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 (!TM.Options.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 (!TM.Options.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 (!TM.Options.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 (TM.Options.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 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
517 // f32 and f64 use SSE.
518 // Set up the FP register classes.
519 addRegisterClass(MVT::f32, &X86::FR32RegClass);
520 addRegisterClass(MVT::f64, &X86::FR64RegClass);
522 // Use ANDPD to simulate FABS.
523 setOperationAction(ISD::FABS , MVT::f64, Custom);
524 setOperationAction(ISD::FABS , MVT::f32, Custom);
526 // Use XORP to simulate FNEG.
527 setOperationAction(ISD::FNEG , MVT::f64, Custom);
528 setOperationAction(ISD::FNEG , MVT::f32, Custom);
530 // Use ANDPD and ORPD to simulate FCOPYSIGN.
531 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
532 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
534 // Lower this to FGETSIGNx86 plus an AND.
535 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
536 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
538 // We don't support sin/cos/fmod
539 setOperationAction(ISD::FSIN , MVT::f64, Expand);
540 setOperationAction(ISD::FCOS , MVT::f64, Expand);
541 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
542 setOperationAction(ISD::FSIN , MVT::f32, Expand);
543 setOperationAction(ISD::FCOS , MVT::f32, Expand);
544 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
546 // Expand FP immediates into loads from the stack, except for the special
548 addLegalFPImmediate(APFloat(+0.0)); // xorpd
549 addLegalFPImmediate(APFloat(+0.0f)); // xorps
550 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
551 // Use SSE for f32, x87 for f64.
552 // Set up the FP register classes.
553 addRegisterClass(MVT::f32, &X86::FR32RegClass);
554 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
556 // Use ANDPS to simulate FABS.
557 setOperationAction(ISD::FABS , MVT::f32, Custom);
559 // Use XORP to simulate FNEG.
560 setOperationAction(ISD::FNEG , MVT::f32, Custom);
562 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
564 // Use ANDPS and ORPS to simulate FCOPYSIGN.
565 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
566 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
568 // We don't support sin/cos/fmod
569 setOperationAction(ISD::FSIN , MVT::f32, Expand);
570 setOperationAction(ISD::FCOS , MVT::f32, Expand);
571 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
573 // Special cases we handle for FP constants.
574 addLegalFPImmediate(APFloat(+0.0f)); // xorps
575 addLegalFPImmediate(APFloat(+0.0)); // FLD0
576 addLegalFPImmediate(APFloat(+1.0)); // FLD1
577 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
578 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
580 if (!TM.Options.UnsafeFPMath) {
581 setOperationAction(ISD::FSIN , MVT::f64, Expand);
582 setOperationAction(ISD::FCOS , MVT::f64, Expand);
583 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
585 } else if (!TM.Options.UseSoftFloat) {
586 // f32 and f64 in x87.
587 // Set up the FP register classes.
588 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
589 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
591 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
592 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
593 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
594 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
596 if (!TM.Options.UnsafeFPMath) {
597 setOperationAction(ISD::FSIN , MVT::f64, Expand);
598 setOperationAction(ISD::FSIN , MVT::f32, Expand);
599 setOperationAction(ISD::FCOS , MVT::f64, Expand);
600 setOperationAction(ISD::FCOS , MVT::f32, Expand);
601 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
602 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
604 addLegalFPImmediate(APFloat(+0.0)); // FLD0
605 addLegalFPImmediate(APFloat(+1.0)); // FLD1
606 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
607 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
608 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
614 // We don't support FMA.
615 setOperationAction(ISD::FMA, MVT::f64, Expand);
616 setOperationAction(ISD::FMA, MVT::f32, Expand);
618 // Long double always uses X87.
619 if (!TM.Options.UseSoftFloat) {
620 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
621 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
622 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
624 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
625 addLegalFPImmediate(TmpFlt); // FLD0
627 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
630 APFloat TmpFlt2(+1.0);
631 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
633 addLegalFPImmediate(TmpFlt2); // FLD1
634 TmpFlt2.changeSign();
635 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
638 if (!TM.Options.UnsafeFPMath) {
639 setOperationAction(ISD::FSIN , MVT::f80, Expand);
640 setOperationAction(ISD::FCOS , MVT::f80, Expand);
641 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
644 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
645 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
646 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
647 setOperationAction(ISD::FRINT, MVT::f80, Expand);
648 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
649 setOperationAction(ISD::FMA, MVT::f80, Expand);
652 // Always use a library call for pow.
653 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
654 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
655 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
657 setOperationAction(ISD::FLOG, MVT::f80, Expand);
658 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
659 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
660 setOperationAction(ISD::FEXP, MVT::f80, Expand);
661 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
662 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
663 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
665 // First set operation action for all vector types to either promote
666 // (for widening) or expand (for scalarization). Then we will selectively
667 // turn on ones that can be effectively codegen'd.
668 for (MVT VT : MVT::vector_valuetypes()) {
669 setOperationAction(ISD::ADD , VT, Expand);
670 setOperationAction(ISD::SUB , VT, Expand);
671 setOperationAction(ISD::FADD, VT, Expand);
672 setOperationAction(ISD::FNEG, VT, Expand);
673 setOperationAction(ISD::FSUB, VT, Expand);
674 setOperationAction(ISD::MUL , VT, Expand);
675 setOperationAction(ISD::FMUL, VT, Expand);
676 setOperationAction(ISD::SDIV, VT, Expand);
677 setOperationAction(ISD::UDIV, VT, Expand);
678 setOperationAction(ISD::FDIV, VT, Expand);
679 setOperationAction(ISD::SREM, VT, Expand);
680 setOperationAction(ISD::UREM, VT, Expand);
681 setOperationAction(ISD::LOAD, VT, Expand);
682 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
683 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
685 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
686 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
687 setOperationAction(ISD::FABS, VT, Expand);
688 setOperationAction(ISD::FSIN, VT, Expand);
689 setOperationAction(ISD::FSINCOS, VT, Expand);
690 setOperationAction(ISD::FCOS, VT, Expand);
691 setOperationAction(ISD::FSINCOS, VT, Expand);
692 setOperationAction(ISD::FREM, VT, Expand);
693 setOperationAction(ISD::FMA, VT, Expand);
694 setOperationAction(ISD::FPOWI, VT, Expand);
695 setOperationAction(ISD::FSQRT, VT, Expand);
696 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
697 setOperationAction(ISD::FFLOOR, VT, Expand);
698 setOperationAction(ISD::FCEIL, VT, Expand);
699 setOperationAction(ISD::FTRUNC, VT, Expand);
700 setOperationAction(ISD::FRINT, VT, Expand);
701 setOperationAction(ISD::FNEARBYINT, VT, Expand);
702 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
703 setOperationAction(ISD::MULHS, VT, Expand);
704 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
705 setOperationAction(ISD::MULHU, VT, Expand);
706 setOperationAction(ISD::SDIVREM, VT, Expand);
707 setOperationAction(ISD::UDIVREM, VT, Expand);
708 setOperationAction(ISD::FPOW, VT, Expand);
709 setOperationAction(ISD::CTPOP, VT, Expand);
710 setOperationAction(ISD::CTTZ, VT, Expand);
711 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
712 setOperationAction(ISD::CTLZ, VT, Expand);
713 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
714 setOperationAction(ISD::SHL, VT, Expand);
715 setOperationAction(ISD::SRA, VT, Expand);
716 setOperationAction(ISD::SRL, VT, Expand);
717 setOperationAction(ISD::ROTL, VT, Expand);
718 setOperationAction(ISD::ROTR, VT, Expand);
719 setOperationAction(ISD::BSWAP, VT, Expand);
720 setOperationAction(ISD::SETCC, VT, Expand);
721 setOperationAction(ISD::FLOG, VT, Expand);
722 setOperationAction(ISD::FLOG2, VT, Expand);
723 setOperationAction(ISD::FLOG10, VT, Expand);
724 setOperationAction(ISD::FEXP, VT, Expand);
725 setOperationAction(ISD::FEXP2, VT, Expand);
726 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
727 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
728 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
729 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
730 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
731 setOperationAction(ISD::TRUNCATE, VT, Expand);
732 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
733 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
734 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
735 setOperationAction(ISD::VSELECT, VT, Expand);
736 setOperationAction(ISD::SELECT_CC, VT, Expand);
737 for (MVT InnerVT : MVT::vector_valuetypes()) {
738 setTruncStoreAction(InnerVT, VT, Expand);
740 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
741 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
743 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
744 // types, we have to deal with them whether we ask for Expansion or not.
745 // Setting Expand causes its own optimisation problems though, so leave
747 if (VT.getVectorElementType() == MVT::i1)
748 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
752 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
753 // with -msoft-float, disable use of MMX as well.
754 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
755 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
756 // No operations on x86mmx supported, everything uses intrinsics.
759 // MMX-sized vectors (other than x86mmx) are expected to be expanded
760 // into smaller operations.
761 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
762 setOperationAction(ISD::MULHS, MMXTy, Expand);
763 setOperationAction(ISD::AND, MMXTy, Expand);
764 setOperationAction(ISD::OR, MMXTy, Expand);
765 setOperationAction(ISD::XOR, MMXTy, Expand);
766 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
767 setOperationAction(ISD::SELECT, MMXTy, Expand);
768 setOperationAction(ISD::BITCAST, MMXTy, Expand);
770 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
772 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
773 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
775 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
776 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
777 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
778 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
779 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
780 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
781 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
782 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
783 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
785 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
787 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
788 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
791 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
792 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
794 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
795 // registers cannot be used even for integer operations.
796 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
797 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
798 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
799 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
801 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
802 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
803 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
804 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
805 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
806 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
807 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
808 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
809 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
810 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
811 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
812 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
813 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
814 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
815 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
816 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
817 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
818 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
819 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
820 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
821 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
822 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
823 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
825 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
826 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
827 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
828 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
830 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
831 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
832 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
833 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
834 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
836 // Only provide customized ctpop vector bit twiddling for vector types we
837 // know to perform better than using the popcnt instructions on each vector
838 // element. If popcnt isn't supported, always provide the custom version.
839 if (!Subtarget->hasPOPCNT()) {
840 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
841 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
844 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
845 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
846 MVT VT = (MVT::SimpleValueType)i;
847 // Do not attempt to custom lower non-power-of-2 vectors
848 if (!isPowerOf2_32(VT.getVectorNumElements()))
850 // Do not attempt to custom lower non-128-bit vectors
851 if (!VT.is128BitVector())
853 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
854 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
855 setOperationAction(ISD::VSELECT, VT, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
859 // We support custom legalizing of sext and anyext loads for specific
860 // memory vector types which we can load as a scalar (or sequence of
861 // scalars) and extend in-register to a legal 128-bit vector type. For sext
862 // loads these must work with a single scalar load.
863 for (MVT VT : MVT::integer_vector_valuetypes()) {
864 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
865 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
866 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
867 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
868 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
869 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
870 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
871 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
872 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
875 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
876 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
877 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
878 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
879 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
880 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
881 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
882 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
884 if (Subtarget->is64Bit()) {
885 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
886 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
889 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
890 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
891 MVT VT = (MVT::SimpleValueType)i;
893 // Do not attempt to promote non-128-bit vectors
894 if (!VT.is128BitVector())
897 setOperationAction(ISD::AND, VT, Promote);
898 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
899 setOperationAction(ISD::OR, VT, Promote);
900 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
901 setOperationAction(ISD::XOR, VT, Promote);
902 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
903 setOperationAction(ISD::LOAD, VT, Promote);
904 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
905 setOperationAction(ISD::SELECT, VT, Promote);
906 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
909 // Custom lower v2i64 and v2f64 selects.
910 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
911 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
912 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
913 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
915 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
916 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
918 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
919 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
920 // As there is no 64-bit GPR available, we need build a special custom
921 // sequence to convert from v2i32 to v2f32.
922 if (!Subtarget->is64Bit())
923 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
925 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
926 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
928 for (MVT VT : MVT::fp_vector_valuetypes())
929 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
931 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
932 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
933 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
936 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
937 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
938 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
939 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
940 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
941 setOperationAction(ISD::FRINT, RoundedTy, Legal);
942 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
945 // FIXME: Do we need to handle scalar-to-vector here?
946 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
948 // We directly match byte blends in the backend as they match the VSELECT
950 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
952 // SSE41 brings specific instructions for doing vector sign extend even in
953 // cases where we don't have SRA.
954 for (MVT VT : MVT::integer_vector_valuetypes()) {
955 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
956 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
957 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
960 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
961 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
962 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
963 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
964 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
965 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
966 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
968 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
969 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
970 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
971 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
972 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
973 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
975 // i8 and i16 vectors are custom because the source register and source
976 // source memory operand types are not the same width. f32 vectors are
977 // custom since the immediate controlling the insert encodes additional
979 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
980 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
981 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
982 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
984 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
985 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
986 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
987 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
989 // FIXME: these should be Legal, but that's only for the case where
990 // the index is constant. For now custom expand to deal with that.
991 if (Subtarget->is64Bit()) {
992 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
993 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
997 if (Subtarget->hasSSE2()) {
998 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
999 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1001 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1002 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1004 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1005 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1007 // In the customized shift lowering, the legal cases in AVX2 will be
1009 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1010 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1012 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1013 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1015 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1018 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1019 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1020 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1021 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1022 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1023 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1024 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1026 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1027 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1028 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1030 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1031 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1032 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1033 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1034 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1035 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1036 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1037 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1038 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1039 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1040 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1041 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1043 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1044 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1045 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1046 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1047 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1048 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1049 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1050 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1051 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1052 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1053 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1054 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1056 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1057 // even though v8i16 is a legal type.
1058 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1059 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1060 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1062 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1063 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1064 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1066 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1067 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1069 for (MVT VT : MVT::fp_vector_valuetypes())
1070 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1072 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1073 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1075 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1076 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1078 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1079 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1081 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1082 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1083 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1084 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1086 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1087 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1088 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1090 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1091 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1092 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1093 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1094 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1095 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1096 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1097 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1098 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1099 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1100 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1101 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1103 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1104 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1105 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1106 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1107 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1108 setOperationAction(ISD::FMA, MVT::f32, Legal);
1109 setOperationAction(ISD::FMA, MVT::f64, Legal);
1112 if (Subtarget->hasInt256()) {
1113 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1114 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1115 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1116 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1118 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1119 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1120 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1121 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1123 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1124 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1125 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1126 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1128 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1129 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1130 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1131 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1133 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1134 // when we have a 256bit-wide blend with immediate.
1135 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1137 // Only provide customized ctpop vector bit twiddling for vector types we
1138 // know to perform better than using the popcnt instructions on each
1139 // vector element. If popcnt isn't supported, always provide the custom
1141 if (!Subtarget->hasPOPCNT())
1142 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1144 // Custom CTPOP always performs better on natively supported v8i32
1145 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1147 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1148 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1149 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1150 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1151 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1152 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1153 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1155 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1156 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1157 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1158 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1159 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1160 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1162 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1163 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1164 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1165 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1167 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1168 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1169 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1170 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1172 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1173 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1174 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1175 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1178 // In the customized shift lowering, the legal cases in AVX2 will be
1180 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1181 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1183 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1184 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1186 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1188 // Custom lower several nodes for 256-bit types.
1189 for (MVT VT : MVT::vector_valuetypes()) {
1190 if (VT.getScalarSizeInBits() >= 32) {
1191 setOperationAction(ISD::MLOAD, VT, Legal);
1192 setOperationAction(ISD::MSTORE, VT, Legal);
1194 // Extract subvector is special because the value type
1195 // (result) is 128-bit but the source is 256-bit wide.
1196 if (VT.is128BitVector()) {
1197 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1199 // Do not attempt to custom lower other non-256-bit vectors
1200 if (!VT.is256BitVector())
1203 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1204 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1205 setOperationAction(ISD::VSELECT, VT, Custom);
1206 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1207 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1208 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1209 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1210 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1213 if (Subtarget->hasInt256())
1214 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1217 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1218 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1219 MVT VT = (MVT::SimpleValueType)i;
1221 // Do not attempt to promote non-256-bit vectors
1222 if (!VT.is256BitVector())
1225 setOperationAction(ISD::AND, VT, Promote);
1226 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1227 setOperationAction(ISD::OR, VT, Promote);
1228 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1229 setOperationAction(ISD::XOR, VT, Promote);
1230 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1231 setOperationAction(ISD::LOAD, VT, Promote);
1232 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1233 setOperationAction(ISD::SELECT, VT, Promote);
1234 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1238 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1239 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1240 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1241 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1242 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1244 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1245 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1246 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1248 for (MVT VT : MVT::fp_vector_valuetypes())
1249 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1251 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1252 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1253 setOperationAction(ISD::XOR, MVT::i1, Legal);
1254 setOperationAction(ISD::OR, MVT::i1, Legal);
1255 setOperationAction(ISD::AND, MVT::i1, Legal);
1256 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1257 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1258 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1259 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1260 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1262 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1263 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1264 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1265 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1266 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1267 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1269 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1270 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1271 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1272 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1273 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1274 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1275 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1276 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1278 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1279 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1280 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1281 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1282 if (Subtarget->is64Bit()) {
1283 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1284 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1285 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1286 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1288 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1289 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1290 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1291 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1292 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1293 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1294 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1295 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1296 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1297 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1298 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1299 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1300 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1301 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1303 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1304 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1305 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1306 if (Subtarget->hasDQI()) {
1307 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1308 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1310 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1311 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1312 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1313 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1314 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1315 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1316 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1317 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1318 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1319 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1320 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1321 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1322 if (Subtarget->hasDQI()) {
1323 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1324 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1326 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1327 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1328 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1329 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1330 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1331 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1332 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1333 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1334 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1335 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1337 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1338 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1339 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1340 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1341 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1343 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1344 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1346 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1348 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1349 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1350 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1351 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1352 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1353 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1354 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1355 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1356 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1358 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1359 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1361 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1362 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1364 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1366 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1367 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1369 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1370 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1372 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1373 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1375 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1376 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1377 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1378 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1379 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1380 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1382 if (Subtarget->hasCDI()) {
1383 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1384 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1386 if (Subtarget->hasDQI()) {
1387 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1388 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1389 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1391 // Custom lower several nodes.
1392 for (MVT VT : MVT::vector_valuetypes()) {
1393 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1394 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1395 setOperationAction(ISD::MGATHER, VT, Custom);
1396 setOperationAction(ISD::MSCATTER, VT, Custom);
1398 // Extract subvector is special because the value type
1399 // (result) is 256/128-bit but the source is 512-bit wide.
1400 if (VT.is128BitVector() || VT.is256BitVector()) {
1401 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1403 if (VT.getVectorElementType() == MVT::i1)
1404 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1406 // Do not attempt to custom lower other non-512-bit vectors
1407 if (!VT.is512BitVector())
1410 if (EltSize >= 32) {
1411 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1412 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1413 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1414 setOperationAction(ISD::VSELECT, VT, Legal);
1415 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1416 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1417 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1418 setOperationAction(ISD::MLOAD, VT, Legal);
1419 setOperationAction(ISD::MSTORE, VT, Legal);
1422 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1423 MVT VT = (MVT::SimpleValueType)i;
1425 // Do not attempt to promote non-512-bit vectors.
1426 if (!VT.is512BitVector())
1429 setOperationAction(ISD::SELECT, VT, Promote);
1430 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1434 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1435 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1436 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1438 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1439 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1441 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1442 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1443 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1444 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1445 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1446 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1447 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1448 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1449 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1450 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1451 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1452 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1453 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1455 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1456 const MVT VT = (MVT::SimpleValueType)i;
1458 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1460 // Do not attempt to promote non-512-bit vectors.
1461 if (!VT.is512BitVector())
1465 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1466 setOperationAction(ISD::VSELECT, VT, Legal);
1471 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1472 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1473 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1475 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1476 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1477 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1478 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1479 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1480 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1482 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1483 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1484 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1485 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1486 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1487 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1490 // We want to custom lower some of our intrinsics.
1491 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1492 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1493 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1494 if (!Subtarget->is64Bit())
1495 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1497 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1498 // handle type legalization for these operations here.
1500 // FIXME: We really should do custom legalization for addition and
1501 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1502 // than generic legalization for 64-bit multiplication-with-overflow, though.
1503 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1504 // Add/Sub/Mul with overflow operations are custom lowered.
1506 setOperationAction(ISD::SADDO, VT, Custom);
1507 setOperationAction(ISD::UADDO, VT, Custom);
1508 setOperationAction(ISD::SSUBO, VT, Custom);
1509 setOperationAction(ISD::USUBO, VT, Custom);
1510 setOperationAction(ISD::SMULO, VT, Custom);
1511 setOperationAction(ISD::UMULO, VT, Custom);
1515 if (!Subtarget->is64Bit()) {
1516 // These libcalls are not available in 32-bit.
1517 setLibcallName(RTLIB::SHL_I128, nullptr);
1518 setLibcallName(RTLIB::SRL_I128, nullptr);
1519 setLibcallName(RTLIB::SRA_I128, nullptr);
1522 // Combine sin / cos into one node or libcall if possible.
1523 if (Subtarget->hasSinCos()) {
1524 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1525 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1526 if (Subtarget->isTargetDarwin()) {
1527 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1528 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1529 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1530 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1534 if (Subtarget->isTargetWin64()) {
1535 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1536 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1537 setOperationAction(ISD::SREM, MVT::i128, Custom);
1538 setOperationAction(ISD::UREM, MVT::i128, Custom);
1539 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1540 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1543 // We have target-specific dag combine patterns for the following nodes:
1544 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1545 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1546 setTargetDAGCombine(ISD::BITCAST);
1547 setTargetDAGCombine(ISD::VSELECT);
1548 setTargetDAGCombine(ISD::SELECT);
1549 setTargetDAGCombine(ISD::SHL);
1550 setTargetDAGCombine(ISD::SRA);
1551 setTargetDAGCombine(ISD::SRL);
1552 setTargetDAGCombine(ISD::OR);
1553 setTargetDAGCombine(ISD::AND);
1554 setTargetDAGCombine(ISD::ADD);
1555 setTargetDAGCombine(ISD::FADD);
1556 setTargetDAGCombine(ISD::FSUB);
1557 setTargetDAGCombine(ISD::FMA);
1558 setTargetDAGCombine(ISD::SUB);
1559 setTargetDAGCombine(ISD::LOAD);
1560 setTargetDAGCombine(ISD::MLOAD);
1561 setTargetDAGCombine(ISD::STORE);
1562 setTargetDAGCombine(ISD::MSTORE);
1563 setTargetDAGCombine(ISD::ZERO_EXTEND);
1564 setTargetDAGCombine(ISD::ANY_EXTEND);
1565 setTargetDAGCombine(ISD::SIGN_EXTEND);
1566 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1567 setTargetDAGCombine(ISD::TRUNCATE);
1568 setTargetDAGCombine(ISD::SINT_TO_FP);
1569 setTargetDAGCombine(ISD::SETCC);
1570 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1571 setTargetDAGCombine(ISD::BUILD_VECTOR);
1572 setTargetDAGCombine(ISD::MUL);
1573 setTargetDAGCombine(ISD::XOR);
1575 computeRegisterProperties(Subtarget->getRegisterInfo());
1577 // On Darwin, -Os means optimize for size without hurting performance,
1578 // do not reduce the limit.
1579 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1580 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1581 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1582 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1583 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1584 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1585 setPrefLoopAlignment(4); // 2^4 bytes.
1587 // Predictable cmov don't hurt on atom because it's in-order.
1588 PredictableSelectIsExpensive = !Subtarget->isAtom();
1589 EnableExtLdPromotion = true;
1590 setPrefFunctionAlignment(4); // 2^4 bytes.
1592 verifyIntrinsicTables();
1595 // This has so far only been implemented for 64-bit MachO.
1596 bool X86TargetLowering::useLoadStackGuardNode() const {
1597 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1600 TargetLoweringBase::LegalizeTypeAction
1601 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1602 if (ExperimentalVectorWideningLegalization &&
1603 VT.getVectorNumElements() != 1 &&
1604 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1605 return TypeWidenVector;
1607 return TargetLoweringBase::getPreferredVectorAction(VT);
1610 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1612 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1614 const unsigned NumElts = VT.getVectorNumElements();
1615 const EVT EltVT = VT.getVectorElementType();
1616 if (VT.is512BitVector()) {
1617 if (Subtarget->hasAVX512())
1618 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1619 EltVT == MVT::f32 || EltVT == MVT::f64)
1621 case 8: return MVT::v8i1;
1622 case 16: return MVT::v16i1;
1624 if (Subtarget->hasBWI())
1625 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1627 case 32: return MVT::v32i1;
1628 case 64: return MVT::v64i1;
1632 if (VT.is256BitVector() || VT.is128BitVector()) {
1633 if (Subtarget->hasVLX())
1634 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1635 EltVT == MVT::f32 || EltVT == MVT::f64)
1637 case 2: return MVT::v2i1;
1638 case 4: return MVT::v4i1;
1639 case 8: return MVT::v8i1;
1641 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1642 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1644 case 8: return MVT::v8i1;
1645 case 16: return MVT::v16i1;
1646 case 32: return MVT::v32i1;
1650 return VT.changeVectorElementTypeToInteger();
1653 /// Helper for getByValTypeAlignment to determine
1654 /// the desired ByVal argument alignment.
1655 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1658 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1659 if (VTy->getBitWidth() == 128)
1661 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1662 unsigned EltAlign = 0;
1663 getMaxByValAlign(ATy->getElementType(), EltAlign);
1664 if (EltAlign > MaxAlign)
1665 MaxAlign = EltAlign;
1666 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1667 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1668 unsigned EltAlign = 0;
1669 getMaxByValAlign(STy->getElementType(i), EltAlign);
1670 if (EltAlign > MaxAlign)
1671 MaxAlign = EltAlign;
1678 /// Return the desired alignment for ByVal aggregate
1679 /// function arguments in the caller parameter area. For X86, aggregates
1680 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1681 /// are at 4-byte boundaries.
1682 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1683 if (Subtarget->is64Bit()) {
1684 // Max of 8 and alignment of type.
