);
19362 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19363 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19364 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19365 unsigned SP = RegInfo->getStackRegister();
19367 MachineInstrBuilder MIB;
19369 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19370 const int64_t SPOffset = 2 * PVT.getStoreSize();
19372 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19373 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19376 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19377 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19378 MIB.addOperand(MI->getOperand(i));
19379 MIB.setMemRefs(MMOBegin, MMOEnd);
19381 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19382 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19383 if (i == X86::AddrDisp)
19384 MIB.addDisp(MI->getOperand(i), LabelOffset);
19386 MIB.addOperand(MI->getOperand(i));
19388 MIB.setMemRefs(MMOBegin, MMOEnd);
19390 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19391 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19392 if (i == X86::AddrDisp)
19393 MIB.addDisp(MI->getOperand(i), SPOffset);
19395 MIB.addOperand(MI->getOperand(i));
19397 MIB.setMemRefs(MMOBegin, MMOEnd);
19399 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19401 MI->eraseFromParent();
19405 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19406 // accumulator loops. Writing back to the accumulator allows the coalescer
19407 // to remove extra copies in the loop.
19408 MachineBasicBlock *
19409 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19410 MachineBasicBlock *MBB) const {
19411 MachineOperand &AddendOp = MI->getOperand(3);
19413 // Bail out early if the addend isn't a register - we can't switch these.
19414 if (!AddendOp.isReg())
19417 MachineFunction &MF = *MBB->getParent();
19418 MachineRegisterInfo &MRI = MF.getRegInfo();
19420 // Check whether the addend is defined by a PHI:
19421 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19422 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19423 if (!AddendDef.isPHI())
19426 // Look for the following pattern:
19428 // %addend = phi [%entry, 0], [%loop, %result]
19430 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19434 // %addend = phi [%entry, 0], [%loop, %result]
19436 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19438 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19439 assert(AddendDef.getOperand(i).isReg());
19440 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19441 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19442 if (&PHISrcInst == MI) {
19443 // Found a matching instruction.
19444 unsigned NewFMAOpc = 0;
19445 switch (MI->getOpcode()) {
19446 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19447 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19448 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19449 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19450 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19451 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19452 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19453 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19454 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19455 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19456 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19457 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19458 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19459 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19460 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19461 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19462 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
19463 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
19464 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
19465 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
19467 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19468 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19469 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19470 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19471 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19472 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19473 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19474 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19475 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
19476 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
19477 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
19478 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
19479 default: llvm_unreachable("Unrecognized FMA variant.");
19482 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
19483 MachineInstrBuilder MIB =
19484 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19485 .addOperand(MI->getOperand(0))
19486 .addOperand(MI->getOperand(3))
19487 .addOperand(MI->getOperand(2))
19488 .addOperand(MI->getOperand(1));
19489 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19490 MI->eraseFromParent();
19497 MachineBasicBlock *
19498 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19499 MachineBasicBlock *BB) const {
19500 switch (MI->getOpcode()) {
19501 default: llvm_unreachable("Unexpected instr type to insert");
19502 case X86::TAILJMPd64:
19503 case X86::TAILJMPr64:
19504 case X86::TAILJMPm64:
19505 case X86::TAILJMPd64_REX:
19506 case X86::TAILJMPr64_REX:
19507 case X86::TAILJMPm64_REX:
19508 llvm_unreachable("TAILJMP64 would not be touched here.");
19509 case X86::TCRETURNdi64:
19510 case X86::TCRETURNri64:
19511 case X86::TCRETURNmi64:
19513 case X86::WIN_ALLOCA:
19514 return EmitLoweredWinAlloca(MI, BB);
19515 case X86::SEG_ALLOCA_32:
19516 case X86::SEG_ALLOCA_64:
19517 return EmitLoweredSegAlloca(MI, BB);
19518 case X86::TLSCall_32:
19519 case X86::TLSCall_64:
19520 return EmitLoweredTLSCall(MI, BB);
19521 case X86::CMOV_GR8:
19522 case X86::CMOV_FR32:
19523 case X86::CMOV_FR64:
19524 case X86::CMOV_V4F32:
19525 case X86::CMOV_V2F64:
19526 case X86::CMOV_V2I64:
19527 case X86::CMOV_V8F32:
19528 case X86::CMOV_V4F64:
19529 case X86::CMOV_V4I64:
19530 case X86::CMOV_V16F32:
19531 case X86::CMOV_V8F64:
19532 case X86::CMOV_V8I64:
19533 case X86::CMOV_GR16:
19534 case X86::CMOV_GR32:
19535 case X86::CMOV_RFP32:
19536 case X86::CMOV_RFP64:
19537 case X86::CMOV_RFP80:
19538 case X86::CMOV_V8I1:
19539 case X86::CMOV_V16I1:
19540 case X86::CMOV_V32I1:
19541 case X86::CMOV_V64I1:
19542 return EmitLoweredSelect(MI, BB);
19544 case X86::FP32_TO_INT16_IN_MEM:
19545 case X86::FP32_TO_INT32_IN_MEM:
19546 case X86::FP32_TO_INT64_IN_MEM:
19547 case X86::FP64_TO_INT16_IN_MEM:
19548 case X86::FP64_TO_INT32_IN_MEM:
19549 case X86::FP64_TO_INT64_IN_MEM:
19550 case X86::FP80_TO_INT16_IN_MEM:
19551 case X86::FP80_TO_INT32_IN_MEM:
19552 case X86::FP80_TO_INT64_IN_MEM: {
19553 MachineFunction *F = BB->getParent();
19554 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19555 DebugLoc DL = MI->getDebugLoc();
19557 // Change the floating point control register to use "round towards zero"
19558 // mode when truncating to an integer value.
19559 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19560 addFrameReference(BuildMI(*BB, MI, DL,
19561 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19563 // Load the old value of the high byte of the control word...
19565 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
19566 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
19569 // Set the high part to be round to zero...
19570 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
19573 // Reload the modified control word now...
19574 addFrameReference(BuildMI(*BB, MI, DL,
19575 TII->get(X86::FLDCW16m)), CWFrameIdx);
19577 // Restore the memory image of control word to original value
19578 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
19581 // Get the X86 opcode to use.
19583 switch (MI->getOpcode()) {
19584 default: llvm_unreachable("illegal opcode!");
19585 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
19586 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
19587 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
19588 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
19589 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
19590 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
19591 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
19592 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
19593 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
19597 MachineOperand &Op = MI->getOperand(0);
19599 AM.BaseType = X86AddressMode::RegBase;
19600 AM.Base.Reg = Op.getReg();
19602 AM.BaseType = X86AddressMode::FrameIndexBase;
19603 AM.Base.FrameIndex = Op.getIndex();
19605 Op = MI->getOperand(1);
19607 AM.Scale = Op.getImm();
19608 Op = MI->getOperand(2);
19610 AM.IndexReg = Op.getImm();
19611 Op = MI->getOperand(3);
19612 if (Op.isGlobal()) {
19613 AM.GV = Op.getGlobal();
19615 AM.Disp = Op.getImm();
19617 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
19618 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
19620 // Reload the original control word now.
19621 addFrameReference(BuildMI(*BB, MI, DL,
19622 TII->get(X86::FLDCW16m)), CWFrameIdx);
19624 MI->eraseFromParent(); // The pseudo instruction is gone now.
19627 // String/text processing lowering.
19628 case X86::PCMPISTRM128REG:
19629 case X86::VPCMPISTRM128REG:
19630 case X86::PCMPISTRM128MEM:
19631 case X86::VPCMPISTRM128MEM:
19632 case X86::PCMPESTRM128REG:
19633 case X86::VPCMPESTRM128REG:
19634 case X86::PCMPESTRM128MEM:
19635 case X86::VPCMPESTRM128MEM:
19636 assert(Subtarget->hasSSE42() &&
19637 "Target must have SSE4.2 or AVX features enabled");
19638 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
19640 // String/text processing lowering.
19641 case X86::PCMPISTRIREG:
19642 case X86::VPCMPISTRIREG:
19643 case X86::PCMPISTRIMEM:
19644 case X86::VPCMPISTRIMEM:
19645 case X86::PCMPESTRIREG:
19646 case X86::VPCMPESTRIREG:
19647 case X86::PCMPESTRIMEM:
19648 case X86::VPCMPESTRIMEM:
19649 assert(Subtarget->hasSSE42() &&
19650 "Target must have SSE4.2 or AVX features enabled");
19651 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
19653 // Thread synchronization.
19655 return EmitMonitor(MI, BB, Subtarget);
19659 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
19661 case X86::VASTART_SAVE_XMM_REGS:
19662 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
19664 case X86::VAARG_64:
19665 return EmitVAARG64WithCustomInserter(MI, BB);
19667 case X86::EH_SjLj_SetJmp32:
19668 case X86::EH_SjLj_SetJmp64:
19669 return emitEHSjLjSetJmp(MI, BB);
19671 case X86::EH_SjLj_LongJmp32:
19672 case X86::EH_SjLj_LongJmp64:
19673 return emitEHSjLjLongJmp(MI, BB);
19675 case TargetOpcode::STATEPOINT:
19676 // As an implementation detail, STATEPOINT shares the STACKMAP format at
19677 // this point in the process. We diverge later.
19678 return emitPatchPoint(MI, BB);
19680 case TargetOpcode::STACKMAP:
19681 case TargetOpcode::PATCHPOINT:
19682 return emitPatchPoint(MI, BB);
19684 case X86::VFMADDPDr213r:
19685 case X86::VFMADDPSr213r:
19686 case X86::VFMADDSDr213r:
19687 case X86::VFMADDSSr213r:
19688 case X86::VFMSUBPDr213r:
19689 case X86::VFMSUBPSr213r:
19690 case X86::VFMSUBSDr213r:
19691 case X86::VFMSUBSSr213r:
19692 case X86::VFNMADDPDr213r:
19693 case X86::VFNMADDPSr213r:
19694 case X86::VFNMADDSDr213r:
19695 case X86::VFNMADDSSr213r:
19696 case X86::VFNMSUBPDr213r:
19697 case X86::VFNMSUBPSr213r:
19698 case X86::VFNMSUBSDr213r:
19699 case X86::VFNMSUBSSr213r:
19700 case X86::VFMADDSUBPDr213r:
19701 case X86::VFMADDSUBPSr213r:
19702 case X86::VFMSUBADDPDr213r:
19703 case X86::VFMSUBADDPSr213r:
19704 case X86::VFMADDPDr213rY:
19705 case X86::VFMADDPSr213rY:
19706 case X86::VFMSUBPDr213rY:
19707 case X86::VFMSUBPSr213rY:
19708 case X86::VFNMADDPDr213rY:
19709 case X86::VFNMADDPSr213rY:
19710 case X86::VFNMSUBPDr213rY:
19711 case X86::VFNMSUBPSr213rY:
19712 case X86::VFMADDSUBPDr213rY:
19713 case X86::VFMADDSUBPSr213rY:
19714 case X86::VFMSUBADDPDr213rY:
19715 case X86::VFMSUBADDPSr213rY:
19716 return emitFMA3Instr(MI, BB);
19720 //===----------------------------------------------------------------------===//
19721 // X86 Optimization Hooks
19722 //===----------------------------------------------------------------------===//
19724 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
19727 const SelectionDAG &DAG,
19728 unsigned Depth) const {
19729 unsigned BitWidth = KnownZero.getBitWidth();
19730 unsigned Opc = Op.getOpcode();
19731 assert((Opc >= ISD::BUILTIN_OP_END ||
19732 Opc == ISD::INTRINSIC_WO_CHAIN ||
19733 Opc == ISD::INTRINSIC_W_CHAIN ||
19734 Opc == ISD::INTRINSIC_VOID) &&
19735 "Should use MaskedValueIsZero if you don't know whether Op"
19736 " is a target node!");
19738 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
19752 // These nodes' second result is a boolean.
19753 if (Op.getResNo() == 0)
19756 case X86ISD::SETCC:
19757 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
19759 case ISD::INTRINSIC_WO_CHAIN: {
19760 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
19761 unsigned NumLoBits = 0;
19764 case Intrinsic::x86_sse_movmsk_ps:
19765 case Intrinsic::x86_avx_movmsk_ps_256:
19766 case Intrinsic::x86_sse2_movmsk_pd:
19767 case Intrinsic::x86_avx_movmsk_pd_256:
19768 case Intrinsic::x86_mmx_pmovmskb:
19769 case Intrinsic::x86_sse2_pmovmskb_128:
19770 case Intrinsic::x86_avx2_pmovmskb: {
19771 // High bits of movmskp{s|d}, pmovmskb are known zero.
19773 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
19774 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
19775 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
19776 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
19777 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
19778 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
19779 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
19780 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
19782 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
19791 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
19793 const SelectionDAG &,
19794 unsigned Depth) const {
19795 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
19796 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
19797 return Op.getValueType().getScalarType().getSizeInBits();
19803 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
19804 /// node is a GlobalAddress + offset.
19805 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
19806 const GlobalValue* &GA,
19807 int64_t &Offset) const {
19808 if (N->getOpcode() == X86ISD::Wrapper) {
19809 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
19810 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
19811 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
19815 return TargetLowering::isGAPlusOffset(N, GA, Offset);
19818 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
19819 /// same as extracting the high 128-bit part of 256-bit vector and then
19820 /// inserting the result into the low part of a new 256-bit vector
19821 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
19822 EVT VT = SVOp->getValueType(0);
19823 unsigned NumElems = VT.getVectorNumElements();
19825 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19826 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
19827 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19828 SVOp->getMaskElt(j) >= 0)
19834 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
19835 /// same as extracting the low 128-bit part of 256-bit vector and then
19836 /// inserting the result into the high part of a new 256-bit vector
19837 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
19838 EVT VT = SVOp->getValueType(0);
19839 unsigned NumElems = VT.getVectorNumElements();
19841 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19842 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
19843 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19844 SVOp->getMaskElt(j) >= 0)
19850 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
19851 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
19852 TargetLowering::DAGCombinerInfo &DCI,
19853 const X86Subtarget* Subtarget) {
19855 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19856 SDValue V1 = SVOp->getOperand(0);
19857 SDValue V2 = SVOp->getOperand(1);
19858 EVT VT = SVOp->getValueType(0);
19859 unsigned NumElems = VT.getVectorNumElements();
19861 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19862 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19866 // V UNDEF BUILD_VECTOR UNDEF
19868 // CONCAT_VECTOR CONCAT_VECTOR
19871 // RESULT: V + zero extended
19873 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19874 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19875 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19878 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19881 // To match the shuffle mask, the first half of the mask should
19882 // be exactly the first vector, and all the rest a splat with the
19883 // first element of the second one.
19884 for (unsigned i = 0; i != NumElems/2; ++i)
19885 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19886 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19889 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19890 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19891 if (Ld->hasNUsesOfValue(1, 0)) {
19892 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19893 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19895 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19897 Ld->getPointerInfo(),
19898 Ld->getAlignment(),
19899 false/*isVolatile*/, true/*ReadMem*/,
19900 false/*WriteMem*/);
19902 // Make sure the newly-created LOAD is in the same position as Ld in
19903 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19904 // and update uses of Ld's output chain to use the TokenFactor.
19905 if (Ld->hasAnyUseOfValue(1)) {
19906 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19907 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19908 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19909 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19910 SDValue(ResNode.getNode(), 1));
19913 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19917 // Emit a zeroed vector and insert the desired subvector on its
19919 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19920 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19921 return DCI.CombineTo(N, InsV);
19924 //===--------------------------------------------------------------------===//
19925 // Combine some shuffles into subvector extracts and inserts:
19928 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19929 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19930 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19931 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19932 return DCI.CombineTo(N, InsV);
19935 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19936 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19937 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19938 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19939 return DCI.CombineTo(N, InsV);
19945 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19948 /// This is the leaf of the recursive combinine below. When we have found some
19949 /// chain of single-use x86 shuffle instructions and accumulated the combined
19950 /// shuffle mask represented by them, this will try to pattern match that mask
19951 /// into either a single instruction if there is a special purpose instruction
19952 /// for this operation, or into a PSHUFB instruction which is a fully general
19953 /// instruction but should only be used to replace chains over a certain depth.
19954 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19955 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19956 TargetLowering::DAGCombinerInfo &DCI,
19957 const X86Subtarget *Subtarget) {
19958 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19960 // Find the operand that enters the chain. Note that multiple uses are OK
19961 // here, we're not going to remove the operand we find.
19962 SDValue Input = Op.getOperand(0);
19963 while (Input.getOpcode() == ISD::BITCAST)
19964 Input = Input.getOperand(0);
19966 MVT VT = Input.getSimpleValueType();
19967 MVT RootVT = Root.getSimpleValueType();
19970 // Just remove no-op shuffle masks.
19971 if (Mask.size() == 1) {
19972 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19977 // Use the float domain if the operand type is a floating point type.
19978 bool FloatDomain = VT.isFloatingPoint();
19980 // For floating point shuffles, we don't have free copies in the shuffle
19981 // instructions or the ability to load as part of the instruction, so
19982 // canonicalize their shuffles to UNPCK or MOV variants.
19984 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19985 // vectors because it can have a load folded into it that UNPCK cannot. This
19986 // doesn't preclude something switching to the shorter encoding post-RA.
19988 // FIXME: Should teach these routines about AVX vector widths.
19989 if (FloatDomain && VT.getSizeInBits() == 128) {
19990 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
19991 bool Lo = Mask.equals({0, 0});
19994 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19995 // is no slower than UNPCKLPD but has the option to fold the input operand
19996 // into even an unaligned memory load.
19997 if (Lo && Subtarget->hasSSE3()) {
19998 Shuffle = X86ISD::MOVDDUP;
19999 ShuffleVT = MVT::v2f64;
20001 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
20002 // than the UNPCK variants.
20003 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
20004 ShuffleVT = MVT::v4f32;
20006 if (Depth == 1 && Root->getOpcode() == Shuffle)
20007 return false; // Nothing to do!
20008 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20009 DCI.AddToWorklist(Op.getNode());
20010 if (Shuffle == X86ISD::MOVDDUP)
20011 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20013 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20014 DCI.AddToWorklist(Op.getNode());
20015 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20019 if (Subtarget->hasSSE3() &&
20020 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
20021 bool Lo = Mask.equals({0, 0, 2, 2});
20022 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
20023 MVT ShuffleVT = MVT::v4f32;
20024 if (Depth == 1 && Root->getOpcode() == Shuffle)
20025 return false; // Nothing to do!
20026 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20027 DCI.AddToWorklist(Op.getNode());
20028 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20029 DCI.AddToWorklist(Op.getNode());
20030 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20034 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
20035 bool Lo = Mask.equals({0, 0, 1, 1});
20036 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20037 MVT ShuffleVT = MVT::v4f32;
20038 if (Depth == 1 && Root->getOpcode() == Shuffle)
20039 return false; // Nothing to do!