1685 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1692 if (Subtarget->hasSSE1())
1693 getMaxByValAlign(Ty, Align);
1697 /// Returns the target specific optimal type for load
1698 /// and store operations as a result of memset, memcpy, and memmove
1699 /// lowering. If DstAlign is zero that means it's safe to destination
1700 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1701 /// means there isn't a need to check it against alignment requirement,
1702 /// probably because the source does not need to be loaded. If 'IsMemset' is
1703 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1704 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1705 /// source is constant so it does not need to be loaded.
1706 /// It returns EVT::Other if the type should be determined using generic
1707 /// target-independent logic.
1709 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1710 unsigned DstAlign, unsigned SrcAlign,
1711 bool IsMemset, bool ZeroMemset,
1713 MachineFunction &MF) const {
1714 const Function *F = MF.getFunction();
1715 if ((!IsMemset || ZeroMemset) &&
1716 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1718 (Subtarget->isUnalignedMemAccessFast() ||
1719 ((DstAlign == 0 || DstAlign >= 16) &&
1720 (SrcAlign == 0 || SrcAlign >= 16)))) {
1722 if (Subtarget->hasInt256())
1724 if (Subtarget->hasFp256())
1727 if (Subtarget->hasSSE2())
1729 if (Subtarget->hasSSE1())
1731 } else if (!MemcpyStrSrc && Size >= 8 &&
1732 !Subtarget->is64Bit() &&
1733 Subtarget->hasSSE2()) {
1734 // Do not use f64 to lower memcpy if source is string constant. It's
1735 // better to use i32 to avoid the loads.
1739 if (Subtarget->is64Bit() && Size >= 8)
1744 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1746 return X86ScalarSSEf32;
1747 else if (VT == MVT::f64)
1748 return X86ScalarSSEf64;
1753 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1758 *Fast = Subtarget->isUnalignedMemAccessFast();
1762 /// Return the entry encoding for a jump table in the
1763 /// current function. The returned value is a member of the
1764 /// MachineJumpTableInfo::JTEntryKind enum.
1765 unsigned X86TargetLowering::getJumpTableEncoding() const {
1766 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1768 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1769 Subtarget->isPICStyleGOT())
1770 return MachineJumpTableInfo::EK_Custom32;
1772 // Otherwise, use the normal jump table encoding heuristics.
1773 return TargetLowering::getJumpTableEncoding();
1777 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1778 const MachineBasicBlock *MBB,
1779 unsigned uid,MCContext &Ctx) const{
1780 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1781 Subtarget->isPICStyleGOT());
1782 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1784 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1785 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1788 /// Returns relocation base for the given PIC jumptable.
1789 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1790 SelectionDAG &DAG) const {
1791 if (!Subtarget->is64Bit())
1792 // This doesn't have SDLoc associated with it, but is not really the
1793 // same as a Register.
1794 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1798 /// This returns the relocation base for the given PIC jumptable,
1799 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1800 const MCExpr *X86TargetLowering::
1801 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1802 MCContext &Ctx) const {
1803 // X86-64 uses RIP relative addressing based on the jump table label.
1804 if (Subtarget->isPICStyleRIPRel())
1805 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1807 // Otherwise, the reference is relative to the PIC base.
1808 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1811 std::pair<const TargetRegisterClass *, uint8_t>
1812 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1814 const TargetRegisterClass *RRC = nullptr;
1816 switch (VT.SimpleTy) {
1818 return TargetLowering::findRepresentativeClass(TRI, VT);
1819 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1820 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1823 RRC = &X86::VR64RegClass;
1825 case MVT::f32: case MVT::f64:
1826 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1827 case MVT::v4f32: case MVT::v2f64:
1828 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1830 RRC = &X86::VR128RegClass;
1833 return std::make_pair(RRC, Cost);
1836 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1837 unsigned &Offset) const {
1838 if (!Subtarget->isTargetLinux())
1841 if (Subtarget->is64Bit()) {
1842 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1844 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1856 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1857 unsigned DestAS) const {
1858 assert(SrcAS != DestAS && "Expected different address spaces!");
1860 return SrcAS < 256 && DestAS < 256;
1863 //===----------------------------------------------------------------------===//
1864 // Return Value Calling Convention Implementation
1865 //===----------------------------------------------------------------------===//
1867 #include "X86GenCallingConv.inc"
1870 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1871 MachineFunction &MF, bool isVarArg,
1872 const SmallVectorImpl<ISD::OutputArg> &Outs,
1873 LLVMContext &Context) const {
1874 SmallVector<CCValAssign, 16> RVLocs;
1875 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1876 return CCInfo.CheckReturn(Outs, RetCC_X86);
1879 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1880 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1885 X86TargetLowering::LowerReturn(SDValue Chain,
1886 CallingConv::ID CallConv, bool isVarArg,
1887 const SmallVectorImpl<ISD::OutputArg> &Outs,
1888 const SmallVectorImpl<SDValue> &OutVals,
1889 SDLoc dl, SelectionDAG &DAG) const {
1890 MachineFunction &MF = DAG.getMachineFunction();
1891 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1893 SmallVector<CCValAssign, 16> RVLocs;
1894 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1895 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1898 SmallVector<SDValue, 6> RetOps;
1899 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1900 // Operand #1 = Bytes To Pop
1901 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
1904 // Copy the result values into the output registers.
1905 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1906 CCValAssign &VA = RVLocs[i];
1907 assert(VA.isRegLoc() && "Can only return in registers!");
1908 SDValue ValToCopy = OutVals[i];
1909 EVT ValVT = ValToCopy.getValueType();
1911 // Promote values to the appropriate types.
1912 if (VA.getLocInfo() == CCValAssign::SExt)
1913 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1914 else if (VA.getLocInfo() == CCValAssign::ZExt)
1915 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1916 else if (VA.getLocInfo() == CCValAssign::AExt) {
1917 if (ValVT.getScalarType() == MVT::i1)
1918 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1920 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1922 else if (VA.getLocInfo() == CCValAssign::BCvt)
1923 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1925 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1926 "Unexpected FP-extend for return value.");
1928 // If this is x86-64, and we disabled SSE, we can't return FP values,
1929 // or SSE or MMX vectors.
1930 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1931 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1932 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1933 report_fatal_error("SSE register return with SSE disabled");
1935 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1936 // llvm-gcc has never done it right and no one has noticed, so this
1937 // should be OK for now.
1938 if (ValVT == MVT::f64 &&
1939 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1940 report_fatal_error("SSE2 register return with SSE2 disabled");
1942 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1943 // the RET instruction and handled by the FP Stackifier.
1944 if (VA.getLocReg() == X86::FP0 ||
1945 VA.getLocReg() == X86::FP1) {
1946 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1947 // change the value to the FP stack register class.
1948 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1949 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1950 RetOps.push_back(ValToCopy);
1951 // Don't emit a copytoreg.
1955 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1956 // which is returned in RAX / RDX.
1957 if (Subtarget->is64Bit()) {
1958 if (ValVT == MVT::x86mmx) {
1959 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1960 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1961 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1963 // If we don't have SSE2 available, convert to v4f32 so the generated
1964 // register is legal.
1965 if (!Subtarget->hasSSE2())
1966 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1971 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1972 Flag = Chain.getValue(1);
1973 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1976 // The x86-64 ABIs require that for returning structs by value we copy
1977 // the sret argument into %rax/%eax (depending on ABI) for the return.
1978 // Win32 requires us to put the sret argument to %eax as well.
1979 // We saved the argument into a virtual register in the entry block,
1980 // so now we copy the value out and into %rax/%eax.
1982 // Checking Function.hasStructRetAttr() here is insufficient because the IR
1983 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
1984 // false, then an sret argument may be implicitly inserted in the SelDAG. In
1985 // either case FuncInfo->setSRetReturnReg() will have been called.
1986 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
1987 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
1988 "No need for an sret register");
1989 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
1992 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1993 X86::RAX : X86::EAX;
1994 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1995 Flag = Chain.getValue(1);
1997 // RAX/EAX now acts like a return value.
1998 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2001 RetOps[0] = Chain; // Update chain.
2003 // Add the flag if we have it.
2005 RetOps.push_back(Flag);
2007 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2010 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2011 if (N->getNumValues() != 1)
2013 if (!N->hasNUsesOfValue(1, 0))
2016 SDValue TCChain = Chain;
2017 SDNode *Copy = *N->use_begin();
2018 if (Copy->getOpcode() == ISD::CopyToReg) {
2019 // If the copy has a glue operand, we conservatively assume it isn't safe to
2020 // perform a tail call.
2021 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2023 TCChain = Copy->getOperand(0);
2024 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2027 bool HasRet = false;
2028 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2030 if (UI->getOpcode() != X86ISD::RET_FLAG)
2032 // If we are returning more than one value, we can definitely
2033 // not make a tail call see PR19530
2034 if (UI->getNumOperands() > 4)
2036 if (UI->getNumOperands() == 4 &&
2037 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2050 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2051 ISD::NodeType ExtendKind) const {
2053 // TODO: Is this also valid on 32-bit?
2054 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2055 ReturnMVT = MVT::i8;
2057 ReturnMVT = MVT::i32;
2059 EVT MinVT = getRegisterType(Context, ReturnMVT);
2060 return VT.bitsLT(MinVT) ? MinVT : VT;
2063 /// Lower the result values of a call into the
2064 /// appropriate copies out of appropriate physical registers.
2067 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2068 CallingConv::ID CallConv, bool isVarArg,
2069 const SmallVectorImpl<ISD::InputArg> &Ins,
2070 SDLoc dl, SelectionDAG &DAG,
2071 SmallVectorImpl<SDValue> &InVals) const {
2073 // Assign locations to each value returned by this call.
2074 SmallVector<CCValAssign, 16> RVLocs;
2075 bool Is64Bit = Subtarget->is64Bit();
2076 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2078 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2080 // Copy all of the result registers out of their specified physreg.
2081 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2082 CCValAssign &VA = RVLocs[i];
2083 EVT CopyVT = VA.getLocVT();
2085 // If this is x86-64, and we disabled SSE, we can't return FP values
2086 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2087 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2088 report_fatal_error("SSE register return with SSE disabled");
2091 // If we prefer to use the value in xmm registers, copy it out as f80 and
2092 // use a truncate to move it from fp stack reg to xmm reg.
2093 bool RoundAfterCopy = false;
2094 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2095 isScalarFPTypeInSSEReg(VA.getValVT())) {
2097 RoundAfterCopy = (CopyVT != VA.getLocVT());
2100 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2101 CopyVT, InFlag).getValue(1);
2102 SDValue Val = Chain.getValue(0);
2105 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2106 // This truncation won't change the value.
2107 DAG.getIntPtrConstant(1, dl));
2109 InFlag = Chain.getValue(2);
2110 InVals.push_back(Val);
2116 //===----------------------------------------------------------------------===//
2117 // C & StdCall & Fast Calling Convention implementation
2118 //===----------------------------------------------------------------------===//
2119 // StdCall calling convention seems to be standard for many Windows' API
2120 // routines and around. It differs from C calling convention just a little:
2121 // callee should clean up the stack, not caller. Symbols should be also
2122 // decorated in some fancy way :) It doesn't support any vector arguments.
2123 // For info on fast calling convention see Fast Calling Convention (tail call)
2124 // implementation LowerX86_32FastCCCallTo.
2126 /// CallIsStructReturn - Determines whether a call uses struct return
2128 enum StructReturnType {
2133 static StructReturnType
2134 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2136 return NotStructReturn;
2138 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2139 if (!Flags.isSRet())
2140 return NotStructReturn;
2141 if (Flags.isInReg())
2142 return RegStructReturn;
2143 return StackStructReturn;
2146 /// Determines whether a function uses struct return semantics.
2147 static StructReturnType
2148 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2150 return NotStructReturn;
2152 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2153 if (!Flags.isSRet())
2154 return NotStructReturn;
2155 if (Flags.isInReg())
2156 return RegStructReturn;
2157 return StackStructReturn;
2160 /// Make a copy of an aggregate at address specified by "Src" to address
2161 /// "Dst" with size and alignment information specified by the specific
2162 /// parameter attribute. The copy will be passed as a byval function parameter.
2164 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2165 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2167 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2169 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2170 /*isVolatile*/false, /*AlwaysInline=*/true,
2171 /*isTailCall*/false,
2172 MachinePointerInfo(), MachinePointerInfo());
2175 /// Return true if the calling convention is one that
2176 /// supports tail call optimization.
2177 static bool IsTailCallConvention(CallingConv::ID CC) {
2178 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2179 CC == CallingConv::HiPE);
2182 /// \brief Return true if the calling convention is a C calling convention.
2183 static bool IsCCallConvention(CallingConv::ID CC) {
2184 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2185 CC == CallingConv::X86_64_SysV);
2188 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2189 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2193 CallingConv::ID CalleeCC = CS.getCallingConv();
2194 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2200 /// Return true if the function is being made into
2201 /// a tailcall target by changing its ABI.
2202 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2203 bool GuaranteedTailCallOpt) {
2204 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2208 X86TargetLowering::LowerMemArgument(SDValue Chain,
2209 CallingConv::ID CallConv,
2210 const SmallVectorImpl<ISD::InputArg> &Ins,
2211 SDLoc dl, SelectionDAG &DAG,
2212 const CCValAssign &VA,
2213 MachineFrameInfo *MFI,
2215 // Create the nodes corresponding to a load from this parameter slot.
2216 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2217 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2218 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2219 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2222 // If value is passed by pointer we have address passed instead of the value
2224 if (VA.getLocInfo() == CCValAssign::Indirect)
2225 ValVT = VA.getLocVT();
2227 ValVT = VA.getValVT();
2229 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2230 // changed with more analysis.
2231 // In case of tail call optimization mark all arguments mutable. Since they
2232 // could be overwritten by lowering of arguments in case of a tail call.
2233 if (Flags.isByVal()) {
2234 unsigned Bytes = Flags.getByValSize();
2235 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2236 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2237 return DAG.getFrameIndex(FI, getPointerTy());
2239 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2240 VA.getLocMemOffset(), isImmutable);
2241 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2242 return DAG.getLoad(ValVT, dl, Chain, FIN,
2243 MachinePointerInfo::getFixedStack(FI),
2244 false, false, false, 0);
2248 // FIXME: Get this from tablegen.
2249 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2250 const X86Subtarget *Subtarget) {
2251 assert(Subtarget->is64Bit());
2253 if (Subtarget->isCallingConvWin64(CallConv)) {
2254 static const MCPhysReg GPR64ArgRegsWin64[] = {
2255 X86::RCX, X86::RDX, X86::R8, X86::R9
2257 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2260 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2261 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2263 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2266 // FIXME: Get this from tablegen.
2267 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2268 CallingConv::ID CallConv,
2269 const X86Subtarget *Subtarget) {
2270 assert(Subtarget->is64Bit());
2271 if (Subtarget->isCallingConvWin64(CallConv)) {
2272 // The XMM registers which might contain var arg parameters are shadowed
2273 // in their paired GPR. So we only need to save the GPR to their home
2275 // TODO: __vectorcall will change this.
2279 const Function *Fn = MF.getFunction();
2280 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2281 bool isSoftFloat = MF.getTarget().Options.UseSoftFloat;
2282 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2283 "SSE register cannot be used when SSE is disabled!");
2284 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2285 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2289 static const MCPhysReg XMMArgRegs64Bit[] = {
2290 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2291 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2293 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2297 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2298 CallingConv::ID CallConv,
2300 const SmallVectorImpl<ISD::InputArg> &Ins,
2303 SmallVectorImpl<SDValue> &InVals)
2305 MachineFunction &MF = DAG.getMachineFunction();
2306 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2307 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2309 const Function* Fn = MF.getFunction();
2310 if (Fn->hasExternalLinkage() &&
2311 Subtarget->isTargetCygMing() &&
2312 Fn->getName() == "main")
2313 FuncInfo->setForceFramePointer(true);
2315 MachineFrameInfo *MFI = MF.getFrameInfo();
2316 bool Is64Bit = Subtarget->is64Bit();
2317 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2319 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2320 "Var args not supported with calling convention fastcc, ghc or hipe");
2322 // Assign locations to all of the incoming arguments.
2323 SmallVector<CCValAssign, 16> ArgLocs;
2324 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2326 // Allocate shadow area for Win64
2328 CCInfo.AllocateStack(32, 8);
2330 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2332 unsigned LastVal = ~0U;
2334 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2335 CCValAssign &VA = ArgLocs[i];
2336 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2338 assert(VA.getValNo() != LastVal &&
2339 "Don't support value assigned to multiple locs yet");
2341 LastVal = VA.getValNo();
2343 if (VA.isRegLoc()) {
2344 EVT RegVT = VA.getLocVT();
2345 const TargetRegisterClass *RC;
2346 if (RegVT == MVT::i32)
2347 RC = &X86::GR32RegClass;
2348 else if (Is64Bit && RegVT == MVT::i64)
2349 RC = &X86::GR64RegClass;
2350 else if (RegVT == MVT::f32)
2351 RC = &X86::FR32RegClass;
2352 else if (RegVT == MVT::f64)
2353 RC = &X86::FR64RegClass;
2354 else if (RegVT.is512BitVector())
2355 RC = &X86::VR512RegClass;
2356 else if (RegVT.is256BitVector())
2357 RC = &X86::VR256RegClass;
2358 else if (RegVT.is128BitVector())
2359 RC = &X86::VR128RegClass;
2360 else if (RegVT == MVT::x86mmx)
2361 RC = &X86::VR64RegClass;
2362 else if (RegVT == MVT::i1)
2363 RC = &X86::VK1RegClass;
2364 else if (RegVT == MVT::v8i1)
2365 RC = &X86::VK8RegClass;
2366 else if (RegVT == MVT::v16i1)
2367 RC = &X86::VK16RegClass;
2368 else if (RegVT == MVT::v32i1)
2369 RC = &X86::VK32RegClass;
2370 else if (RegVT == MVT::v64i1)
2371 RC = &X86::VK64RegClass;
2373 llvm_unreachable("Unknown argument type!");
2375 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2376 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2378 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2379 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2381 if (VA.getLocInfo() == CCValAssign::SExt)
2382 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2383 DAG.getValueType(VA.getValVT()));
2384 else if (VA.getLocInfo() == CCValAssign::ZExt)
2385 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2386 DAG.getValueType(VA.getValVT()));
2387 else if (VA.getLocInfo() == CCValAssign::BCvt)
2388 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2390 if (VA.isExtInLoc()) {
2391 // Handle MMX values passed in XMM regs.
2392 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2393 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2395 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2398 assert(VA.isMemLoc());
2399 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2402 // If value is passed via pointer - do a load.
2403 if (VA.getLocInfo() == CCValAssign::Indirect)
2404 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2405 MachinePointerInfo(), false, false, false, 0);
2407 InVals.push_back(ArgValue);
2410 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2411 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2412 // The x86-64 ABIs require that for returning structs by value we copy
2413 // the sret argument into %rax/%eax (depending on ABI) for the return.
2414 // Win32 requires us to put the sret argument to %eax as well.
2415 // Save the argument into a virtual register so that we can access it
2416 // from the return points.
2417 if (Ins[i].Flags.isSRet()) {
2418 unsigned Reg = FuncInfo->getSRetReturnReg();
2420 MVT PtrTy = getPointerTy();
2421 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2422 FuncInfo->setSRetReturnReg(Reg);
2424 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2425 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2431 unsigned StackSize = CCInfo.getNextStackOffset();
2432 // Align stack specially for tail calls.
2433 if (FuncIsMadeTailCallSafe(CallConv,
2434 MF.getTarget().Options.GuaranteedTailCallOpt))
2435 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2437 // If the function takes variable number of arguments, make a frame index for
2438 // the start of the first vararg value... for expansion of llvm.va_start. We
2439 // can skip this if there are no va_start calls.
2440 if (MFI->hasVAStart() &&
2441 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2442 CallConv != CallingConv::X86_ThisCall))) {
2443 FuncInfo->setVarArgsFrameIndex(
2444 MFI->CreateFixedObject(1, StackSize, true));
2447 MachineModuleInfo &MMI = MF.getMMI();
2448 const Function *WinEHParent = nullptr;
2449 if (IsWin64 && MMI.hasWinEHFuncInfo(Fn))
2450 WinEHParent = MMI.getWinEHParent(Fn);
2451 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2452 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2454 // Figure out if XMM registers are in use.
2455 assert(!(MF.getTarget().Options.UseSoftFloat &&
2456 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2457 "SSE register cannot be used when SSE is disabled!");
2459 // 64-bit calling conventions support varargs and register parameters, so we
2460 // have to do extra work to spill them in the prologue.
2461 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2462 // Find the first unallocated argument registers.
2463 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2464 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2465 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2466 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2467 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2468 "SSE register cannot be used when SSE is disabled!");
2470 // Gather all the live in physical registers.
2471 SmallVector<SDValue, 6> LiveGPRs;
2472 SmallVector<SDValue, 8> LiveXMMRegs;
2474 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2475 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2477 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2479 if (!ArgXMMs.empty()) {
2480 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2481 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2482 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2483 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2484 LiveXMMRegs.push_back(
2485 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2490 // Get to the caller-allocated home save location. Add 8 to account
2491 // for the return address.
2492 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2493 FuncInfo->setRegSaveFrameIndex(
2494 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2495 // Fixup to set vararg frame on shadow area (4 x i64).
2497 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2499 // For X86-64, if there are vararg parameters that are passed via
2500 // registers, then we must store them to their spots on the stack so
2501 // they may be loaded by deferencing the result of va_next.
2502 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2503 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2504 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2505 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2508 // Store the integer parameter registers.
2509 SmallVector<SDValue, 8> MemOps;
2510 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2512 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2513 for (SDValue Val : LiveGPRs) {
2514 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2515 DAG.getIntPtrConstant(Offset, dl));
2517 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2518 MachinePointerInfo::getFixedStack(
2519 FuncInfo->getRegSaveFrameIndex(), Offset),
2521 MemOps.push_back(Store);
2525 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2526 // Now store the XMM (fp + vector) parameter registers.
2527 SmallVector<SDValue, 12> SaveXMMOps;
2528 SaveXMMOps.push_back(Chain);
2529 SaveXMMOps.push_back(ALVal);
2530 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2531 FuncInfo->getRegSaveFrameIndex(), dl));
2532 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2533 FuncInfo->getVarArgsFPOffset(), dl));
2534 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2536 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2537 MVT::Other, SaveXMMOps));
2540 if (!MemOps.empty())
2541 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2542 } else if (IsWinEHOutlined) {
2543 // Get to the caller-allocated home save location. Add 8 to account
2544 // for the return address.
2545 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2546 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2547 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2549 MMI.getWinEHFuncInfo(Fn)
2550 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2551 FuncInfo->getRegSaveFrameIndex();
2553 // Store the second integer parameter (rdx) into rsp+16 relative to the
2554 // stack pointer at the entry of the function.
2556 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), getPointerTy());
2557 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2558 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2559 Chain = DAG.getStore(
2560 Val.getValue(1), dl, Val, RSFIN,
2561 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2562 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2565 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2566 // Find the largest legal vector type.
2567 MVT VecVT = MVT::Other;
2568 // FIXME: Only some x86_32 calling conventions support AVX512.
2569 if (Subtarget->hasAVX512() &&
2570 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2571 CallConv == CallingConv::Intel_OCL_BI)))
2572 VecVT = MVT::v16f32;
2573 else if (Subtarget->hasAVX())
2575 else if (Subtarget->hasSSE2())
2578 // We forward some GPRs and some vector types.
2579 SmallVector<MVT, 2> RegParmTypes;
2580 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2581 RegParmTypes.push_back(IntVT);
2582 if (VecVT != MVT::Other)
2583 RegParmTypes.push_back(VecVT);
2585 // Compute the set of forwarded registers. The rest are scratch.
2586 SmallVectorImpl<ForwardedRegister> &Forwards =
2587 FuncInfo->getForwardedMustTailRegParms();
2588 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2590 // Conservatively forward AL on x86_64, since it might be used for varargs.
2591 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2592 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2593 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2596 // Copy all forwards from physical to virtual registers.
2597 for (ForwardedRegister &F : Forwards) {
2598 // FIXME: Can we use a less constrained schedule?
2599 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2600 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2601 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2605 // Some CCs need callee pop.
2606 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2607 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2608 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2610 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2611 // If this is an sret function, the return should pop the hidden pointer.
2612 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2613 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2614 argsAreStructReturn(Ins) == StackStructReturn)
2615 FuncInfo->setBytesToPopOnReturn(4);
2619 // RegSaveFrameIndex is X86-64 only.
2620 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2621 if (CallConv == CallingConv::X86_FastCall ||
2622 CallConv == CallingConv::X86_ThisCall)
2623 // fastcc functions can't have varargs.
2624 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2627 FuncInfo->setArgumentStackSize(StackSize);
2629 if (IsWinEHParent) {
2630 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2631 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2632 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2633 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2634 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2635 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2636 /*isVolatile=*/true,
2637 /*isNonTemporal=*/false, /*Alignment=*/0);
2644 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2645 SDValue StackPtr, SDValue Arg,
2646 SDLoc dl, SelectionDAG &DAG,
2647 const CCValAssign &VA,
2648 ISD::ArgFlagsTy Flags) const {
2649 unsigned LocMemOffset = VA.getLocMemOffset();
2650 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2651 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2652 if (Flags.isByVal())
2653 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2655 return DAG.getStore(Chain, dl, Arg, PtrOff,
2656 MachinePointerInfo::getStack(LocMemOffset),
2660 /// Emit a load of return address if tail call
2661 /// optimization is performed and it is required.
2663 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2664 SDValue &OutRetAddr, SDValue Chain,
2665 bool IsTailCall, bool Is64Bit,
2666 int FPDiff, SDLoc dl) const {
2667 // Adjust the Return address stack slot.
2668 EVT VT = getPointerTy();
2669 OutRetAddr = getReturnAddressFrameIndex(DAG);
2671 // Load the "old" Return address.
2672 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2673 false, false, false, 0);
2674 return SDValue(OutRetAddr.getNode(), 1);
2677 /// Emit a store of the return address if tail call
2678 /// optimization is performed and it is required (FPDiff!=0).
2679 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2680 SDValue Chain, SDValue RetAddrFrIdx,
2681 EVT PtrVT, unsigned SlotSize,
2682 int FPDiff, SDLoc dl) {
2683 // Store the return address to the appropriate stack slot.
2684 if (!FPDiff) return Chain;
2685 // Calculate the new stack slot for the return address.