20040 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20041 DCI.AddToWorklist(Op.getNode());
20042 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20043 DCI.AddToWorklist(Op.getNode());
20044 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20050 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
20051 // variants as none of these have single-instruction variants that are
20052 // superior to the UNPCK formulation.
20053 if (!FloatDomain && VT.getSizeInBits() == 128 &&
20054 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20055 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
20056 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
20058 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
20059 bool Lo = Mask[0] == 0;
20060 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20061 if (Depth == 1 && Root->getOpcode() == Shuffle)
20062 return false; // Nothing to do!
20064 switch (Mask.size()) {
20066 ShuffleVT = MVT::v8i16;
20069 ShuffleVT = MVT::v16i8;
20072 llvm_unreachable("Impossible mask size!");
20074 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20075 DCI.AddToWorklist(Op.getNode());
20076 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20077 DCI.AddToWorklist(Op.getNode());
20078 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20083 // Don't try to re-form single instruction chains under any circumstances now
20084 // that we've done encoding canonicalization for them.
20088 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
20089 // can replace them with a single PSHUFB instruction profitably. Intel's
20090 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
20091 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
20092 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
20093 SmallVector<SDValue, 16> PSHUFBMask;
20094 int NumBytes = VT.getSizeInBits() / 8;
20095 int Ratio = NumBytes / Mask.size();
20096 for (int i = 0; i < NumBytes; ++i) {
20097 if (Mask[i / Ratio] == SM_SentinelUndef) {
20098 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
20101 int M = Mask[i / Ratio] != SM_SentinelZero
20102 ? Ratio * Mask[i / Ratio] + i % Ratio
20104 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
20106 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
20107 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Input);
20108 DCI.AddToWorklist(Op.getNode());
20109 SDValue PSHUFBMaskOp =
20110 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
20111 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
20112 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
20113 DCI.AddToWorklist(Op.getNode());
20114 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20119 // Failed to find any combines.
20123 /// \brief Fully generic combining of x86 shuffle instructions.
20125 /// This should be the last combine run over the x86 shuffle instructions. Once
20126 /// they have been fully optimized, this will recursively consider all chains
20127 /// of single-use shuffle instructions, build a generic model of the cumulative
20128 /// shuffle operation, and check for simpler instructions which implement this
20129 /// operation. We use this primarily for two purposes:
20131 /// 1) Collapse generic shuffles to specialized single instructions when
20132 /// equivalent. In most cases, this is just an encoding size win, but
20133 /// sometimes we will collapse multiple generic shuffles into a single
20134 /// special-purpose shuffle.
20135 /// 2) Look for sequences of shuffle instructions with 3 or more total
20136 /// instructions, and replace them with the slightly more expensive SSSE3
20137 /// PSHUFB instruction if available. We do this as the last combining step
20138 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
20139 /// a suitable short sequence of other instructions. The PHUFB will either
20140 /// use a register or have to read from memory and so is slightly (but only
20141 /// slightly) more expensive than the other shuffle instructions.
20143 /// Because this is inherently a quadratic operation (for each shuffle in
20144 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
20145 /// This should never be an issue in practice as the shuffle lowering doesn't
20146 /// produce sequences of more than 8 instructions.
20148 /// FIXME: We will currently miss some cases where the redundant shuffling
20149 /// would simplify under the threshold for PSHUFB formation because of
20150 /// combine-ordering. To fix this, we should do the redundant instruction
20151 /// combining in this recursive walk.
20152 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
20153 ArrayRef<int> RootMask,
20154 int Depth, bool HasPSHUFB,
20156 TargetLowering::DAGCombinerInfo &DCI,
20157 const X86Subtarget *Subtarget) {
20158 // Bound the depth of our recursive combine because this is ultimately
20159 // quadratic in nature.
20163 // Directly rip through bitcasts to find the underlying operand.
20164 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
20165 Op = Op.getOperand(0);
20167 MVT VT = Op.getSimpleValueType();
20168 if (!VT.isVector())
20169 return false; // Bail if we hit a non-vector.
20171 assert(Root.getSimpleValueType().isVector() &&
20172 "Shuffles operate on vector types!");
20173 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
20174 "Can only combine shuffles of the same vector register size.");
20176 if (!isTargetShuffle(Op.getOpcode()))
20178 SmallVector<int, 16> OpMask;
20180 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
20181 // We only can combine unary shuffles which we can decode the mask for.
20182 if (!HaveMask || !IsUnary)
20185 assert(VT.getVectorNumElements() == OpMask.size() &&
20186 "Different mask size from vector size!");
20187 assert(((RootMask.size() > OpMask.size() &&
20188 RootMask.size() % OpMask.size() == 0) ||
20189 (OpMask.size() > RootMask.size() &&
20190 OpMask.size() % RootMask.size() == 0) ||
20191 OpMask.size() == RootMask.size()) &&
20192 "The smaller number of elements must divide the larger.");
20193 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
20194 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
20195 assert(((RootRatio == 1 && OpRatio == 1) ||
20196 (RootRatio == 1) != (OpRatio == 1)) &&
20197 "Must not have a ratio for both incoming and op masks!");
20199 SmallVector<int, 16> Mask;
20200 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
20202 // Merge this shuffle operation's mask into our accumulated mask. Note that
20203 // this shuffle's mask will be the first applied to the input, followed by the
20204 // root mask to get us all the way to the root value arrangement. The reason
20205 // for this order is that we are recursing up the operation chain.
20206 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
20207 int RootIdx = i / RootRatio;
20208 if (RootMask[RootIdx] < 0) {
20209 // This is a zero or undef lane, we're done.
20210 Mask.push_back(RootMask[RootIdx]);
20214 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
20215 int OpIdx = RootMaskedIdx / OpRatio;
20216 if (OpMask[OpIdx] < 0) {
20217 // The incoming lanes are zero or undef, it doesn't matter which ones we
20219 Mask.push_back(OpMask[OpIdx]);
20223 // Ok, we have non-zero lanes, map them through.
20224 Mask.push_back(OpMask[OpIdx] * OpRatio +
20225 RootMaskedIdx % OpRatio);
20228 // See if we can recurse into the operand to combine more things.
20229 switch (Op.getOpcode()) {
20230 case X86ISD::PSHUFB:
20232 case X86ISD::PSHUFD:
20233 case X86ISD::PSHUFHW:
20234 case X86ISD::PSHUFLW:
20235 if (Op.getOperand(0).hasOneUse() &&
20236 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20237 HasPSHUFB, DAG, DCI, Subtarget))
20241 case X86ISD::UNPCKL:
20242 case X86ISD::UNPCKH:
20243 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
20244 // We can't check for single use, we have to check that this shuffle is the only user.
20245 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
20246 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20247 HasPSHUFB, DAG, DCI, Subtarget))
20252 // Minor canonicalization of the accumulated shuffle mask to make it easier
20253 // to match below. All this does is detect masks with squential pairs of
20254 // elements, and shrink them to the half-width mask. It does this in a loop
20255 // so it will reduce the size of the mask to the minimal width mask which
20256 // performs an equivalent shuffle.
20257 SmallVector<int, 16> WidenedMask;
20258 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
20259 Mask = std::move(WidenedMask);
20260 WidenedMask.clear();
20263 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20267 /// \brief Get the PSHUF-style mask from PSHUF node.
20269 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20270 /// PSHUF-style masks that can be reused with such instructions.
20271 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20272 MVT VT = N.getSimpleValueType();
20273 SmallVector<int, 4> Mask;
20275 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
20279 // If we have more than 128-bits, only the low 128-bits of shuffle mask
20280 // matter. Check that the upper masks are repeats and remove them.
20281 if (VT.getSizeInBits() > 128) {
20282 int LaneElts = 128 / VT.getScalarSizeInBits();
20284 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
20285 for (int j = 0; j < LaneElts; ++j)
20286 assert(Mask[j] == Mask[i * LaneElts + j] - LaneElts &&
20287 "Mask doesn't repeat in high 128-bit lanes!");
20289 Mask.resize(LaneElts);
20292 switch (N.getOpcode()) {
20293 case X86ISD::PSHUFD:
20295 case X86ISD::PSHUFLW:
20298 case X86ISD::PSHUFHW:
20299 Mask.erase(Mask.begin(), Mask.begin() + 4);
20300 for (int &M : Mask)
20304 llvm_unreachable("No valid shuffle instruction found!");
20308 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20310 /// We walk up the chain and look for a combinable shuffle, skipping over
20311 /// shuffles that we could hoist this shuffle's transformation past without
20312 /// altering anything.
20314 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20316 TargetLowering::DAGCombinerInfo &DCI) {
20317 assert(N.getOpcode() == X86ISD::PSHUFD &&
20318 "Called with something other than an x86 128-bit half shuffle!");
20321 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20322 // of the shuffles in the chain so that we can form a fresh chain to replace
20324 SmallVector<SDValue, 8> Chain;
20325 SDValue V = N.getOperand(0);
20326 for (; V.hasOneUse(); V = V.getOperand(0)) {
20327 switch (V.getOpcode()) {
20329 return SDValue(); // Nothing combined!
20332 // Skip bitcasts as we always know the type for the target specific
20336 case X86ISD::PSHUFD:
20337 // Found another dword shuffle.
20340 case X86ISD::PSHUFLW:
20341 // Check that the low words (being shuffled) are the identity in the
20342 // dword shuffle, and the high words are self-contained.
20343 if (Mask[0] != 0 || Mask[1] != 1 ||
20344 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20347 Chain.push_back(V);
20350 case X86ISD::PSHUFHW:
20351 // Check that the high words (being shuffled) are the identity in the
20352 // dword shuffle, and the low words are self-contained.
20353 if (Mask[2] != 2 || Mask[3] != 3 ||
20354 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20357 Chain.push_back(V);
20360 case X86ISD::UNPCKL:
20361 case X86ISD::UNPCKH:
20362 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20363 // shuffle into a preceding word shuffle.
20364 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
20365 V.getSimpleValueType().getScalarType() != MVT::i16)
20368 // Search for a half-shuffle which we can combine with.
20369 unsigned CombineOp =
20370 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20371 if (V.getOperand(0) != V.getOperand(1) ||
20372 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20374 Chain.push_back(V);
20375 V = V.getOperand(0);
20377 switch (V.getOpcode()) {
20379 return SDValue(); // Nothing to combine.
20381 case X86ISD::PSHUFLW:
20382 case X86ISD::PSHUFHW:
20383 if (V.getOpcode() == CombineOp)
20386 Chain.push_back(V);
20390 V = V.getOperand(0);
20394 } while (V.hasOneUse());
20397 // Break out of the loop if we break out of the switch.
20401 if (!V.hasOneUse())
20402 // We fell out of the loop without finding a viable combining instruction.
20405 // Merge this node's mask and our incoming mask.
20406 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20407 for (int &M : Mask)
20409 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20410 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20412 // Rebuild the chain around this new shuffle.
20413 while (!Chain.empty()) {
20414 SDValue W = Chain.pop_back_val();
20416 if (V.getValueType() != W.getOperand(0).getValueType())
20417 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
20419 switch (W.getOpcode()) {
20421 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20423 case X86ISD::UNPCKL:
20424 case X86ISD::UNPCKH:
20425 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20428 case X86ISD::PSHUFD:
20429 case X86ISD::PSHUFLW:
20430 case X86ISD::PSHUFHW:
20431 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20435 if (V.getValueType() != N.getValueType())
20436 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
20438 // Return the new chain to replace N.
20442 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20444 /// We walk up the chain, skipping shuffles of the other half and looking
20445 /// through shuffles which switch halves trying to find a shuffle of the same
20446 /// pair of dwords.
20447 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20449 TargetLowering::DAGCombinerInfo &DCI) {
20451 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20452 "Called with something other than an x86 128-bit half shuffle!");
20454 unsigned CombineOpcode = N.getOpcode();
20456 // Walk up a single-use chain looking for a combinable shuffle.
20457 SDValue V = N.getOperand(0);
20458 for (; V.hasOneUse(); V = V.getOperand(0)) {
20459 switch (V.getOpcode()) {
20461 return false; // Nothing combined!
20464 // Skip bitcasts as we always know the type for the target specific
20468 case X86ISD::PSHUFLW:
20469 case X86ISD::PSHUFHW:
20470 if (V.getOpcode() == CombineOpcode)
20473 // Other-half shuffles are no-ops.
20476 // Break out of the loop if we break out of the switch.
20480 if (!V.hasOneUse())
20481 // We fell out of the loop without finding a viable combining instruction.
20484 // Combine away the bottom node as its shuffle will be accumulated into
20485 // a preceding shuffle.
20486 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20488 // Record the old value.
20491 // Merge this node's mask and our incoming mask (adjusted to account for all
20492 // the pshufd instructions encountered).
20493 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20494 for (int &M : Mask)
20496 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20497 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20499 // Check that the shuffles didn't cancel each other out. If not, we need to
20500 // combine to the new one.
20502 // Replace the combinable shuffle with the combined one, updating all users
20503 // so that we re-evaluate the chain here.
20504 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20509 /// \brief Try to combine x86 target specific shuffles.
20510 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20511 TargetLowering::DAGCombinerInfo &DCI,
20512 const X86Subtarget *Subtarget) {
20514 MVT VT = N.getSimpleValueType();
20515 SmallVector<int, 4> Mask;
20517 switch (N.getOpcode()) {
20518 case X86ISD::PSHUFD:
20519 case X86ISD::PSHUFLW:
20520 case X86ISD::PSHUFHW:
20521 Mask = getPSHUFShuffleMask(N);
20522 assert(Mask.size() == 4);
20528 // Nuke no-op shuffles that show up after combining.
20529 if (isNoopShuffleMask(Mask))
20530 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20532 // Look for simplifications involving one or two shuffle instructions.
20533 SDValue V = N.getOperand(0);
20534 switch (N.getOpcode()) {
20537 case X86ISD::PSHUFLW:
20538 case X86ISD::PSHUFHW:
20539 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
20541 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20542 return SDValue(); // We combined away this shuffle, so we're done.
20544 // See if this reduces to a PSHUFD which is no more expensive and can
20545 // combine with more operations. Note that it has to at least flip the
20546 // dwords as otherwise it would have been removed as a no-op.
20547 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
20548 int DMask[] = {0, 1, 2, 3};
20549 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20550 DMask[DOffset + 0] = DOffset + 1;
20551 DMask[DOffset + 1] = DOffset + 0;
20552 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
20553 V = DAG.getNode(ISD::BITCAST, DL, DVT, V);
20554 DCI.AddToWorklist(V.getNode());
20555 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
20556 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
20557 DCI.AddToWorklist(V.getNode());
20558 return DAG.getNode(ISD::BITCAST, DL, VT, V);
20561 // Look for shuffle patterns which can be implemented as a single unpack.
20562 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20563 // only works when we have a PSHUFD followed by two half-shuffles.
20564 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20565 (V.getOpcode() == X86ISD::PSHUFLW ||
20566 V.getOpcode() == X86ISD::PSHUFHW) &&
20567 V.getOpcode() != N.getOpcode() &&
20569 SDValue D = V.getOperand(0);
20570 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20571 D = D.getOperand(0);
20572 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20573 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20574 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20575 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20576 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20578 for (int i = 0; i < 4; ++i) {
20579 WordMask[i + NOffset] = Mask[i] + NOffset;
20580 WordMask[i + VOffset] = VMask[i] + VOffset;
20582 // Map the word mask through the DWord mask.
20584 for (int i = 0; i < 8; ++i)
20585 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20586 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20587 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
20588 // We can replace all three shuffles with an unpack.
20589 V = DAG.getNode(ISD::BITCAST, DL, VT, D.getOperand(0));
20590 DCI.AddToWorklist(V.getNode());
20591 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
20600 case X86ISD::PSHUFD:
20601 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
20610 /// \brief Try to combine a shuffle into a target-specific add-sub node.
20612 /// We combine this directly on the abstract vector shuffle nodes so it is
20613 /// easier to generically match. We also insert dummy vector shuffle nodes for
20614 /// the operands which explicitly discard the lanes which are unused by this
20615 /// operation to try to flow through the rest of the combiner the fact that
20616 /// they're unused.
20617 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
20619 EVT VT = N->getValueType(0);
20621 // We only handle target-independent shuffles.
20622 // FIXME: It would be easy and harmless to use the target shuffle mask
20623 // extraction tool to support more.
20624 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
20627 auto *SVN = cast<ShuffleVectorSDNode>(N);
20628 ArrayRef<int> Mask = SVN->getMask();
20629 SDValue V1 = N->getOperand(0);
20630 SDValue V2 = N->getOperand(1);
20632 // We require the first shuffle operand to be the SUB node, and the second to
20633 // be the ADD node.
20634 // FIXME: We should support the commuted patterns.
20635 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
20638 // If there are other uses of these operations we can't fold them.
20639 if (!V1->hasOneUse() || !V2->hasOneUse())
20642 // Ensure that both operations have the same operands. Note that we can
20643 // commute the FADD operands.
20644 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
20645 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
20646 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
20649 // We're looking for blends between FADD and FSUB nodes. We insist on these
20650 // nodes being lined up in a specific expected pattern.
20651 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
20652 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
20653 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
20656 // Only specific types are legal at this point, assert so we notice if and
20657 // when these change.
20658 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
20659 VT == MVT::v4f64) &&
20660 "Unknown vector type encountered!");
20662 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
20665 /// PerformShuffleCombine - Performs several different shuffle combines.
20666 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
20667 TargetLowering::DAGCombinerInfo &DCI,
20668 const X86Subtarget *Subtarget) {
20670 SDValue N0 = N->getOperand(0);
20671 SDValue N1 = N->getOperand(1);
20672 EVT VT = N->getValueType(0);
20674 // Don't create instructions with illegal types after legalize types has run.
20675 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20676 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
20679 // If we have legalized the vector types, look for blends of FADD and FSUB
20680 // nodes that we can fuse into an ADDSUB node.
20681 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
20682 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
20685 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
20686 if (Subtarget->hasFp256() && VT.is256BitVector() &&
20687 N->getOpcode() == ISD::VECTOR_SHUFFLE)
20688 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
20690 // During Type Legalization, when promoting illegal vector types,
20691 // the backend might introduce new shuffle dag nodes and bitcasts.
20693 // This code performs the following transformation:
20694 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
20695 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
20697 // We do this only if both the bitcast and the BINOP dag nodes have
20698 // one use. Also, perform this transformation only if the new binary
20699 // operation is legal. This is to avoid introducing dag nodes that
20700 // potentially need to be further expanded (or custom lowered) into a
20701 // less optimal sequence of dag nodes.
20702 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
20703 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
20704 N0.getOpcode() == ISD::BITCAST) {
20705 SDValue BC0 = N0.getOperand(0);
20706 EVT SVT = BC0.getValueType();
20707 unsigned Opcode = BC0.getOpcode();
20708 unsigned NumElts = VT.getVectorNumElements();
20710 if (BC0.hasOneUse() && SVT.isVector() &&
20711 SVT.getVectorNumElements() * 2 == NumElts &&
20712 TLI.isOperationLegal(Opcode, VT)) {
20713 bool CanFold = false;
20725 unsigned SVTNumElts = SVT.getVectorNumElements();
20726 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20727 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
20728 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
20729 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
20730 CanFold = SVOp->getMaskElt(i) < 0;
20733 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
20734 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
20735 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
20736 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
20741 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
20742 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
20743 // consecutive, non-overlapping, and in the right order.
20744 SmallVector<SDValue, 16> Elts;
20745 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
20746 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
20748 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
20752 if (isTargetShuffle(N->getOpcode())) {
20754 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
20755 if (Shuffle.getNode())
20758 // Try recursively combining arbitrary sequences of x86 shuffle
20759 // instructions into higher-order shuffles. We do this after combining
20760 // specific PSHUF instruction sequences into their minimal form so that we
20761 // can evaluate how many specialized shuffle instructions are involved in
20762 // a particular chain.
20763 SmallVector<int, 1> NonceMask; // Just a placeholder.
20764 NonceMask.push_back(0);
20765 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
20766 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
20768 return SDValue(); // This routine will use CombineTo to replace N.
20774 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
20775 /// specific shuffle of a load can be folded into a single element load.
20776 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
20777 /// shuffles have been custom lowered so we need to handle those here.
20778 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
20779 TargetLowering::DAGCombinerInfo &DCI) {
20780 if (DCI.isBeforeLegalizeOps())
20783 SDValue InVec = N->getOperand(0);
20784 SDValue EltNo = N->getOperand(1);
20786 if (!isa<ConstantSDNode>(EltNo))
20789 EVT OriginalVT = InVec.getValueType();
20791 if (InVec.getOpcode() == ISD::BITCAST) {
20792 // Don't duplicate a load with other uses.
20793 if (!InVec.hasOneUse())
20795 EVT BCVT = InVec.getOperand(0).getValueType();
20796 if (!BCVT.isVector() ||
20797 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
20799 InVec = InVec.getOperand(0);
20802 EVT CurrentVT = InVec.getValueType();
20804 if (!isTargetShuffle(InVec.getOpcode()))
20807 // Don't duplicate a load with other uses.
20808 if (!InVec.hasOneUse())
20811 SmallVector<int, 16> ShuffleMask;
20813 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
20814 ShuffleMask, UnaryShuffle))
20817 // Select the input vector, guarding against out of range extract vector.
20818 unsigned NumElems = CurrentVT.getVectorNumElements();
20819 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
20820 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
20821 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
20822 : InVec.getOperand(1);
20824 // If inputs to shuffle are the same for both ops, then allow 2 uses
20825 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
20826 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
20828 if (LdNode.getOpcode() == ISD::BITCAST) {
20829 // Don't duplicate a load with other uses.
20830 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
20833 AllowedUses = 1; // only allow 1 load use if we have a bitcast
20834 LdNode = LdNode.getOperand(0);
20837 if (!ISD::isNormalLoad(LdNode.getNode()))
20840 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
20842 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
20845 EVT EltVT = N->getValueType(0);
20846 // If there's a bitcast before the shuffle, check if the load type and
20847 // alignment is valid.
20848 unsigned Align = LN0->getAlignment();
20849 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20850 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
20851 EltVT.getTypeForEVT(*DAG.getContext()));
20853 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
20856 // All checks match so transform back to vector_shuffle so that DAG combiner
20857 // can finish the job
20860 // Create shuffle node taking into account the case that its a unary shuffle
20861 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20862 : InVec.getOperand(1);
20863 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20864 InVec.getOperand(0), Shuffle,
20866 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20867 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20871 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20872 /// special and don't usually play with other vector types, it's better to
20873 /// handle them early to be sure we emit efficient code by avoiding
20874 /// store-load conversions.
20875 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20876 if (N->getValueType(0) != MVT::x86mmx ||
20877 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20878 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20881 SDValue V = N->getOperand(0);
20882 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20883 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20884 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20885 N->getValueType(0), V.getOperand(0));
20890 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20891 /// generation and convert it from being a bunch of shuffles and extracts
20892 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20893 /// storing the value and loading scalars back, while for x64 we should
20894 /// use 64-bit extracts and shifts.
20895 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20896 TargetLowering::DAGCombinerInfo &DCI) {
20897 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20898 if (NewOp.getNode())
20901 SDValue InputVector = N->getOperand(0);
20902 SDLoc dl(InputVector);
20903 // Detect mmx to i32 conversion through a v2i32 elt extract.
20904 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20905 N->getValueType(0) == MVT::i32 &&
20906 InputVector.getValueType() == MVT::v2i32) {
20908 // The bitcast source is a direct mmx result.
20909 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20910 if (MMXSrc.getValueType() == MVT::x86mmx)
20911 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20912 N->getValueType(0),
20913 InputVector.getNode()->getOperand(0));
20915 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20916 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20917 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20918 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20919 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20920 MMXSrcOp.getValueType() == MVT::v1i64 &&
20921 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20922 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20923 N->getValueType(0),
20924 MMXSrcOp.getOperand(0));
20927 EVT VT = N->getValueType(0);
20929 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
20930 InputVector.getOpcode() == ISD::BITCAST &&
20931 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
20932 uint64_t ExtractedElt =
20933 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20934 uint64_t InputValue =
20935 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
20936 uint64_t Res = (InputValue >> ExtractedElt) & 1;
20937 return DAG.getConstant(Res, dl, MVT::i1);
20939 // Only operate on vectors of 4 elements, where the alternative shuffling
20940 // gets to be more expensive.
20941 if (InputVector.getValueType() != MVT::v4i32)
20944 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20945 // single use which is a sign-extend or zero-extend, and all elements are
20947 SmallVector<SDNode *, 4> Uses;
20948 unsigned ExtractedElements = 0;
20949 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20950 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20951 if (UI.getUse().getResNo() != InputVector.getResNo())
20954 SDNode *Extract = *UI;
20955 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20958 if (Extract->getValueType(0) != MVT::i32)
20960 if (!Extract->hasOneUse())
20962 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20963 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20965 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20968 // Record which element was extracted.
20969 ExtractedElements |=
20970 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20972 Uses.push_back(Extract);
20975 // If not all the elements were used, this may not be worthwhile.
20976 if (ExtractedElements != 15)
20979 // Ok, we've now decided to do the transformation.
20980 // If 64-bit shifts are legal, use the extract-shift sequence,
20981 // otherwise bounce the vector off the cache.
20982 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20985 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20986 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20987 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20988 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20989 DAG.getConstant(0, dl, VecIdxTy));
20990 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20991 DAG.getConstant(1, dl, VecIdxTy));
20993 SDValue ShAmt = DAG.getConstant(32, dl,
20994 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20995 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20996 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20997 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20998 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20999 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
21000 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
21002 // Store the value to a temporary stack slot.
21003 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
21004 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
21005 MachinePointerInfo(), false, false, 0);
21007 EVT ElementType = InputVector.getValueType().getVectorElementType();
21008 unsigned EltSize = ElementType.getSizeInBits() / 8;
21010 // Replace each use (extract) with a load of the appropriate element.
21011 for (unsigned i = 0; i < 4; ++i) {
21012 uint64_t Offset = EltSize * i;
21013 SDValue OffsetVal = DAG.getConstant(Offset, dl, TLI.getPointerTy());
21015 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
21016 StackPtr, OffsetVal);
21018 // Load the scalar.
21019 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
21020 ScalarAddr, MachinePointerInfo(),
21021 false, false, false, 0);
21026 // Replace the extracts
21027 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
21028 UE = Uses.end(); UI != UE; ++UI) {
21029 SDNode *Extract = *UI;
21031 SDValue Idx = Extract->getOperand(1);
21032 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
21033 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
21036 // The replacement was made in place; don't return anything.
21040 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
21041 static std::pair<unsigned, bool>
21042 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
21043 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
21044 if (!VT.isVector())
21045 return std::make_pair(0, false);
21047 bool NeedSplit = false;
21048 switch (VT.getSimpleVT().SimpleTy) {
21049 default: return std::make_pair(0, false);
21052 if (!Subtarget->hasVLX())
21053 return std::make_pair(0, false);
21057 if (!Subtarget->hasBWI())
21058 return std::make_pair(0, false);
21062 if (!Subtarget->hasAVX512())
21063 return std::make_pair(0, false);
21068 if (!Subtarget->hasAVX2())
21070 if (!Subtarget->hasAVX())
21071 return std::make_pair(0, false);
21076 if (!Subtarget->hasSSE2())
21077 return std::make_pair(0, false);
21080 // SSE2 has only a small subset of the operations.
21081 bool hasUnsigned = Subtarget->hasSSE41() ||
21082 (Subtarget->hasSSE2() && VT == MVT::v16i8);
21083 bool hasSigned = Subtarget->hasSSE41() ||
21084 (Subtarget->hasSSE2() && VT == MVT::v8i16);
21086 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21089 // Check for x CC y ? x : y.
21090 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21091 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21096 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21099 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21102 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21105 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21107 // Check for x CC y ? y : x -- a min/max with reversed arms.
21108 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21109 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21114 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21117 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21120 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21123 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21127 return std::make_pair(Opc, NeedSplit);
21131 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
21132 const X86Subtarget *Subtarget) {
21134 SDValue Cond = N->getOperand(0);
21135 SDValue LHS = N->getOperand(1);
21136 SDValue RHS = N->getOperand(2);
21138 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
21139 SDValue CondSrc = Cond->getOperand(0);
21140 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
21141 Cond = CondSrc->getOperand(0);
21144 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
21147 // A vselect where all conditions and data are constants can be optimized into
21148 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
21149 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
21150 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
21153 unsigned MaskValue = 0;
21154 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
21157 MVT VT = N->getSimpleValueType(0);
21158 unsigned NumElems = VT.getVectorNumElements();
21159 SmallVector<int, 8> ShuffleMask(NumElems, -1);
21160 for (unsigned i = 0; i < NumElems; ++i) {
21161 // Be sure we emit undef where we can.
21162 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
21163 ShuffleMask[i] = -1;
21165 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
21168 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21169 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
21171 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
21174 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
21176 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
21177 TargetLowering::DAGCombinerInfo &DCI,
21178 const X86Subtarget *Subtarget) {
21180 SDValue Cond = N->getOperand(0);
21181 // Get the LHS/RHS of the select.
21182 SDValue LHS = N->getOperand(1);
21183 SDValue RHS = N->getOperand(2);
21184 EVT VT = LHS.getValueType();
21185 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21187 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
21188 // instructions match the semantics of the common C idiom x<y?x:y but not
21189 // x<=y?x:y, because of how they handle negative zero (which can be
21190 // ignored in unsafe-math mode).
21191 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
21192 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
21193 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
21194 (Subtarget->hasSSE2() ||
21195 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
21196 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21198 unsigned Opcode = 0;
21199 // Check for x CC y ? x : y.
21200 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21201 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21205 // Converting this to a min would handle NaNs incorrectly, and swapping
21206 // the operands would cause it to handle comparisons between positive
21207 // and negative zero incorrectly.
21208 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21209 if (!DAG.getTarget().Options.UnsafeFPMath &&
21210 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21212 std::swap(LHS, RHS);
21214 Opcode = X86ISD::FMIN;
21217 // Converting this to a min would handle comparisons between positive
21218 // and negative zero incorrectly.
21219 if (!DAG.getTarget().Options.UnsafeFPMath &&
21220 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21222 Opcode = X86ISD::FMIN;
21225 // Converting this to a min would handle both negative zeros and NaNs
21226 // incorrectly, but we can swap the operands to fix both.
21227 std::swap(LHS, RHS);
21231 Opcode = X86ISD::FMIN;
21235 // Converting this to a max would handle comparisons between positive
21236 // and negative zero incorrectly.
21237 if (!DAG.getTarget().Options.UnsafeFPMath &&
21238 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21240 Opcode = X86ISD::FMAX;
21243 // Converting this to a max would handle NaNs incorrectly, and swapping
21244 // the operands would cause it to handle comparisons between positive
21245 // and negative zero incorrectly.
21246 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21247 if (!DAG.getTarget().Options.UnsafeFPMath &&
21248 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21250 std::swap(LHS, RHS);
21252 Opcode = X86ISD::FMAX;
21255 // Converting this to a max would handle both negative zeros and NaNs
21256 // incorrectly, but we can swap the operands to fix both.
21257 std::swap(LHS, RHS);
21261 Opcode = X86ISD::FMAX;
21264 // Check for x CC y ? y : x -- a min/max with reversed arms.
21265 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21266 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21270 // Converting this to a min would handle comparisons between positive
21271 // and negative zero incorrectly, and swapping the operands would
21272 // cause it to handle NaNs incorrectly.
21273 if (!DAG.getTarget().Options.UnsafeFPMath &&
21274 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21275 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21277 std::swap(LHS, RHS);
21279 Opcode = X86ISD::FMIN;
21282 // Converting this to a min would handle NaNs incorrectly.
21283 if (!DAG.getTarget().Options.UnsafeFPMath &&
21284 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21286 Opcode = X86ISD::FMIN;
21289 // Converting this to a min would handle both negative zeros and NaNs
21290 // incorrectly, but we can swap the operands to fix both.
21291 std::swap(LHS, RHS);
21295 Opcode = X86ISD::FMIN;
21299 // Converting this to a max would handle NaNs incorrectly.
21300 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21302 Opcode = X86ISD::FMAX;
21305 // Converting this to a max would handle comparisons between positive
21306 // and negative zero incorrectly, and swapping the operands would
21307 // cause it to handle NaNs incorrectly.
21308 if (!DAG.getTarget().Options.UnsafeFPMath &&
21309 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21310 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21312 std::swap(LHS, RHS);
21314 Opcode = X86ISD::FMAX;
21317 // Converting this to a max would handle both negative zeros and NaNs
21318 // incorrectly, but we can swap the operands to fix both.
21319 std::swap(LHS, RHS);
21323 Opcode = X86ISD::FMAX;
21329 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21332 EVT CondVT = Cond.getValueType();
21333 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21334 CondVT.getVectorElementType() == MVT::i1) {
21335 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21336 // lowering on KNL. In this case we convert it to
21337 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21338 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21339 // Since SKX these selects have a proper lowering.
21340 EVT OpVT = LHS.getValueType();
21341 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21342 (OpVT.getVectorElementType() == MVT::i8 ||
21343 OpVT.getVectorElementType() == MVT::i16) &&
21344 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21345 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21346 DCI.AddToWorklist(Cond.getNode());
21347 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21350 // If this is a select between two integer constants, try to do some
21352 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21353 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21354 // Don't do this for crazy integer types.
21355 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21356 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21357 // so that TrueC (the true value) is larger than FalseC.
21358 bool NeedsCondInvert = false;
21360 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21361 // Efficiently invertible.
21362 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21363 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21364 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21365 NeedsCondInvert = true;
21366 std::swap(TrueC, FalseC);
21369 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21370 if (FalseC->getAPIntValue() == 0 &&
21371 TrueC->getAPIntValue().isPowerOf2()) {
21372 if (NeedsCondInvert) // Invert the condition if needed.
21373 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21374 DAG.getConstant(1, DL, Cond.getValueType()));
21376 // Zero extend the condition if needed.
21377 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21379 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21380 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21381 DAG.getConstant(ShAmt, DL, MVT::i8));
21384 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21385 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21386 if (NeedsCondInvert) // Invert the condition if needed.
21387 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21388 DAG.getConstant(1, DL, Cond.getValueType()));
21390 // Zero extend the condition if needed.
21391 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21392 FalseC->getValueType(0), Cond);
21393 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21394 SDValue(FalseC, 0));
21397 // Optimize cases that will turn into an LEA instruction. This requires
21398 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21399 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21400 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21401 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21403 bool isFastMultiplier = false;
21405 switch ((unsigned char)Diff) {
21407 case 1: // result = add base, cond
21408 case 2: // result = lea base( , cond*2)
21409 case 3: // result = lea base(cond, cond*2)
21410 case 4: // result = lea base( , cond*4)
21411 case 5: // result = lea base(cond, cond*4)
21412 case 8: // result = lea base( , cond*8)
21413 case 9: // result = lea base(cond, cond*8)
21414 isFastMultiplier = true;
21419 if (isFastMultiplier) {
21420 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21421 if (NeedsCondInvert) // Invert the condition if needed.
21422 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21423 DAG.getConstant(1, DL, Cond.getValueType()));
21425 // Zero extend the condition if needed.
21426 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21428 // Scale the condition by the difference.
21430 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21431 DAG.getConstant(Diff, DL,
21432 Cond.getValueType()));
21434 // Add the base if non-zero.
21435 if (FalseC->getAPIntValue() != 0)
21436 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21437 SDValue(FalseC, 0));
21444 // Canonicalize max and min:
21445 // (x > y) ? x : y -> (x >= y) ? x : y
21446 // (x < y) ? x : y -> (x <= y) ? x : y
21447 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21448 // the need for an extra compare
21449 // against zero. e.g.
21450 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21452 // testl %edi, %edi
21454 // cmovgl %edi, %eax
21458 // cmovsl %eax, %edi
21459 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21460 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21461 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21462 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21467 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21468 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21469 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21470 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21475 // Early exit check
21476 if (!TLI.isTypeLegal(VT))
21479 // Match VSELECTs into subs with unsigned saturation.
21480 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21481 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21482 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21483 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21484 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21486 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21487 // left side invert the predicate to simplify logic below.
21489 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21491 CC = ISD::getSetCCInverse(CC, true);
21492 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21496 if (Other.getNode() && Other->getNumOperands() == 2 &&
21497 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21498 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21499 SDValue CondRHS = Cond->getOperand(1);
21501 // Look for a general sub with unsigned saturation first.
21502 // x >= y ? x-y : 0 --> subus x, y
21503 // x > y ? x-y : 0 --> subus x, y
21504 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21505 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21506 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21508 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21509 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21510 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21511 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21512 // If the RHS is a constant we have to reverse the const
21513 // canonicalization.
21514 // x > C-1 ? x+-C : 0 --> subus x, C
21515 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21516 CondRHSConst->getAPIntValue() ==
21517 (-OpRHSConst->getAPIntValue() - 1))
21518 return DAG.getNode(
21519 X86ISD::SUBUS, DL, VT, OpLHS,
21520 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
21522 // Another special case: If C was a sign bit, the sub has been
21523 // canonicalized into a xor.
21524 // FIXME: Would it be better to use computeKnownBits to determine
21525 // whether it's safe to decanonicalize the xor?
21526 // x s< 0 ? x^C : 0 --> subus x, C
21527 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21528 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21529 OpRHSConst->getAPIntValue().isSignBit())
21530 // Note that we have to rebuild the RHS constant here to ensure we
21531 // don't rely on particular values of undef lanes.