2686 int NewReturnAddrFI =
2687 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2689 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2690 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2691 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2697 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2698 SmallVectorImpl<SDValue> &InVals) const {
2699 SelectionDAG &DAG = CLI.DAG;
2701 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2702 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2703 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2704 SDValue Chain = CLI.Chain;
2705 SDValue Callee = CLI.Callee;
2706 CallingConv::ID CallConv = CLI.CallConv;
2707 bool &isTailCall = CLI.IsTailCall;
2708 bool isVarArg = CLI.IsVarArg;
2710 MachineFunction &MF = DAG.getMachineFunction();
2711 bool Is64Bit = Subtarget->is64Bit();
2712 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2713 StructReturnType SR = callIsStructReturn(Outs);
2714 bool IsSibcall = false;
2715 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2717 if (MF.getTarget().Options.DisableTailCalls)
2720 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2722 // Force this to be a tail call. The verifier rules are enough to ensure
2723 // that we can lower this successfully without moving the return address
2726 } else if (isTailCall) {
2727 // Check if it's really possible to do a tail call.
2728 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2729 isVarArg, SR != NotStructReturn,
2730 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2731 Outs, OutVals, Ins, DAG);
2733 // Sibcalls are automatically detected tailcalls which do not require
2735 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2742 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2743 "Var args not supported with calling convention fastcc, ghc or hipe");
2745 // Analyze operands of the call, assigning locations to each operand.
2746 SmallVector<CCValAssign, 16> ArgLocs;
2747 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2749 // Allocate shadow area for Win64
2751 CCInfo.AllocateStack(32, 8);
2753 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2755 // Get a count of how many bytes are to be pushed on the stack.
2756 unsigned NumBytes = CCInfo.getNextStackOffset();
2758 // This is a sibcall. The memory operands are available in caller's
2759 // own caller's stack.
2761 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2762 IsTailCallConvention(CallConv))
2763 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2766 if (isTailCall && !IsSibcall && !IsMustTail) {
2767 // Lower arguments at fp - stackoffset + fpdiff.
2768 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2770 FPDiff = NumBytesCallerPushed - NumBytes;
2772 // Set the delta of movement of the returnaddr stackslot.
2773 // But only set if delta is greater than previous delta.
2774 if (FPDiff < X86Info->getTCReturnAddrDelta())
2775 X86Info->setTCReturnAddrDelta(FPDiff);
2778 unsigned NumBytesToPush = NumBytes;
2779 unsigned NumBytesToPop = NumBytes;
2781 // If we have an inalloca argument, all stack space has already been allocated
2782 // for us and be right at the top of the stack. We don't support multiple
2783 // arguments passed in memory when using inalloca.
2784 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2786 if (!ArgLocs.back().isMemLoc())
2787 report_fatal_error("cannot use inalloca attribute on a register "
2789 if (ArgLocs.back().getLocMemOffset() != 0)
2790 report_fatal_error("any parameter with the inalloca attribute must be "
2791 "the only memory argument");
2795 Chain = DAG.getCALLSEQ_START(
2796 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
2798 SDValue RetAddrFrIdx;
2799 // Load return address for tail calls.
2800 if (isTailCall && FPDiff)
2801 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2802 Is64Bit, FPDiff, dl);
2804 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2805 SmallVector<SDValue, 8> MemOpChains;
2808 // Walk the register/memloc assignments, inserting copies/loads. In the case
2809 // of tail call optimization arguments are handle later.
2810 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2811 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2812 // Skip inalloca arguments, they have already been written.
2813 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2814 if (Flags.isInAlloca())
2817 CCValAssign &VA = ArgLocs[i];
2818 EVT RegVT = VA.getLocVT();
2819 SDValue Arg = OutVals[i];
2820 bool isByVal = Flags.isByVal();
2822 // Promote the value if needed.
2823 switch (VA.getLocInfo()) {
2824 default: llvm_unreachable("Unknown loc info!");
2825 case CCValAssign::Full: break;
2826 case CCValAssign::SExt:
2827 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2829 case CCValAssign::ZExt:
2830 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2832 case CCValAssign::AExt:
2833 if (Arg.getValueType().getScalarType() == MVT::i1)
2834 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2835 else if (RegVT.is128BitVector()) {
2836 // Special case: passing MMX values in XMM registers.
2837 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2838 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2839 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2841 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2843 case CCValAssign::BCvt:
2844 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2846 case CCValAssign::Indirect: {
2847 // Store the argument.
2848 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2849 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2850 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2851 MachinePointerInfo::getFixedStack(FI),
2858 if (VA.isRegLoc()) {
2859 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2860 if (isVarArg && IsWin64) {
2861 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2862 // shadow reg if callee is a varargs function.
2863 unsigned ShadowReg = 0;
2864 switch (VA.getLocReg()) {
2865 case X86::XMM0: ShadowReg = X86::RCX; break;
2866 case X86::XMM1: ShadowReg = X86::RDX; break;
2867 case X86::XMM2: ShadowReg = X86::R8; break;
2868 case X86::XMM3: ShadowReg = X86::R9; break;
2871 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2873 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2874 assert(VA.isMemLoc());
2875 if (!StackPtr.getNode())
2876 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2878 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2879 dl, DAG, VA, Flags));
2883 if (!MemOpChains.empty())
2884 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2886 if (Subtarget->isPICStyleGOT()) {
2887 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2890 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2891 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2893 // If we are tail calling and generating PIC/GOT style code load the
2894 // address of the callee into ECX. The value in ecx is used as target of
2895 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2896 // for tail calls on PIC/GOT architectures. Normally we would just put the
2897 // address of GOT into ebx and then call target@PLT. But for tail calls
2898 // ebx would be restored (since ebx is callee saved) before jumping to the
2901 // Note: The actual moving to ECX is done further down.
2902 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2903 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2904 !G->getGlobal()->hasProtectedVisibility())
2905 Callee = LowerGlobalAddress(Callee, DAG);
2906 else if (isa<ExternalSymbolSDNode>(Callee))
2907 Callee = LowerExternalSymbol(Callee, DAG);
2911 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2912 // From AMD64 ABI document:
2913 // For calls that may call functions that use varargs or stdargs
2914 // (prototype-less calls or calls to functions containing ellipsis (...) in
2915 // the declaration) %al is used as hidden argument to specify the number
2916 // of SSE registers used. The contents of %al do not need to match exactly
2917 // the number of registers, but must be an ubound on the number of SSE
2918 // registers used and is in the range 0 - 8 inclusive.
2920 // Count the number of XMM registers allocated.
2921 static const MCPhysReg XMMArgRegs[] = {
2922 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2923 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2925 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2926 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2927 && "SSE registers cannot be used when SSE is disabled");
2929 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2930 DAG.getConstant(NumXMMRegs, dl,
2934 if (isVarArg && IsMustTail) {
2935 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
2936 for (const auto &F : Forwards) {
2937 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2938 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
2942 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2943 // don't need this because the eligibility check rejects calls that require
2944 // shuffling arguments passed in memory.
2945 if (!IsSibcall && isTailCall) {
2946 // Force all the incoming stack arguments to be loaded from the stack
2947 // before any new outgoing arguments are stored to the stack, because the
2948 // outgoing stack slots may alias the incoming argument stack slots, and
2949 // the alias isn't otherwise explicit. This is slightly more conservative
2950 // than necessary, because it means that each store effectively depends
2951 // on every argument instead of just those arguments it would clobber.
2952 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2954 SmallVector<SDValue, 8> MemOpChains2;
2957 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2958 CCValAssign &VA = ArgLocs[i];
2961 assert(VA.isMemLoc());
2962 SDValue Arg = OutVals[i];
2963 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2964 // Skip inalloca arguments. They don't require any work.
2965 if (Flags.isInAlloca())
2967 // Create frame index.
2968 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2969 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2970 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2971 FIN = DAG.getFrameIndex(FI, getPointerTy());
2973 if (Flags.isByVal()) {
2974 // Copy relative to framepointer.
2975 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
2976 if (!StackPtr.getNode())
2977 StackPtr = DAG.getCopyFromReg(Chain, dl,
2978 RegInfo->getStackRegister(),
2980 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2982 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2986 // Store relative to framepointer.
2987 MemOpChains2.push_back(
2988 DAG.getStore(ArgChain, dl, Arg, FIN,
2989 MachinePointerInfo::getFixedStack(FI),
2994 if (!MemOpChains2.empty())
2995 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
2997 // Store the return address to the appropriate stack slot.
2998 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2999 getPointerTy(), RegInfo->getSlotSize(),
3003 // Build a sequence of copy-to-reg nodes chained together with token chain
3004 // and flag operands which copy the outgoing args into registers.
3006 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3007 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3008 RegsToPass[i].second, InFlag);
3009 InFlag = Chain.getValue(1);
3012 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3013 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3014 // In the 64-bit large code model, we have to make all calls
3015 // through a register, since the call instruction's 32-bit
3016 // pc-relative offset may not be large enough to hold the whole
3018 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3019 // If the callee is a GlobalAddress node (quite common, every direct call
3020 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3022 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3024 // We should use extra load for direct calls to dllimported functions in
3026 const GlobalValue *GV = G->getGlobal();
3027 if (!GV->hasDLLImportStorageClass()) {
3028 unsigned char OpFlags = 0;
3029 bool ExtraLoad = false;
3030 unsigned WrapperKind = ISD::DELETED_NODE;
3032 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3033 // external symbols most go through the PLT in PIC mode. If the symbol
3034 // has hidden or protected visibility, or if it is static or local, then
3035 // we don't need to use the PLT - we can directly call it.
3036 if (Subtarget->isTargetELF() &&
3037 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3038 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3039 OpFlags = X86II::MO_PLT;
3040 } else if (Subtarget->isPICStyleStubAny() &&
3041 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3042 (!Subtarget->getTargetTriple().isMacOSX() ||
3043 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3044 // PC-relative references to external symbols should go through $stub,
3045 // unless we're building with the leopard linker or later, which
3046 // automatically synthesizes these stubs.
3047 OpFlags = X86II::MO_DARWIN_STUB;
3048 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3049 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3050 // If the function is marked as non-lazy, generate an indirect call
3051 // which loads from the GOT directly. This avoids runtime overhead
3052 // at the cost of eager binding (and one extra byte of encoding).
3053 OpFlags = X86II::MO_GOTPCREL;
3054 WrapperKind = X86ISD::WrapperRIP;
3058 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3059 G->getOffset(), OpFlags);
3061 // Add a wrapper if needed.
3062 if (WrapperKind != ISD::DELETED_NODE)
3063 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3064 // Add extra indirection if needed.
3066 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3067 MachinePointerInfo::getGOT(),
3068 false, false, false, 0);
3070 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3071 unsigned char OpFlags = 0;
3073 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3074 // external symbols should go through the PLT.
3075 if (Subtarget->isTargetELF() &&
3076 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3077 OpFlags = X86II::MO_PLT;
3078 } else if (Subtarget->isPICStyleStubAny() &&
3079 (!Subtarget->getTargetTriple().isMacOSX() ||
3080 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3081 // PC-relative references to external symbols should go through $stub,
3082 // unless we're building with the leopard linker or later, which
3083 // automatically synthesizes these stubs.
3084 OpFlags = X86II::MO_DARWIN_STUB;
3087 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3089 } else if (Subtarget->isTarget64BitILP32() &&
3090 Callee->getValueType(0) == MVT::i32) {
3091 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3092 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3095 // Returns a chain & a flag for retval copy to use.
3096 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3097 SmallVector<SDValue, 8> Ops;
3099 if (!IsSibcall && isTailCall) {
3100 Chain = DAG.getCALLSEQ_END(Chain,
3101 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3102 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3103 InFlag = Chain.getValue(1);
3106 Ops.push_back(Chain);
3107 Ops.push_back(Callee);
3110 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3112 // Add argument registers to the end of the list so that they are known live
3114 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3115 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3116 RegsToPass[i].second.getValueType()));
3118 // Add a register mask operand representing the call-preserved registers.
3119 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3120 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
3121 assert(Mask && "Missing call preserved mask for calling convention");
3122 Ops.push_back(DAG.getRegisterMask(Mask));
3124 if (InFlag.getNode())
3125 Ops.push_back(InFlag);
3129 //// If this is the first return lowered for this function, add the regs
3130 //// to the liveout set for the function.
3131 // This isn't right, although it's probably harmless on x86; liveouts
3132 // should be computed from returns not tail calls. Consider a void
3133 // function making a tail call to a function returning int.
3134 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3137 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3138 InFlag = Chain.getValue(1);
3140 // Create the CALLSEQ_END node.
3141 unsigned NumBytesForCalleeToPop;
3142 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3143 DAG.getTarget().Options.GuaranteedTailCallOpt))
3144 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3145 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3146 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3147 SR == StackStructReturn)
3148 // If this is a call to a struct-return function, the callee
3149 // pops the hidden struct pointer, so we have to push it back.
3150 // This is common for Darwin/X86, Linux & Mingw32 targets.
3151 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3152 NumBytesForCalleeToPop = 4;
3154 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3156 // Returns a flag for retval copy to use.
3158 Chain = DAG.getCALLSEQ_END(Chain,
3159 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3160 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3163 InFlag = Chain.getValue(1);
3166 // Handle result values, copying them out of physregs into vregs that we
3168 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3169 Ins, dl, DAG, InVals);
3172 //===----------------------------------------------------------------------===//
3173 // Fast Calling Convention (tail call) implementation
3174 //===----------------------------------------------------------------------===//
3176 // Like std call, callee cleans arguments, convention except that ECX is
3177 // reserved for storing the tail called function address. Only 2 registers are
3178 // free for argument passing (inreg). Tail call optimization is performed
3180 // * tailcallopt is enabled
3181 // * caller/callee are fastcc
3182 // On X86_64 architecture with GOT-style position independent code only local
3183 // (within module) calls are supported at the moment.
3184 // To keep the stack aligned according to platform abi the function
3185 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3186 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3187 // If a tail called function callee has more arguments than the caller the
3188 // caller needs to make sure that there is room to move the RETADDR to. This is
3189 // achieved by reserving an area the size of the argument delta right after the
3190 // original RETADDR, but before the saved framepointer or the spilled registers
3191 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3203 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3204 /// for a 16 byte align requirement.
3206 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3207 SelectionDAG& DAG) const {
3208 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3209 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3210 unsigned StackAlignment = TFI.getStackAlignment();
3211 uint64_t AlignMask = StackAlignment - 1;
3212 int64_t Offset = StackSize;
3213 unsigned SlotSize = RegInfo->getSlotSize();
3214 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3215 // Number smaller than 12 so just add the difference.
3216 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3218 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3219 Offset = ((~AlignMask) & Offset) + StackAlignment +
3220 (StackAlignment-SlotSize);
3225 /// MatchingStackOffset - Return true if the given stack call argument is
3226 /// already available in the same position (relatively) of the caller's
3227 /// incoming argument stack.
3229 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3230 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3231 const X86InstrInfo *TII) {
3232 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3234 if (Arg.getOpcode() == ISD::CopyFromReg) {
3235 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3236 if (!TargetRegisterInfo::isVirtualRegister(VR))
3238 MachineInstr *Def = MRI->getVRegDef(VR);
3241 if (!Flags.isByVal()) {
3242 if (!TII->isLoadFromStackSlot(Def, FI))
3245 unsigned Opcode = Def->getOpcode();
3246 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3247 Opcode == X86::LEA64_32r) &&
3248 Def->getOperand(1).isFI()) {
3249 FI = Def->getOperand(1).getIndex();
3250 Bytes = Flags.getByValSize();
3254 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3255 if (Flags.isByVal())
3256 // ByVal argument is passed in as a pointer but it's now being
3257 // dereferenced. e.g.
3258 // define @foo(%struct.X* %A) {
3259 // tail call @bar(%struct.X* byval %A)
3262 SDValue Ptr = Ld->getBasePtr();
3263 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3266 FI = FINode->getIndex();
3267 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3268 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3269 FI = FINode->getIndex();
3270 Bytes = Flags.getByValSize();
3274 assert(FI != INT_MAX);
3275 if (!MFI->isFixedObjectIndex(FI))
3277 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3280 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3281 /// for tail call optimization. Targets which want to do tail call
3282 /// optimization should implement this function.
3284 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3285 CallingConv::ID CalleeCC,
3287 bool isCalleeStructRet,
3288 bool isCallerStructRet,
3290 const SmallVectorImpl<ISD::OutputArg> &Outs,
3291 const SmallVectorImpl<SDValue> &OutVals,
3292 const SmallVectorImpl<ISD::InputArg> &Ins,
3293 SelectionDAG &DAG) const {
3294 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3297 // If -tailcallopt is specified, make fastcc functions tail-callable.
3298 const MachineFunction &MF = DAG.getMachineFunction();
3299 const Function *CallerF = MF.getFunction();
3301 // If the function return type is x86_fp80 and the callee return type is not,
3302 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3303 // perform a tailcall optimization here.
3304 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3307 CallingConv::ID CallerCC = CallerF->getCallingConv();
3308 bool CCMatch = CallerCC == CalleeCC;
3309 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3310 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3312 // Win64 functions have extra shadow space for argument homing. Don't do the
3313 // sibcall if the caller and callee have mismatched expectations for this
3315 if (IsCalleeWin64 != IsCallerWin64)
3318 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3319 if (IsTailCallConvention(CalleeCC) && CCMatch)
3324 // Look for obvious safe cases to perform tail call optimization that do not
3325 // require ABI changes. This is what gcc calls sibcall.
3327 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3328 // emit a special epilogue.
3329 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3330 if (RegInfo->needsStackRealignment(MF))
3333 // Also avoid sibcall optimization if either caller or callee uses struct
3334 // return semantics.
3335 if (isCalleeStructRet || isCallerStructRet)
3338 // An stdcall/thiscall caller is expected to clean up its arguments; the
3339 // callee isn't going to do that.
3340 // FIXME: this is more restrictive than needed. We could produce a tailcall
3341 // when the stack adjustment matches. For example, with a thiscall that takes
3342 // only one argument.
3343 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3344 CallerCC == CallingConv::X86_ThisCall))
3347 // Do not sibcall optimize vararg calls unless all arguments are passed via
3349 if (isVarArg && !Outs.empty()) {
3351 // Optimizing for varargs on Win64 is unlikely to be safe without
3352 // additional testing.
3353 if (IsCalleeWin64 || IsCallerWin64)
3356 SmallVector<CCValAssign, 16> ArgLocs;
3357 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3360 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3361 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3362 if (!ArgLocs[i].isRegLoc())
3366 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3367 // stack. Therefore, if it's not used by the call it is not safe to optimize
3368 // this into a sibcall.
3369 bool Unused = false;
3370 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3377 SmallVector<CCValAssign, 16> RVLocs;
3378 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3380 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3381 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3382 CCValAssign &VA = RVLocs[i];
3383 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3388 // If the calling conventions do not match, then we'd better make sure the
3389 // results are returned in the same way as what the caller expects.
3391 SmallVector<CCValAssign, 16> RVLocs1;
3392 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3394 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3396 SmallVector<CCValAssign, 16> RVLocs2;
3397 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3399 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3401 if (RVLocs1.size() != RVLocs2.size())
3403 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3404 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3406 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3408 if (RVLocs1[i].isRegLoc()) {
3409 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3412 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3418 // If the callee takes no arguments then go on to check the results of the
3420 if (!Outs.empty()) {
3421 // Check if stack adjustment is needed. For now, do not do this if any
3422 // argument is passed on the stack.
3423 SmallVector<CCValAssign, 16> ArgLocs;
3424 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3427 // Allocate shadow area for Win64
3429 CCInfo.AllocateStack(32, 8);
3431 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3432 if (CCInfo.getNextStackOffset()) {
3433 MachineFunction &MF = DAG.getMachineFunction();
3434 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3437 // Check if the arguments are already laid out in the right way as
3438 // the caller's fixed stack objects.
3439 MachineFrameInfo *MFI = MF.getFrameInfo();
3440 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3441 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3442 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3443 CCValAssign &VA = ArgLocs[i];
3444 SDValue Arg = OutVals[i];
3445 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3446 if (VA.getLocInfo() == CCValAssign::Indirect)
3448 if (!VA.isRegLoc()) {
3449 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3456 // If the tailcall address may be in a register, then make sure it's
3457 // possible to register allocate for it. In 32-bit, the call address can
3458 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3459 // callee-saved registers are restored. These happen to be the same
3460 // registers used to pass 'inreg' arguments so watch out for those.
3461 if (!Subtarget->is64Bit() &&
3462 ((!isa<GlobalAddressSDNode>(Callee) &&
3463 !isa<ExternalSymbolSDNode>(Callee)) ||
3464 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3465 unsigned NumInRegs = 0;
3466 // In PIC we need an extra register to formulate the address computation
3468 unsigned MaxInRegs =
3469 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3471 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3472 CCValAssign &VA = ArgLocs[i];
3475 unsigned Reg = VA.getLocReg();
3478 case X86::EAX: case X86::EDX: case X86::ECX:
3479 if (++NumInRegs == MaxInRegs)
3491 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3492 const TargetLibraryInfo *libInfo) const {
3493 return X86::createFastISel(funcInfo, libInfo);
3496 //===----------------------------------------------------------------------===//
3497 // Other Lowering Hooks
3498 //===----------------------------------------------------------------------===//
3500 static bool MayFoldLoad(SDValue Op) {
3501 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3504 static bool MayFoldIntoStore(SDValue Op) {
3505 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3508 static bool isTargetShuffle(unsigned Opcode) {
3510 default: return false;
3511 case X86ISD::BLENDI:
3512 case X86ISD::PSHUFB:
3513 case X86ISD::PSHUFD:
3514 case X86ISD::PSHUFHW:
3515 case X86ISD::PSHUFLW:
3517 case X86ISD::PALIGNR:
3518 case X86ISD::MOVLHPS:
3519 case X86ISD::MOVLHPD:
3520 case X86ISD::MOVHLPS:
3521 case X86ISD::MOVLPS:
3522 case X86ISD::MOVLPD:
3523 case X86ISD::MOVSHDUP:
3524 case X86ISD::MOVSLDUP:
3525 case X86ISD::MOVDDUP:
3528 case X86ISD::UNPCKL:
3529 case X86ISD::UNPCKH:
3530 case X86ISD::VPERMILPI:
3531 case X86ISD::VPERM2X128:
3532 case X86ISD::VPERMI:
3537 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3538 SDValue V1, unsigned TargetMask,
3539 SelectionDAG &DAG) {
3541 default: llvm_unreachable("Unknown x86 shuffle node");
3542 case X86ISD::PSHUFD:
3543 case X86ISD::PSHUFHW:
3544 case X86ISD::PSHUFLW:
3545 case X86ISD::VPERMILPI:
3546 case X86ISD::VPERMI:
3547 return DAG.getNode(Opc, dl, VT, V1,
3548 DAG.getConstant(TargetMask, dl, MVT::i8));
3552 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3553 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3555 default: llvm_unreachable("Unknown x86 shuffle node");
3556 case X86ISD::MOVLHPS:
3557 case X86ISD::MOVLHPD:
3558 case X86ISD::MOVHLPS:
3559 case X86ISD::MOVLPS:
3560 case X86ISD::MOVLPD:
3563 case X86ISD::UNPCKL:
3564 case X86ISD::UNPCKH:
3565 return DAG.getNode(Opc, dl, VT, V1, V2);
3569 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3570 MachineFunction &MF = DAG.getMachineFunction();
3571 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3572 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3573 int ReturnAddrIndex = FuncInfo->getRAIndex();
3575 if (ReturnAddrIndex == 0) {
3576 // Set up a frame object for the return address.
3577 unsigned SlotSize = RegInfo->getSlotSize();
3578 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3581 FuncInfo->setRAIndex(ReturnAddrIndex);
3584 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3587 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3588 bool hasSymbolicDisplacement) {
3589 // Offset should fit into 32 bit immediate field.
3590 if (!isInt<32>(Offset))
3593 // If we don't have a symbolic displacement - we don't have any extra
3595 if (!hasSymbolicDisplacement)
3598 // FIXME: Some tweaks might be needed for medium code model.
3599 if (M != CodeModel::Small && M != CodeModel::Kernel)
3602 // For small code model we assume that latest object is 16MB before end of 31
3603 // bits boundary. We may also accept pretty large negative constants knowing
3604 // that all objects are in the positive half of address space.
3605 if (M == CodeModel::Small && Offset < 16*1024*1024)
3608 // For kernel code model we know that all object resist in the negative half
3609 // of 32bits address space. We may not accept negative offsets, since they may
3610 // be just off and we may accept pretty large positive ones.
3611 if (M == CodeModel::Kernel && Offset >= 0)
3617 /// isCalleePop - Determines whether the callee is required to pop its
3618 /// own arguments. Callee pop is necessary to support tail calls.
3619 bool X86::isCalleePop(CallingConv::ID CallingConv,
3620 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3621 switch (CallingConv) {
3624 case CallingConv::X86_StdCall:
3625 case CallingConv::X86_FastCall:
3626 case CallingConv::X86_ThisCall:
3628 case CallingConv::Fast:
3629 case CallingConv::GHC:
3630 case CallingConv::HiPE:
3637 /// \brief Return true if the condition is an unsigned comparison operation.
3638 static bool isX86CCUnsigned(unsigned X86CC) {
3640 default: llvm_unreachable("Invalid integer condition!");
3641 case X86::COND_E: return true;
3642 case X86::COND_G: return false;
3643 case X86::COND_GE: return false;
3644 case X86::COND_L: return false;
3645 case X86::COND_LE: return false;
3646 case X86::COND_NE: return true;
3647 case X86::COND_B: return true;
3648 case X86::COND_A: return true;
3649 case X86::COND_BE: return true;
3650 case X86::COND_AE: return true;
3652 llvm_unreachable("covered switch fell through?!");
3655 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3656 /// specific condition code, returning the condition code and the LHS/RHS of the
3657 /// comparison to make.
3658 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3659 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3661 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3662 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3663 // X > -1 -> X == 0, jump !sign.
3664 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3665 return X86::COND_NS;
3667 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3668 // X < 0 -> X == 0, jump on sign.
3671 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3673 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3674 return X86::COND_LE;
3678 switch (SetCCOpcode) {
3679 default: llvm_unreachable("Invalid integer condition!");
3680 case ISD::SETEQ: return X86::COND_E;
3681 case ISD::SETGT: return X86::COND_G;
3682 case ISD::SETGE: return X86::COND_GE;
3683 case ISD::SETLT: return X86::COND_L;
3684 case ISD::SETLE: return X86::COND_LE;
3685 case ISD::SETNE: return X86::COND_NE;
3686 case ISD::SETULT: return X86::COND_B;
3687 case ISD::SETUGT: return X86::COND_A;
3688 case ISD::SETULE: return X86::COND_BE;
3689 case ISD::SETUGE: return X86::COND_AE;
3693 // First determine if it is required or is profitable to flip the operands.
3695 // If LHS is a foldable load, but RHS is not, flip the condition.
3696 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3697 !ISD::isNON_EXTLoad(RHS.getNode())) {
3698 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3699 std::swap(LHS, RHS);
3702 switch (SetCCOpcode) {
3708 std::swap(LHS, RHS);
3712 // On a floating point condition, the flags are set as follows:
3714 // 0 | 0 | 0 | X > Y
3715 // 0 | 0 | 1 | X < Y
3716 // 1 | 0 | 0 | X == Y
3717 // 1 | 1 | 1 | unordered
3718 switch (SetCCOpcode) {
3719 default: llvm_unreachable("Condcode should be pre-legalized away");
3721 case ISD::SETEQ: return X86::COND_E;
3722 case ISD::SETOLT: // flipped
3724 case ISD::SETGT: return X86::COND_A;
3725 case ISD::SETOLE: // flipped
3727 case ISD::SETGE: return X86::COND_AE;
3728 case ISD::SETUGT: // flipped
3730 case ISD::SETLT: return X86::COND_B;
3731 case ISD::SETUGE: // flipped
3733 case ISD::SETLE: return X86::COND_BE;
3735 case ISD::SETNE: return X86::COND_NE;
3736 case ISD::SETUO: return X86::COND_P;
3737 case ISD::SETO: return X86::COND_NP;
3739 case ISD::SETUNE: return X86::COND_INVALID;
3743 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3744 /// code. Current x86 isa includes the following FP cmov instructions:
3745 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3746 static bool hasFPCMov(unsigned X86CC) {
3762 /// isFPImmLegal - Returns true if the target can instruction select the
3763 /// specified FP immediate natively. If false, the legalizer will
3764 /// materialize the FP immediate as a load from a constant pool.
3765 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3766 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3767 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3773 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3774 ISD::LoadExtType ExtTy,
3776 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3777 // relocation target a movq or addq instruction: don't let the load shrink.
3778 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3779 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3780 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3781 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3785 /// \brief Returns true if it is beneficial to convert a load of a constant
3786 /// to just the constant itself.
3787 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3789 assert(Ty->isIntegerTy());
3791 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3792 if (BitSize == 0 || BitSize > 64)
3797 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3798 unsigned Index) const {
3799 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3802 return (Index == 0 || Index == ResVT.getVectorNumElements());
3805 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3806 // Speculate cttz only if we can directly use TZCNT.
3807 return Subtarget->hasBMI();
3810 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3811 // Speculate ctlz only if we can directly use LZCNT.
3812 return Subtarget->hasLZCNT();
3815 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3816 /// the specified range (L, H].
3817 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3818 return (Val < 0) || (Val >= Low && Val < Hi);
3821 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3822 /// specified value.
3823 static bool isUndefOrEqual(int Val, int CmpVal) {
3824 return (Val < 0 || Val == CmpVal);
3827 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3828 /// from position Pos and ending in Pos+Size, falls within the specified
3829 /// sequential range (Low, Low+Size]. or is undef.
3830 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3831 unsigned Pos, unsigned Size, int Low) {
3832 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3833 if (!isUndefOrEqual(Mask[i], Low))
3838 /// isVEXTRACTIndex - Return true if the specified
3839 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3840 /// suitable for instruction that extract 128 or 256 bit vectors
3841 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3842 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3843 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3846 // The index should be aligned on a vecWidth-bit boundary.
3848 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3850 MVT VT = N->getSimpleValueType(0);
3851 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3852 bool Result = (Index * ElSize) % vecWidth == 0;
3857 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3858 /// operand specifies a subvector insert that is suitable for input to
3859 /// insertion of 128 or 256-bit subvectors
3860 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3861 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3862 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3864 // The index should be aligned on a vecWidth-bit boundary.
3866 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3868 MVT VT = N->getSimpleValueType(0);
3869 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3870 bool Result = (Index * ElSize) % vecWidth == 0;
3875 bool X86::isVINSERT128Index(SDNode *N) {
3876 return isVINSERTIndex(N, 128);
3879 bool X86::isVINSERT256Index(SDNode *N) {
3880 return isVINSERTIndex(N, 256);
3883 bool X86::isVEXTRACT128Index(SDNode *N) {
3884 return isVEXTRACTIndex(N, 128);
3887 bool X86::isVEXTRACT256Index(SDNode *N) {
3888 return isVEXTRACTIndex(N, 256);
3891 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3892 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3893 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3894 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3897 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3899 MVT VecVT = N->getOperand(0).getSimpleValueType();
3900 MVT ElVT = VecVT.getVectorElementType();
3902 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3903 return Index / NumElemsPerChunk;
3906 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3907 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3908 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3909 llvm_unreachable("Illegal insert subvector for VINSERT");
3912 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3914 MVT VecVT = N->getSimpleValueType(0);
3915 MVT ElVT = VecVT.getVectorElementType();
3917 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3918 return Index / NumElemsPerChunk;
3921 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
3922 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3923 /// and VINSERTI128 instructions.
3924 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
3925 return getExtractVEXTRACTImmediate(N, 128);
3928 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
3929 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
3930 /// and VINSERTI64x4 instructions.
3931 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
3932 return getExtractVEXTRACTImmediate(N, 256);
3935 /// getInsertVINSERT128Immediate - Return the appropriate immediate
3936 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3937 /// and VINSERTI128 instructions.
3938 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
3939 return getInsertVINSERTImmediate(N, 128);
3942 /// getInsertVINSERT256Immediate - Return the appropriate immediate
3943 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
3944 /// and VINSERTI64x4 instructions.
3945 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
3946 return getInsertVINSERTImmediate(N, 256);
3949 /// isZero - Returns true if Elt is a constant integer zero
3950 static bool isZero(SDValue V) {
3951 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
3952 return C && C->isNullValue();
3955 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3957 bool X86::isZeroNode(SDValue Elt) {
3960 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
3961 return CFP->getValueAPF().isPosZero();
3965 /// getZeroVector - Returns a vector of specified type with all zero elements.
3967 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
3968 SelectionDAG &DAG, SDLoc dl) {
3969 assert(VT.isVector() && "Expected a vector type");
3971 // Always build SSE zero vectors as <4 x i32> bitcasted
3972 // to their dest type. This ensures they get CSE'd.
3974 if (VT.is128BitVector()) { // SSE
3975 if (Subtarget->hasSSE2()) { // SSE2
3976 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
3977 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3979 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
3980 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3982 } else if (VT.is256BitVector()) { // AVX
3983 if (Subtarget->hasInt256()) { // AVX2
3984 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
3985 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3986 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
3988 // 256-bit logic and arithmetic instructions in AVX are all
3989 // floating-point, no support for integer ops. Emit fp zeroed vectors.
3990 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
3991 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3992 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
3994 } else if (VT.is512BitVector()) { // AVX-512
3995 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
3996 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
3997 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3998 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
3999 } else if (VT.getScalarType() == MVT::i1) {
4001 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4002 && "Unexpected vector type");
4003 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4004 && "Unexpected vector type");
4005 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4006 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4007 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4009 llvm_unreachable("Unexpected vector type");
4011 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4014 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4015 SelectionDAG &DAG, SDLoc dl,
4016 unsigned vectorWidth) {
4017 assert((vectorWidth == 128 || vectorWidth == 256) &&
4018 "Unsupported vector width");
4019 EVT VT = Vec.getValueType();
4020 EVT ElVT = VT.getVectorElementType();
4021 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4022 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4023 VT.getVectorNumElements()/Factor);
4025 // Extract from UNDEF is UNDEF.
4026 if (Vec.getOpcode() == ISD::UNDEF)
4027 return DAG.getUNDEF(ResultVT);
4029 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4030 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4032 // This is the index of the first element of the vectorWidth-bit chunk
4034 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4037 // If the input is a buildvector just emit a smaller one.
4038 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4039 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4040 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4043 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4044 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4047 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4048 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4049 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4050 /// instructions or a simple subregister reference. Idx is an index in the
4051 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4052 /// lowering EXTRACT_VECTOR_ELT operations easier.
4053 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4054 SelectionDAG &DAG, SDLoc dl) {
4055 assert((Vec.getValueType().is256BitVector() ||
4056 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4057 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4060 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4061 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4062 SelectionDAG &DAG, SDLoc dl) {
4063 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4064 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4067 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4068 unsigned IdxVal, SelectionDAG &DAG,
4069 SDLoc dl, unsigned vectorWidth) {
4070 assert((vectorWidth == 128 || vectorWidth == 256) &&
4071 "Unsupported vector width");
4072 // Inserting UNDEF is Result
4073 if (Vec.getOpcode() == ISD::UNDEF)
4075 EVT VT = Vec.getValueType();
4076 EVT ElVT = VT.getVectorElementType();
4077 EVT ResultVT = Result.getValueType();
4079 // Insert the relevant vectorWidth bits.
4080 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4082 // This is the index of the first element of the vectorWidth-bit chunk
4084 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4087 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4088 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4091 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4092 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4093 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4094 /// simple superregister reference. Idx is an index in the 128 bits
4095 /// we want. It need not be aligned to a 128-bit boundary. That makes
4096 /// lowering INSERT_VECTOR_ELT operations easier.
4097 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4098 SelectionDAG &DAG, SDLoc dl) {
4099 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4101 // For insertion into the zero index (low half) of a 256-bit vector, it is
4102 // more efficient to generate a blend with immediate instead of an insert*128.
4103 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4104 // extend the subvector to the size of the result vector. Make sure that
4105 // we are not recursing on that node by checking for undef here.
4106 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4107 Result.getOpcode() != ISD::UNDEF) {
4108 EVT ResultVT = Result.getValueType();
4109 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4110 SDValue Undef = DAG.getUNDEF(ResultVT);
4111 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4114 // The blend instruction, and therefore its mask, depend on the data type.
4115 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4116 if (ScalarType.isFloatingPoint()) {
4117 // Choose either vblendps (float) or vblendpd (double).
4118 unsigned ScalarSize = ScalarType.getSizeInBits();
4119 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4120 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4121 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4122 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4125 const X86Subtarget &Subtarget =
4126 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4128 // AVX2 is needed for 256-bit integer blend support.
4129 // Integers must be cast to 32-bit because there is only vpblendd;
4130 // vpblendw can't be used for this because it has a handicapped mask.
4132 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4133 // is still more efficient than using the wrong domain vinsertf128 that
4134 // will be created by InsertSubVector().
4135 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4137 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4138 Vec256 = DAG.getNode(ISD::BITCAST, dl, CastVT, Vec256);
4139 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4140 return DAG.getNode(ISD::BITCAST, dl, ResultVT, Vec256);
4143 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4146 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4147 SelectionDAG &DAG, SDLoc dl) {
4148 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4149 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4152 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4153 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4154 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4155 /// large BUILD_VECTORS.
4156 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4157 unsigned NumElems, SelectionDAG &DAG,
4159 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4160 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4163 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4164 unsigned NumElems, SelectionDAG &DAG,
4166 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4167 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4170 /// getOnesVector - Returns a vector of specified type with all bits set.
4171 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4172 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4173 /// Then bitcast to their original type, ensuring they get CSE'd.
4174 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4176 assert(VT.isVector() && "Expected a vector type");
4178 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4180 if (VT.is256BitVector()) {
4181 if (HasInt256) { // AVX2
4182 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4183 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4185 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4186 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4188 } else if (VT.is128BitVector()) {
4189 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4191 llvm_unreachable("Unexpected vector type");
4193 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4196 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4197 /// operation of specified width.
4198 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4200 unsigned NumElems = VT.getVectorNumElements();
4201 SmallVector<int, 8> Mask;
4202 Mask.push_back(NumElems);
4203 for (unsigned i = 1; i != NumElems; ++i)
4205 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4208 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4209 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4211 unsigned NumElems = VT.getVectorNumElements();
4212 SmallVector<int, 8> Mask;
4213 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4215 Mask.push_back(i + NumElems);
4217 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4220 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4221 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4223 unsigned NumElems = VT.getVectorNumElements();
4224 SmallVector<int, 8> Mask;
4225 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4226 Mask.push_back(i + Half);
4227 Mask.push_back(i + NumElems + Half);
4229 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4232 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4233 /// vector of zero or undef vector. This produces a shuffle where the low
4234 /// element of V2 is swizzled into the zero/undef vector, landing at element
4235 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4236 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4238 const X86Subtarget *Subtarget,
4239 SelectionDAG &DAG) {
4240 MVT VT = V2.getSimpleValueType();
4242 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4243 unsigned NumElems = VT.getVectorNumElements();
4244 SmallVector<int, 16> MaskVec;
4245 for (unsigned i = 0; i != NumElems; ++i)
4246 // If this is the insertion idx, put the low elt of V2 here.
4247 MaskVec.push_back(i == Idx ? NumElems : i);
4248 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4251 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4252 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4253 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4254 /// shuffles which use a single input multiple times, and in those cases it will
4255 /// adjust the mask to only have indices within that single input.
4256 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4257 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4258 unsigned NumElems = VT.getVectorNumElements();
4262 bool IsFakeUnary = false;
4263 switch(N->getOpcode()) {
4264 case X86ISD::BLENDI:
4265 ImmN = N->getOperand(N->getNumOperands()-1);
4266 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4269 ImmN = N->getOperand(N->getNumOperands()-1);
4270 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4271 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4273 case X86ISD::UNPCKH:
4274 DecodeUNPCKHMask(VT, Mask);
4275 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4277 case X86ISD::UNPCKL:
4278 DecodeUNPCKLMask(VT, Mask);
4279 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4281 case X86ISD::MOVHLPS:
4282 DecodeMOVHLPSMask(NumElems, Mask);
4283 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4285 case X86ISD::MOVLHPS:
4286 DecodeMOVLHPSMask(NumElems, Mask);
4287 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4289 case X86ISD::PALIGNR:
4290 ImmN = N->getOperand(N->getNumOperands()-1);
4291 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4293 case X86ISD::PSHUFD:
4294 case X86ISD::VPERMILPI:
4295 ImmN = N->getOperand(N->getNumOperands()-1);
4296 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4299 case X86ISD::PSHUFHW:
4300 ImmN = N->getOperand(N->getNumOperands()-1);
4301 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4304 case X86ISD::PSHUFLW:
4305 ImmN = N->getOperand(N->getNumOperands()-1);
4306 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4309 case X86ISD::PSHUFB: {
4311 SDValue MaskNode = N->getOperand(1);
4312 while (MaskNode->getOpcode() == ISD::BITCAST)
4313 MaskNode = MaskNode->getOperand(0);
4315 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4316 // If we have a build-vector, then things are easy.
4317 EVT VT = MaskNode.getValueType();
4318 assert(VT.isVector() &&
4319 "Can't produce a non-vector with a build_vector!");
4320 if (!VT.isInteger())
4323 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4325 SmallVector<uint64_t, 32> RawMask;
4326 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4327 SDValue Op = MaskNode->getOperand(i);
4328 if (Op->getOpcode() == ISD::UNDEF) {
4329 RawMask.push_back((uint64_t)SM_SentinelUndef);
4332 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4335 APInt MaskElement = CN->getAPIntValue();
4337 // We now have to decode the element which could be any integer size and
4338 // extract each byte of it.
4339 for (int j = 0; j < NumBytesPerElement; ++j) {
4340 // Note that this is x86 and so always little endian: the low byte is
4341 // the first byte of the mask.
4342 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4343 MaskElement = MaskElement.lshr(8);
4346 DecodePSHUFBMask(RawMask, Mask);
4350 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4354 SDValue Ptr = MaskLoad->getBasePtr();
4355 if (Ptr->getOpcode() == X86ISD::Wrapper)
4356 Ptr = Ptr->getOperand(0);
4358 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4359 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4362 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4363 DecodePSHUFBMask(C, Mask);
4371 case X86ISD::VPERMI:
4372 ImmN = N->getOperand(N->getNumOperands()-1);
4373 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4378 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4380 case X86ISD::VPERM2X128:
4381 ImmN = N->getOperand(N->getNumOperands()-1);
4382 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4383 if (Mask.empty()) return false;
4385 case X86ISD::MOVSLDUP:
4386 DecodeMOVSLDUPMask(VT, Mask);
4389 case X86ISD::MOVSHDUP:
4390 DecodeMOVSHDUPMask(VT, Mask);
4393 case X86ISD::MOVDDUP:
4394 DecodeMOVDDUPMask(VT, Mask);
4397 case X86ISD::MOVLHPD:
4398 case X86ISD::MOVLPD:
4399 case X86ISD::MOVLPS:
4400 // Not yet implemented
4402 default: llvm_unreachable("unknown target shuffle node");
4405 // If we have a fake unary shuffle, the shuffle mask is spread across two
4406 // inputs that are actually the same node. Re-map the mask to always point
4407 // into the first input.
4410 if (M >= (int)Mask.size())
4416 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4417 /// element of the result of the vector shuffle.
4418 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4421 return SDValue(); // Limit search depth.
4423 SDValue V = SDValue(N, 0);
4424 EVT VT = V.getValueType();
4425 unsigned Opcode = V.getOpcode();
4427 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4428 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4429 int Elt = SV->getMaskElt(Index);
4432 return DAG.getUNDEF(VT.getVectorElementType());
4434 unsigned NumElems = VT.getVectorNumElements();
4435 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4436 : SV->getOperand(1);
4437 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4440 // Recurse into target specific vector shuffles to find scalars.
4441 if (isTargetShuffle(Opcode)) {
4442 MVT ShufVT = V.getSimpleValueType();
4443 unsigned NumElems = ShufVT.getVectorNumElements();
4444 SmallVector<int, 16> ShuffleMask;
4447 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4450 int Elt = ShuffleMask[Index];
4452 return DAG.getUNDEF(ShufVT.getVectorElementType());
4454 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4456 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4460 // Actual nodes that may contain scalar elements
4461 if (Opcode == ISD::BITCAST) {
4462 V = V.getOperand(0);
4463 EVT SrcVT = V.getValueType();
4464 unsigned NumElems = VT.getVectorNumElements();
4466 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4470 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4471 return (Index == 0) ? V.getOperand(0)
4472 : DAG.getUNDEF(VT.getVectorElementType());
4474 if (V.getOpcode() == ISD::BUILD_VECTOR)
4475 return V.getOperand(Index);
4480 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4482 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4483 unsigned NumNonZero, unsigned NumZero,
4485 const X86Subtarget* Subtarget,
4486 const TargetLowering &TLI) {
4494 // SSE4.1 - use PINSRB to insert each byte directly.
4495 if (Subtarget->hasSSE41()) {
4496 for (unsigned i = 0; i < 16; ++i) {
4497 bool isNonZero = (NonZeros & (1 << i)) != 0;
4501 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4503 V = DAG.getUNDEF(MVT::v16i8);
4506 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4507 MVT::v16i8, V, Op.getOperand(i),
4508 DAG.getIntPtrConstant(i, dl));
4515 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4516 for (unsigned i = 0; i < 16; ++i) {
4517 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4518 if (ThisIsNonZero && First) {
4520 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4522 V = DAG.getUNDEF(MVT::v8i16);
4527 SDValue ThisElt, LastElt;
4528 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4529 if (LastIsNonZero) {
4530 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4531 MVT::i16, Op.getOperand(i-1));
4533 if (ThisIsNonZero) {
4534 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4535 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4536 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4538 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4542 if (ThisElt.getNode())
4543 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4544 DAG.getIntPtrConstant(i/2, dl));
4548 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4551 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4553 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4554 unsigned NumNonZero, unsigned NumZero,
4556 const X86Subtarget* Subtarget,
4557 const TargetLowering &TLI) {
4564 for (unsigned i = 0; i < 8; ++i) {
4565 bool isNonZero = (NonZeros & (1 << i)) != 0;
4569 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4571 V = DAG.getUNDEF(MVT::v8i16);
4574 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4575 MVT::v8i16, V, Op.getOperand(i),
4576 DAG.getIntPtrConstant(i, dl));
4583 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4584 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4585 const X86Subtarget *Subtarget,
4586 const TargetLowering &TLI) {
4587 // Find all zeroable elements.
4588 std::bitset<4> Zeroable;
4589 for (int i=0; i < 4; ++i) {
4590 SDValue Elt = Op->getOperand(i);
4591 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4593 assert(Zeroable.size() - Zeroable.count() > 1 &&
4594 "We expect at least two non-zero elements!");
4596 // We only know how to deal with build_vector nodes where elements are either
4597 // zeroable or extract_vector_elt with constant index.
4598 SDValue FirstNonZero;
4599 unsigned FirstNonZeroIdx;
4600 for (unsigned i=0; i < 4; ++i) {
4603 SDValue Elt = Op->getOperand(i);
4604 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4605 !isa<ConstantSDNode>(Elt.getOperand(1)))
4607 // Make sure that this node is extracting from a 128-bit vector.
4608 MVT VT = Elt.getOperand(0).getSimpleValueType();
4609 if (!VT.is128BitVector())
4611 if (!FirstNonZero.getNode()) {
4613 FirstNonZeroIdx = i;
4617 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4618 SDValue V1 = FirstNonZero.getOperand(0);
4619 MVT VT = V1.getSimpleValueType();
4621 // See if this build_vector can be lowered as a blend with zero.
4623 unsigned EltMaskIdx, EltIdx;
4625 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4626 if (Zeroable[EltIdx]) {
4627 // The zero vector will be on the right hand side.
4628 Mask[EltIdx] = EltIdx+4;
4632 Elt = Op->getOperand(EltIdx);
4633 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4634 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4635 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4637 Mask[EltIdx] = EltIdx;
4641 // Let the shuffle legalizer deal with blend operations.
4642 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4643 if (V1.getSimpleValueType() != VT)
4644 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4645 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4648 // See if we can lower this build_vector to a INSERTPS.
4649 if (!Subtarget->hasSSE41())
4652 SDValue V2 = Elt.getOperand(0);
4653 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4656 bool CanFold = true;
4657 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4661 SDValue Current = Op->getOperand(i);
4662 SDValue SrcVector = Current->getOperand(0);
4665 CanFold = SrcVector == V1 &&
4666 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4672 assert(V1.getNode() && "Expected at least two non-zero elements!");
4673 if (V1.getSimpleValueType() != MVT::v4f32)
4674 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4675 if (V2.getSimpleValueType() != MVT::v4f32)
4676 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4678 // Ok, we can emit an INSERTPS instruction.
4679 unsigned ZMask = Zeroable.to_ulong();
4681 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4682 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4684 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4685 DAG.getIntPtrConstant(InsertPSMask, DL));
4686 return DAG.getNode(ISD::BITCAST, DL, VT, Result);
4689 /// Return a vector logical shift node.
4690 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4691 unsigned NumBits, SelectionDAG &DAG,
4692 const TargetLowering &TLI, SDLoc dl) {
4693 assert(VT.is128BitVector() && "Unknown type for VShift");
4694 MVT ShVT = MVT::v2i64;
4695 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4696 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4697 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4698 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4699 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4700 return DAG.getNode(ISD::BITCAST, dl, VT,
4701 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4705 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4707 // Check if the scalar load can be widened into a vector load. And if
4708 // the address is "base + cst" see if the cst can be "absorbed" into
4709 // the shuffle mask.
4710 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4711 SDValue Ptr = LD->getBasePtr();
4712 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4714 EVT PVT = LD->getValueType(0);
4715 if (PVT != MVT::i32 && PVT != MVT::f32)
4720 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4721 FI = FINode->getIndex();
4723 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4724 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4725 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4726 Offset = Ptr.getConstantOperandVal(1);
4727 Ptr = Ptr.getOperand(0);
4732 // FIXME: 256-bit vector instructions don't require a strict alignment,
4733 // improve this code to support it better.
4734 unsigned RequiredAlign = VT.getSizeInBits()/8;
4735 SDValue Chain = LD->getChain();
4736 // Make sure the stack object alignment is at least 16 or 32.
4737 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4738 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4739 if (MFI->isFixedObjectIndex(FI)) {
4740 // Can't change the alignment. FIXME: It's possible to compute
4741 // the exact stack offset and reference FI + adjust offset instead.
4742 // If someone *really* cares about this. That's the way to implement it.
4745 MFI->setObjectAlignment(FI, RequiredAlign);
4749 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4750 // Ptr + (Offset & ~15).
4753 if ((Offset % RequiredAlign) & 3)
4755 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4758 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4759 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4762 int EltNo = (Offset - StartOffset) >> 2;
4763 unsigned NumElems = VT.getVectorNumElements();
4765 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4766 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4767 LD->getPointerInfo().getWithOffset(StartOffset),
4768 false, false, false, 0);
4770 SmallVector<int, 8> Mask(NumElems, EltNo);
4772 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4778 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4779 /// elements can be replaced by a single large load which has the same value as
4780 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4782 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4784 /// FIXME: we'd also like to handle the case where the last elements are zero
4785 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4786 /// There's even a handy isZeroNode for that purpose.
4787 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4788 SDLoc &DL, SelectionDAG &DAG,
4789 bool isAfterLegalize) {
4790 unsigned NumElems = Elts.size();
4792 LoadSDNode *LDBase = nullptr;
4793 unsigned LastLoadedElt = -1U;
4795 // For each element in the initializer, see if we've found a load or an undef.
4796 // If we don't find an initial load element, or later load elements are
4797 // non-consecutive, bail out.
4798 for (unsigned i = 0; i < NumElems; ++i) {
4799 SDValue Elt = Elts[i];
4800 // Look through a bitcast.
4801 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4802 Elt = Elt.getOperand(0);
4803 if (!Elt.getNode() ||
4804 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4807 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4809 LDBase = cast<LoadSDNode>(Elt.getNode());
4813 if (Elt.getOpcode() == ISD::UNDEF)
4816 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4817 EVT LdVT = Elt.getValueType();
4818 // Each loaded element must be the correct fractional portion of the
4819 // requested vector load.