21532 return DAG.getNode(
21533 X86ISD::SUBUS, DL, VT, OpLHS,
21534 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
21539 // Try to match a min/max vector operation.
21540 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21541 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21542 unsigned Opc = ret.first;
21543 bool NeedSplit = ret.second;
21545 if (Opc && NeedSplit) {
21546 unsigned NumElems = VT.getVectorNumElements();
21547 // Extract the LHS vectors
21548 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21549 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21551 // Extract the RHS vectors
21552 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21553 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21555 // Create min/max for each subvector
21556 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21557 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21559 // Merge the result
21560 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21562 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21565 // Simplify vector selection if condition value type matches vselect
21567 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
21568 assert(Cond.getValueType().isVector() &&
21569 "vector select expects a vector selector!");
21571 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21572 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21574 // Try invert the condition if true value is not all 1s and false value
21576 if (!TValIsAllOnes && !FValIsAllZeros &&
21577 // Check if the selector will be produced by CMPP*/PCMP*
21578 Cond.getOpcode() == ISD::SETCC &&
21579 // Check if SETCC has already been promoted
21580 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
21581 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21582 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21584 if (TValIsAllZeros || FValIsAllOnes) {
21585 SDValue CC = Cond.getOperand(2);
21586 ISD::CondCode NewCC =
21587 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21588 Cond.getOperand(0).getValueType().isInteger());
21589 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21590 std::swap(LHS, RHS);
21591 TValIsAllOnes = FValIsAllOnes;
21592 FValIsAllZeros = TValIsAllZeros;
21596 if (TValIsAllOnes || FValIsAllZeros) {
21599 if (TValIsAllOnes && FValIsAllZeros)
21601 else if (TValIsAllOnes)
21602 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
21603 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
21604 else if (FValIsAllZeros)
21605 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21606 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
21608 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
21612 // We should generate an X86ISD::BLENDI from a vselect if its argument
21613 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
21614 // constants. This specific pattern gets generated when we split a
21615 // selector for a 512 bit vector in a machine without AVX512 (but with
21616 // 256-bit vectors), during legalization:
21618 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
21620 // Iff we find this pattern and the build_vectors are built from
21621 // constants, we translate the vselect into a shuffle_vector that we
21622 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
21623 if ((N->getOpcode() == ISD::VSELECT ||
21624 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
21625 !DCI.isBeforeLegalize()) {
21626 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
21627 if (Shuffle.getNode())
21631 // If this is a *dynamic* select (non-constant condition) and we can match
21632 // this node with one of the variable blend instructions, restructure the
21633 // condition so that the blends can use the high bit of each element and use
21634 // SimplifyDemandedBits to simplify the condition operand.
21635 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
21636 !DCI.isBeforeLegalize() &&
21637 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
21638 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
21640 // Don't optimize vector selects that map to mask-registers.
21644 // We can only handle the cases where VSELECT is directly legal on the
21645 // subtarget. We custom lower VSELECT nodes with constant conditions and
21646 // this makes it hard to see whether a dynamic VSELECT will correctly
21647 // lower, so we both check the operation's status and explicitly handle the
21648 // cases where a *dynamic* blend will fail even though a constant-condition
21649 // blend could be custom lowered.
21650 // FIXME: We should find a better way to handle this class of problems.
21651 // Potentially, we should combine constant-condition vselect nodes
21652 // pre-legalization into shuffles and not mark as many types as custom
21654 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
21656 // FIXME: We don't support i16-element blends currently. We could and
21657 // should support them by making *all* the bits in the condition be set
21658 // rather than just the high bit and using an i8-element blend.
21659 if (VT.getScalarType() == MVT::i16)
21661 // Dynamic blending was only available from SSE4.1 onward.
21662 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
21664 // Byte blends are only available in AVX2
21665 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
21666 !Subtarget->hasAVX2())
21669 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
21670 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
21672 APInt KnownZero, KnownOne;
21673 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
21674 DCI.isBeforeLegalizeOps());
21675 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
21676 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
21678 // If we changed the computation somewhere in the DAG, this change
21679 // will affect all users of Cond.
21680 // Make sure it is fine and update all the nodes so that we do not
21681 // use the generic VSELECT anymore. Otherwise, we may perform
21682 // wrong optimizations as we messed up with the actual expectation
21683 // for the vector boolean values.
21684 if (Cond != TLO.Old) {
21685 // Check all uses of that condition operand to check whether it will be
21686 // consumed by non-BLEND instructions, which may depend on all bits are
21688 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21690 if (I->getOpcode() != ISD::VSELECT)
21691 // TODO: Add other opcodes eventually lowered into BLEND.
21694 // Update all the users of the condition, before committing the change,
21695 // so that the VSELECT optimizations that expect the correct vector
21696 // boolean value will not be triggered.
21697 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21699 DAG.ReplaceAllUsesOfValueWith(
21701 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
21702 Cond, I->getOperand(1), I->getOperand(2)));
21703 DCI.CommitTargetLoweringOpt(TLO);
21706 // At this point, only Cond is changed. Change the condition
21707 // just for N to keep the opportunity to optimize all other
21708 // users their own way.
21709 DAG.ReplaceAllUsesOfValueWith(
21711 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
21712 TLO.New, N->getOperand(1), N->getOperand(2)));
21720 // Check whether a boolean test is testing a boolean value generated by
21721 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
21724 // Simplify the following patterns:
21725 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
21726 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
21727 // to (Op EFLAGS Cond)
21729 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
21730 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
21731 // to (Op EFLAGS !Cond)
21733 // where Op could be BRCOND or CMOV.
21735 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
21736 // Quit if not CMP and SUB with its value result used.
21737 if (Cmp.getOpcode() != X86ISD::CMP &&
21738 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
21741 // Quit if not used as a boolean value.
21742 if (CC != X86::COND_E && CC != X86::COND_NE)
21745 // Check CMP operands. One of them should be 0 or 1 and the other should be
21746 // an SetCC or extended from it.
21747 SDValue Op1 = Cmp.getOperand(0);
21748 SDValue Op2 = Cmp.getOperand(1);
21751 const ConstantSDNode* C = nullptr;
21752 bool needOppositeCond = (CC == X86::COND_E);
21753 bool checkAgainstTrue = false; // Is it a comparison against 1?
21755 if ((C = dyn_cast<ConstantSDNode>(Op1)))
21757 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
21759 else // Quit if all operands are not constants.
21762 if (C->getZExtValue() == 1) {
21763 needOppositeCond = !needOppositeCond;
21764 checkAgainstTrue = true;
21765 } else if (C->getZExtValue() != 0)
21766 // Quit if the constant is neither 0 or 1.
21769 bool truncatedToBoolWithAnd = false;
21770 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
21771 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
21772 SetCC.getOpcode() == ISD::TRUNCATE ||
21773 SetCC.getOpcode() == ISD::AND) {
21774 if (SetCC.getOpcode() == ISD::AND) {
21776 ConstantSDNode *CS;
21777 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
21778 CS->getZExtValue() == 1)
21780 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
21781 CS->getZExtValue() == 1)
21785 SetCC = SetCC.getOperand(OpIdx);
21786 truncatedToBoolWithAnd = true;
21788 SetCC = SetCC.getOperand(0);
21791 switch (SetCC.getOpcode()) {
21792 case X86ISD::SETCC_CARRY:
21793 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
21794 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
21795 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
21796 // truncated to i1 using 'and'.
21797 if (checkAgainstTrue && !truncatedToBoolWithAnd)
21799 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
21800 "Invalid use of SETCC_CARRY!");
21802 case X86ISD::SETCC:
21803 // Set the condition code or opposite one if necessary.
21804 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
21805 if (needOppositeCond)
21806 CC = X86::GetOppositeBranchCondition(CC);
21807 return SetCC.getOperand(1);
21808 case X86ISD::CMOV: {
21809 // Check whether false/true value has canonical one, i.e. 0 or 1.
21810 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
21811 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
21812 // Quit if true value is not a constant.
21815 // Quit if false value is not a constant.
21817 SDValue Op = SetCC.getOperand(0);
21818 // Skip 'zext' or 'trunc' node.
21819 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
21820 Op.getOpcode() == ISD::TRUNCATE)
21821 Op = Op.getOperand(0);
21822 // A special case for rdrand/rdseed, where 0 is set if false cond is
21824 if ((Op.getOpcode() != X86ISD::RDRAND &&
21825 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
21828 // Quit if false value is not the constant 0 or 1.
21829 bool FValIsFalse = true;
21830 if (FVal && FVal->getZExtValue() != 0) {
21831 if (FVal->getZExtValue() != 1)
21833 // If FVal is 1, opposite cond is needed.
21834 needOppositeCond = !needOppositeCond;
21835 FValIsFalse = false;
21837 // Quit if TVal is not the constant opposite of FVal.
21838 if (FValIsFalse && TVal->getZExtValue() != 1)
21840 if (!FValIsFalse && TVal->getZExtValue() != 0)
21842 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
21843 if (needOppositeCond)
21844 CC = X86::GetOppositeBranchCondition(CC);
21845 return SetCC.getOperand(3);
21852 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
21854 /// (X86or (X86setcc) (X86setcc))
21855 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
21856 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
21857 X86::CondCode &CC1, SDValue &Flags,
21859 if (Cond->getOpcode() == X86ISD::CMP) {
21860 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
21861 if (!CondOp1C || !CondOp1C->isNullValue())
21864 Cond = Cond->getOperand(0);
21869 SDValue SetCC0, SetCC1;
21870 switch (Cond->getOpcode()) {
21871 default: return false;
21878 SetCC0 = Cond->getOperand(0);
21879 SetCC1 = Cond->getOperand(1);
21883 // Make sure we have SETCC nodes, using the same flags value.
21884 if (SetCC0.getOpcode() != X86ISD::SETCC ||
21885 SetCC1.getOpcode() != X86ISD::SETCC ||
21886 SetCC0->getOperand(1) != SetCC1->getOperand(1))
21889 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
21890 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
21891 Flags = SetCC0->getOperand(1);
21895 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
21896 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
21897 TargetLowering::DAGCombinerInfo &DCI,
21898 const X86Subtarget *Subtarget) {
21901 // If the flag operand isn't dead, don't touch this CMOV.
21902 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
21905 SDValue FalseOp = N->getOperand(0);
21906 SDValue TrueOp = N->getOperand(1);
21907 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
21908 SDValue Cond = N->getOperand(3);
21910 if (CC == X86::COND_E || CC == X86::COND_NE) {
21911 switch (Cond.getOpcode()) {
21915 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
21916 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
21917 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21923 Flags = checkBoolTestSetCCCombine(Cond, CC);
21924 if (Flags.getNode() &&
21925 // Extra check as FCMOV only supports a subset of X86 cond.
21926 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21927 SDValue Ops[] = { FalseOp, TrueOp,
21928 DAG.getConstant(CC, DL, MVT::i8), Flags };
21929 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21932 // If this is a select between two integer constants, try to do some
21933 // optimizations. Note that the operands are ordered the opposite of SELECT
21935 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21936 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21937 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21938 // larger than FalseC (the false value).
21939 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21940 CC = X86::GetOppositeBranchCondition(CC);
21941 std::swap(TrueC, FalseC);
21942 std::swap(TrueOp, FalseOp);
21945 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21946 // This is efficient for any integer data type (including i8/i16) and
21948 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21949 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21950 DAG.getConstant(CC, DL, MVT::i8), Cond);
21952 // Zero extend the condition if needed.
21953 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21955 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21956 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21957 DAG.getConstant(ShAmt, DL, MVT::i8));
21958 if (N->getNumValues() == 2) // Dead flag value?
21959 return DCI.CombineTo(N, Cond, SDValue());
21963 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21964 // for any integer data type, including i8/i16.
21965 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21966 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21967 DAG.getConstant(CC, DL, MVT::i8), Cond);
21969 // Zero extend the condition if needed.
21970 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21971 FalseC->getValueType(0), Cond);
21972 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21973 SDValue(FalseC, 0));
21975 if (N->getNumValues() == 2) // Dead flag value?
21976 return DCI.CombineTo(N, Cond, SDValue());
21980 // Optimize cases that will turn into an LEA instruction. This requires
21981 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21982 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21983 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21984 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21986 bool isFastMultiplier = false;
21988 switch ((unsigned char)Diff) {
21990 case 1: // result = add base, cond
21991 case 2: // result = lea base( , cond*2)
21992 case 3: // result = lea base(cond, cond*2)
21993 case 4: // result = lea base( , cond*4)
21994 case 5: // result = lea base(cond, cond*4)
21995 case 8: // result = lea base( , cond*8)
21996 case 9: // result = lea base(cond, cond*8)
21997 isFastMultiplier = true;
22002 if (isFastMultiplier) {
22003 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22004 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22005 DAG.getConstant(CC, DL, MVT::i8), Cond);
22006 // Zero extend the condition if needed.
22007 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22009 // Scale the condition by the difference.
22011 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22012 DAG.getConstant(Diff, DL, Cond.getValueType()));
22014 // Add the base if non-zero.
22015 if (FalseC->getAPIntValue() != 0)
22016 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22017 SDValue(FalseC, 0));
22018 if (N->getNumValues() == 2) // Dead flag value?
22019 return DCI.CombineTo(N, Cond, SDValue());
22026 // Handle these cases:
22027 // (select (x != c), e, c) -> select (x != c), e, x),
22028 // (select (x == c), c, e) -> select (x == c), x, e)
22029 // where the c is an integer constant, and the "select" is the combination
22030 // of CMOV and CMP.
22032 // The rationale for this change is that the conditional-move from a constant
22033 // needs two instructions, however, conditional-move from a register needs
22034 // only one instruction.
22036 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
22037 // some instruction-combining opportunities. This opt needs to be
22038 // postponed as late as possible.
22040 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
22041 // the DCI.xxxx conditions are provided to postpone the optimization as
22042 // late as possible.
22044 ConstantSDNode *CmpAgainst = nullptr;
22045 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
22046 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
22047 !isa<ConstantSDNode>(Cond.getOperand(0))) {
22049 if (CC == X86::COND_NE &&
22050 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
22051 CC = X86::GetOppositeBranchCondition(CC);
22052 std::swap(TrueOp, FalseOp);
22055 if (CC == X86::COND_E &&
22056 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
22057 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
22058 DAG.getConstant(CC, DL, MVT::i8), Cond };
22059 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
22064 // Fold and/or of setcc's to double CMOV:
22065 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
22066 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
22068 // This combine lets us generate:
22069 // cmovcc1 (jcc1 if we don't have CMOV)
22075 // cmovne (jne if we don't have CMOV)
22076 // When we can't use the CMOV instruction, it might increase branch
22078 // When we can use CMOV, or when there is no mispredict, this improves
22079 // throughput and reduces register pressure.
22081 if (CC == X86::COND_NE) {
22083 X86::CondCode CC0, CC1;
22085 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
22087 std::swap(FalseOp, TrueOp);
22088 CC0 = X86::GetOppositeBranchCondition(CC0);
22089 CC1 = X86::GetOppositeBranchCondition(CC1);
22092 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
22094 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
22095 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
22096 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
22097 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
22105 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
22106 const X86Subtarget *Subtarget) {
22107 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
22109 default: return SDValue();
22110 // SSE/AVX/AVX2 blend intrinsics.
22111 case Intrinsic::x86_avx2_pblendvb:
22112 // Don't try to simplify this intrinsic if we don't have AVX2.
22113 if (!Subtarget->hasAVX2())
22116 case Intrinsic::x86_avx_blendv_pd_256:
22117 case Intrinsic::x86_avx_blendv_ps_256:
22118 // Don't try to simplify this intrinsic if we don't have AVX.
22119 if (!Subtarget->hasAVX())
22122 case Intrinsic::x86_sse41_blendvps:
22123 case Intrinsic::x86_sse41_blendvpd:
22124 case Intrinsic::x86_sse41_pblendvb: {
22125 SDValue Op0 = N->getOperand(1);
22126 SDValue Op1 = N->getOperand(2);
22127 SDValue Mask = N->getOperand(3);
22129 // Don't try to simplify this intrinsic if we don't have SSE4.1.
22130 if (!Subtarget->hasSSE41())
22133 // fold (blend A, A, Mask) -> A
22136 // fold (blend A, B, allZeros) -> A
22137 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
22139 // fold (blend A, B, allOnes) -> B
22140 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
22143 // Simplify the case where the mask is a constant i32 value.
22144 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
22145 if (C->isNullValue())
22147 if (C->isAllOnesValue())
22154 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
22155 case Intrinsic::x86_sse2_psrai_w:
22156 case Intrinsic::x86_sse2_psrai_d:
22157 case Intrinsic::x86_avx2_psrai_w:
22158 case Intrinsic::x86_avx2_psrai_d:
22159 case Intrinsic::x86_sse2_psra_w:
22160 case Intrinsic::x86_sse2_psra_d:
22161 case Intrinsic::x86_avx2_psra_w:
22162 case Intrinsic::x86_avx2_psra_d: {
22163 SDValue Op0 = N->getOperand(1);
22164 SDValue Op1 = N->getOperand(2);
22165 EVT VT = Op0.getValueType();
22166 assert(VT.isVector() && "Expected a vector type!");
22168 if (isa<BuildVectorSDNode>(Op1))
22169 Op1 = Op1.getOperand(0);
22171 if (!isa<ConstantSDNode>(Op1))
22174 EVT SVT = VT.getVectorElementType();
22175 unsigned SVTBits = SVT.getSizeInBits();
22177 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
22178 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
22179 uint64_t ShAmt = C.getZExtValue();
22181 // Don't try to convert this shift into a ISD::SRA if the shift
22182 // count is bigger than or equal to the element size.
22183 if (ShAmt >= SVTBits)
22186 // Trivial case: if the shift count is zero, then fold this
22187 // into the first operand.
22191 // Replace this packed shift intrinsic with a target independent
22194 SDValue Splat = DAG.getConstant(C, DL, VT);
22195 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
22200 /// PerformMulCombine - Optimize a single multiply with constant into two
22201 /// in order to implement it with two cheaper instructions, e.g.
22202 /// LEA + SHL, LEA + LEA.
22203 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
22204 TargetLowering::DAGCombinerInfo &DCI) {
22205 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
22208 EVT VT = N->getValueType(0);
22209 if (VT != MVT::i64 && VT != MVT::i32)
22212 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
22215 uint64_t MulAmt = C->getZExtValue();
22216 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
22219 uint64_t MulAmt1 = 0;
22220 uint64_t MulAmt2 = 0;
22221 if ((MulAmt % 9) == 0) {
22223 MulAmt2 = MulAmt / 9;
22224 } else if ((MulAmt % 5) == 0) {
22226 MulAmt2 = MulAmt / 5;
22227 } else if ((MulAmt % 3) == 0) {
22229 MulAmt2 = MulAmt / 3;
22232 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
22235 if (isPowerOf2_64(MulAmt2) &&
22236 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
22237 // If second multiplifer is pow2, issue it first. We want the multiply by
22238 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
22240 std::swap(MulAmt1, MulAmt2);
22243 if (isPowerOf2_64(MulAmt1))
22244 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
22245 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
22247 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
22248 DAG.getConstant(MulAmt1, DL, VT));
22250 if (isPowerOf2_64(MulAmt2))
22251 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
22252 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
22254 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
22255 DAG.getConstant(MulAmt2, DL, VT));
22257 // Do not add new nodes to DAG combiner worklist.