4820 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4822 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4827 // If we have found an entire vector of loads and undefs, then return a large
4828 // load of the entire vector width starting at the base pointer. If we found
4829 // consecutive loads for the low half, generate a vzext_load node.
4830 if (LastLoadedElt == NumElems - 1) {
4831 assert(LDBase && "Did not find base load for merging consecutive loads");
4832 EVT EltVT = LDBase->getValueType(0);
4833 // Ensure that the input vector size for the merged loads matches the
4834 // cumulative size of the input elements.
4835 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4838 if (isAfterLegalize &&
4839 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4842 SDValue NewLd = SDValue();
4844 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4845 LDBase->getPointerInfo(), LDBase->isVolatile(),
4846 LDBase->isNonTemporal(), LDBase->isInvariant(),
4847 LDBase->getAlignment());
4849 if (LDBase->hasAnyUseOfValue(1)) {
4850 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4852 SDValue(NewLd.getNode(), 1));
4853 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4854 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4855 SDValue(NewLd.getNode(), 1));
4861 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4862 //of a v4i32 / v4f32. It's probably worth generalizing.
4863 EVT EltVT = VT.getVectorElementType();
4864 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4865 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4866 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4867 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4869 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4870 LDBase->getPointerInfo(),
4871 LDBase->getAlignment(),
4872 false/*isVolatile*/, true/*ReadMem*/,
4875 // Make sure the newly-created LOAD is in the same position as LDBase in
4876 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4877 // update uses of LDBase's output chain to use the TokenFactor.
4878 if (LDBase->hasAnyUseOfValue(1)) {
4879 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4880 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4881 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4882 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4883 SDValue(ResNode.getNode(), 1));
4886 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4891 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4892 /// to generate a splat value for the following cases:
4893 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4894 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4895 /// a scalar load, or a constant.
4896 /// The VBROADCAST node is returned when a pattern is found,
4897 /// or SDValue() otherwise.
4898 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4899 SelectionDAG &DAG) {
4900 // VBROADCAST requires AVX.
4901 // TODO: Splats could be generated for non-AVX CPUs using SSE
4902 // instructions, but there's less potential gain for only 128-bit vectors.
4903 if (!Subtarget->hasAVX())
4906 MVT VT = Op.getSimpleValueType();
4909 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4910 "Unsupported vector type for broadcast.");
4915 switch (Op.getOpcode()) {
4917 // Unknown pattern found.
4920 case ISD::BUILD_VECTOR: {
4921 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4922 BitVector UndefElements;
4923 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4925 // We need a splat of a single value to use broadcast, and it doesn't
4926 // make any sense if the value is only in one element of the vector.
4927 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4931 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4932 Ld.getOpcode() == ISD::ConstantFP);
4934 // Make sure that all of the users of a non-constant load are from the
4935 // BUILD_VECTOR node.
4936 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4941 case ISD::VECTOR_SHUFFLE: {
4942 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4944 // Shuffles must have a splat mask where the first element is
4946 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4949 SDValue Sc = Op.getOperand(0);
4950 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4951 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4953 if (!Subtarget->hasInt256())
4956 // Use the register form of the broadcast instruction available on AVX2.
4957 if (VT.getSizeInBits() >= 256)
4958 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4959 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
4962 Ld = Sc.getOperand(0);
4963 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4964 Ld.getOpcode() == ISD::ConstantFP);
4966 // The scalar_to_vector node and the suspected
4967 // load node must have exactly one user.
4968 // Constants may have multiple users.
4970 // AVX-512 has register version of the broadcast
4971 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
4972 Ld.getValueType().getSizeInBits() >= 32;
4973 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
4980 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
4981 bool IsGE256 = (VT.getSizeInBits() >= 256);
4983 // When optimizing for size, generate up to 5 extra bytes for a broadcast
4984 // instruction to save 8 or more bytes of constant pool data.
4985 // TODO: If multiple splats are generated to load the same constant,
4986 // it may be detrimental to overall size. There needs to be a way to detect
4987 // that condition to know if this is truly a size win.
4988 const Function *F = DAG.getMachineFunction().getFunction();
4989 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
4991 // Handle broadcasting a single constant scalar from the constant pool
4993 // On Sandybridge (no AVX2), it is still better to load a constant vector
4994 // from the constant pool and not to broadcast it from a scalar.
4995 // But override that restriction when optimizing for size.
4996 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
4997 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
4998 EVT CVT = Ld.getValueType();
4999 assert(!CVT.isVector() && "Must not broadcast a vector type");
5001 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5002 // For size optimization, also splat v2f64 and v2i64, and for size opt
5003 // with AVX2, also splat i8 and i16.
5004 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5005 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5006 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5007 const Constant *C = nullptr;
5008 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5009 C = CI->getConstantIntValue();
5010 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5011 C = CF->getConstantFPValue();
5013 assert(C && "Invalid constant type");
5015 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5016 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5017 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5018 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5019 MachinePointerInfo::getConstantPool(),
5020 false, false, false, Alignment);
5022 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5026 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5028 // Handle AVX2 in-register broadcasts.
5029 if (!IsLoad && Subtarget->hasInt256() &&
5030 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5031 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5033 // The scalar source must be a normal load.
5037 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5038 (Subtarget->hasVLX() && ScalarSize == 64))
5039 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5041 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5042 // double since there is no vbroadcastsd xmm
5043 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5044 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5045 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5048 // Unsupported broadcast.
5052 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5053 /// underlying vector and index.
5055 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5057 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5059 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5060 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5063 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5065 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5067 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5068 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5071 // In this case the vector is the extract_subvector expression and the index
5072 // is 2, as specified by the shuffle.
5073 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5074 SDValue ShuffleVec = SVOp->getOperand(0);
5075 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5076 assert(ShuffleVecVT.getVectorElementType() ==
5077 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5079 int ShuffleIdx = SVOp->getMaskElt(Idx);
5080 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5081 ExtractedFromVec = ShuffleVec;
5087 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5088 MVT VT = Op.getSimpleValueType();
5090 // Skip if insert_vec_elt is not supported.
5091 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5092 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5096 unsigned NumElems = Op.getNumOperands();
5100 SmallVector<unsigned, 4> InsertIndices;
5101 SmallVector<int, 8> Mask(NumElems, -1);
5103 for (unsigned i = 0; i != NumElems; ++i) {
5104 unsigned Opc = Op.getOperand(i).getOpcode();
5106 if (Opc == ISD::UNDEF)
5109 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5110 // Quit if more than 1 elements need inserting.
5111 if (InsertIndices.size() > 1)
5114 InsertIndices.push_back(i);
5118 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5119 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5120 // Quit if non-constant index.
5121 if (!isa<ConstantSDNode>(ExtIdx))
5123 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5125 // Quit if extracted from vector of different type.
5126 if (ExtractedFromVec.getValueType() != VT)
5129 if (!VecIn1.getNode())
5130 VecIn1 = ExtractedFromVec;
5131 else if (VecIn1 != ExtractedFromVec) {
5132 if (!VecIn2.getNode())
5133 VecIn2 = ExtractedFromVec;
5134 else if (VecIn2 != ExtractedFromVec)
5135 // Quit if more than 2 vectors to shuffle
5139 if (ExtractedFromVec == VecIn1)
5141 else if (ExtractedFromVec == VecIn2)
5142 Mask[i] = Idx + NumElems;
5145 if (!VecIn1.getNode())
5148 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5149 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5150 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5151 unsigned Idx = InsertIndices[i];
5152 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5153 DAG.getIntPtrConstant(Idx, DL));
5159 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5161 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5163 MVT VT = Op.getSimpleValueType();
5164 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5165 "Unexpected type in LowerBUILD_VECTORvXi1!");
5168 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5169 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5170 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5171 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5174 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5175 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5176 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5177 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5180 bool AllContants = true;
5181 uint64_t Immediate = 0;
5182 int NonConstIdx = -1;
5183 bool IsSplat = true;
5184 unsigned NumNonConsts = 0;
5185 unsigned NumConsts = 0;
5186 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5187 SDValue In = Op.getOperand(idx);
5188 if (In.getOpcode() == ISD::UNDEF)
5190 if (!isa<ConstantSDNode>(In)) {
5191 AllContants = false;
5196 if (cast<ConstantSDNode>(In)->getZExtValue())
5197 Immediate |= (1ULL << idx);
5199 if (In != Op.getOperand(0))
5204 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5205 DAG.getConstant(Immediate, dl, MVT::i16));
5206 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5207 DAG.getIntPtrConstant(0, dl));
5210 if (NumNonConsts == 1 && NonConstIdx != 0) {
5213 SDValue VecAsImm = DAG.getConstant(Immediate, dl,
5214 MVT::getIntegerVT(VT.getSizeInBits()));
5215 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
5218 DstVec = DAG.getUNDEF(VT);
5219 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5220 Op.getOperand(NonConstIdx),
5221 DAG.getIntPtrConstant(NonConstIdx, dl));
5223 if (!IsSplat && (NonConstIdx != 0))
5224 llvm_unreachable("Unsupported BUILD_VECTOR operation");
5225 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
5228 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5229 DAG.getConstant(-1, dl, SelectVT),
5230 DAG.getConstant(0, dl, SelectVT));
5232 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5233 DAG.getConstant((Immediate | 1), dl, SelectVT),
5234 DAG.getConstant(Immediate, dl, SelectVT));
5235 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
5238 /// \brief Return true if \p N implements a horizontal binop and return the
5239 /// operands for the horizontal binop into V0 and V1.
5241 /// This is a helper function of LowerToHorizontalOp().
5242 /// This function checks that the build_vector \p N in input implements a
5243 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5244 /// operation to match.
5245 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5246 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5247 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5250 /// This function only analyzes elements of \p N whose indices are
5251 /// in range [BaseIdx, LastIdx).
5252 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5254 unsigned BaseIdx, unsigned LastIdx,
5255 SDValue &V0, SDValue &V1) {
5256 EVT VT = N->getValueType(0);
5258 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5259 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5260 "Invalid Vector in input!");
5262 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5263 bool CanFold = true;
5264 unsigned ExpectedVExtractIdx = BaseIdx;
5265 unsigned NumElts = LastIdx - BaseIdx;
5266 V0 = DAG.getUNDEF(VT);
5267 V1 = DAG.getUNDEF(VT);
5269 // Check if N implements a horizontal binop.
5270 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5271 SDValue Op = N->getOperand(i + BaseIdx);
5274 if (Op->getOpcode() == ISD::UNDEF) {
5275 // Update the expected vector extract index.
5276 if (i * 2 == NumElts)
5277 ExpectedVExtractIdx = BaseIdx;
5278 ExpectedVExtractIdx += 2;
5282 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5287 SDValue Op0 = Op.getOperand(0);
5288 SDValue Op1 = Op.getOperand(1);
5290 // Try to match the following pattern:
5291 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5292 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5293 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5294 Op0.getOperand(0) == Op1.getOperand(0) &&
5295 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5296 isa<ConstantSDNode>(Op1.getOperand(1)));
5300 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5301 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5303 if (i * 2 < NumElts) {
5304 if (V0.getOpcode() == ISD::UNDEF) {
5305 V0 = Op0.getOperand(0);
5306 if (V0.getValueType() != VT)
5310 if (V1.getOpcode() == ISD::UNDEF) {
5311 V1 = Op0.getOperand(0);
5312 if (V1.getValueType() != VT)
5315 if (i * 2 == NumElts)
5316 ExpectedVExtractIdx = BaseIdx;
5319 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5320 if (I0 == ExpectedVExtractIdx)
5321 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5322 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5323 // Try to match the following dag sequence:
5324 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5325 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5329 ExpectedVExtractIdx += 2;
5335 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5336 /// a concat_vector.
5338 /// This is a helper function of LowerToHorizontalOp().
5339 /// This function expects two 256-bit vectors called V0 and V1.
5340 /// At first, each vector is split into two separate 128-bit vectors.
5341 /// Then, the resulting 128-bit vectors are used to implement two
5342 /// horizontal binary operations.
5344 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5346 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5347 /// the two new horizontal binop.
5348 /// When Mode is set, the first horizontal binop dag node would take as input
5349 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5350 /// horizontal binop dag node would take as input the lower 128-bit of V1
5351 /// and the upper 128-bit of V1.
5353 /// HADD V0_LO, V0_HI
5354 /// HADD V1_LO, V1_HI
5356 /// Otherwise, the first horizontal binop dag node takes as input the lower
5357 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5358 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5360 /// HADD V0_LO, V1_LO
5361 /// HADD V0_HI, V1_HI
5363 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5364 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5365 /// the upper 128-bits of the result.
5366 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5367 SDLoc DL, SelectionDAG &DAG,
5368 unsigned X86Opcode, bool Mode,
5369 bool isUndefLO, bool isUndefHI) {
5370 EVT VT = V0.getValueType();
5371 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5372 "Invalid nodes in input!");
5374 unsigned NumElts = VT.getVectorNumElements();
5375 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5376 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5377 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5378 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5379 EVT NewVT = V0_LO.getValueType();
5381 SDValue LO = DAG.getUNDEF(NewVT);
5382 SDValue HI = DAG.getUNDEF(NewVT);
5385 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5386 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5387 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5388 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5389 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5391 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5392 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5393 V1_LO->getOpcode() != ISD::UNDEF))
5394 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5396 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5397 V1_HI->getOpcode() != ISD::UNDEF))
5398 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5401 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5404 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5406 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5407 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5408 EVT VT = BV->getValueType(0);
5409 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5410 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5414 unsigned NumElts = VT.getVectorNumElements();
5415 SDValue InVec0 = DAG.getUNDEF(VT);
5416 SDValue InVec1 = DAG.getUNDEF(VT);
5418 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5419 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5421 // Odd-numbered elements in the input build vector are obtained from
5422 // adding two integer/float elements.
5423 // Even-numbered elements in the input build vector are obtained from
5424 // subtracting two integer/float elements.
5425 unsigned ExpectedOpcode = ISD::FSUB;
5426 unsigned NextExpectedOpcode = ISD::FADD;
5427 bool AddFound = false;
5428 bool SubFound = false;
5430 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5431 SDValue Op = BV->getOperand(i);
5433 // Skip 'undef' values.
5434 unsigned Opcode = Op.getOpcode();
5435 if (Opcode == ISD::UNDEF) {
5436 std::swap(ExpectedOpcode, NextExpectedOpcode);
5440 // Early exit if we found an unexpected opcode.
5441 if (Opcode != ExpectedOpcode)
5444 SDValue Op0 = Op.getOperand(0);
5445 SDValue Op1 = Op.getOperand(1);
5447 // Try to match the following pattern:
5448 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5449 // Early exit if we cannot match that sequence.
5450 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5451 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5452 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5453 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5454 Op0.getOperand(1) != Op1.getOperand(1))
5457 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5461 // We found a valid add/sub node. Update the information accordingly.
5467 // Update InVec0 and InVec1.
5468 if (InVec0.getOpcode() == ISD::UNDEF) {
5469 InVec0 = Op0.getOperand(0);
5470 if (InVec0.getValueType() != VT)
5473 if (InVec1.getOpcode() == ISD::UNDEF) {
5474 InVec1 = Op1.getOperand(0);
5475 if (InVec1.getValueType() != VT)
5479 // Make sure that operands in input to each add/sub node always
5480 // come from a same pair of vectors.
5481 if (InVec0 != Op0.getOperand(0)) {
5482 if (ExpectedOpcode == ISD::FSUB)
5485 // FADD is commutable. Try to commute the operands
5486 // and then test again.
5487 std::swap(Op0, Op1);
5488 if (InVec0 != Op0.getOperand(0))
5492 if (InVec1 != Op1.getOperand(0))
5495 // Update the pair of expected opcodes.
5496 std::swap(ExpectedOpcode, NextExpectedOpcode);
5499 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5500 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5501 InVec1.getOpcode() != ISD::UNDEF)
5502 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5507 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5508 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5509 const X86Subtarget *Subtarget,
5510 SelectionDAG &DAG) {
5511 EVT VT = BV->getValueType(0);
5512 unsigned NumElts = VT.getVectorNumElements();
5513 unsigned NumUndefsLO = 0;
5514 unsigned NumUndefsHI = 0;
5515 unsigned Half = NumElts/2;
5517 // Count the number of UNDEF operands in the build_vector in input.
5518 for (unsigned i = 0, e = Half; i != e; ++i)
5519 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5522 for (unsigned i = Half, e = NumElts; i != e; ++i)
5523 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5526 // Early exit if this is either a build_vector of all UNDEFs or all the
5527 // operands but one are UNDEF.
5528 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5532 SDValue InVec0, InVec1;
5533 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5534 // Try to match an SSE3 float HADD/HSUB.
5535 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5536 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5538 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5539 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5540 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5541 // Try to match an SSSE3 integer HADD/HSUB.
5542 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5543 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5545 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5546 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5549 if (!Subtarget->hasAVX())
5552 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5553 // Try to match an AVX horizontal add/sub of packed single/double
5554 // precision floating point values from 256-bit vectors.
5555 SDValue InVec2, InVec3;
5556 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5557 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5558 ((InVec0.getOpcode() == ISD::UNDEF ||
5559 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5560 ((InVec1.getOpcode() == ISD::UNDEF ||
5561 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5562 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5564 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5565 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5566 ((InVec0.getOpcode() == ISD::UNDEF ||
5567 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5568 ((InVec1.getOpcode() == ISD::UNDEF ||
5569 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5570 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5571 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5572 // Try to match an AVX2 horizontal add/sub of signed integers.
5573 SDValue InVec2, InVec3;
5575 bool CanFold = true;
5577 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5578 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5579 ((InVec0.getOpcode() == ISD::UNDEF ||
5580 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5581 ((InVec1.getOpcode() == ISD::UNDEF ||
5582 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5583 X86Opcode = X86ISD::HADD;
5584 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5585 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5586 ((InVec0.getOpcode() == ISD::UNDEF ||
5587 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5588 ((InVec1.getOpcode() == ISD::UNDEF ||
5589 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5590 X86Opcode = X86ISD::HSUB;
5595 // Fold this build_vector into a single horizontal add/sub.
5596 // Do this only if the target has AVX2.
5597 if (Subtarget->hasAVX2())
5598 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5600 // Do not try to expand this build_vector into a pair of horizontal
5601 // add/sub if we can emit a pair of scalar add/sub.
5602 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5605 // Convert this build_vector into a pair of horizontal binop followed by
5607 bool isUndefLO = NumUndefsLO == Half;
5608 bool isUndefHI = NumUndefsHI == Half;
5609 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5610 isUndefLO, isUndefHI);
5614 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5615 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5617 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5618 X86Opcode = X86ISD::HADD;
5619 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5620 X86Opcode = X86ISD::HSUB;
5621 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5622 X86Opcode = X86ISD::FHADD;
5623 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5624 X86Opcode = X86ISD::FHSUB;
5628 // Don't try to expand this build_vector into a pair of horizontal add/sub
5629 // if we can simply emit a pair of scalar add/sub.
5630 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5633 // Convert this build_vector into two horizontal add/sub followed by
5635 bool isUndefLO = NumUndefsLO == Half;
5636 bool isUndefHI = NumUndefsHI == Half;
5637 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5638 isUndefLO, isUndefHI);
5645 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5648 MVT VT = Op.getSimpleValueType();
5649 MVT ExtVT = VT.getVectorElementType();
5650 unsigned NumElems = Op.getNumOperands();
5652 // Generate vectors for predicate vectors.
5653 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5654 return LowerBUILD_VECTORvXi1(Op, DAG);
5656 // Vectors containing all zeros can be matched by pxor and xorps later
5657 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5658 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5659 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5660 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5663 return getZeroVector(VT, Subtarget, DAG, dl);
5666 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5667 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5668 // vpcmpeqd on 256-bit vectors.
5669 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5670 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5673 if (!VT.is512BitVector())
5674 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5677 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5678 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5680 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5681 return HorizontalOp;
5682 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5685 unsigned EVTBits = ExtVT.getSizeInBits();
5687 unsigned NumZero = 0;
5688 unsigned NumNonZero = 0;
5689 unsigned NonZeros = 0;
5690 bool IsAllConstants = true;
5691 SmallSet<SDValue, 8> Values;
5692 for (unsigned i = 0; i < NumElems; ++i) {
5693 SDValue Elt = Op.getOperand(i);
5694 if (Elt.getOpcode() == ISD::UNDEF)
5697 if (Elt.getOpcode() != ISD::Constant &&
5698 Elt.getOpcode() != ISD::ConstantFP)
5699 IsAllConstants = false;
5700 if (X86::isZeroNode(Elt))
5703 NonZeros |= (1 << i);
5708 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5709 if (NumNonZero == 0)
5710 return DAG.getUNDEF(VT);
5712 // Special case for single non-zero, non-undef, element.
5713 if (NumNonZero == 1) {
5714 unsigned Idx = countTrailingZeros(NonZeros);
5715 SDValue Item = Op.getOperand(Idx);
5717 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5718 // the value are obviously zero, truncate the value to i32 and do the
5719 // insertion that way. Only do this if the value is non-constant or if the
5720 // value is a constant being inserted into element 0. It is cheaper to do
5721 // a constant pool load than it is to do a movd + shuffle.
5722 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5723 (!IsAllConstants || Idx == 0)) {
5724 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5726 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5727 EVT VecVT = MVT::v4i32;
5729 // Truncate the value (which may itself be a constant) to i32, and
5730 // convert it to a vector with movd (S2V+shuffle to zero extend).
5731 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5732 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5734 ISD::BITCAST, dl, VT,
5735 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5739 // If we have a constant or non-constant insertion into the low element of
5740 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5741 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5742 // depending on what the source datatype is.
5745 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5747 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5748 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5749 if (VT.is512BitVector()) {
5750 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5751 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5752 Item, DAG.getIntPtrConstant(0, dl));
5754 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5755 "Expected an SSE value type!");
5756 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5757 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5758 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5761 // We can't directly insert an i8 or i16 into a vector, so zero extend
5763 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5764 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5765 if (VT.is256BitVector()) {
5766 if (Subtarget->hasAVX()) {
5767 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5768 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5770 // Without AVX, we need to extend to a 128-bit vector and then
5771 // insert into the 256-bit vector.
5772 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5773 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5774 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5777 assert(VT.is128BitVector() && "Expected an SSE value type!");
5778 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5779 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5781 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5785 // Is it a vector logical left shift?
5786 if (NumElems == 2 && Idx == 1 &&
5787 X86::isZeroNode(Op.getOperand(0)) &&
5788 !X86::isZeroNode(Op.getOperand(1))) {
5789 unsigned NumBits = VT.getSizeInBits();
5790 return getVShift(true, VT,
5791 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5792 VT, Op.getOperand(1)),
5793 NumBits/2, DAG, *this, dl);
5796 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5799 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5800 // is a non-constant being inserted into an element other than the low one,
5801 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5802 // movd/movss) to move this into the low element, then shuffle it into
5804 if (EVTBits == 32) {
5805 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5806 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5810 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5811 if (Values.size() == 1) {
5812 if (EVTBits == 32) {
5813 // Instead of a shuffle like this:
5814 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5815 // Check if it's possible to issue this instead.
5816 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5817 unsigned Idx = countTrailingZeros(NonZeros);
5818 SDValue Item = Op.getOperand(Idx);
5819 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5820 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5825 // A vector full of immediates; various special cases are already
5826 // handled, so this is best done with a single constant-pool load.
5830 // For AVX-length vectors, see if we can use a vector load to get all of the
5831 // elements, otherwise build the individual 128-bit pieces and use
5832 // shuffles to put them in place.
5833 if (VT.is256BitVector() || VT.is512BitVector()) {
5834 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5836 // Check for a build vector of consecutive loads.
5837 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5840 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5842 // Build both the lower and upper subvector.
5843 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5844 makeArrayRef(&V[0], NumElems/2));
5845 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5846 makeArrayRef(&V[NumElems / 2], NumElems/2));
5848 // Recreate the wider vector with the lower and upper part.
5849 if (VT.is256BitVector())
5850 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5851 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5854 // Let legalizer expand 2-wide build_vectors.
5855 if (EVTBits == 64) {
5856 if (NumNonZero == 1) {
5857 // One half is zero or undef.
5858 unsigned Idx = countTrailingZeros(NonZeros);
5859 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5860 Op.getOperand(Idx));
5861 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5866 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5867 if (EVTBits == 8 && NumElems == 16)
5868 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5872 if (EVTBits == 16 && NumElems == 8)
5873 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5877 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5878 if (EVTBits == 32 && NumElems == 4)
5879 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
5882 // If element VT is == 32 bits, turn it into a number of shuffles.
5883 SmallVector<SDValue, 8> V(NumElems);
5884 if (NumElems == 4 && NumZero > 0) {
5885 for (unsigned i = 0; i < 4; ++i) {
5886 bool isZero = !(NonZeros & (1 << i));
5888 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5890 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5893 for (unsigned i = 0; i < 2; ++i) {
5894 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5897 V[i] = V[i*2]; // Must be a zero vector.
5900 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5903 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5906 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5911 bool Reverse1 = (NonZeros & 0x3) == 2;
5912 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5916 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5917 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5919 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5922 if (Values.size() > 1 && VT.is128BitVector()) {
5923 // Check for a build vector of consecutive loads.
5924 for (unsigned i = 0; i < NumElems; ++i)
5925 V[i] = Op.getOperand(i);
5927 // Check for elements which are consecutive loads.
5928 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5931 // Check for a build vector from mostly shuffle plus few inserting.
5932 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
5935 // For SSE 4.1, use insertps to put the high elements into the low element.
5936 if (Subtarget->hasSSE41()) {
5938 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5939 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5941 Result = DAG.getUNDEF(VT);
5943 for (unsigned i = 1; i < NumElems; ++i) {
5944 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5945 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5946 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
5951 // Otherwise, expand into a number of unpckl*, start by extending each of
5952 // our (non-undef) elements to the full vector width with the element in the
5953 // bottom slot of the vector (which generates no code for SSE).
5954 for (unsigned i = 0; i < NumElems; ++i) {
5955 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5956 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5958 V[i] = DAG.getUNDEF(VT);
5961 // Next, we iteratively mix elements, e.g. for v4f32:
5962 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5963 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5964 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5965 unsigned EltStride = NumElems >> 1;
5966 while (EltStride != 0) {
5967 for (unsigned i = 0; i < EltStride; ++i) {
5968 // If V[i+EltStride] is undef and this is the first round of mixing,
5969 // then it is safe to just drop this shuffle: V[i] is already in the
5970 // right place, the one element (since it's the first round) being
5971 // inserted as undef can be dropped. This isn't safe for successive
5972 // rounds because they will permute elements within both vectors.
5973 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5974 EltStride == NumElems/2)
5977 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5986 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5987 // to create 256-bit vectors from two other 128-bit ones.
5988 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5990 MVT ResVT = Op.getSimpleValueType();
5992 assert((ResVT.is256BitVector() ||
5993 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
5995 SDValue V1 = Op.getOperand(0);
5996 SDValue V2 = Op.getOperand(1);
5997 unsigned NumElems = ResVT.getVectorNumElements();
5998 if (ResVT.is256BitVector())
5999 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6001 if (Op.getNumOperands() == 4) {
6002 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6003 ResVT.getVectorNumElements()/2);
6004 SDValue V3 = Op.getOperand(2);
6005 SDValue V4 = Op.getOperand(3);
6006 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6007 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6009 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6012 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6013 const X86Subtarget *Subtarget,
6014 SelectionDAG & DAG) {
6016 MVT ResVT = Op.getSimpleValueType();
6017 unsigned NumOfOperands = Op.getNumOperands();
6019 assert(isPowerOf2_32(NumOfOperands) &&
6020 "Unexpected number of operands in CONCAT_VECTORS");
6022 if (NumOfOperands > 2) {
6023 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6024 ResVT.getVectorNumElements()/2);
6025 SmallVector<SDValue, 2> Ops;
6026 for (unsigned i = 0; i < NumOfOperands/2; i++)
6027 Ops.push_back(Op.getOperand(i));
6028 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6030 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6031 Ops.push_back(Op.getOperand(i));
6032 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6033 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6036 SDValue V1 = Op.getOperand(0);
6037 SDValue V2 = Op.getOperand(1);
6038 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6039 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6041 if (IsZeroV1 && IsZeroV2)
6042 return getZeroVector(ResVT, Subtarget, DAG, dl);
6044 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6045 SDValue Undef = DAG.getUNDEF(ResVT);
6046 unsigned NumElems = ResVT.getVectorNumElements();
6047 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6049 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6050 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6054 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6055 // Zero the upper bits of V1
6056 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6057 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6060 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6063 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6064 const X86Subtarget *Subtarget,
6065 SelectionDAG &DAG) {
6066 MVT VT = Op.getSimpleValueType();
6067 if (VT.getVectorElementType() == MVT::i1)
6068 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6070 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6071 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6072 Op.getNumOperands() == 4)));
6074 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6075 // from two other 128-bit ones.
6077 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6078 return LowerAVXCONCAT_VECTORS(Op, DAG);
6082 //===----------------------------------------------------------------------===//
6083 // Vector shuffle lowering
6085 // This is an experimental code path for lowering vector shuffles on x86. It is
6086 // designed to handle arbitrary vector shuffles and blends, gracefully
6087 // degrading performance as necessary. It works hard to recognize idiomatic
6088 // shuffles and lower them to optimal instruction patterns without leaving
6089 // a framework that allows reasonably efficient handling of all vector shuffle
6091 //===----------------------------------------------------------------------===//
6093 /// \brief Tiny helper function to identify a no-op mask.
6095 /// This is a somewhat boring predicate function. It checks whether the mask
6096 /// array input, which is assumed to be a single-input shuffle mask of the kind
6097 /// used by the X86 shuffle instructions (not a fully general
6098 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6099 /// in-place shuffle are 'no-op's.
6100 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6101 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6102 if (Mask[i] != -1 && Mask[i] != i)
6107 /// \brief Helper function to classify a mask as a single-input mask.
6109 /// This isn't a generic single-input test because in the vector shuffle
6110 /// lowering we canonicalize single inputs to be the first input operand. This
6111 /// means we can more quickly test for a single input by only checking whether
6112 /// an input from the second operand exists. We also assume that the size of
6113 /// mask corresponds to the size of the input vectors which isn't true in the
6114 /// fully general case.
6115 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6117 if (M >= (int)Mask.size())
6122 /// \brief Test whether there are elements crossing 128-bit lanes in this
6125 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6126 /// and we routinely test for these.
6127 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6128 int LaneSize = 128 / VT.getScalarSizeInBits();
6129 int Size = Mask.size();
6130 for (int i = 0; i < Size; ++i)
6131 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6136 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6138 /// This checks a shuffle mask to see if it is performing the same
6139 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6140 /// that it is also not lane-crossing. It may however involve a blend from the
6141 /// same lane of a second vector.
6143 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6144 /// non-trivial to compute in the face of undef lanes. The representation is
6145 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6146 /// entries from both V1 and V2 inputs to the wider mask.
6148 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6149 SmallVectorImpl<int> &RepeatedMask) {
6150 int LaneSize = 128 / VT.getScalarSizeInBits();
6151 RepeatedMask.resize(LaneSize, -1);
6152 int Size = Mask.size();
6153 for (int i = 0; i < Size; ++i) {
6156 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6157 // This entry crosses lanes, so there is no way to model this shuffle.
6160 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6161 if (RepeatedMask[i % LaneSize] == -1)
6162 // This is the first non-undef entry in this slot of a 128-bit lane.
6163 RepeatedMask[i % LaneSize] =
6164 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6165 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6166 // Found a mismatch with the repeated mask.
6172 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6175 /// This is a fast way to test a shuffle mask against a fixed pattern:
6177 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6179 /// It returns true if the mask is exactly as wide as the argument list, and
6180 /// each element of the mask is either -1 (signifying undef) or the value given
6181 /// in the argument.
6182 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6183 ArrayRef<int> ExpectedMask) {
6184 if (Mask.size() != ExpectedMask.size())
6187 int Size = Mask.size();
6189 // If the values are build vectors, we can look through them to find
6190 // equivalent inputs that make the shuffles equivalent.
6191 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6192 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6194 for (int i = 0; i < Size; ++i)
6195 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6196 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6197 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6198 if (!MaskBV || !ExpectedBV ||
6199 MaskBV->getOperand(Mask[i] % Size) !=
6200 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6207 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6209 /// This helper function produces an 8-bit shuffle immediate corresponding to
6210 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6211 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6214 /// NB: We rely heavily on "undef" masks preserving the input lane.
6215 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6216 SelectionDAG &DAG) {
6217 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6218 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6219 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6220 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6221 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6224 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6225 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6226 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6227 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6228 return DAG.getConstant(Imm, DL, MVT::i8);
6231 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6233 /// This is used as a fallback approach when first class blend instructions are
6234 /// unavailable. Currently it is only suitable for integer vectors, but could
6235 /// be generalized for floating point vectors if desirable.
6236 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6237 SDValue V2, ArrayRef<int> Mask,
6238 SelectionDAG &DAG) {
6239 assert(VT.isInteger() && "Only supports integer vector types!");
6240 MVT EltVT = VT.getScalarType();
6241 int NumEltBits = EltVT.getSizeInBits();
6242 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6243 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6245 SmallVector<SDValue, 16> MaskOps;
6246 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6247 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6248 return SDValue(); // Shuffled input!
6249 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6252 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6253 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6254 // We have to cast V2 around.
6255 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6256 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6257 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6258 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6259 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6260 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6263 /// \brief Try to emit a blend instruction for a shuffle.
6265 /// This doesn't do any checks for the availability of instructions for blending
6266 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6267 /// be matched in the backend with the type given. What it does check for is
6268 /// that the shuffle mask is in fact a blend.
6269 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6270 SDValue V2, ArrayRef<int> Mask,
6271 const X86Subtarget *Subtarget,
6272 SelectionDAG &DAG) {
6273 unsigned BlendMask = 0;
6274 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6275 if (Mask[i] >= Size) {
6276 if (Mask[i] != i + Size)
6277 return SDValue(); // Shuffled V2 input!
6278 BlendMask |= 1u << i;
6281 if (Mask[i] >= 0 && Mask[i] != i)
6282 return SDValue(); // Shuffled V1 input!
6284 switch (VT.SimpleTy) {
6289 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6290 DAG.getConstant(BlendMask, DL, MVT::i8));
6294 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6298 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6299 // that instruction.
6300 if (Subtarget->hasAVX2()) {
6301 // Scale the blend by the number of 32-bit dwords per element.
6302 int Scale = VT.getScalarSizeInBits() / 32;
6304 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6305 if (Mask[i] >= Size)
6306 for (int j = 0; j < Scale; ++j)
6307 BlendMask |= 1u << (i * Scale + j);
6309 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6310 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6311 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6312 return DAG.getNode(ISD::BITCAST, DL, VT,
6313 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6314 DAG.getConstant(BlendMask, DL, MVT::i8)));
6318 // For integer shuffles we need to expand the mask and cast the inputs to
6319 // v8i16s prior to blending.
6320 int Scale = 8 / VT.getVectorNumElements();
6322 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6323 if (Mask[i] >= Size)
6324 for (int j = 0; j < Scale; ++j)
6325 BlendMask |= 1u << (i * Scale + j);
6327 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6328 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6329 return DAG.getNode(ISD::BITCAST, DL, VT,
6330 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6331 DAG.getConstant(BlendMask, DL, MVT::i8)));
6335 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6336 SmallVector<int, 8> RepeatedMask;
6337 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6338 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6339 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6341 for (int i = 0; i < 8; ++i)
6342 if (RepeatedMask[i] >= 16)
6343 BlendMask |= 1u << i;
6344 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6345 DAG.getConstant(BlendMask, DL, MVT::i8));
6351 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6352 "256-bit byte-blends require AVX2 support!");
6354 // Scale the blend by the number of bytes per element.
6355 int Scale = VT.getScalarSizeInBits() / 8;
6357 // This form of blend is always done on bytes. Compute the byte vector
6359 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6361 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6362 // mix of LLVM's code generator and the x86 backend. We tell the code
6363 // generator that boolean values in the elements of an x86 vector register
6364 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6365 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6366 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6367 // of the element (the remaining are ignored) and 0 in that high bit would
6368 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6369 // the LLVM model for boolean values in vector elements gets the relevant
6370 // bit set, it is set backwards and over constrained relative to x86's
6372 SmallVector<SDValue, 32> VSELECTMask;
6373 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6374 for (int j = 0; j < Scale; ++j)
6375 VSELECTMask.push_back(
6376 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6377 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6380 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6381 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6383 ISD::BITCAST, DL, VT,
6384 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6385 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6390 llvm_unreachable("Not a supported integer vector type!");
6394 /// \brief Try to lower as a blend of elements from two inputs followed by
6395 /// a single-input permutation.
6397 /// This matches the pattern where we can blend elements from two inputs and
6398 /// then reduce the shuffle to a single-input permutation.
6399 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6402 SelectionDAG &DAG) {
6403 // We build up the blend mask while checking whether a blend is a viable way
6404 // to reduce the shuffle.
6405 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6406 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6408 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6412 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6414 if (BlendMask[Mask[i] % Size] == -1)
6415 BlendMask[Mask[i] % Size] = Mask[i];
6416 else if (BlendMask[Mask[i] % Size] != Mask[i])
6417 return SDValue(); // Can't blend in the needed input!
6419 PermuteMask[i] = Mask[i] % Size;
6422 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6423 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6426 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6427 /// blends and permutes.
6429 /// This matches the extremely common pattern for handling combined
6430 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6431 /// operations. It will try to pick the best arrangement of shuffles and
6433 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6437 SelectionDAG &DAG) {
6438 // Shuffle the input elements into the desired positions in V1 and V2 and
6439 // blend them together.
6440 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6441 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6442 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6443 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6444 if (Mask[i] >= 0 && Mask[i] < Size) {
6445 V1Mask[i] = Mask[i];
6447 } else if (Mask[i] >= Size) {
6448 V2Mask[i] = Mask[i] - Size;
6449 BlendMask[i] = i + Size;
6452 // Try to lower with the simpler initial blend strategy unless one of the
6453 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6454 // shuffle may be able to fold with a load or other benefit. However, when
6455 // we'll have to do 2x as many shuffles in order to achieve this, blending
6456 // first is a better strategy.
6457 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6458 if (SDValue BlendPerm =
6459 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6462 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6463 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6464 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6467 /// \brief Try to lower a vector shuffle as a byte rotation.
6469 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6470 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6471 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6472 /// try to generically lower a vector shuffle through such an pattern. It
6473 /// does not check for the profitability of lowering either as PALIGNR or
6474 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6475 /// This matches shuffle vectors that look like:
6477 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6479 /// Essentially it concatenates V1 and V2, shifts right by some number of
6480 /// elements, and takes the low elements as the result. Note that while this is
6481 /// specified as a *right shift* because x86 is little-endian, it is a *left
6482 /// rotate* of the vector lanes.
6483 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6486 const X86Subtarget *Subtarget,
6487 SelectionDAG &DAG) {
6488 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6490 int NumElts = Mask.size();
6491 int NumLanes = VT.getSizeInBits() / 128;
6492 int NumLaneElts = NumElts / NumLanes;
6494 // We need to detect various ways of spelling a rotation:
6495 // [11, 12, 13, 14, 15, 0, 1, 2]
6496 // [-1, 12, 13, 14, -1, -1, 1, -1]
6497 // [-1, -1, -1, -1, -1, -1, 1, 2]
6498 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6499 // [-1, 4, 5, 6, -1, -1, 9, -1]
6500 // [-1, 4, 5, 6, -1, -1, -1, -1]
6503 for (int l = 0; l < NumElts; l += NumLaneElts) {
6504 for (int i = 0; i < NumLaneElts; ++i) {
6505 if (Mask[l + i] == -1)
6507 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6509 // Get the mod-Size index and lane correct it.
6510 int LaneIdx = (Mask[l + i] % NumElts) - l;
6511 // Make sure it was in this lane.
6512 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6515 // Determine where a rotated vector would have started.
6516 int StartIdx = i - LaneIdx;
6518 // The identity rotation isn't interesting, stop.
6521 // If we found the tail of a vector the rotation must be the missing
6522 // front. If we found the head of a vector, it must be how much of the
6524 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6527 Rotation = CandidateRotation;
6528 else if (Rotation != CandidateRotation)
6529 // The rotations don't match, so we can't match this mask.
6532 // Compute which value this mask is pointing at.
6533 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6535 // Compute which of the two target values this index should be assigned
6536 // to. This reflects whether the high elements are remaining or the low
6537 // elements are remaining.
6538 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6540 // Either set up this value if we've not encountered it before, or check
6541 // that it remains consistent.
6544 else if (TargetV != MaskV)
6545 // This may be a rotation, but it pulls from the inputs in some
6546 // unsupported interleaving.
6551 // Check that we successfully analyzed the mask, and normalize the results.
6552 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6553 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6559 // The actual rotate instruction rotates bytes, so we need to scale the
6560 // rotation based on how many bytes are in the vector lane.
6561 int Scale = 16 / NumLaneElts;
6563 // SSSE3 targets can use the palignr instruction.
6564 if (Subtarget->hasSSSE3()) {
6565 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6566 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6567 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6568 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6570 return DAG.getNode(ISD::BITCAST, DL, VT,
6571 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6572 DAG.getConstant(Rotation * Scale, DL,
6576 assert(VT.getSizeInBits() == 128 &&
6577 "Rotate-based lowering only supports 128-bit lowering!");
6578 assert(Mask.size() <= 16 &&
6579 "Can shuffle at most 16 bytes in a 128-bit vector!");
6581 // Default SSE2 implementation
6582 int LoByteShift = 16 - Rotation * Scale;
6583 int HiByteShift = Rotation * Scale;
6585 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6586 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6587 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6589 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6590 DAG.getConstant(LoByteShift, DL, MVT::i8));
6591 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6592 DAG.getConstant(HiByteShift, DL, MVT::i8));
6593 return DAG.getNode(ISD::BITCAST, DL, VT,
6594 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6597 /// \brief Compute whether each element of a shuffle is zeroable.
6599 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6600 /// Either it is an undef element in the shuffle mask, the element of the input
6601 /// referenced is undef, or the element of the input referenced is known to be
6602 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6603 /// as many lanes with this technique as possible to simplify the remaining
6605 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6606 SDValue V1, SDValue V2) {
6607 SmallBitVector Zeroable(Mask.size(), false);
6609 while (V1.getOpcode() == ISD::BITCAST)
6610 V1 = V1->getOperand(0);
6611 while (V2.getOpcode() == ISD::BITCAST)
6612 V2 = V2->getOperand(0);
6614 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6615 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6617 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6619 // Handle the easy cases.
6620 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6625 // If this is an index into a build_vector node (which has the same number
6626 // of elements), dig out the input value and use it.
6627 SDValue V = M < Size ? V1 : V2;
6628 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6631 SDValue Input = V.getOperand(M % Size);
6632 // The UNDEF opcode check really should be dead code here, but not quite
6633 // worth asserting on (it isn't invalid, just unexpected).
6634 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6641 /// \brief Try to emit a bitmask instruction for a shuffle.
6643 /// This handles cases where we can model a blend exactly as a bitmask due to
6644 /// one of the inputs being zeroable.
6645 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6646 SDValue V2, ArrayRef<int> Mask,
6647 SelectionDAG &DAG) {
6648 MVT EltVT = VT.getScalarType();
6649 int NumEltBits = EltVT.getSizeInBits();
6650 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6651 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6652 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6654 if (EltVT.isFloatingPoint()) {
6655 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6656 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6658 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6659 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6661 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6664 if (Mask[i] % Size != i)
6665 return SDValue(); // Not a blend.
6667 V = Mask[i] < Size ? V1 : V2;
6668 else if (V != (Mask[i] < Size ? V1 : V2))
6669 return SDValue(); // Can only let one input through the mask.
6671 VMaskOps[i] = AllOnes;
6674 return SDValue(); // No non-zeroable elements!
6676 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6677 V = DAG.getNode(VT.isFloatingPoint()
6678 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6683 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6685 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6686 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6687 /// matches elements from one of the input vectors shuffled to the left or
6688 /// right with zeroable elements 'shifted in'. It handles both the strictly
6689 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6692 /// PSHL : (little-endian) left bit shift.
6693 /// [ zz, 0, zz, 2 ]
6694 /// [ -1, 4, zz, -1 ]
6695 /// PSRL : (little-endian) right bit shift.
6697 /// [ -1, -1, 7, zz]
6698 /// PSLLDQ : (little-endian) left byte shift
6699 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6700 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6701 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6702 /// PSRLDQ : (little-endian) right byte shift
6703 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6704 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6705 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6706 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6707 SDValue V2, ArrayRef<int> Mask,
6708 SelectionDAG &DAG) {
6709 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6711 int Size = Mask.size();
6712 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6714 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6715 for (int i = 0; i < Size; i += Scale)
6716 for (int j = 0; j < Shift; ++j)
6717 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6723 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6724 for (int i = 0; i != Size; i += Scale) {
6725 unsigned Pos = Left ? i + Shift : i;
6726 unsigned Low = Left ? i : i + Shift;
6727 unsigned Len = Scale - Shift;
6728 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6729 Low + (V == V1 ? 0 : Size)))
6733 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6734 bool ByteShift = ShiftEltBits > 64;
6735 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6736 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6737 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6739 // Normalize the scale for byte shifts to still produce an i64 element
6741 Scale = ByteShift ? Scale / 2 : Scale;
6743 // We need to round trip through the appropriate type for the shift.
6744 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6745 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6746 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6747 "Illegal integer vector type");
6748 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6750 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6751 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6752 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6755 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6756 // keep doubling the size of the integer elements up to that. We can
6757 // then shift the elements of the integer vector by whole multiples of
6758 // their width within the elements of the larger integer vector. Test each
6759 // multiple to see if we can find a match with the moved element indices
6760 // and that the shifted in elements are all zeroable.
6761 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6762 for (int Shift = 1; Shift != Scale; ++Shift)
6763 for (bool Left : {true, false})
6764 if (CheckZeros(Shift, Scale, Left))
6765 for (SDValue V : {V1, V2})
6766 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6773 /// \brief Lower a vector shuffle as a zero or any extension.
6775 /// Given a specific number of elements, element bit width, and extension
6776 /// stride, produce either a zero or any extension based on the available
6777 /// features of the subtarget.
6778 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6779 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6780 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6781 assert(Scale > 1 && "Need a scale to extend.");
6782 int NumElements = VT.getVectorNumElements();
6783 int EltBits = VT.getScalarSizeInBits();
6784 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6785 "Only 8, 16, and 32 bit elements can be extended.");
6786 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6788 // Found a valid zext mask! Try various lowering strategies based on the
6789 // input type and available ISA extensions.
6790 if (Subtarget->hasSSE41()) {
6791 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6792 NumElements / Scale);
6793 return DAG.getNode(ISD::BITCAST, DL, VT,
6794 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6797 // For any extends we can cheat for larger element sizes and use shuffle
6798 // instructions that can fold with a load and/or copy.
6799 if (AnyExt && EltBits == 32) {
6800 int PSHUFDMask[4] = {0, -1, 1, -1};
6802 ISD::BITCAST, DL, VT,
6803 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6804 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6805 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
6807 if (AnyExt && EltBits == 16 && Scale > 2) {
6808 int PSHUFDMask[4] = {0, -1, 0, -1};
6809 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6810 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6811 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
6812 int PSHUFHWMask[4] = {1, -1, -1, -1};
6814 ISD::BITCAST, DL, VT,
6815 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6816 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6817 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
6820 // If this would require more than 2 unpack instructions to expand, use
6821 // pshufb when available. We can only use more than 2 unpack instructions
6822 // when zero extending i8 elements which also makes it easier to use pshufb.
6823 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6824 assert(NumElements == 16 && "Unexpected byte vector width!");
6825 SDValue PSHUFBMask[16];
6826 for (int i = 0; i < 16; ++i)
6828 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
6829 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6830 return DAG.getNode(ISD::BITCAST, DL, VT,
6831 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6832 DAG.getNode(ISD::BUILD_VECTOR, DL,
6833 MVT::v16i8, PSHUFBMask)));
6836 // Otherwise emit a sequence of unpacks.
6838 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6839 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6840 : getZeroVector(InputVT, Subtarget, DAG, DL);
6841 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6842 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6846 } while (Scale > 1);
6847 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6850 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6852 /// This routine will try to do everything in its power to cleverly lower
6853 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6854 /// check for the profitability of this lowering, it tries to aggressively
6855 /// match this pattern. It will use all of the micro-architectural details it
6856 /// can to emit an efficient lowering. It handles both blends with all-zero
6857 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6858 /// masking out later).
6860 /// The reason we have dedicated lowering for zext-style shuffles is that they
6861 /// are both incredibly common and often quite performance sensitive.
6862 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6863 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6864 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6865 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6867 int Bits = VT.getSizeInBits();
6868 int NumElements = VT.getVectorNumElements();
6869 assert(VT.getScalarSizeInBits() <= 32 &&
6870 "Exceeds 32-bit integer zero extension limit");
6871 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6873 // Define a helper function to check a particular ext-scale and lower to it if
6875 auto Lower = [&](int Scale) -> SDValue {
6878 for (int i = 0; i < NumElements; ++i) {
6880 continue; // Valid anywhere but doesn't tell us anything.
6881 if (i % Scale != 0) {
6882 // Each of the extended elements need to be zeroable.
6886 // We no longer are in the anyext case.
6891 // Each of the base elements needs to be consecutive indices into the
6892 // same input vector.
6893 SDValue V = Mask[i] < NumElements ? V1 : V2;
6896 else if (InputV != V)
6897 return SDValue(); // Flip-flopping inputs.
6899 if (Mask[i] % NumElements != i / Scale)
6900 return SDValue(); // Non-consecutive strided elements.
6903 // If we fail to find an input, we have a zero-shuffle which should always
6904 // have already been handled.
6905 // FIXME: Maybe handle this here in case during blending we end up with one?
6909 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6910 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6913 // The widest scale possible for extending is to a 64-bit integer.
6914 assert(Bits % 64 == 0 &&
6915 "The number of bits in a vector must be divisible by 64 on x86!");
6916 int NumExtElements = Bits / 64;
6918 // Each iteration, try extending the elements half as much, but into twice as
6920 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6921 assert(NumElements % NumExtElements == 0 &&
6922 "The input vector size must be divisible by the extended size.");
6923 if (SDValue V = Lower(NumElements / NumExtElements))
6927 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6931 // Returns one of the source operands if the shuffle can be reduced to a
6932 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6933 auto CanZExtLowHalf = [&]() {
6934 for (int i = NumElements / 2; i != NumElements; ++i)
6937 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6939 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
6944 if (SDValue V = CanZExtLowHalf()) {
6945 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
6946 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
6947 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6950 // No viable ext lowering found.
6954 /// \brief Try to get a scalar value for a specific element of a vector.
6956 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
6957 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
6958 SelectionDAG &DAG) {
6959 MVT VT = V.getSimpleValueType();
6960 MVT EltVT = VT.getVectorElementType();
6961 while (V.getOpcode() == ISD::BITCAST)
6962 V = V.getOperand(0);
6963 // If the bitcasts shift the element size, we can't extract an equivalent
6965 MVT NewVT = V.getSimpleValueType();
6966 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
6969 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6970 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
6971 // Ensure the scalar operand is the same size as the destination.
6972 // FIXME: Add support for scalar truncation where possible.
6973 SDValue S = V.getOperand(Idx);
6974 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
6975 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
6981 /// \brief Helper to test for a load that can be folded with x86 shuffles.
6983 /// This is particularly important because the set of instructions varies
6984 /// significantly based on whether the operand is a load or not.
6985 static bool isShuffleFoldableLoad(SDValue V) {
6986 while (V.getOpcode() == ISD::BITCAST)
6987 V = V.getOperand(0);
6989 return ISD::isNON_EXTLoad(V.getNode());
6992 /// \brief Try to lower insertion of a single element into a zero vector.
6994 /// This is a common pattern that we have especially efficient patterns to lower
6995 /// across all subtarget feature sets.