22258 DCI.CombineTo(N, NewMul, false);
22263 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22264 SDValue N0 = N->getOperand(0);
22265 SDValue N1 = N->getOperand(1);
22266 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22267 EVT VT = N0.getValueType();
22269 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22270 // since the result of setcc_c is all zero's or all ones.
22271 if (VT.isInteger() && !VT.isVector() &&
22272 N1C && N0.getOpcode() == ISD::AND &&
22273 N0.getOperand(1).getOpcode() == ISD::Constant) {
22274 SDValue N00 = N0.getOperand(0);
22275 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22276 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22277 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22278 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22279 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22280 APInt ShAmt = N1C->getAPIntValue();
22281 Mask = Mask.shl(ShAmt);
22284 return DAG.getNode(ISD::AND, DL, VT,
22285 N00, DAG.getConstant(Mask, DL, VT));
22290 // Hardware support for vector shifts is sparse which makes us scalarize the
22291 // vector operations in many cases. Also, on sandybridge ADD is faster than
22293 // (shl V, 1) -> add V,V
22294 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22295 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22296 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22297 // We shift all of the values by one. In many cases we do not have
22298 // hardware support for this operation. This is better expressed as an ADD
22300 if (N1SplatC->getZExtValue() == 1)
22301 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22307 /// \brief Returns a vector of 0s if the node in input is a vector logical
22308 /// shift by a constant amount which is known to be bigger than or equal
22309 /// to the vector element size in bits.
22310 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22311 const X86Subtarget *Subtarget) {
22312 EVT VT = N->getValueType(0);
22314 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22315 (!Subtarget->hasInt256() ||
22316 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22319 SDValue Amt = N->getOperand(1);
22321 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22322 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22323 APInt ShiftAmt = AmtSplat->getAPIntValue();
22324 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22326 // SSE2/AVX2 logical shifts always return a vector of 0s
22327 // if the shift amount is bigger than or equal to
22328 // the element size. The constant shift amount will be
22329 // encoded as a 8-bit immediate.
22330 if (ShiftAmt.trunc(8).uge(MaxAmount))
22331 return getZeroVector(VT, Subtarget, DAG, DL);
22337 /// PerformShiftCombine - Combine shifts.
22338 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22339 TargetLowering::DAGCombinerInfo &DCI,
22340 const X86Subtarget *Subtarget) {
22341 if (N->getOpcode() == ISD::SHL) {
22342 SDValue V = PerformSHLCombine(N, DAG);
22343 if (V.getNode()) return V;
22346 if (N->getOpcode() != ISD::SRA) {
22347 // Try to fold this logical shift into a zero vector.
22348 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22349 if (V.getNode()) return V;
22355 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22356 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22357 // and friends. Likewise for OR -> CMPNEQSS.
22358 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22359 TargetLowering::DAGCombinerInfo &DCI,
22360 const X86Subtarget *Subtarget) {
22363 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22364 // we're requiring SSE2 for both.
22365 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22366 SDValue N0 = N->getOperand(0);
22367 SDValue N1 = N->getOperand(1);
22368 SDValue CMP0 = N0->getOperand(1);
22369 SDValue CMP1 = N1->getOperand(1);
22372 // The SETCCs should both refer to the same CMP.
22373 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22376 SDValue CMP00 = CMP0->getOperand(0);
22377 SDValue CMP01 = CMP0->getOperand(1);
22378 EVT VT = CMP00.getValueType();
22380 if (VT == MVT::f32 || VT == MVT::f64) {
22381 bool ExpectingFlags = false;
22382 // Check for any users that want flags:
22383 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22384 !ExpectingFlags && UI != UE; ++UI)
22385 switch (UI->getOpcode()) {
22390 ExpectingFlags = true;
22392 case ISD::CopyToReg:
22393 case ISD::SIGN_EXTEND:
22394 case ISD::ZERO_EXTEND:
22395 case ISD::ANY_EXTEND:
22399 if (!ExpectingFlags) {
22400 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22401 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22403 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22404 X86::CondCode tmp = cc0;
22409 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22410 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22411 // FIXME: need symbolic constants for these magic numbers.
22412 // See X86ATTInstPrinter.cpp:printSSECC().
22413 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22414 if (Subtarget->hasAVX512()) {
22415 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22417 DAG.getConstant(x86cc, DL, MVT::i8));
22418 if (N->getValueType(0) != MVT::i1)
22419 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22423 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22424 CMP00.getValueType(), CMP00, CMP01,
22425 DAG.getConstant(x86cc, DL,
22428 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22429 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22431 if (is64BitFP && !Subtarget->is64Bit()) {
22432 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22433 // 64-bit integer, since that's not a legal type. Since
22434 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22435 // bits, but can do this little dance to extract the lowest 32 bits
22436 // and work with those going forward.
22437 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22439 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
22441 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22442 Vector32, DAG.getIntPtrConstant(0, DL));
22446 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
22448 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22449 DAG.getConstant(1, DL, IntVT));
22450 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
22452 return OneBitOfTruth;
22460 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22461 /// so it can be folded inside ANDNP.
22462 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22463 EVT VT = N->getValueType(0);
22465 // Match direct AllOnes for 128 and 256-bit vectors
22466 if (ISD::isBuildVectorAllOnes(N))
22469 // Look through a bit convert.
22470 if (N->getOpcode() == ISD::BITCAST)
22471 N = N->getOperand(0).getNode();
22473 // Sometimes the operand may come from a insert_subvector building a 256-bit
22475 if (VT.is256BitVector() &&
22476 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22477 SDValue V1 = N->getOperand(0);
22478 SDValue V2 = N->getOperand(1);
22480 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22481 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22482 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22483 ISD::isBuildVectorAllOnes(V2.getNode()))
22490 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22491 // register. In most cases we actually compare or select YMM-sized registers
22492 // and mixing the two types creates horrible code. This method optimizes
22493 // some of the transition sequences.
22494 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22495 TargetLowering::DAGCombinerInfo &DCI,
22496 const X86Subtarget *Subtarget) {
22497 EVT VT = N->getValueType(0);
22498 if (!VT.is256BitVector())
22501 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22502 N->getOpcode() == ISD::ZERO_EXTEND ||
22503 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22505 SDValue Narrow = N->getOperand(0);
22506 EVT NarrowVT = Narrow->getValueType(0);
22507 if (!NarrowVT.is128BitVector())
22510 if (Narrow->getOpcode() != ISD::XOR &&
22511 Narrow->getOpcode() != ISD::AND &&
22512 Narrow->getOpcode() != ISD::OR)
22515 SDValue N0 = Narrow->getOperand(0);
22516 SDValue N1 = Narrow->getOperand(1);
22519 // The Left side has to be a trunc.
22520 if (N0.getOpcode() != ISD::TRUNCATE)
22523 // The type of the truncated inputs.
22524 EVT WideVT = N0->getOperand(0)->getValueType(0);
22528 // The right side has to be a 'trunc' or a constant vector.
22529 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22530 ConstantSDNode *RHSConstSplat = nullptr;
22531 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22532 RHSConstSplat = RHSBV->getConstantSplatNode();
22533 if (!RHSTrunc && !RHSConstSplat)
22536 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22538 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22541 // Set N0 and N1 to hold the inputs to the new wide operation.
22542 N0 = N0->getOperand(0);
22543 if (RHSConstSplat) {
22544 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22545 SDValue(RHSConstSplat, 0));
22546 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22547 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22548 } else if (RHSTrunc) {
22549 N1 = N1->getOperand(0);
22552 // Generate the wide operation.
22553 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22554 unsigned Opcode = N->getOpcode();
22556 case ISD::ANY_EXTEND:
22558 case ISD::ZERO_EXTEND: {
22559 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22560 APInt Mask = APInt::getAllOnesValue(InBits);
22561 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22562 return DAG.getNode(ISD::AND, DL, VT,
22563 Op, DAG.getConstant(Mask, DL, VT));
22565 case ISD::SIGN_EXTEND:
22566 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22567 Op, DAG.getValueType(NarrowVT));
22569 llvm_unreachable("Unexpected opcode");
22573 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
22574 TargetLowering::DAGCombinerInfo &DCI,
22575 const X86Subtarget *Subtarget) {
22576 SDValue N0 = N->getOperand(0);
22577 SDValue N1 = N->getOperand(1);
22580 // A vector zext_in_reg may be represented as a shuffle,
22581 // feeding into a bitcast (this represents anyext) feeding into
22582 // an and with a mask.
22583 // We'd like to try to combine that into a shuffle with zero
22584 // plus a bitcast, removing the and.
22585 if (N0.getOpcode() != ISD::BITCAST ||
22586 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
22589 // The other side of the AND should be a splat of 2^C, where C
22590 // is the number of bits in the source type.
22591 if (N1.getOpcode() == ISD::BITCAST)
22592 N1 = N1.getOperand(0);
22593 if (N1.getOpcode() != ISD::BUILD_VECTOR)
22595 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
22597 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
22598 EVT SrcType = Shuffle->getValueType(0);
22600 // We expect a single-source shuffle
22601 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
22604 unsigned SrcSize = SrcType.getScalarSizeInBits();
22606 APInt SplatValue, SplatUndef;
22607 unsigned SplatBitSize;
22609 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
22610 SplatBitSize, HasAnyUndefs))
22613 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
22614 // Make sure the splat matches the mask we expect
22615 if (SplatBitSize > ResSize ||
22616 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
22619 // Make sure the input and output size make sense
22620 if (SrcSize >= ResSize || ResSize % SrcSize)
22623 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
22624 // The number of u's between each two values depends on the ratio between
22625 // the source and dest type.
22626 unsigned ZextRatio = ResSize / SrcSize;
22627 bool IsZext = true;
22628 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
22629 if (i % ZextRatio) {
22630 if (Shuffle->getMaskElt(i) > 0) {
22636 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
22637 // Expected element number
22647 // Ok, perform the transformation - replace the shuffle with
22648 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
22649 // (instead of undef) where the k elements come from the zero vector.
22650 SmallVector<int, 8> Mask;
22651 unsigned NumElems = SrcType.getVectorNumElements();
22652 for (unsigned i = 0; i < NumElems; ++i)
22654 Mask.push_back(NumElems);
22656 Mask.push_back(i / ZextRatio);
22658 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
22659 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
22660 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
22663 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22664 TargetLowering::DAGCombinerInfo &DCI,
22665 const X86Subtarget *Subtarget) {
22666 if (DCI.isBeforeLegalizeOps())
22669 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
22672 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
22675 EVT VT = N->getValueType(0);
22676 SDValue N0 = N->getOperand(0);
22677 SDValue N1 = N->getOperand(1);
22680 // Create BEXTR instructions
22681 // BEXTR is ((X >> imm) & (2**size-1))
22682 if (VT == MVT::i32 || VT == MVT::i64) {
22683 // Check for BEXTR.
22684 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22685 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22686 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22687 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22688 if (MaskNode && ShiftNode) {
22689 uint64_t Mask = MaskNode->getZExtValue();
22690 uint64_t Shift = ShiftNode->getZExtValue();
22691 if (isMask_64(Mask)) {
22692 uint64_t MaskSize = countPopulation(Mask);
22693 if (Shift + MaskSize <= VT.getSizeInBits())
22694 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22695 DAG.getConstant(Shift | (MaskSize << 8), DL,
22704 // Want to form ANDNP nodes:
22705 // 1) In the hopes of then easily combining them with OR and AND nodes
22706 // to form PBLEND/PSIGN.
22707 // 2) To match ANDN packed intrinsics
22708 if (VT != MVT::v2i64 && VT != MVT::v4i64)
22711 // Check LHS for vnot
22712 if (N0.getOpcode() == ISD::XOR &&
22713 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
22714 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
22715 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
22717 // Check RHS for vnot
22718 if (N1.getOpcode() == ISD::XOR &&
22719 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
22720 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
22721 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
22726 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
22727 TargetLowering::DAGCombinerInfo &DCI,
22728 const X86Subtarget *Subtarget) {
22729 if (DCI.isBeforeLegalizeOps())
22732 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22736 SDValue N0 = N->getOperand(0);
22737 SDValue N1 = N->getOperand(1);
22738 EVT VT = N->getValueType(0);
22740 // look for psign/blend
22741 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
22742 if (!Subtarget->hasSSSE3() ||
22743 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
22746 // Canonicalize pandn to RHS
22747 if (N0.getOpcode() == X86ISD::ANDNP)
22749 // or (and (m, y), (pandn m, x))
22750 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
22751 SDValue Mask = N1.getOperand(0);
22752 SDValue X = N1.getOperand(1);
22754 if (N0.getOperand(0) == Mask)
22755 Y = N0.getOperand(1);
22756 if (N0.getOperand(1) == Mask)
22757 Y = N0.getOperand(0);
22759 // Check to see if the mask appeared in both the AND and ANDNP and
22763 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
22764 // Look through mask bitcast.
22765 if (Mask.getOpcode() == ISD::BITCAST)
22766 Mask = Mask.getOperand(0);
22767 if (X.getOpcode() == ISD::BITCAST)
22768 X = X.getOperand(0);
22769 if (Y.getOpcode() == ISD::BITCAST)
22770 Y = Y.getOperand(0);
22772 EVT MaskVT = Mask.getValueType();
22774 // Validate that the Mask operand is a vector sra node.
22775 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
22776 // there is no psrai.b
22777 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
22778 unsigned SraAmt = ~0;
22779 if (Mask.getOpcode() == ISD::SRA) {
22780 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
22781 if (auto *AmtConst = AmtBV->getConstantSplatNode())
22782 SraAmt = AmtConst->getZExtValue();
22783 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
22784 SDValue SraC = Mask.getOperand(1);
22785 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
22787 if ((SraAmt + 1) != EltBits)
22792 // Now we know we at least have a plendvb with the mask val. See if
22793 // we can form a psignb/w/d.
22794 // psign = x.type == y.type == mask.type && y = sub(0, x);
22795 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
22796 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
22797 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
22798 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
22799 "Unsupported VT for PSIGN");
22800 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
22801 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22803 // PBLENDVB only available on SSE 4.1
22804 if (!Subtarget->hasSSE41())
22807 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
22809 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
22810 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
22811 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
22812 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
22813 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22817 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
22820 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
22821 MachineFunction &MF = DAG.getMachineFunction();
22823 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
22825 // SHLD/SHRD instructions have lower register pressure, but on some
22826 // platforms they have higher latency than the equivalent
22827 // series of shifts/or that would otherwise be generated.
22828 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
22829 // have higher latencies and we are not optimizing for size.
22830 if (!OptForSize && Subtarget->isSHLDSlow())
22833 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
22835 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
22837 if (!N0.hasOneUse() || !N1.hasOneUse())
22840 SDValue ShAmt0 = N0.getOperand(1);
22841 if (ShAmt0.getValueType() != MVT::i8)
22843 SDValue ShAmt1 = N1.getOperand(1);
22844 if (ShAmt1.getValueType() != MVT::i8)
22846 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
22847 ShAmt0 = ShAmt0.getOperand(0);
22848 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
22849 ShAmt1 = ShAmt1.getOperand(0);
22852 unsigned Opc = X86ISD::SHLD;
22853 SDValue Op0 = N0.getOperand(0);
22854 SDValue Op1 = N1.getOperand(0);
22855 if (ShAmt0.getOpcode() == ISD::SUB) {
22856 Opc = X86ISD::SHRD;
22857 std::swap(Op0, Op1);
22858 std::swap(ShAmt0, ShAmt1);
22861 unsigned Bits = VT.getSizeInBits();
22862 if (ShAmt1.getOpcode() == ISD::SUB) {
22863 SDValue Sum = ShAmt1.getOperand(0);
22864 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
22865 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
22866 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
22867 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
22868 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
22869 return DAG.getNode(Opc, DL, VT,
22871 DAG.getNode(ISD::TRUNCATE, DL,
22874 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
22875 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
22877 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
22878 return DAG.getNode(Opc, DL, VT,
22879 N0.getOperand(0), N1.getOperand(0),
22880 DAG.getNode(ISD::TRUNCATE, DL,
22887 // Generate NEG and CMOV for integer abs.
22888 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
22889 EVT VT = N->getValueType(0);
22891 // Since X86 does not have CMOV for 8-bit integer, we don't convert
22892 // 8-bit integer abs to NEG and CMOV.
22893 if (VT.isInteger() && VT.getSizeInBits() == 8)
22896 SDValue N0 = N->getOperand(0);
22897 SDValue N1 = N->getOperand(1);
22900 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
22901 // and change it to SUB and CMOV.
22902 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
22903 N0.getOpcode() == ISD::ADD &&
22904 N0.getOperand(1) == N1 &&
22905 N1.getOpcode() == ISD::SRA &&
22906 N1.getOperand(0) == N0.getOperand(0))
22907 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
22908 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
22909 // Generate SUB & CMOV.
22910 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
22911 DAG.getConstant(0, DL, VT), N0.getOperand(0));
22913 SDValue Ops[] = { N0.getOperand(0), Neg,
22914 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
22915 SDValue(Neg.getNode(), 1) };
22916 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
22921 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
22922 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
22923 TargetLowering::DAGCombinerInfo &DCI,
22924 const X86Subtarget *Subtarget) {
22925 if (DCI.isBeforeLegalizeOps())
22928 if (Subtarget->hasCMov()) {
22929 SDValue RV = performIntegerAbsCombine(N, DAG);
22937 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
22938 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
22939 TargetLowering::DAGCombinerInfo &DCI,
22940 const X86Subtarget *Subtarget) {
22941 LoadSDNode *Ld = cast<LoadSDNode>(N);
22942 EVT RegVT = Ld->getValueType(0);
22943 EVT MemVT = Ld->getMemoryVT();
22945 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22947 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
22948 // into two 16-byte operations.