6996 static SDValue lowerVectorShuffleAsElementInsertion(
6997 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6998 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6999 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7001 MVT EltVT = VT.getVectorElementType();
7003 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7004 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7006 bool IsV1Zeroable = true;
7007 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7008 if (i != V2Index && !Zeroable[i]) {
7009 IsV1Zeroable = false;
7013 // Check for a single input from a SCALAR_TO_VECTOR node.
7014 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7015 // all the smarts here sunk into that routine. However, the current
7016 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7017 // vector shuffle lowering is dead.
7018 if (SDValue V2S = getScalarValueForVectorElement(
7019 V2, Mask[V2Index] - Mask.size(), DAG)) {
7020 // We need to zext the scalar if it is smaller than an i32.
7021 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
7022 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7023 // Using zext to expand a narrow element won't work for non-zero
7028 // Zero-extend directly to i32.
7030 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7032 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7033 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7034 EltVT == MVT::i16) {
7035 // Either not inserting from the low element of the input or the input
7036 // element size is too small to use VZEXT_MOVL to clear the high bits.
7040 if (!IsV1Zeroable) {
7041 // If V1 can't be treated as a zero vector we have fewer options to lower
7042 // this. We can't support integer vectors or non-zero targets cheaply, and
7043 // the V1 elements can't be permuted in any way.
7044 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7045 if (!VT.isFloatingPoint() || V2Index != 0)
7047 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7048 V1Mask[V2Index] = -1;
7049 if (!isNoopShuffleMask(V1Mask))
7051 // This is essentially a special case blend operation, but if we have
7052 // general purpose blend operations, they are always faster. Bail and let
7053 // the rest of the lowering handle these as blends.
7054 if (Subtarget->hasSSE41())
7057 // Otherwise, use MOVSD or MOVSS.
7058 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7059 "Only two types of floating point element types to handle!");
7060 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7064 // This lowering only works for the low element with floating point vectors.
7065 if (VT.isFloatingPoint() && V2Index != 0)
7068 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7070 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7073 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7074 // the desired position. Otherwise it is more efficient to do a vector
7075 // shift left. We know that we can do a vector shift left because all
7076 // the inputs are zero.
7077 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7078 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7079 V2Shuffle[V2Index] = 0;
7080 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7082 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
7084 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7086 V2Index * EltVT.getSizeInBits()/8, DL,
7087 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7088 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7094 /// \brief Try to lower broadcast of a single element.
7096 /// For convenience, this code also bundles all of the subtarget feature set
7097 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7098 /// a convenient way to factor it out.
7099 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7101 const X86Subtarget *Subtarget,
7102 SelectionDAG &DAG) {
7103 if (!Subtarget->hasAVX())
7105 if (VT.isInteger() && !Subtarget->hasAVX2())
7108 // Check that the mask is a broadcast.
7109 int BroadcastIdx = -1;
7111 if (M >= 0 && BroadcastIdx == -1)
7113 else if (M >= 0 && M != BroadcastIdx)
7116 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7117 "a sorted mask where the broadcast "
7120 // Go up the chain of (vector) values to find a scalar load that we can
7121 // combine with the broadcast.
7123 switch (V.getOpcode()) {
7124 case ISD::CONCAT_VECTORS: {
7125 int OperandSize = Mask.size() / V.getNumOperands();
7126 V = V.getOperand(BroadcastIdx / OperandSize);
7127 BroadcastIdx %= OperandSize;
7131 case ISD::INSERT_SUBVECTOR: {
7132 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7133 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7137 int BeginIdx = (int)ConstantIdx->getZExtValue();
7139 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7140 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7141 BroadcastIdx -= BeginIdx;
7152 // Check if this is a broadcast of a scalar. We special case lowering
7153 // for scalars so that we can more effectively fold with loads.
7154 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7155 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7156 V = V.getOperand(BroadcastIdx);
7158 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7159 // Only AVX2 has register broadcasts.
7160 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7162 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7163 // We can't broadcast from a vector register without AVX2, and we can only
7164 // broadcast from the zero-element of a vector register.
7168 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7171 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7172 // INSERTPS when the V1 elements are already in the correct locations
7173 // because otherwise we can just always use two SHUFPS instructions which
7174 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7175 // perform INSERTPS if a single V1 element is out of place and all V2
7176 // elements are zeroable.
7177 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7179 SelectionDAG &DAG) {
7180 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7181 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7182 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7183 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7185 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7188 int V1DstIndex = -1;
7189 int V2DstIndex = -1;
7190 bool V1UsedInPlace = false;
7192 for (int i = 0; i < 4; ++i) {
7193 // Synthesize a zero mask from the zeroable elements (includes undefs).
7199 // Flag if we use any V1 inputs in place.
7201 V1UsedInPlace = true;
7205 // We can only insert a single non-zeroable element.
7206 if (V1DstIndex != -1 || V2DstIndex != -1)
7210 // V1 input out of place for insertion.
7213 // V2 input for insertion.
7218 // Don't bother if we have no (non-zeroable) element for insertion.
7219 if (V1DstIndex == -1 && V2DstIndex == -1)
7222 // Determine element insertion src/dst indices. The src index is from the
7223 // start of the inserted vector, not the start of the concatenated vector.
7224 unsigned V2SrcIndex = 0;
7225 if (V1DstIndex != -1) {
7226 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7227 // and don't use the original V2 at all.
7228 V2SrcIndex = Mask[V1DstIndex];
7229 V2DstIndex = V1DstIndex;
7232 V2SrcIndex = Mask[V2DstIndex] - 4;
7235 // If no V1 inputs are used in place, then the result is created only from
7236 // the zero mask and the V2 insertion - so remove V1 dependency.
7238 V1 = DAG.getUNDEF(MVT::v4f32);
7240 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7241 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7243 // Insert the V2 element into the desired position.
7245 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7246 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7249 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7250 /// UNPCK instruction.
7252 /// This specifically targets cases where we end up with alternating between
7253 /// the two inputs, and so can permute them into something that feeds a single
7254 /// UNPCK instruction. Note that this routine only targets integer vectors
7255 /// because for floating point vectors we have a generalized SHUFPS lowering
7256 /// strategy that handles everything that doesn't *exactly* match an unpack,
7257 /// making this clever lowering unnecessary.
7258 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7259 SDValue V2, ArrayRef<int> Mask,
7260 SelectionDAG &DAG) {
7261 assert(!VT.isFloatingPoint() &&
7262 "This routine only supports integer vectors.");
7263 assert(!isSingleInputShuffleMask(Mask) &&
7264 "This routine should only be used when blending two inputs.");
7265 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7267 int Size = Mask.size();
7269 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7270 return M >= 0 && M % Size < Size / 2;
7272 int NumHiInputs = std::count_if(
7273 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7275 bool UnpackLo = NumLoInputs >= NumHiInputs;
7277 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7278 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7279 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7281 for (int i = 0; i < Size; ++i) {
7285 // Each element of the unpack contains Scale elements from this mask.
7286 int UnpackIdx = i / Scale;
7288 // We only handle the case where V1 feeds the first slots of the unpack.
7289 // We rely on canonicalization to ensure this is the case.
7290 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7293 // Setup the mask for this input. The indexing is tricky as we have to
7294 // handle the unpack stride.
7295 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7296 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7300 // If we will have to shuffle both inputs to use the unpack, check whether
7301 // we can just unpack first and shuffle the result. If so, skip this unpack.
7302 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7303 !isNoopShuffleMask(V2Mask))
7306 // Shuffle the inputs into place.
7307 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7308 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7310 // Cast the inputs to the type we will use to unpack them.
7311 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7312 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7314 // Unpack the inputs and cast the result back to the desired type.
7315 return DAG.getNode(ISD::BITCAST, DL, VT,
7316 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7317 DL, UnpackVT, V1, V2));
7320 // We try each unpack from the largest to the smallest to try and find one
7321 // that fits this mask.
7322 int OrigNumElements = VT.getVectorNumElements();
7323 int OrigScalarSize = VT.getScalarSizeInBits();
7324 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7325 int Scale = ScalarSize / OrigScalarSize;
7326 int NumElements = OrigNumElements / Scale;
7327 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7328 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7332 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7334 if (NumLoInputs == 0 || NumHiInputs == 0) {
7335 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7336 "We have to have *some* inputs!");
7337 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7339 // FIXME: We could consider the total complexity of the permute of each
7340 // possible unpacking. Or at the least we should consider how many
7341 // half-crossings are created.
7342 // FIXME: We could consider commuting the unpacks.
7344 SmallVector<int, 32> PermMask;
7345 PermMask.assign(Size, -1);
7346 for (int i = 0; i < Size; ++i) {
7350 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7353 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7355 return DAG.getVectorShuffle(
7356 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7358 DAG.getUNDEF(VT), PermMask);
7364 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7366 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7367 /// support for floating point shuffles but not integer shuffles. These
7368 /// instructions will incur a domain crossing penalty on some chips though so
7369 /// it is better to avoid lowering through this for integer vectors where
7371 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7372 const X86Subtarget *Subtarget,
7373 SelectionDAG &DAG) {
7375 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7376 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7377 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7378 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7379 ArrayRef<int> Mask = SVOp->getMask();
7380 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7382 if (isSingleInputShuffleMask(Mask)) {
7383 // Use low duplicate instructions for masks that match their pattern.
7384 if (Subtarget->hasSSE3())
7385 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7386 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7388 // Straight shuffle of a single input vector. Simulate this by using the
7389 // single input as both of the "inputs" to this instruction..
7390 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7392 if (Subtarget->hasAVX()) {
7393 // If we have AVX, we can use VPERMILPS which will allow folding a load
7394 // into the shuffle.
7395 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7396 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7399 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7400 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7402 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7403 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7405 // If we have a single input, insert that into V1 if we can do so cheaply.
7406 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7407 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7408 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7410 // Try inverting the insertion since for v2 masks it is easy to do and we
7411 // can't reliably sort the mask one way or the other.
7412 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7413 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7414 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7415 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7419 // Try to use one of the special instruction patterns to handle two common
7420 // blend patterns if a zero-blend above didn't work.
7421 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7422 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7423 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7424 // We can either use a special instruction to load over the low double or
7425 // to move just the low double.
7427 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7429 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7431 if (Subtarget->hasSSE41())
7432 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7436 // Use dedicated unpack instructions for masks that match their pattern.
7437 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7438 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7439 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7440 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7442 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7443 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7444 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7447 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7449 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7450 /// the integer unit to minimize domain crossing penalties. However, for blends
7451 /// it falls back to the floating point shuffle operation with appropriate bit
7453 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7454 const X86Subtarget *Subtarget,
7455 SelectionDAG &DAG) {
7457 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7458 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7459 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7460 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7461 ArrayRef<int> Mask = SVOp->getMask();
7462 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7464 if (isSingleInputShuffleMask(Mask)) {
7465 // Check for being able to broadcast a single element.
7466 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7467 Mask, Subtarget, DAG))
7470 // Straight shuffle of a single input vector. For everything from SSE2
7471 // onward this has a single fast instruction with no scary immediates.
7472 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7473 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7474 int WidenedMask[4] = {
7475 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7476 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7478 ISD::BITCAST, DL, MVT::v2i64,
7479 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7480 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7482 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7483 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7484 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7485 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7487 // If we have a blend of two PACKUS operations an the blend aligns with the
7488 // low and half halves, we can just merge the PACKUS operations. This is
7489 // particularly important as it lets us merge shuffles that this routine itself
7491 auto GetPackNode = [](SDValue V) {
7492 while (V.getOpcode() == ISD::BITCAST)
7493 V = V.getOperand(0);
7495 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7497 if (SDValue V1Pack = GetPackNode(V1))
7498 if (SDValue V2Pack = GetPackNode(V2))
7499 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7500 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7501 Mask[0] == 0 ? V1Pack.getOperand(0)
7502 : V1Pack.getOperand(1),
7503 Mask[1] == 2 ? V2Pack.getOperand(0)
7504 : V2Pack.getOperand(1)));
7506 // Try to use shift instructions.
7508 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7511 // When loading a scalar and then shuffling it into a vector we can often do
7512 // the insertion cheaply.
7513 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7514 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7516 // Try inverting the insertion since for v2 masks it is easy to do and we
7517 // can't reliably sort the mask one way or the other.
7518 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7519 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7520 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7523 // We have different paths for blend lowering, but they all must use the
7524 // *exact* same predicate.
7525 bool IsBlendSupported = Subtarget->hasSSE41();
7526 if (IsBlendSupported)
7527 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7531 // Use dedicated unpack instructions for masks that match their pattern.
7532 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7533 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7534 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7535 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7537 // Try to use byte rotation instructions.
7538 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7539 if (Subtarget->hasSSSE3())
7540 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7541 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7544 // If we have direct support for blends, we should lower by decomposing into
7545 // a permute. That will be faster than the domain cross.
7546 if (IsBlendSupported)
7547 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7550 // We implement this with SHUFPD which is pretty lame because it will likely
7551 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7552 // However, all the alternatives are still more cycles and newer chips don't
7553 // have this problem. It would be really nice if x86 had better shuffles here.
7554 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7555 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7556 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7557 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7560 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7562 /// This is used to disable more specialized lowerings when the shufps lowering
7563 /// will happen to be efficient.
7564 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7565 // This routine only handles 128-bit shufps.
7566 assert(Mask.size() == 4 && "Unsupported mask size!");
7568 // To lower with a single SHUFPS we need to have the low half and high half
7569 // each requiring a single input.
7570 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7572 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7578 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7580 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7581 /// It makes no assumptions about whether this is the *best* lowering, it simply
7583 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7584 ArrayRef<int> Mask, SDValue V1,
7585 SDValue V2, SelectionDAG &DAG) {
7586 SDValue LowV = V1, HighV = V2;
7587 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7590 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7592 if (NumV2Elements == 1) {
7594 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7597 // Compute the index adjacent to V2Index and in the same half by toggling
7599 int V2AdjIndex = V2Index ^ 1;
7601 if (Mask[V2AdjIndex] == -1) {
7602 // Handles all the cases where we have a single V2 element and an undef.
7603 // This will only ever happen in the high lanes because we commute the
7604 // vector otherwise.
7606 std::swap(LowV, HighV);
7607 NewMask[V2Index] -= 4;
7609 // Handle the case where the V2 element ends up adjacent to a V1 element.
7610 // To make this work, blend them together as the first step.
7611 int V1Index = V2AdjIndex;
7612 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7613 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7614 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7616 // Now proceed to reconstruct the final blend as we have the necessary
7617 // high or low half formed.
7624 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7625 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7627 } else if (NumV2Elements == 2) {
7628 if (Mask[0] < 4 && Mask[1] < 4) {
7629 // Handle the easy case where we have V1 in the low lanes and V2 in the
7633 } else if (Mask[2] < 4 && Mask[3] < 4) {
7634 // We also handle the reversed case because this utility may get called
7635 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7636 // arrange things in the right direction.
7642 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7643 // trying to place elements directly, just blend them and set up the final
7644 // shuffle to place them.
7646 // The first two blend mask elements are for V1, the second two are for
7648 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7649 Mask[2] < 4 ? Mask[2] : Mask[3],
7650 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7651 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7652 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7653 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7655 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7658 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7659 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7660 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7661 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7664 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7665 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
7668 /// \brief Lower 4-lane 32-bit floating point shuffles.
7670 /// Uses instructions exclusively from the floating point unit to minimize
7671 /// domain crossing penalties, as these are sufficient to implement all v4f32
7673 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7674 const X86Subtarget *Subtarget,
7675 SelectionDAG &DAG) {
7677 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7678 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7679 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7680 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7681 ArrayRef<int> Mask = SVOp->getMask();
7682 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7685 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7687 if (NumV2Elements == 0) {
7688 // Check for being able to broadcast a single element.
7689 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7690 Mask, Subtarget, DAG))
7693 // Use even/odd duplicate instructions for masks that match their pattern.
7694 if (Subtarget->hasSSE3()) {
7695 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7696 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7697 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7698 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7701 if (Subtarget->hasAVX()) {
7702 // If we have AVX, we can use VPERMILPS which will allow folding a load
7703 // into the shuffle.
7704 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7705 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7708 // Otherwise, use a straight shuffle of a single input vector. We pass the
7709 // input vector to both operands to simulate this with a SHUFPS.
7710 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7711 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7714 // There are special ways we can lower some single-element blends. However, we
7715 // have custom ways we can lower more complex single-element blends below that
7716 // we defer to if both this and BLENDPS fail to match, so restrict this to
7717 // when the V2 input is targeting element 0 of the mask -- that is the fast
7719 if (NumV2Elements == 1 && Mask[0] >= 4)
7720 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7721 Mask, Subtarget, DAG))
7724 if (Subtarget->hasSSE41()) {
7725 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7729 // Use INSERTPS if we can complete the shuffle efficiently.
7730 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7733 if (!isSingleSHUFPSMask(Mask))
7734 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7735 DL, MVT::v4f32, V1, V2, Mask, DAG))
7739 // Use dedicated unpack instructions for masks that match their pattern.
7740 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7741 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7742 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7743 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7744 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7745 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7746 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7747 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7749 // Otherwise fall back to a SHUFPS lowering strategy.
7750 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7753 /// \brief Lower 4-lane i32 vector shuffles.
7755 /// We try to handle these with integer-domain shuffles where we can, but for
7756 /// blends we use the floating point domain blend instructions.
7757 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7758 const X86Subtarget *Subtarget,
7759 SelectionDAG &DAG) {
7761 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7762 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7763 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7764 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7765 ArrayRef<int> Mask = SVOp->getMask();
7766 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7768 // Whenever we can lower this as a zext, that instruction is strictly faster
7769 // than any alternative. It also allows us to fold memory operands into the
7770 // shuffle in many cases.
7771 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7772 Mask, Subtarget, DAG))
7776 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7778 if (NumV2Elements == 0) {
7779 // Check for being able to broadcast a single element.
7780 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7781 Mask, Subtarget, DAG))
7784 // Straight shuffle of a single input vector. For everything from SSE2
7785 // onward this has a single fast instruction with no scary immediates.
7786 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7787 // but we aren't actually going to use the UNPCK instruction because doing
7788 // so prevents folding a load into this instruction or making a copy.
7789 const int UnpackLoMask[] = {0, 0, 1, 1};
7790 const int UnpackHiMask[] = {2, 2, 3, 3};
7791 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7792 Mask = UnpackLoMask;
7793 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7794 Mask = UnpackHiMask;
7796 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7797 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7800 // Try to use shift instructions.
7802 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7805 // There are special ways we can lower some single-element blends.
7806 if (NumV2Elements == 1)
7807 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7808 Mask, Subtarget, DAG))
7811 // We have different paths for blend lowering, but they all must use the
7812 // *exact* same predicate.
7813 bool IsBlendSupported = Subtarget->hasSSE41();
7814 if (IsBlendSupported)
7815 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7819 if (SDValue Masked =
7820 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7823 // Use dedicated unpack instructions for masks that match their pattern.
7824 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7825 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7826 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7827 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7828 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7829 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7830 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7831 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7833 // Try to use byte rotation instructions.
7834 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7835 if (Subtarget->hasSSSE3())
7836 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7837 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7840 // If we have direct support for blends, we should lower by decomposing into
7841 // a permute. That will be faster than the domain cross.
7842 if (IsBlendSupported)
7843 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7846 // Try to lower by permuting the inputs into an unpack instruction.
7847 if (SDValue Unpack =
7848 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7851 // We implement this with SHUFPS because it can blend from two vectors.
7852 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7853 // up the inputs, bypassing domain shift penalties that we would encur if we
7854 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7856 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7857 DAG.getVectorShuffle(
7859 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7860 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7863 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7864 /// shuffle lowering, and the most complex part.
7866 /// The lowering strategy is to try to form pairs of input lanes which are
7867 /// targeted at the same half of the final vector, and then use a dword shuffle
7868 /// to place them onto the right half, and finally unpack the paired lanes into
7869 /// their final position.
7871 /// The exact breakdown of how to form these dword pairs and align them on the
7872 /// correct sides is really tricky. See the comments within the function for
7873 /// more of the details.
7875 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
7876 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
7877 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
7878 /// vector, form the analogous 128-bit 8-element Mask.
7879 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7880 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
7881 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7882 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
7883 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
7885 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
7886 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7887 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7889 SmallVector<int, 4> LoInputs;
7890 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7891 [](int M) { return M >= 0; });
7892 std::sort(LoInputs.begin(), LoInputs.end());
7893 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7894 SmallVector<int, 4> HiInputs;
7895 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7896 [](int M) { return M >= 0; });
7897 std::sort(HiInputs.begin(), HiInputs.end());
7898 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7900 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7901 int NumHToL = LoInputs.size() - NumLToL;
7903 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7904 int NumHToH = HiInputs.size() - NumLToH;
7905 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7906 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7907 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7908 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7910 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7911 // such inputs we can swap two of the dwords across the half mark and end up
7912 // with <=2 inputs to each half in each half. Once there, we can fall through
7913 // to the generic code below. For example:
7915 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7916 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7918 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7919 // and an existing 2-into-2 on the other half. In this case we may have to
7920 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7921 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7922 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7923 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7924 // half than the one we target for fixing) will be fixed when we re-enter this
7925 // path. We will also combine away any sequence of PSHUFD instructions that
7926 // result into a single instruction. Here is an example of the tricky case:
7928 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7929 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7931 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7933 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7934 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7936 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7937 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7939 // The result is fine to be handled by the generic logic.
7940 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
7941 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
7942 int AOffset, int BOffset) {
7943 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
7944 "Must call this with A having 3 or 1 inputs from the A half.");
7945 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
7946 "Must call this with B having 1 or 3 inputs from the B half.");
7947 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
7948 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
7950 // Compute the index of dword with only one word among the three inputs in
7951 // a half by taking the sum of the half with three inputs and subtracting
7952 // the sum of the actual three inputs. The difference is the remaining
7955 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
7956 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
7957 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
7958 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
7959 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
7960 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
7961 int TripleNonInputIdx =
7962 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
7963 TripleDWord = TripleNonInputIdx / 2;
7965 // We use xor with one to compute the adjacent DWord to whichever one the
7967 OneInputDWord = (OneInput / 2) ^ 1;
7969 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
7970 // and BToA inputs. If there is also such a problem with the BToB and AToB
7971 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
7972 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
7973 // is essential that we don't *create* a 3<-1 as then we might oscillate.
7974 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
7975 // Compute how many inputs will be flipped by swapping these DWords. We
7977 // to balance this to ensure we don't form a 3-1 shuffle in the other
7979 int NumFlippedAToBInputs =
7980 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
7981 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
7982 int NumFlippedBToBInputs =
7983 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
7984 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
7985 if ((NumFlippedAToBInputs == 1 &&
7986 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
7987 (NumFlippedBToBInputs == 1 &&
7988 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
7989 // We choose whether to fix the A half or B half based on whether that
7990 // half has zero flipped inputs. At zero, we may not be able to fix it
7991 // with that half. We also bias towards fixing the B half because that
7992 // will more commonly be the high half, and we have to bias one way.
7993 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
7994 ArrayRef<int> Inputs) {
7995 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
7996 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
7997 PinnedIdx ^ 1) != Inputs.end();
7998 // Determine whether the free index is in the flipped dword or the
7999 // unflipped dword based on where the pinned index is. We use this bit
8000 // in an xor to conditionally select the adjacent dword.
8001 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8002 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8003 FixFreeIdx) != Inputs.end();
8004 if (IsFixIdxInput == IsFixFreeIdxInput)
8006 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8007 FixFreeIdx) != Inputs.end();
8008 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8009 "We need to be changing the number of flipped inputs!");
8010 int PSHUFHalfMask[] = {0, 1, 2, 3};
8011 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8012 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8014 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8017 if (M != -1 && M == FixIdx)
8019 else if (M != -1 && M == FixFreeIdx)
8022 if (NumFlippedBToBInputs != 0) {
8024 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8025 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8027 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8029 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8030 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8035 int PSHUFDMask[] = {0, 1, 2, 3};
8036 PSHUFDMask[ADWord] = BDWord;
8037 PSHUFDMask[BDWord] = ADWord;
8038 V = DAG.getNode(ISD::BITCAST, DL, VT,
8039 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8040 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8041 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8044 // Adjust the mask to match the new locations of A and B.
8046 if (M != -1 && M/2 == ADWord)
8047 M = 2 * BDWord + M % 2;
8048 else if (M != -1 && M/2 == BDWord)
8049 M = 2 * ADWord + M % 2;
8051 // Recurse back into this routine to re-compute state now that this isn't
8052 // a 3 and 1 problem.
8053 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8056 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8057 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8058 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8059 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8061 // At this point there are at most two inputs to the low and high halves from
8062 // each half. That means the inputs can always be grouped into dwords and
8063 // those dwords can then be moved to the correct half with a dword shuffle.
8064 // We use at most one low and one high word shuffle to collect these paired
8065 // inputs into dwords, and finally a dword shuffle to place them.
8066 int PSHUFLMask[4] = {-1, -1, -1, -1};
8067 int PSHUFHMask[4] = {-1, -1, -1, -1};
8068 int PSHUFDMask[4] = {-1, -1, -1, -1};
8070 // First fix the masks for all the inputs that are staying in their
8071 // original halves. This will then dictate the targets of the cross-half
8073 auto fixInPlaceInputs =
8074 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8075 MutableArrayRef<int> SourceHalfMask,
8076 MutableArrayRef<int> HalfMask, int HalfOffset) {
8077 if (InPlaceInputs.empty())
8079 if (InPlaceInputs.size() == 1) {
8080 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8081 InPlaceInputs[0] - HalfOffset;
8082 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8085 if (IncomingInputs.empty()) {
8086 // Just fix all of the in place inputs.
8087 for (int Input : InPlaceInputs) {
8088 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8089 PSHUFDMask[Input / 2] = Input / 2;
8094 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8095 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8096 InPlaceInputs[0] - HalfOffset;
8097 // Put the second input next to the first so that they are packed into
8098 // a dword. We find the adjacent index by toggling the low bit.
8099 int AdjIndex = InPlaceInputs[0] ^ 1;
8100 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8101 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8102 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8104 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8105 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8107 // Now gather the cross-half inputs and place them into a free dword of
8108 // their target half.
8109 // FIXME: This operation could almost certainly be simplified dramatically to
8110 // look more like the 3-1 fixing operation.