22949 ISD::LoadExtType Ext = Ld->getExtensionType();
22950 unsigned Alignment = Ld->getAlignment();
22951 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
22952 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22953 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
22954 unsigned NumElems = RegVT.getVectorNumElements();
22958 SDValue Ptr = Ld->getBasePtr();
22959 SDValue Increment = DAG.getConstant(16, dl, TLI.getPointerTy());
22961 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
22963 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22964 Ld->getPointerInfo(), Ld->isVolatile(),
22965 Ld->isNonTemporal(), Ld->isInvariant(),
22967 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22968 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22969 Ld->getPointerInfo(), Ld->isVolatile(),
22970 Ld->isNonTemporal(), Ld->isInvariant(),
22971 std::min(16U, Alignment));
22972 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22974 Load2.getValue(1));
22976 SDValue NewVec = DAG.getUNDEF(RegVT);
22977 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22978 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22979 return DCI.CombineTo(N, NewVec, TF, true);
22985 /// PerformMLOADCombine - Resolve extending loads
22986 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22987 TargetLowering::DAGCombinerInfo &DCI,
22988 const X86Subtarget *Subtarget) {
22989 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22990 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22993 EVT VT = Mld->getValueType(0);
22994 unsigned NumElems = VT.getVectorNumElements();
22995 EVT LdVT = Mld->getMemoryVT();
22998 assert(LdVT != VT && "Cannot extend to the same type");
22999 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
23000 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
23001 // From, To sizes and ElemCount must be pow of two
23002 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23003 "Unexpected size for extending masked load");
23005 unsigned SizeRatio = ToSz / FromSz;
23006 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
23008 // Create a type on which we perform the shuffle
23009 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23010 LdVT.getScalarType(), NumElems*SizeRatio);
23011 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23013 // Convert Src0 value
23014 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
23015 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
23016 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23017 for (unsigned i = 0; i != NumElems; ++i)
23018 ShuffleVec[i] = i * SizeRatio;
23020 // Can't shuffle using an illegal type.
23021 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23022 && "WideVecVT should be legal");
23023 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
23024 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
23026 // Prepare the new mask
23028 SDValue Mask = Mld->getMask();
23029 if (Mask.getValueType() == VT) {
23030 // Mask and original value have the same type
23031 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
23032 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23033 for (unsigned i = 0; i != NumElems; ++i)
23034 ShuffleVec[i] = i * SizeRatio;
23035 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23036 ShuffleVec[i] = NumElems*SizeRatio;
23037 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23038 DAG.getConstant(0, dl, WideVecVT),
23042 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23043 unsigned WidenNumElts = NumElems*SizeRatio;
23044 unsigned MaskNumElts = VT.getVectorNumElements();
23045 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23048 unsigned NumConcat = WidenNumElts / MaskNumElts;
23049 SmallVector<SDValue, 16> Ops(NumConcat);
23050 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23052 for (unsigned i = 1; i != NumConcat; ++i)
23055 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23058 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
23059 Mld->getBasePtr(), NewMask, WideSrc0,
23060 Mld->getMemoryVT(), Mld->getMemOperand(),
23062 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
23063 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
23066 /// PerformMSTORECombine - Resolve truncating stores
23067 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
23068 const X86Subtarget *Subtarget) {
23069 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
23070 if (!Mst->isTruncatingStore())
23073 EVT VT = Mst->getValue().getValueType();
23074 unsigned NumElems = VT.getVectorNumElements();
23075 EVT StVT = Mst->getMemoryVT();
23078 assert(StVT != VT && "Cannot truncate to the same type");
23079 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23080 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23082 // From, To sizes and ElemCount must be pow of two
23083 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23084 "Unexpected size for truncating masked store");
23085 // We are going to use the original vector elt for storing.
23086 // Accumulated smaller vector elements must be a multiple of the store size.
23087 assert (((NumElems * FromSz) % ToSz) == 0 &&
23088 "Unexpected ratio for truncating masked store");
23090 unsigned SizeRatio = FromSz / ToSz;
23091 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23093 // Create a type on which we perform the shuffle
23094 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23095 StVT.getScalarType(), NumElems*SizeRatio);
23097 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23099 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
23100 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23101 for (unsigned i = 0; i != NumElems; ++i)
23102 ShuffleVec[i] = i * SizeRatio;
23104 // Can't shuffle using an illegal type.
23105 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23106 && "WideVecVT should be legal");
23108 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23109 DAG.getUNDEF(WideVecVT),
23113 SDValue Mask = Mst->getMask();
23114 if (Mask.getValueType() == VT) {
23115 // Mask and original value have the same type
23116 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
23117 for (unsigned i = 0; i != NumElems; ++i)
23118 ShuffleVec[i] = i * SizeRatio;
23119 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23120 ShuffleVec[i] = NumElems*SizeRatio;
23121 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23122 DAG.getConstant(0, dl, WideVecVT),
23126 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23127 unsigned WidenNumElts = NumElems*SizeRatio;
23128 unsigned MaskNumElts = VT.getVectorNumElements();
23129 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23132 unsigned NumConcat = WidenNumElts / MaskNumElts;
23133 SmallVector<SDValue, 16> Ops(NumConcat);
23134 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23136 for (unsigned i = 1; i != NumConcat; ++i)
23139 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23142 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
23143 NewMask, StVT, Mst->getMemOperand(), false);
23145 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
23146 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
23147 const X86Subtarget *Subtarget) {
23148 StoreSDNode *St = cast<StoreSDNode>(N);
23149 EVT VT = St->getValue().getValueType();
23150 EVT StVT = St->getMemoryVT();
23152 SDValue StoredVal = St->getOperand(1);
23153 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23155 // If we are saving a concatenation of two XMM registers and 32-byte stores
23156 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
23157 unsigned Alignment = St->getAlignment();
23158 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
23159 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
23160 StVT == VT && !IsAligned) {
23161 unsigned NumElems = VT.getVectorNumElements();
23165 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
23166 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
23168 SDValue Stride = DAG.getConstant(16, dl, TLI.getPointerTy());
23169 SDValue Ptr0 = St->getBasePtr();
23170 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
23172 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
23173 St->getPointerInfo(), St->isVolatile(),
23174 St->isNonTemporal(), Alignment);
23175 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
23176 St->getPointerInfo(), St->isVolatile(),
23177 St->isNonTemporal(),
23178 std::min(16U, Alignment));
23179 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
23182 // Optimize trunc store (of multiple scalars) to shuffle and store.
23183 // First, pack all of the elements in one place. Next, store to memory
23184 // in fewer chunks.
23185 if (St->isTruncatingStore() && VT.isVector()) {
23186 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23187 unsigned NumElems = VT.getVectorNumElements();
23188 assert(StVT != VT && "Cannot truncate to the same type");
23189 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23190 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23192 // From, To sizes and ElemCount must be pow of two
23193 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
23194 // We are going to use the original vector elt for storing.
23195 // Accumulated smaller vector elements must be a multiple of the store size.
23196 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
23198 unsigned SizeRatio = FromSz / ToSz;
23200 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23202 // Create a type on which we perform the shuffle
23203 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23204 StVT.getScalarType(), NumElems*SizeRatio);
23206 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23208 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
23209 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
23210 for (unsigned i = 0; i != NumElems; ++i)
23211 ShuffleVec[i] = i * SizeRatio;
23213 // Can't shuffle using an illegal type.
23214 if (!TLI.isTypeLegal(WideVecVT))
23217 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23218 DAG.getUNDEF(WideVecVT),
23220 // At this point all of the data is stored at the bottom of the
23221 // register. We now need to save it to mem.
23223 // Find the largest store unit
23224 MVT StoreType = MVT::i8;
23225 for (MVT Tp : MVT::integer_valuetypes()) {
23226 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
23230 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
23231 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
23232 (64 <= NumElems * ToSz))
23233 StoreType = MVT::f64;
23235 // Bitcast the original vector into a vector of store-size units
23236 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
23237 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
23238 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
23239 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
23240 SmallVector<SDValue, 8> Chains;
23241 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, dl,
23242 TLI.getPointerTy());
23243 SDValue Ptr = St->getBasePtr();
23245 // Perform one or more big stores into memory.
23246 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
23247 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
23248 StoreType, ShuffWide,
23249 DAG.getIntPtrConstant(i, dl));
23250 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
23251 St->getPointerInfo(), St->isVolatile(),
23252 St->isNonTemporal(), St->getAlignment());
23253 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23254 Chains.push_back(Ch);
23257 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
23260 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
23261 // the FP state in cases where an emms may be missing.
23262 // A preferable solution to the general problem is to figure out the right
23263 // places to insert EMMS. This qualifies as a quick hack.
23265 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
23266 if (VT.getSizeInBits() != 64)
23269 const Function *F = DAG.getMachineFunction().getFunction();
23270 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
23272 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
23273 if ((VT.isVector() ||
23274 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23275 isa<LoadSDNode>(St->getValue()) &&
23276 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23277 St->getChain().hasOneUse() && !St->isVolatile()) {
23278 SDNode* LdVal = St->getValue().getNode();
23279 LoadSDNode *Ld = nullptr;
23280 int TokenFactorIndex = -1;
23281 SmallVector<SDValue, 8> Ops;
23282 SDNode* ChainVal = St->getChain().getNode();
23283 // Must be a store of a load. We currently handle two cases: the load
23284 // is a direct child, and it's under an intervening TokenFactor. It is
23285 // possible to dig deeper under nested TokenFactors.
23286 if (ChainVal == LdVal)
23287 Ld = cast<LoadSDNode>(St->getChain());
23288 else if (St->getValue().hasOneUse() &&
23289 ChainVal->getOpcode() == ISD::TokenFactor) {
23290 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23291 if (ChainVal->getOperand(i).getNode() == LdVal) {
23292 TokenFactorIndex = i;
23293 Ld = cast<LoadSDNode>(St->getValue());
23295 Ops.push_back(ChainVal->getOperand(i));
23299 if (!Ld || !ISD::isNormalLoad(Ld))
23302 // If this is not the MMX case, i.e. we are just turning i64 load/store
23303 // into f64 load/store, avoid the transformation if there are multiple
23304 // uses of the loaded value.
23305 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23310 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23311 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23313 if (Subtarget->is64Bit() || F64IsLegal) {
23314 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23315 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23316 Ld->getPointerInfo(), Ld->isVolatile(),
23317 Ld->isNonTemporal(), Ld->isInvariant(),
23318 Ld->getAlignment());
23319 SDValue NewChain = NewLd.getValue(1);
23320 if (TokenFactorIndex != -1) {
23321 Ops.push_back(NewChain);
23322 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23324 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23325 St->getPointerInfo(),
23326 St->isVolatile(), St->isNonTemporal(),
23327 St->getAlignment());
23330 // Otherwise, lower to two pairs of 32-bit loads / stores.
23331 SDValue LoAddr = Ld->getBasePtr();
23332 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23333 DAG.getConstant(4, LdDL, MVT::i32));
23335 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23336 Ld->getPointerInfo(),
23337 Ld->isVolatile(), Ld->isNonTemporal(),
23338 Ld->isInvariant(), Ld->getAlignment());
23339 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23340 Ld->getPointerInfo().getWithOffset(4),
23341 Ld->isVolatile(), Ld->isNonTemporal(),
23343 MinAlign(Ld->getAlignment(), 4));
23345 SDValue NewChain = LoLd.getValue(1);
23346 if (TokenFactorIndex != -1) {
23347 Ops.push_back(LoLd);
23348 Ops.push_back(HiLd);
23349 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23352 LoAddr = St->getBasePtr();
23353 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23354 DAG.getConstant(4, StDL, MVT::i32));
23356 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23357 St->getPointerInfo(),
23358 St->isVolatile(), St->isNonTemporal(),
23359 St->getAlignment());
23360 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23361 St->getPointerInfo().getWithOffset(4),
23363 St->isNonTemporal(),
23364 MinAlign(St->getAlignment(), 4));
23365 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23368 // This is similar to the above case, but here we handle a scalar 64-bit
23369 // integer store that is extracted from a vector on a 32-bit target.
23370 // If we have SSE2, then we can treat it like a floating-point double
23371 // to get past legalization. The execution dependencies fixup pass will
23372 // choose the optimal machine instruction for the store if this really is
23373 // an integer or v2f32 rather than an f64.
23374 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
23375 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
23376 SDValue OldExtract = St->getOperand(1);
23377 SDValue ExtOp0 = OldExtract.getOperand(0);
23378 unsigned VecSize = ExtOp0.getValueSizeInBits();
23379 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
23380 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtOp0);
23381 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
23382 BitCast, OldExtract.getOperand(1));
23383 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
23384 St->getPointerInfo(), St->isVolatile(),
23385 St->isNonTemporal(), St->getAlignment());
23391 /// Return 'true' if this vector operation is "horizontal"
23392 /// and return the operands for the horizontal operation in LHS and RHS. A
23393 /// horizontal operation performs the binary operation on successive elements
23394 /// of its first operand, then on successive elements of its second operand,
23395 /// returning the resulting values in a vector. For example, if
23396 /// A = < float a0, float a1, float a2, float a3 >
23398 /// B = < float b0, float b1, float b2, float b3 >
23399 /// then the result of doing a horizontal operation on A and B is
23400 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23401 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23402 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23403 /// set to A, RHS to B, and the routine returns 'true'.
23404 /// Note that the binary operation should have the property that if one of the
23405 /// operands is UNDEF then the result is UNDEF.
23406 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23407 // Look for the following pattern: if
23408 // A = < float a0, float a1, float a2, float a3 >
23409 // B = < float b0, float b1, float b2, float b3 >
23411 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23412 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23413 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23414 // which is A horizontal-op B.
23416 // At least one of the operands should be a vector shuffle.
23417 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23418 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23421 MVT VT = LHS.getSimpleValueType();
23423 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23424 "Unsupported vector type for horizontal add/sub");
23426 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23427 // operate independently on 128-bit lanes.
23428 unsigned NumElts = VT.getVectorNumElements();
23429 unsigned NumLanes = VT.getSizeInBits()/128;
23430 unsigned NumLaneElts = NumElts / NumLanes;
23431 assert((NumLaneElts % 2 == 0) &&
23432 "Vector type should have an even number of elements in each lane");
23433 unsigned HalfLaneElts = NumLaneElts/2;
23435 // View LHS in the form
23436 // LHS = VECTOR_SHUFFLE A, B, LMask
23437 // If LHS is not a shuffle then pretend it is the shuffle
23438 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23439 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23442 SmallVector<int, 16> LMask(NumElts);
23443 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23444 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23445 A = LHS.getOperand(0);
23446 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23447 B = LHS.getOperand(1);
23448 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23449 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23451 if (LHS.getOpcode() != ISD::UNDEF)
23453 for (unsigned i = 0; i != NumElts; ++i)
23457 // Likewise, view RHS in the form
23458 // RHS = VECTOR_SHUFFLE C, D, RMask
23460 SmallVector<int, 16> RMask(NumElts);
23461 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23462 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23463 C = RHS.getOperand(0);
23464 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23465 D = RHS.getOperand(1);
23466 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23467 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23469 if (RHS.getOpcode() != ISD::UNDEF)
23471 for (unsigned i = 0; i != NumElts; ++i)
23475 // Check that the shuffles are both shuffling the same vectors.
23476 if (!(A == C && B == D) && !(A == D && B == C))
23479 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23480 if (!A.getNode() && !B.getNode())
23483 // If A and B occur in reverse order in RHS, then "swap" them (which means
23484 // rewriting the mask).
23486 ShuffleVectorSDNode::commuteMask(RMask);
23488 // At this point LHS and RHS are equivalent to
23489 // LHS = VECTOR_SHUFFLE A, B, LMask
23490 // RHS = VECTOR_SHUFFLE A, B, RMask
23491 // Check that the masks correspond to performing a horizontal operation.
23492 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23493 for (unsigned i = 0; i != NumLaneElts; ++i) {
23494 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23496 // Ignore any UNDEF components.
23497 if (LIdx < 0 || RIdx < 0 ||
23498 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23499 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23502 // Check that successive elements are being operated on. If not, this is
23503 // not a horizontal operation.
23504 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23505 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23506 if (!(LIdx == Index && RIdx == Index + 1) &&
23507 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23512 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23513 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23517 /// Do target-specific dag combines on floating point adds.
23518 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23519 const X86Subtarget *Subtarget) {
23520 EVT VT = N->getValueType(0);
23521 SDValue LHS = N->getOperand(0);
23522 SDValue RHS = N->getOperand(1);
23524 // Try to synthesize horizontal adds from adds of shuffles.
23525 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23526 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23527 isHorizontalBinOp(LHS, RHS, true))
23528 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23532 /// Do target-specific dag combines on floating point subs.
23533 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23534 const X86Subtarget *Subtarget) {
23535 EVT VT = N->getValueType(0);
23536 SDValue LHS = N->getOperand(0);
23537 SDValue RHS = N->getOperand(1);
23539 // Try to synthesize horizontal subs from subs of shuffles.
23540 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23541 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23542 isHorizontalBinOp(LHS, RHS, false))
23543 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23547 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
23548 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23549 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23551 // F[X]OR(0.0, x) -> x
23552 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23553 if (C->getValueAPF().isPosZero())
23554 return N->getOperand(1);
23556 // F[X]OR(x, 0.0) -> x
23557 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23558 if (C->getValueAPF().isPosZero())
23559 return N->getOperand(0);
23563 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
23564 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23565 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23567 // Only perform optimizations if UnsafeMath is used.
23568 if (!DAG.getTarget().Options.UnsafeFPMath)
23571 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23572 // into FMINC and FMAXC, which are Commutative operations.
23573 unsigned NewOp = 0;
23574 switch (N->getOpcode()) {
23575 default: llvm_unreachable("unknown opcode");
23576 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23577 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23580 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23581 N->getOperand(0), N->getOperand(1));
23584 /// Do target-specific dag combines on X86ISD::FAND nodes.
23585 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23586 // FAND(0.0, x) -> 0.0
23587 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23588 if (C->getValueAPF().isPosZero())
23589 return N->getOperand(0);
23591 // FAND(x, 0.0) -> 0.0
23592 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23593 if (C->getValueAPF().isPosZero())
23594 return N->getOperand(1);
23599 /// Do target-specific dag combines on X86ISD::FANDN nodes
23600 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23601 // FANDN(0.0, x) -> x
23602 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23603 if (C->getValueAPF().isPosZero())
23604 return N->getOperand(1);
23606 // FANDN(x, 0.0) -> 0.0
23607 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23608 if (C->getValueAPF().isPosZero())
23609 return N->getOperand(1);
23614 static SDValue PerformBTCombine(SDNode *N,
23616 TargetLowering::DAGCombinerInfo &DCI) {
23617 // BT ignores high bits in the bit index operand.
23618 SDValue Op1 = N->getOperand(1);
23619 if (Op1.hasOneUse()) {
23620 unsigned BitWidth = Op1.getValueSizeInBits();
23621 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23622 APInt KnownZero, KnownOne;
23623 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23624 !DCI.isBeforeLegalizeOps());
23625 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23626 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23627 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23628 DCI.CommitTargetLoweringOpt(TLO);
23633 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23634 SDValue Op = N->getOperand(0);
23635 if (Op.getOpcode() == ISD::BITCAST)
23636 Op = Op.getOperand(0);
23637 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23638 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23639 VT.getVectorElementType().getSizeInBits() ==
23640 OpVT.getVectorElementType().getSizeInBits()) {
23641 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23646 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23647 const X86Subtarget *Subtarget) {
23648 EVT VT = N->getValueType(0);
23649 if (!VT.isVector())
23652 SDValue N0 = N->getOperand(0);
23653 SDValue N1 = N->getOperand(1);
23654 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23657 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23658 // both SSE and AVX2 since there is no sign-extended shift right
23659 // operation on a vector with 64-bit elements.
23660 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23661 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23662 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23663 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23664 SDValue N00 = N0.getOperand(0);
23666 // EXTLOAD has a better solution on AVX2,
23667 // it may be replaced with X86ISD::VSEXT node.
23668 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23669 if (!ISD::isNormalLoad(N00.getNode()))
23672 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23673 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23675 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23681 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23682 TargetLowering::DAGCombinerInfo &DCI,
23683 const X86Subtarget *Subtarget) {
23684 SDValue N0 = N->getOperand(0);
23685 EVT VT = N->getValueType(0);
23688 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
23689 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
23690 // This exposes the sext to the sdivrem lowering, so that it directly extends
23691 // from AH (which we otherwise need to do contortions to access).
23692 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
23693 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
23694 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23695 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
23696 N0.getOperand(0), N0.getOperand(1));
23697 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23698 return R.getValue(1);
23701 if (!DCI.isBeforeLegalizeOps()) {
23702 if (N0.getValueType() == MVT::i1) {
23703 SDValue Zero = DAG.getConstant(0, dl, VT);
23705 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl, VT);
23706 return DAG.getNode(ISD::SELECT, dl, VT, N0, AllOnes, Zero);
23711 if (!Subtarget->hasFp256())
23714 if (VT.isVector() && VT.getSizeInBits() == 256) {
23715 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23723 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
23724 const X86Subtarget* Subtarget) {
23726 EVT VT = N->getValueType(0);
23728 // Let legalize expand this if it isn't a legal type yet.
23729 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
23732 EVT ScalarVT = VT.getScalarType();
23733 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
23734 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
23737 SDValue A = N->getOperand(0);
23738 SDValue B = N->getOperand(1);
23739 SDValue C = N->getOperand(2);
23741 bool NegA = (A.getOpcode() == ISD::FNEG);
23742 bool NegB = (B.getOpcode() == ISD::FNEG);
23743 bool NegC = (C.getOpcode() == ISD::FNEG);
23745 // Negative multiplication when NegA xor NegB
23746 bool NegMul = (NegA != NegB);
23748 A = A.getOperand(0);
23750 B = B.getOperand(0);
23752 C = C.getOperand(0);
23756 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
23758 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
23760 return DAG.getNode(Opcode, dl, VT, A, B, C);
23763 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
23764 TargetLowering::DAGCombinerInfo &DCI,
23765 const X86Subtarget *Subtarget) {
23766 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
23767 // (and (i32 x86isd::setcc_carry), 1)
23768 // This eliminates the zext. This transformation is necessary because
23769 // ISD::SETCC is always legalized to i8.
23771 SDValue N0 = N->getOperand(0);
23772 EVT VT = N->getValueType(0);
23774 if (N0.getOpcode() == ISD::AND &&
23776 N0.getOperand(0).hasOneUse()) {
23777 SDValue N00 = N0.getOperand(0);
23778 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23779 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23780 if (!C || C->getZExtValue() != 1)
23782 return DAG.getNode(ISD::AND, dl, VT,
23783 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23784 N00.getOperand(0), N00.getOperand(1)),
23785 DAG.getConstant(1, dl, VT));
23789 if (N0.getOpcode() == ISD::TRUNCATE &&
23791 N0.getOperand(0).hasOneUse()) {
23792 SDValue N00 = N0.getOperand(0);
23793 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23794 return DAG.getNode(ISD::AND, dl, VT,
23795 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23796 N00.getOperand(0), N00.getOperand(1)),
23797 DAG.getConstant(1, dl, VT));
23800 if (VT.is256BitVector()) {
23801 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23806 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
23807 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
23808 // This exposes the zext to the udivrem lowering, so that it directly extends
23809 // from AH (which we otherwise need to do contortions to access).
23810 if (N0.getOpcode() == ISD::UDIVREM &&
23811 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
23812 (VT == MVT::i32 || VT == MVT::i64)) {
23813 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23814 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
23815 N0.getOperand(0), N0.getOperand(1));
23816 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23817 return R.getValue(1);
23823 // Optimize x == -y --> x+y == 0
23824 // x != -y --> x+y != 0
23825 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
23826 const X86Subtarget* Subtarget) {
23827 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
23828 SDValue LHS = N->getOperand(0);
23829 SDValue RHS = N->getOperand(1);
23830 EVT VT = N->getValueType(0);
23833 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
23834 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
23835 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
23836 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
23837 LHS.getOperand(1));
23838 return DAG.getSetCC(DL, N->getValueType(0), addV,
23839 DAG.getConstant(0, DL, addV.getValueType()), CC);
23841 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
23842 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
23843 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
23844 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
23845 RHS.getOperand(1));
23846 return DAG.getSetCC(DL, N->getValueType(0), addV,
23847 DAG.getConstant(0, DL, addV.getValueType()), CC);
23850 if (VT.getScalarType() == MVT::i1 &&
23851 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
23853 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23854 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23855 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23857 if (!IsSEXT0 || !IsVZero1) {
23858 // Swap the operands and update the condition code.
23859 std::swap(LHS, RHS);
23860 CC = ISD::getSetCCSwappedOperands(CC);
23862 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23863 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23864 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23867 if (IsSEXT0 && IsVZero1) {
23868 assert(VT == LHS.getOperand(0).getValueType() &&
23869 "Uexpected operand type");
23870 if (CC == ISD::SETGT)
23871 return DAG.getConstant(0, DL, VT);
23872 if (CC == ISD::SETLE)
23873 return DAG.getConstant(1, DL, VT);
23874 if (CC == ISD::SETEQ || CC == ISD::SETGE)
23875 return DAG.getNOT(DL, LHS.getOperand(0), VT);
23877 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
23878 "Unexpected condition code!");
23879 return LHS.getOperand(0);
23886 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
23887 SelectionDAG &DAG) {
23889 MVT VT = Load->getSimpleValueType(0);
23890 MVT EVT = VT.getVectorElementType();
23891 SDValue Addr = Load->getOperand(1);
23892 SDValue NewAddr = DAG.getNode(
23893 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
23894 DAG.getConstant(Index * EVT.getStoreSize(), dl,
23895 Addr.getSimpleValueType()));
23898 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
23899 DAG.getMachineFunction().getMachineMemOperand(
23900 Load->getMemOperand(), 0, EVT.getStoreSize()));
23904 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
23905 const X86Subtarget *Subtarget) {
23907 MVT VT = N->getOperand(1)->getSimpleValueType(0);
23908 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
23909 "X86insertps is only defined for v4x32");
23911 SDValue Ld = N->getOperand(1);
23912 if (MayFoldLoad(Ld)) {
23913 // Extract the countS bits from the immediate so we can get the proper
23914 // address when narrowing the vector load to a specific element.
23915 // When the second source op is a memory address, insertps doesn't use
23916 // countS and just gets an f32 from that address.
23917 unsigned DestIndex =
23918 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
23920 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
23922 // Create this as a scalar to vector to match the instruction pattern.
23923 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
23924 // countS bits are ignored when loading from memory on insertps, which
23925 // means we don't need to explicitly set them to 0.
23926 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
23927 LoadScalarToVector, N->getOperand(2));
23932 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
23933 SDValue V0 = N->getOperand(0);
23934 SDValue V1 = N->getOperand(1);
23936 EVT VT = N->getValueType(0);
23938 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23939 // operands and changing the mask to 1. This saves us a bunch of
23940 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23941 // x86InstrInfo knows how to commute this back after instruction selection
23942 // if it would help register allocation.
23944 // TODO: If optimizing for size or a processor that doesn't suffer from
23945 // partial register update stalls, this should be transformed into a MOVSD
23946 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23948 if (VT == MVT::v2f64)
23949 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23950 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23951 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
23952 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23958 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
23959 // as "sbb reg,reg", since it can be extended without zext and produces
23960 // an all-ones bit which is more useful than 0/1 in some cases.
23961 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
23964 return DAG.getNode(ISD::AND, DL, VT,
23965 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23966 DAG.getConstant(X86::COND_B, DL, MVT::i8),
23968 DAG.getConstant(1, DL, VT));
23969 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
23970 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
23971 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23972 DAG.getConstant(X86::COND_B, DL, MVT::i8),
23976 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
23977 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
23978 TargetLowering::DAGCombinerInfo &DCI,
23979 const X86Subtarget *Subtarget) {
23981 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
23982 SDValue EFLAGS = N->getOperand(1);
23984 if (CC == X86::COND_A) {
23985 // Try to convert COND_A into COND_B in an attempt to facilitate
23986 // materializing "setb reg".
23988 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
23989 // cannot take an immediate as its first operand.
23991 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
23992 EFLAGS.getValueType().isInteger() &&
23993 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
23994 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23995 EFLAGS.getNode()->getVTList(),
23996 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23997 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23998 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
24002 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
24003 // a zext and produces an all-ones bit which is more useful than 0/1 in some
24005 if (CC == X86::COND_B)
24006 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
24010 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24011 if (Flags.getNode()) {
24012 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24013 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
24019 // Optimize branch condition evaluation.
24021 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
24022 TargetLowering::DAGCombinerInfo &DCI,
24023 const X86Subtarget *Subtarget) {
24025 SDValue Chain = N->getOperand(0);
24026 SDValue Dest = N->getOperand(1);
24027 SDValue EFLAGS = N->getOperand(3);
24028 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
24032 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24033 if (Flags.getNode()) {
24034 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24035 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
24042 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
24043 SelectionDAG &DAG) {
24044 // Take advantage of vector comparisons producing 0 or -1 in each lane to
24045 // optimize away operation when it's from a constant.
24047 // The general transformation is:
24048 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
24049 // AND(VECTOR_CMP(x,y), constant2)
24050 // constant2 = UNARYOP(constant)
24052 // Early exit if this isn't a vector operation, the operand of the
24053 // unary operation isn't a bitwise AND, or if the sizes of the operations
24054 // aren't the same.
24055 EVT VT = N->getValueType(0);
24056 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
24057 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
24058 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
24061 // Now check that the other operand of the AND is a constant. We could
24062 // make the transformation for non-constant splats as well, but it's unclear
24063 // that would be a benefit as it would not eliminate any operations, just
24064 // perform one more step in scalar code before moving to the vector unit.
24065 if (BuildVectorSDNode *BV =
24066 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
24067 // Bail out if the vector isn't a constant.
24068 if (!BV->isConstant())
24071 // Everything checks out. Build up the new and improved node.
24073 EVT IntVT = BV->getValueType(0);
24074 // Create a new constant of the appropriate type for the transformed
24076 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
24077 // The AND node needs bitcasts to/from an integer vector type around it.
24078 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
24079 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
24080 N->getOperand(0)->getOperand(0), MaskConst);
24081 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
24088 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
24089 const X86Subtarget *Subtarget) {
24090 // First try to optimize away the conversion entirely when it's
24091 // conditionally from a constant. Vectors only.
24092 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
24093 if (Res != SDValue())
24096 // Now move on to more general possibilities.
24097 SDValue Op0 = N->getOperand(0);
24098 EVT InVT = Op0->getValueType(0);
24100 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
24101 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
24103 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
24104 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
24105 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
24108 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
24109 // a 32-bit target where SSE doesn't support i64->FP operations.
24110 if (Op0.getOpcode() == ISD::LOAD) {
24111 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
24112 EVT VT = Ld->getValueType(0);
24114 // This transformation is not supported if the result type is f16
24115 if (N->getValueType(0) == MVT::f16)
24118 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
24119 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
24120 !Subtarget->is64Bit() && VT == MVT::i64) {
24121 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
24122 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
24123 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
24130 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
24131 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
24132 X86TargetLowering::DAGCombinerInfo &DCI) {
24133 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
24134 // the result is either zero or one (depending on the input carry bit).
24135 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
24136 if (X86::isZeroNode(N->getOperand(0)) &&
24137 X86::isZeroNode(N->getOperand(1)) &&
24138 // We don't have a good way to replace an EFLAGS use, so only do this when
24140 SDValue(N, 1).use_empty()) {
24142 EVT VT = N->getValueType(0);
24143 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
24144 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
24145 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
24146 DAG.getConstant(X86::COND_B, DL,
24149 DAG.getConstant(1, DL, VT));
24150 return DCI.CombineTo(N, Res1, CarryOut);
24156 // fold (add Y, (sete X, 0)) -> adc 0, Y
24157 // (add Y, (setne X, 0)) -> sbb -1, Y
24158 // (sub (sete X, 0), Y) -> sbb 0, Y
24159 // (sub (setne X, 0), Y) -> adc -1, Y
24160 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
24163 // Look through ZExts.
24164 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
24165 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
24168 SDValue SetCC = Ext.getOperand(0);
24169 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
24172 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
24173 if (CC != X86::COND_E && CC != X86::COND_NE)
24176 SDValue Cmp = SetCC.getOperand(1);
24177 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
24178 !X86::isZeroNode(Cmp.getOperand(1)) ||
24179 !Cmp.getOperand(0).getValueType().isInteger())
24182 SDValue CmpOp0 = Cmp.getOperand(0);
24183 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
24184 DAG.getConstant(1, DL, CmpOp0.getValueType()));
24186 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
24187 if (CC == X86::COND_NE)
24188 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
24189 DL, OtherVal.getValueType(), OtherVal,
24190 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
24192 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
24193 DL, OtherVal.getValueType(), OtherVal,
24194 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
24197 /// PerformADDCombine - Do target-specific dag combines on integer adds.
24198 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
24199 const X86Subtarget *Subtarget) {
24200 EVT VT = N->getValueType(0);
24201 SDValue Op0 = N->getOperand(0);
24202 SDValue Op1 = N->getOperand(1);
24204 // Try to synthesize horizontal adds from adds of shuffles.
24205 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24206 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24207 isHorizontalBinOp(Op0, Op1, true))
24208 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
24210 return OptimizeConditionalInDecrement(N, DAG);
24213 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
24214 const X86Subtarget *Subtarget) {
24215 SDValue Op0 = N->getOperand(0);
24216 SDValue Op1 = N->getOperand(1);
24218 // X86 can't encode an immediate LHS of a sub. See if we can push the
24219 // negation into a preceding instruction.
24220 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
24221 // If the RHS of the sub is a XOR with one use and a constant, invert the
24222 // immediate. Then add one to the LHS of the sub so we can turn
24223 // X-Y -> X+~Y+1, saving one register.
24224 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
24225 isa<ConstantSDNode>(Op1.getOperand(1))) {
24226 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
24227 EVT VT = Op0.getValueType();
24228 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
24230 DAG.getConstant(~XorC, SDLoc(Op1), VT));
24231 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
24232 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
24236 // Try to synthesize horizontal adds from adds of shuffles.
24237 EVT VT = N->getValueType(0);
24238 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24239 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24240 isHorizontalBinOp(Op0, Op1, true))
24241 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
24243 return OptimizeConditionalInDecrement(N, DAG);
24246 /// performVZEXTCombine - Performs build vector combines
24247 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
24248 TargetLowering::DAGCombinerInfo &DCI,
24249 const X86Subtarget *Subtarget) {
24251 MVT VT = N->getSimpleValueType(0);
24252 SDValue Op = N->getOperand(0);
24253 MVT OpVT = Op.getSimpleValueType();
24254 MVT OpEltVT = OpVT.getVectorElementType();
24255 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
24257 // (vzext (bitcast (vzext (x)) -> (vzext x)
24259 while (V.getOpcode() == ISD::BITCAST)
24260 V = V.getOperand(0);
24262 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
24263 MVT InnerVT = V.getSimpleValueType();
24264 MVT InnerEltVT = InnerVT.getVectorElementType();
24266 // If the element sizes match exactly, we can just do one larger vzext. This
24267 // is always an exact type match as vzext operates on integer types.
24268 if (OpEltVT == InnerEltVT) {
24269 assert(OpVT == InnerVT && "Types must match for vzext!");
24270 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
24273 // The only other way we can combine them is if only a single element of the
24274 // inner vzext is used in the input to the outer vzext.
24275 if (InnerEltVT.getSizeInBits() < InputBits)
24278 // In this case, the inner vzext is completely dead because we're going to
24279 // only look at bits inside of the low element. Just do the outer vzext on
24280 // a bitcast of the input to the inner.
24281 return DAG.getNode(X86ISD::VZEXT, DL, VT,
24282 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
24285 // Check if we can bypass extracting and re-inserting an element of an input
24286 // vector. Essentialy:
24287 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
24288 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
24289 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
24290 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
24291 SDValue ExtractedV = V.getOperand(0);
24292 SDValue OrigV = ExtractedV.getOperand(0);
24293 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
24294 if (ExtractIdx->getZExtValue() == 0) {
24295 MVT OrigVT = OrigV.getSimpleValueType();
24296 // Extract a subvector if necessary...
24297 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
24298 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
24299 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
24300 OrigVT.getVectorNumElements() / Ratio);
24301 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
24302 DAG.getIntPtrConstant(0, DL));
24304 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
24305 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
24312 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24313 DAGCombinerInfo &DCI) const {
24314 SelectionDAG &DAG = DCI.DAG;
24315 switch (N->getOpcode()) {
24317 case ISD::EXTRACT_VECTOR_ELT:
24318 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24321 case X86ISD::SHRUNKBLEND:
24322 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24323 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
24324 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24325 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24326 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24327 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24328 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24331 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24332 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24333 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24334 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24335 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24336 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
24337 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24338 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
24339 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
24340 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24341 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24343 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24345 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24346 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24347 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24348 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24349 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24350 case ISD::ANY_EXTEND:
24351 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24352 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24353 case ISD::SIGN_EXTEND_INREG:
24354 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24355 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24356 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24357 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24358 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24359 case X86ISD::SHUFP: // Handle all target specific shuffles
24360 case X86ISD::PALIGNR:
24361 case X86ISD::UNPCKH:
24362 case X86ISD::UNPCKL:
24363 case X86ISD::MOVHLPS:
24364 case X86ISD::MOVLHPS:
24365 case X86ISD::PSHUFB:
24366 case X86ISD::PSHUFD:
24367 case X86ISD::PSHUFHW:
24368 case X86ISD::PSHUFLW:
24369 case X86ISD::MOVSS:
24370 case X86ISD::MOVSD:
24371 case X86ISD::VPERMILPI:
24372 case X86ISD::VPERM2X128:
24373 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24374 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24375 case ISD::INTRINSIC_WO_CHAIN:
24376 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24377 case X86ISD::INSERTPS: {
24378 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
24379 return PerformINSERTPSCombine(N, DAG, Subtarget);
24382 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
24388 /// isTypeDesirableForOp - Return true if the target has native support for
24389 /// the specified value type and it is 'desirable' to use the type for the
24390 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24391 /// instruction encodings are longer and some i16 instructions are slow.
24392 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24393 if (!isTypeLegal(VT))
24395 if (VT != MVT::i16)
24402 case ISD::SIGN_EXTEND:
24403 case ISD::ZERO_EXTEND:
24404 case ISD::ANY_EXTEND:
24417 /// IsDesirableToPromoteOp - This method query the target whether it is
24418 /// beneficial for dag combiner to promote the specified node. If true, it
24419 /// should return the desired promotion type by reference.
24420 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24421 EVT VT = Op.getValueType();
24422 if (VT != MVT::i16)
24425 bool Promote = false;
24426 bool Commute = false;
24427 switch (Op.getOpcode()) {
24430 LoadSDNode *LD = cast<LoadSDNode>(Op);
24431 // If the non-extending load has a single use and it's not live out, then it
24432 // might be folded.
24433 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24434 Op.hasOneUse()*/) {
24435 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24436 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24437 // The only case where we'd want to promote LOAD (rather then it being
24438 // promoted as an operand is when it's only use is liveout.
24439 if (UI->getOpcode() != ISD::CopyToReg)
24446 case ISD::SIGN_EXTEND:
24447 case ISD::ZERO_EXTEND:
24448 case ISD::ANY_EXTEND:
24453 SDValue N0 = Op.getOperand(0);
24454 // Look out for (store (shl (load), x)).
24455 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24468 SDValue N0 = Op.getOperand(0);
24469 SDValue N1 = Op.getOperand(1);
24470 if (!Commute && MayFoldLoad(N1))
24472 // Avoid disabling potential load folding opportunities.
24473 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24475 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24485 //===----------------------------------------------------------------------===//
24486 // X86 Inline Assembly Support
24487 //===----------------------------------------------------------------------===//
24489 // Helper to match a string separated by whitespace.
24490 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
24491 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
24493 for (StringRef Piece : Pieces) {
24494 if (!S.startswith(Piece)) // Check if the piece matches.
24497 S = S.substr(Piece.size());
24498 StringRef::size_type Pos = S.find_first_not_of(" \t");
24499 if (Pos == 0) // We matched a prefix.
24508 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24510 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24511 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24512 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24513 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24515 if (AsmPieces.size() == 3)
24517 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24524 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24525 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24527 std::string AsmStr = IA->getAsmString();
24529 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24530 if (!Ty || Ty->getBitWidth() % 16 != 0)
24533 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24534 SmallVector<StringRef, 4> AsmPieces;
24535 SplitString(AsmStr, AsmPieces, ";\n");
24537 switch (AsmPieces.size()) {
24538 default: return false;
24540 // FIXME: this should verify that we are targeting a 486 or better. If not,
24541 // we will turn this bswap into something that will be lowered to logical
24542 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24543 // lower so don't worry about this.
24545 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
24546 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
24547 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
24548 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
24549 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
24550 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
24551 // No need to check constraints, nothing other than the equivalent of
24552 // "=r,0" would be valid here.
24553 return IntrinsicLowering::LowerToByteSwap(CI);
24556 // rorw $$8, ${0:w} --> llvm.bswap.i16
24557 if (CI->getType()->isIntegerTy(16) &&
24558 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24559 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
24560 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
24562 const std::string &ConstraintsStr = IA->getConstraintString();
24563 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24564 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24565 if (clobbersFlagRegisters(AsmPieces))
24566 return IntrinsicLowering::LowerToByteSwap(CI);
24570 if (CI->getType()->isIntegerTy(32) &&
24571 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24572 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
24573 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
24574 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
24576 const std::string &ConstraintsStr = IA->getConstraintString();
24577 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24578 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24579 if (clobbersFlagRegisters(AsmPieces))
24580 return IntrinsicLowering::LowerToByteSwap(CI);
24583 if (CI->getType()->isIntegerTy(64)) {
24584 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24585 if (Constraints.size() >= 2 &&
24586 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24587 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24588 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24589 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
24590 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
24591 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
24592 return IntrinsicLowering::LowerToByteSwap(CI);
24600 /// getConstraintType - Given a constraint letter, return the type of
24601 /// constraint it is for this target.
24602 X86TargetLowering::ConstraintType
24603 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24604 if (Constraint.size() == 1) {
24605 switch (Constraint[0]) {
24616 return C_RegisterClass;
24640 return TargetLowering::getConstraintType(Constraint);
24643 /// Examine constraint type and operand type and determine a weight value.
24644 /// This object must already have been set up with the operand type
24645 /// and the current alternative constraint selected.
24646 TargetLowering::ConstraintWeight
24647 X86TargetLowering::getSingleConstraintMatchWeight(
24648 AsmOperandInfo &info, const char *constraint) const {
24649 ConstraintWeight weight = CW_Invalid;
24650 Value *CallOperandVal = info.CallOperandVal;
24651 // If we don't have a value, we can't do a match,
24652 // but allow it at the lowest weight.
24653 if (!CallOperandVal)
24655 Type *type = CallOperandVal->getType();
24656 // Look at the constraint type.
24657 switch (*constraint) {
24659 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
24670 if (CallOperandVal->getType()->isIntegerTy())
24671 weight = CW_SpecificReg;
24676 if (type->isFloatingPointTy())
24677 weight = CW_SpecificReg;
24680 if (type->isX86_MMXTy() && Subtarget->hasMMX())
24681 weight = CW_SpecificReg;
24685 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
24686 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
24687 weight = CW_Register;
24690 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
24691 if (C->getZExtValue() <= 31)
24692 weight = CW_Constant;
24696 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24697 if (C->getZExtValue() <= 63)
24698 weight = CW_Constant;
24702 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24703 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
24704 weight = CW_Constant;
24708 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24709 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
24710 weight = CW_Constant;
24714 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24715 if (C->getZExtValue() <= 3)
24716 weight = CW_Constant;
24720 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24721 if (C->getZExtValue() <= 0xff)
24722 weight = CW_Constant;
24727 if (isa<ConstantFP>(CallOperandVal)) {
24728 weight = CW_Constant;
24732 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24733 if ((C->getSExtValue() >= -0x80000000LL) &&
24734 (C->getSExtValue() <= 0x7fffffffLL))
24735 weight = CW_Constant;
24739 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24740 if (C->getZExtValue() <= 0xffffffff)
24741 weight = CW_Constant;
24748 /// LowerXConstraint - try to replace an X constraint, which matches anything,
24749 /// with another that has more specific requirements based on the type of the
24750 /// corresponding operand.
24751 const char *X86TargetLowering::
24752 LowerXConstraint(EVT ConstraintVT) const {
24753 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
24754 // 'f' like normal targets.
24755 if (ConstraintVT.isFloatingPoint()) {
24756 if (Subtarget->hasSSE2())
24758 if (Subtarget->hasSSE1())
24762 return TargetLowering::LowerXConstraint(ConstraintVT);
24765 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
24766 /// vector. If it is invalid, don't add anything to Ops.
24767 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
24768 std::string &Constraint,
24769 std::vector<SDValue>&Ops,
24770 SelectionDAG &DAG) const {
24773 // Only support length 1 constraints for now.
24774 if (Constraint.length() > 1) return;
24776 char ConstraintLetter = Constraint[0];
24777 switch (ConstraintLetter) {
24780 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24781 if (C->getZExtValue() <= 31) {
24782 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24783 Op.getValueType());
24789 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24790 if (C->getZExtValue() <= 63) {
24791 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24792 Op.getValueType());
24798 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24799 if (isInt<8>(C->getSExtValue())) {
24800 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24801 Op.getValueType());
24807 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24808 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
24809 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
24810 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
24811 Op.getValueType());
24817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24818 if (C->getZExtValue() <= 3) {
24819 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24820 Op.getValueType());
24826 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24827 if (C->getZExtValue() <= 255) {
24828 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24829 Op.getValueType());
24835 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24836 if (C->getZExtValue() <= 127) {
24837 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24838 Op.getValueType());
24844 // 32-bit signed value
24845 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24846 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24847 C->getSExtValue())) {
24848 // Widen to 64 bits here to get it sign extended.
24849 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
24852 // FIXME gcc accepts some relocatable values here too, but only in certain
24853 // memory models; it's complicated.
24858 // 32-bit unsigned value
24859 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24860 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24861 C->getZExtValue())) {
24862 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24863 Op.getValueType());
24867 // FIXME gcc accepts some relocatable values here too, but only in certain
24868 // memory models; it's complicated.
24872 // Literal immediates are always ok.
24873 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
24874 // Widen to 64 bits here to get it sign extended.
24875 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
24879 // In any sort of PIC mode addresses need to be computed at runtime by
24880 // adding in a register or some sort of table lookup. These can't
24881 // be used as immediates.
24882 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
24885 // If we are in non-pic codegen mode, we allow the address of a global (with
24886 // an optional displacement) to be used with 'i'.
24887 GlobalAddressSDNode *GA = nullptr;
24888 int64_t Offset = 0;
24890 // Match either (GA), (GA+C), (GA+C1+C2), etc.
24892 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
24893 Offset += GA->getOffset();
24895 } else if (Op.getOpcode() == ISD::ADD) {
24896 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24897 Offset += C->getZExtValue();
24898 Op = Op.getOperand(0);
24901 } else if (Op.getOpcode() == ISD::SUB) {
24902 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24903 Offset += -C->getZExtValue();
24904 Op = Op.getOperand(0);
24909 // Otherwise, this isn't something we can handle, reject it.
24913 const GlobalValue *GV = GA->getGlobal();
24914 // If we require an extra load to get this address, as in PIC mode, we
24915 // can't accept it.
24916 if (isGlobalStubReference(
24917 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
24920 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
24921 GA->getValueType(0), Offset);
24926 if (Result.getNode()) {
24927 Ops.push_back(Result);
24930 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
24933 std::pair<unsigned, const TargetRegisterClass *>
24934 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
24935 const std::string &Constraint,
24937 // First, see if this is a constraint that directly corresponds to an LLVM
24939 if (Constraint.size() == 1) {
24940 // GCC Constraint Letters
24941 switch (Constraint[0]) {
24943 // TODO: Slight differences here in allocation order and leaving
24944 // RIP in the class. Do they matter any more here than they do
24945 // in the normal allocation?
24946 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
24947 if (Subtarget->is64Bit()) {
24948 if (VT == MVT::i32 || VT == MVT::f32)
24949 return std::make_pair(0U, &X86::GR32RegClass);
24950 if (VT == MVT::i16)
24951 return std::make_pair(0U, &X86::GR16RegClass);
24952 if (VT == MVT::i8 || VT == MVT::i1)
24953 return std::make_pair(0U, &X86::GR8RegClass);
24954 if (VT == MVT::i64 || VT == MVT::f64)
24955 return std::make_pair(0U, &X86::GR64RegClass);
24958 // 32-bit fallthrough
24959 case 'Q': // Q_REGS
24960 if (VT == MVT::i32 || VT == MVT::f32)
24961 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
24962 if (VT == MVT::i16)
24963 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
24964 if (VT == MVT::i8 || VT == MVT::i1)
24965 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
24966 if (VT == MVT::i64)
24967 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
24969 case 'r': // GENERAL_REGS
24970 case 'l': // INDEX_REGS
24971 if (VT == MVT::i8 || VT == MVT::i1)
24972 return std::make_pair(0U, &X86::GR8RegClass);
24973 if (VT == MVT::i16)
24974 return std::make_pair(0U, &X86::GR16RegClass);
24975 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
24976 return std::make_pair(0U, &X86::GR32RegClass);
24977 return std::make_pair(0U, &X86::GR64RegClass);
24978 case 'R': // LEGACY_REGS
24979 if (VT == MVT::i8 || VT == MVT::i1)
24980 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
24981 if (VT == MVT::i16)
24982 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
24983 if (VT == MVT::i32 || !Subtarget->is64Bit())
24984 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
24985 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
24986 case 'f': // FP Stack registers.
24987 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
24988 // value to the correct fpstack register class.
24989 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
24990 return std::make_pair(0U, &X86::RFP32RegClass);
24991 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
24992 return std::make_pair(0U, &X86::RFP64RegClass);
24993 return std::make_pair(0U, &X86::RFP80RegClass);
24994 case 'y': // MMX_REGS if MMX allowed.
24995 if (!Subtarget->hasMMX()) break;
24996 return std::make_pair(0U, &X86::VR64RegClass);
24997 case 'Y': // SSE_REGS if SSE2 allowed
24998 if (!Subtarget->hasSSE2()) break;
25000 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
25001 if (!Subtarget->hasSSE1()) break;
25003 switch (VT.SimpleTy) {
25005 // Scalar SSE types.
25008 return std::make_pair(0U, &X86::FR32RegClass);
25011 return std::make_pair(0U, &X86::FR64RegClass);
25019 return std::make_pair(0U, &X86::VR128RegClass);
25027 return std::make_pair(0U, &X86::VR256RegClass);
25032 return std::make_pair(0U, &X86::VR512RegClass);
25038 // Use the default implementation in TargetLowering to convert the register
25039 // constraint into a member of a register class.
25040 std::pair<unsigned, const TargetRegisterClass*> Res;
25041 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
25043 // Not found as a standard register?
25045 // Map st(0) -> st(7) -> ST0
25046 if (Constraint.size() == 7 && Constraint[0] == '{' &&
25047 tolower(Constraint[1]) == 's' &&
25048 tolower(Constraint[2]) == 't' &&
25049 Constraint[3] == '(' &&
25050 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
25051 Constraint[5] == ')' &&
25052 Constraint[6] == '}') {
25054 Res.first = X86::FP0+Constraint[4]-'0';
25055 Res.second = &X86::RFP80RegClass;
25059 // GCC allows "st(0)" to be called just plain "st".
25060 if (StringRef("{st}").equals_lower(Constraint)) {
25061 Res.first = X86::FP0;
25062 Res.second = &X86::RFP80RegClass;
25067 if (StringRef("{flags}").equals_lower(Constraint)) {
25068 Res.first = X86::EFLAGS;
25069 Res.second = &X86::CCRRegClass;
25073 // 'A' means EAX + EDX.
25074 if (Constraint == "A") {
25075 Res.first = X86::EAX;
25076 Res.second = &X86::GR32_ADRegClass;
25082 // Otherwise, check to see if this is a register class of the wrong value
25083 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
25084 // turn into {ax},{dx}.
25085 if (Res.second->hasType(VT))
25086 return Res; // Correct type already, nothing to do.
25088 // All of the single-register GCC register classes map their values onto
25089 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
25090 // really want an 8-bit or 32-bit register, map to the appropriate register
25091 // class and return the appropriate register.
25092 if (Res.second == &X86::GR16RegClass) {
25093 if (VT == MVT::i8 || VT == MVT::i1) {
25094 unsigned DestReg = 0;
25095 switch (Res.first) {
25097 case X86::AX: DestReg = X86::AL; break;
25098 case X86::DX: DestReg = X86::DL; break;
25099 case X86::CX: DestReg = X86::CL; break;
25100 case X86::BX: DestReg = X86::BL; break;
25103 Res.first = DestReg;
25104 Res.second = &X86::GR8RegClass;
25106 } else if (VT == MVT::i32 || VT == MVT::f32) {
25107 unsigned DestReg = 0;
25108 switch (Res.first) {
25110 case X86::AX: DestReg = X86::EAX; break;
25111 case X86::DX: DestReg = X86::EDX; break;
25112 case X86::CX: DestReg = X86::ECX; break;
25113 case X86::BX: DestReg = X86::EBX; break;
25114 case X86::SI: DestReg = X86::ESI; break;
25115 case X86::DI: DestReg = X86::EDI; break;
25116 case X86::BP: DestReg = X86::EBP; break;
25117 case X86::SP: DestReg = X86::ESP; break;
25120 Res.first = DestReg;
25121 Res.second = &X86::GR32RegClass;
25123 } else if (VT == MVT::i64 || VT == MVT::f64) {
25124 unsigned DestReg = 0;
25125 switch (Res.first) {
25127 case X86::AX: DestReg = X86::RAX; break;
25128 case X86::DX: DestReg = X86::RDX; break;
25129 case X86::CX: DestReg = X86::RCX; break;
25130 case X86::BX: DestReg = X86::RBX; break;
25131 case X86::SI: DestReg = X86::RSI; break;
25132 case X86::DI: DestReg = X86::RDI; break;
25133 case X86::BP: DestReg = X86::RBP; break;
25134 case X86::SP: DestReg = X86::RSP; break;
25137 Res.first = DestReg;
25138 Res.second = &X86::GR64RegClass;
25141 } else if (Res.second == &X86::FR32RegClass ||
25142 Res.second == &X86::FR64RegClass ||
25143 Res.second == &X86::VR128RegClass ||
25144 Res.second == &X86::VR256RegClass ||
25145 Res.second == &X86::FR32XRegClass ||
25146 Res.second == &X86::FR64XRegClass ||
25147 Res.second == &X86::VR128XRegClass ||
25148 Res.second == &X86::VR256XRegClass ||
25149 Res.second == &X86::VR512RegClass) {
25150 // Handle references to XMM physical registers that got mapped into the
25151 // wrong class. This can happen with constraints like {xmm0} where the
25152 // target independent register mapper will just pick the first match it can
25153 // find, ignoring the required type.
25155 if (VT == MVT::f32 || VT == MVT::i32)
25156 Res.second = &X86::FR32RegClass;
25157 else if (VT == MVT::f64 || VT == MVT::i64)
25158 Res.second = &X86::FR64RegClass;
25159 else if (X86::VR128RegClass.hasType(VT))
25160 Res.second = &X86::VR128RegClass;
25161 else if (X86::VR256RegClass.hasType(VT))
25162 Res.second = &X86::VR256RegClass;
25163 else if (X86::VR512RegClass.hasType(VT))
25164 Res.second = &X86::VR512RegClass;
25170 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
25172 // Scaling factors are not free at all.
25173 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
25174 // will take 2 allocations in the out of order engine instead of 1
25175 // for plain addressing mode, i.e. inst (reg1).
25177 // vaddps (%rsi,%drx), %ymm0, %ymm1
25178 // Requires two allocations (one for the load, one for the computation)
25180 // vaddps (%rsi), %ymm0, %ymm1
25181 // Requires just 1 allocation, i.e., freeing allocations for other operations
25182 // and having less micro operations to execute.
25184 // For some X86 architectures, this is even worse because for instance for
25185 // stores, the complex addressing mode forces the instruction to use the
25186 // "load" ports instead of the dedicated "store" port.
25187 // E.g., on Haswell:
25188 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
25189 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
25190 if (isLegalAddressingMode(AM, Ty))
25191 // Scale represents reg2 * scale, thus account for 1
25192 // as soon as we use a second register.
25193 return AM.Scale != 0;
25197 bool X86TargetLowering::isTargetFTOL() const {
25198 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();