8111 auto moveInputsToRightHalf = [&PSHUFDMask](
8112 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8113 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8114 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8116 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8117 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8119 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8121 int LowWord = Word & ~1;
8122 int HighWord = Word | 1;
8123 return isWordClobbered(SourceHalfMask, LowWord) ||
8124 isWordClobbered(SourceHalfMask, HighWord);
8127 if (IncomingInputs.empty())
8130 if (ExistingInputs.empty()) {
8131 // Map any dwords with inputs from them into the right half.
8132 for (int Input : IncomingInputs) {
8133 // If the source half mask maps over the inputs, turn those into
8134 // swaps and use the swapped lane.
8135 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8136 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8137 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8138 Input - SourceOffset;
8139 // We have to swap the uses in our half mask in one sweep.
8140 for (int &M : HalfMask)
8141 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8143 else if (M == Input)
8144 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8146 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8147 Input - SourceOffset &&
8148 "Previous placement doesn't match!");
8150 // Note that this correctly re-maps both when we do a swap and when
8151 // we observe the other side of the swap above. We rely on that to
8152 // avoid swapping the members of the input list directly.
8153 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8156 // Map the input's dword into the correct half.
8157 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8158 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8160 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8162 "Previous placement doesn't match!");
8165 // And just directly shift any other-half mask elements to be same-half
8166 // as we will have mirrored the dword containing the element into the
8167 // same position within that half.
8168 for (int &M : HalfMask)
8169 if (M >= SourceOffset && M < SourceOffset + 4) {
8170 M = M - SourceOffset + DestOffset;
8171 assert(M >= 0 && "This should never wrap below zero!");
8176 // Ensure we have the input in a viable dword of its current half. This
8177 // is particularly tricky because the original position may be clobbered
8178 // by inputs being moved and *staying* in that half.
8179 if (IncomingInputs.size() == 1) {
8180 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8181 int InputFixed = std::find(std::begin(SourceHalfMask),
8182 std::end(SourceHalfMask), -1) -
8183 std::begin(SourceHalfMask) + SourceOffset;
8184 SourceHalfMask[InputFixed - SourceOffset] =
8185 IncomingInputs[0] - SourceOffset;
8186 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8188 IncomingInputs[0] = InputFixed;
8190 } else if (IncomingInputs.size() == 2) {
8191 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8192 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8193 // We have two non-adjacent or clobbered inputs we need to extract from
8194 // the source half. To do this, we need to map them into some adjacent
8195 // dword slot in the source mask.
8196 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8197 IncomingInputs[1] - SourceOffset};
8199 // If there is a free slot in the source half mask adjacent to one of
8200 // the inputs, place the other input in it. We use (Index XOR 1) to
8201 // compute an adjacent index.
8202 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8203 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8204 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8205 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8206 InputsFixed[1] = InputsFixed[0] ^ 1;
8207 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8208 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8209 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8210 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8211 InputsFixed[0] = InputsFixed[1] ^ 1;
8212 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8213 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8214 // The two inputs are in the same DWord but it is clobbered and the
8215 // adjacent DWord isn't used at all. Move both inputs to the free
8217 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8218 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8219 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8220 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8222 // The only way we hit this point is if there is no clobbering
8223 // (because there are no off-half inputs to this half) and there is no
8224 // free slot adjacent to one of the inputs. In this case, we have to
8225 // swap an input with a non-input.
8226 for (int i = 0; i < 4; ++i)
8227 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8228 "We can't handle any clobbers here!");
8229 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8230 "Cannot have adjacent inputs here!");
8232 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8233 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8235 // We also have to update the final source mask in this case because
8236 // it may need to undo the above swap.
8237 for (int &M : FinalSourceHalfMask)
8238 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8239 M = InputsFixed[1] + SourceOffset;
8240 else if (M == InputsFixed[1] + SourceOffset)
8241 M = (InputsFixed[0] ^ 1) + SourceOffset;
8243 InputsFixed[1] = InputsFixed[0] ^ 1;
8246 // Point everything at the fixed inputs.
8247 for (int &M : HalfMask)
8248 if (M == IncomingInputs[0])
8249 M = InputsFixed[0] + SourceOffset;
8250 else if (M == IncomingInputs[1])
8251 M = InputsFixed[1] + SourceOffset;
8253 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8254 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8257 llvm_unreachable("Unhandled input size!");
8260 // Now hoist the DWord down to the right half.
8261 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8262 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8263 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8264 for (int &M : HalfMask)
8265 for (int Input : IncomingInputs)
8267 M = FreeDWord * 2 + Input % 2;
8269 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8270 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8271 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8272 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8274 // Now enact all the shuffles we've computed to move the inputs into their
8276 if (!isNoopShuffleMask(PSHUFLMask))
8277 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8278 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8279 if (!isNoopShuffleMask(PSHUFHMask))
8280 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8281 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8282 if (!isNoopShuffleMask(PSHUFDMask))
8283 V = DAG.getNode(ISD::BITCAST, DL, VT,
8284 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8285 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8286 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8289 // At this point, each half should contain all its inputs, and we can then
8290 // just shuffle them into their final position.
8291 assert(std::count_if(LoMask.begin(), LoMask.end(),
8292 [](int M) { return M >= 4; }) == 0 &&
8293 "Failed to lift all the high half inputs to the low mask!");
8294 assert(std::count_if(HiMask.begin(), HiMask.end(),
8295 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8296 "Failed to lift all the low half inputs to the high mask!");
8298 // Do a half shuffle for the low mask.
8299 if (!isNoopShuffleMask(LoMask))
8300 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8301 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8303 // Do a half shuffle with the high mask after shifting its values down.
8304 for (int &M : HiMask)
8307 if (!isNoopShuffleMask(HiMask))
8308 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8309 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8314 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8315 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8316 SDValue V2, ArrayRef<int> Mask,
8317 SelectionDAG &DAG, bool &V1InUse,
8319 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8325 int Size = Mask.size();
8326 int Scale = 16 / Size;
8327 for (int i = 0; i < 16; ++i) {
8328 if (Mask[i / Scale] == -1) {
8329 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8331 const int ZeroMask = 0x80;
8332 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8334 int V2Idx = Mask[i / Scale] < Size
8336 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8337 if (Zeroable[i / Scale])
8338 V1Idx = V2Idx = ZeroMask;
8339 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8340 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8341 V1InUse |= (ZeroMask != V1Idx);
8342 V2InUse |= (ZeroMask != V2Idx);
8347 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8348 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8349 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8351 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8352 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8353 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8355 // If we need shuffled inputs from both, blend the two.
8357 if (V1InUse && V2InUse)
8358 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8360 V = V1InUse ? V1 : V2;
8362 // Cast the result back to the correct type.
8363 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8366 /// \brief Generic lowering of 8-lane i16 shuffles.
8368 /// This handles both single-input shuffles and combined shuffle/blends with
8369 /// two inputs. The single input shuffles are immediately delegated to
8370 /// a dedicated lowering routine.
8372 /// The blends are lowered in one of three fundamental ways. If there are few
8373 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8374 /// of the input is significantly cheaper when lowered as an interleaving of
8375 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8376 /// halves of the inputs separately (making them have relatively few inputs)
8377 /// and then concatenate them.
8378 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8379 const X86Subtarget *Subtarget,
8380 SelectionDAG &DAG) {
8382 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8383 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8384 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8385 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8386 ArrayRef<int> OrigMask = SVOp->getMask();
8387 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8388 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8389 MutableArrayRef<int> Mask(MaskStorage);
8391 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8393 // Whenever we can lower this as a zext, that instruction is strictly faster
8394 // than any alternative.
8395 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8396 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8399 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8401 auto isV2 = [](int M) { return M >= 8; };
8403 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8405 if (NumV2Inputs == 0) {
8406 // Check for being able to broadcast a single element.
8407 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8408 Mask, Subtarget, DAG))
8411 // Try to use shift instructions.
8413 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8416 // Use dedicated unpack instructions for masks that match their pattern.
8417 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8418 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8419 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8420 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8422 // Try to use byte rotation instructions.
8423 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8424 Mask, Subtarget, DAG))
8427 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8431 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8432 "All single-input shuffles should be canonicalized to be V1-input "
8435 // Try to use shift instructions.
8437 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8440 // There are special ways we can lower some single-element blends.
8441 if (NumV2Inputs == 1)
8442 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8443 Mask, Subtarget, DAG))
8446 // We have different paths for blend lowering, but they all must use the
8447 // *exact* same predicate.
8448 bool IsBlendSupported = Subtarget->hasSSE41();
8449 if (IsBlendSupported)
8450 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8454 if (SDValue Masked =
8455 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8458 // Use dedicated unpack instructions for masks that match their pattern.
8459 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8460 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8461 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8462 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8464 // Try to use byte rotation instructions.
8465 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8466 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8469 if (SDValue BitBlend =
8470 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8473 if (SDValue Unpack =
8474 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8477 // If we can't directly blend but can use PSHUFB, that will be better as it
8478 // can both shuffle and set up the inefficient blend.
8479 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8480 bool V1InUse, V2InUse;
8481 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8485 // We can always bit-blend if we have to so the fallback strategy is to
8486 // decompose into single-input permutes and blends.
8487 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8491 /// \brief Check whether a compaction lowering can be done by dropping even
8492 /// elements and compute how many times even elements must be dropped.
8494 /// This handles shuffles which take every Nth element where N is a power of
8495 /// two. Example shuffle masks:
8497 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8498 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8499 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8500 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8501 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8502 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8504 /// Any of these lanes can of course be undef.
8506 /// This routine only supports N <= 3.
8507 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8510 /// \returns N above, or the number of times even elements must be dropped if
8511 /// there is such a number. Otherwise returns zero.
8512 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8513 // Figure out whether we're looping over two inputs or just one.
8514 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8516 // The modulus for the shuffle vector entries is based on whether this is
8517 // a single input or not.
8518 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8519 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8520 "We should only be called with masks with a power-of-2 size!");
8522 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8524 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8525 // and 2^3 simultaneously. This is because we may have ambiguity with
8526 // partially undef inputs.
8527 bool ViableForN[3] = {true, true, true};
8529 for (int i = 0, e = Mask.size(); i < e; ++i) {
8530 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8535 bool IsAnyViable = false;
8536 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8537 if (ViableForN[j]) {
8540 // The shuffle mask must be equal to (i * 2^N) % M.
8541 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8544 ViableForN[j] = false;
8546 // Early exit if we exhaust the possible powers of two.
8551 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8555 // Return 0 as there is no viable power of two.
8559 /// \brief Generic lowering of v16i8 shuffles.
8561 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8562 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8563 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8564 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8566 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8567 const X86Subtarget *Subtarget,
8568 SelectionDAG &DAG) {
8570 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8571 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8572 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8573 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8574 ArrayRef<int> Mask = SVOp->getMask();
8575 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8577 // Try to use shift instructions.
8579 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8582 // Try to use byte rotation instructions.
8583 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8584 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8587 // Try to use a zext lowering.
8588 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8589 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8593 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8595 // For single-input shuffles, there are some nicer lowering tricks we can use.
8596 if (NumV2Elements == 0) {
8597 // Check for being able to broadcast a single element.
8598 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8599 Mask, Subtarget, DAG))
8602 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8603 // Notably, this handles splat and partial-splat shuffles more efficiently.
8604 // However, it only makes sense if the pre-duplication shuffle simplifies
8605 // things significantly. Currently, this means we need to be able to
8606 // express the pre-duplication shuffle as an i16 shuffle.
8608 // FIXME: We should check for other patterns which can be widened into an
8609 // i16 shuffle as well.
8610 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8611 for (int i = 0; i < 16; i += 2)
8612 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8617 auto tryToWidenViaDuplication = [&]() -> SDValue {
8618 if (!canWidenViaDuplication(Mask))
8620 SmallVector<int, 4> LoInputs;
8621 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8622 [](int M) { return M >= 0 && M < 8; });
8623 std::sort(LoInputs.begin(), LoInputs.end());
8624 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8626 SmallVector<int, 4> HiInputs;
8627 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8628 [](int M) { return M >= 8; });
8629 std::sort(HiInputs.begin(), HiInputs.end());
8630 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8633 bool TargetLo = LoInputs.size() >= HiInputs.size();
8634 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8635 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8637 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8638 SmallDenseMap<int, int, 8> LaneMap;
8639 for (int I : InPlaceInputs) {
8640 PreDupI16Shuffle[I/2] = I/2;
8643 int j = TargetLo ? 0 : 4, je = j + 4;
8644 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8645 // Check if j is already a shuffle of this input. This happens when
8646 // there are two adjacent bytes after we move the low one.
8647 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8648 // If we haven't yet mapped the input, search for a slot into which
8650 while (j < je && PreDupI16Shuffle[j] != -1)
8654 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8657 // Map this input with the i16 shuffle.
8658 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8661 // Update the lane map based on the mapping we ended up with.
8662 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8665 ISD::BITCAST, DL, MVT::v16i8,
8666 DAG.getVectorShuffle(MVT::v8i16, DL,
8667 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8668 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8670 // Unpack the bytes to form the i16s that will be shuffled into place.
8671 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8672 MVT::v16i8, V1, V1);
8674 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8675 for (int i = 0; i < 16; ++i)
8676 if (Mask[i] != -1) {
8677 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8678 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8679 if (PostDupI16Shuffle[i / 2] == -1)
8680 PostDupI16Shuffle[i / 2] = MappedMask;
8682 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8683 "Conflicting entrties in the original shuffle!");
8686 ISD::BITCAST, DL, MVT::v16i8,
8687 DAG.getVectorShuffle(MVT::v8i16, DL,
8688 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8689 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8691 if (SDValue V = tryToWidenViaDuplication())
8695 // Use dedicated unpack instructions for masks that match their pattern.
8696 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8697 0, 16, 1, 17, 2, 18, 3, 19,
8699 4, 20, 5, 21, 6, 22, 7, 23}))
8700 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8701 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8702 8, 24, 9, 25, 10, 26, 11, 27,
8704 12, 28, 13, 29, 14, 30, 15, 31}))
8705 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8707 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8708 // with PSHUFB. It is important to do this before we attempt to generate any
8709 // blends but after all of the single-input lowerings. If the single input
8710 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8711 // want to preserve that and we can DAG combine any longer sequences into
8712 // a PSHUFB in the end. But once we start blending from multiple inputs,
8713 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8714 // and there are *very* few patterns that would actually be faster than the
8715 // PSHUFB approach because of its ability to zero lanes.
8717 // FIXME: The only exceptions to the above are blends which are exact
8718 // interleavings with direct instructions supporting them. We currently don't
8719 // handle those well here.
8720 if (Subtarget->hasSSSE3()) {
8721 bool V1InUse = false;
8722 bool V2InUse = false;
8724 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8725 DAG, V1InUse, V2InUse);
8727 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8728 // do so. This avoids using them to handle blends-with-zero which is
8729 // important as a single pshufb is significantly faster for that.
8730 if (V1InUse && V2InUse) {
8731 if (Subtarget->hasSSE41())
8732 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8733 Mask, Subtarget, DAG))
8736 // We can use an unpack to do the blending rather than an or in some
8737 // cases. Even though the or may be (very minorly) more efficient, we
8738 // preference this lowering because there are common cases where part of
8739 // the complexity of the shuffles goes away when we do the final blend as
8741 // FIXME: It might be worth trying to detect if the unpack-feeding
8742 // shuffles will both be pshufb, in which case we shouldn't bother with
8744 if (SDValue Unpack =
8745 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8752 // There are special ways we can lower some single-element blends.
8753 if (NumV2Elements == 1)
8754 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8755 Mask, Subtarget, DAG))
8758 if (SDValue BitBlend =
8759 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8762 // Check whether a compaction lowering can be done. This handles shuffles
8763 // which take every Nth element for some even N. See the helper function for
8766 // We special case these as they can be particularly efficiently handled with
8767 // the PACKUSB instruction on x86 and they show up in common patterns of
8768 // rearranging bytes to truncate wide elements.
8769 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8770 // NumEvenDrops is the power of two stride of the elements. Another way of
8771 // thinking about it is that we need to drop the even elements this many
8772 // times to get the original input.
8773 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8775 // First we need to zero all the dropped bytes.
8776 assert(NumEvenDrops <= 3 &&
8777 "No support for dropping even elements more than 3 times.");
8778 // We use the mask type to pick which bytes are preserved based on how many
8779 // elements are dropped.
8780 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8781 SDValue ByteClearMask =
8782 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8783 DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
8784 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8786 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8788 // Now pack things back together.
8789 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8790 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8791 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8792 for (int i = 1; i < NumEvenDrops; ++i) {
8793 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8794 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8800 // Handle multi-input cases by blending single-input shuffles.
8801 if (NumV2Elements > 0)
8802 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8805 // The fallback path for single-input shuffles widens this into two v8i16
8806 // vectors with unpacks, shuffles those, and then pulls them back together
8810 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8811 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8812 for (int i = 0; i < 16; ++i)
8814 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8816 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8818 SDValue VLoHalf, VHiHalf;
8819 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8820 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8822 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8823 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8824 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8825 [](int M) { return M >= 0 && M % 2 == 1; })) {
8826 // Use a mask to drop the high bytes.
8827 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8828 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8829 DAG.getConstant(0x00FF, DL, MVT::v8i16));
8831 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8832 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8834 // Squash the masks to point directly into VLoHalf.
8835 for (int &M : LoBlendMask)
8838 for (int &M : HiBlendMask)
8842 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8843 // VHiHalf so that we can blend them as i16s.
8844 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8845 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8846 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8847 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8850 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8851 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8853 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8856 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8858 /// This routine breaks down the specific type of 128-bit shuffle and
8859 /// dispatches to the lowering routines accordingly.
8860 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8861 MVT VT, const X86Subtarget *Subtarget,
8862 SelectionDAG &DAG) {
8863 switch (VT.SimpleTy) {
8865 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8867 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8869 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8871 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8873 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8875 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8878 llvm_unreachable("Unimplemented!");
8882 /// \brief Helper function to test whether a shuffle mask could be
8883 /// simplified by widening the elements being shuffled.
8885 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8886 /// leaves it in an unspecified state.
8888 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8889 /// shuffle masks. The latter have the special property of a '-2' representing
8890 /// a zero-ed lane of a vector.
8891 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8892 SmallVectorImpl<int> &WidenedMask) {
8893 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8894 // If both elements are undef, its trivial.
8895 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8896 WidenedMask.push_back(SM_SentinelUndef);
8900 // Check for an undef mask and a mask value properly aligned to fit with
8901 // a pair of values. If we find such a case, use the non-undef mask's value.
8902 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8903 WidenedMask.push_back(Mask[i + 1] / 2);
8906 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8907 WidenedMask.push_back(Mask[i] / 2);
8911 // When zeroing, we need to spread the zeroing across both lanes to widen.
8912 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8913 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8914 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8915 WidenedMask.push_back(SM_SentinelZero);
8921 // Finally check if the two mask values are adjacent and aligned with
8923 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8924 WidenedMask.push_back(Mask[i] / 2);
8928 // Otherwise we can't safely widen the elements used in this shuffle.
8931 assert(WidenedMask.size() == Mask.size() / 2 &&
8932 "Incorrect size of mask after widening the elements!");
8937 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8939 /// This routine just extracts two subvectors, shuffles them independently, and
8940 /// then concatenates them back together. This should work effectively with all
8941 /// AVX vector shuffle types.
8942 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8943 SDValue V2, ArrayRef<int> Mask,
8944 SelectionDAG &DAG) {
8945 assert(VT.getSizeInBits() >= 256 &&
8946 "Only for 256-bit or wider vector shuffles!");
8947 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
8948 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
8950 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
8951 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
8953 int NumElements = VT.getVectorNumElements();
8954 int SplitNumElements = NumElements / 2;
8955 MVT ScalarVT = VT.getScalarType();
8956 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
8958 // Rather than splitting build-vectors, just build two narrower build
8959 // vectors. This helps shuffling with splats and zeros.
8960 auto SplitVector = [&](SDValue V) {
8961 while (V.getOpcode() == ISD::BITCAST)
8962 V = V->getOperand(0);
8964 MVT OrigVT = V.getSimpleValueType();
8965 int OrigNumElements = OrigVT.getVectorNumElements();
8966 int OrigSplitNumElements = OrigNumElements / 2;
8967 MVT OrigScalarVT = OrigVT.getScalarType();
8968 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
8972 auto *BV = dyn_cast<BuildVectorSDNode>(V);
8974 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8975 DAG.getIntPtrConstant(0, DL));
8976 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8977 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
8980 SmallVector<SDValue, 16> LoOps, HiOps;
8981 for (int i = 0; i < OrigSplitNumElements; ++i) {
8982 LoOps.push_back(BV->getOperand(i));
8983 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
8985 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
8986 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
8988 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
8989 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
8992 SDValue LoV1, HiV1, LoV2, HiV2;
8993 std::tie(LoV1, HiV1) = SplitVector(V1);
8994 std::tie(LoV2, HiV2) = SplitVector(V2);
8996 // Now create two 4-way blends of these half-width vectors.
8997 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
8998 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
8999 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9000 for (int i = 0; i < SplitNumElements; ++i) {
9001 int M = HalfMask[i];
9002 if (M >= NumElements) {
9003 if (M >= NumElements + SplitNumElements)
9007 V2BlendMask.push_back(M - NumElements);
9008 V1BlendMask.push_back(-1);
9009 BlendMask.push_back(SplitNumElements + i);
9010 } else if (M >= 0) {
9011 if (M >= SplitNumElements)
9015 V2BlendMask.push_back(-1);
9016 V1BlendMask.push_back(M);
9017 BlendMask.push_back(i);
9019 V2BlendMask.push_back(-1);
9020 V1BlendMask.push_back(-1);
9021 BlendMask.push_back(-1);
9025 // Because the lowering happens after all combining takes place, we need to
9026 // manually combine these blend masks as much as possible so that we create
9027 // a minimal number of high-level vector shuffle nodes.
9029 // First try just blending the halves of V1 or V2.
9030 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9031 return DAG.getUNDEF(SplitVT);
9032 if (!UseLoV2 && !UseHiV2)
9033 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9034 if (!UseLoV1 && !UseHiV1)
9035 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9037 SDValue V1Blend, V2Blend;
9038 if (UseLoV1 && UseHiV1) {
9040 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9042 // We only use half of V1 so map the usage down into the final blend mask.
9043 V1Blend = UseLoV1 ? LoV1 : HiV1;
9044 for (int i = 0; i < SplitNumElements; ++i)
9045 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9046 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9048 if (UseLoV2 && UseHiV2) {
9050 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9052 // We only use half of V2 so map the usage down into the final blend mask.
9053 V2Blend = UseLoV2 ? LoV2 : HiV2;
9054 for (int i = 0; i < SplitNumElements; ++i)
9055 if (BlendMask[i] >= SplitNumElements)
9056 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9058 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9060 SDValue Lo = HalfBlend(LoMask);
9061 SDValue Hi = HalfBlend(HiMask);
9062 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9065 /// \brief Either split a vector in halves or decompose the shuffles and the
9068 /// This is provided as a good fallback for many lowerings of non-single-input
9069 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9070 /// between splitting the shuffle into 128-bit components and stitching those
9071 /// back together vs. extracting the single-input shuffles and blending those
9073 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9074 SDValue V2, ArrayRef<int> Mask,
9075 SelectionDAG &DAG) {
9076 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9077 "lower single-input shuffles as it "
9078 "could then recurse on itself.");
9079 int Size = Mask.size();
9081 // If this can be modeled as a broadcast of two elements followed by a blend,
9082 // prefer that lowering. This is especially important because broadcasts can
9083 // often fold with memory operands.
9084 auto DoBothBroadcast = [&] {
9085 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9088 if (V2BroadcastIdx == -1)
9089 V2BroadcastIdx = M - Size;
9090 else if (M - Size != V2BroadcastIdx)
9092 } else if (M >= 0) {
9093 if (V1BroadcastIdx == -1)
9095 else if (M != V1BroadcastIdx)
9100 if (DoBothBroadcast())
9101 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9104 // If the inputs all stem from a single 128-bit lane of each input, then we
9105 // split them rather than blending because the split will decompose to
9106 // unusually few instructions.
9107 int LaneCount = VT.getSizeInBits() / 128;
9108 int LaneSize = Size / LaneCount;
9109 SmallBitVector LaneInputs[2];
9110 LaneInputs[0].resize(LaneCount, false);
9111 LaneInputs[1].resize(LaneCount, false);
9112 for (int i = 0; i < Size; ++i)
9114 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9115 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9116 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9118 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9119 // that the decomposed single-input shuffles don't end up here.
9120 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9123 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9124 /// a permutation and blend of those lanes.
9126 /// This essentially blends the out-of-lane inputs to each lane into the lane
9127 /// from a permuted copy of the vector. This lowering strategy results in four
9128 /// instructions in the worst case for a single-input cross lane shuffle which
9129 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9130 /// of. Special cases for each particular shuffle pattern should be handled
9131 /// prior to trying this lowering.
9132 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9133 SDValue V1, SDValue V2,
9135 SelectionDAG &DAG) {
9136 // FIXME: This should probably be generalized for 512-bit vectors as well.
9137 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9138 int LaneSize = Mask.size() / 2;
9140 // If there are only inputs from one 128-bit lane, splitting will in fact be
9141 // less expensive. The flags track whether the given lane contains an element
9142 // that crosses to another lane.
9143 bool LaneCrossing[2] = {false, false};
9144 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9145 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9146 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9147 if (!LaneCrossing[0] || !LaneCrossing[1])
9148 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9150 if (isSingleInputShuffleMask(Mask)) {
9151 SmallVector<int, 32> FlippedBlendMask;
9152 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9153 FlippedBlendMask.push_back(
9154 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9156 : Mask[i] % LaneSize +
9157 (i / LaneSize) * LaneSize + Size));
9159 // Flip the vector, and blend the results which should now be in-lane. The
9160 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9161 // 5 for the high source. The value 3 selects the high half of source 2 and
9162 // the value 2 selects the low half of source 2. We only use source 2 to
9163 // allow folding it into a memory operand.
9164 unsigned PERMMask = 3 | 2 << 4;
9165 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9166 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9167 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9170 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9171 // will be handled by the above logic and a blend of the results, much like
9172 // other patterns in AVX.
9173 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9176 /// \brief Handle lowering 2-lane 128-bit shuffles.
9177 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9178 SDValue V2, ArrayRef<int> Mask,
9179 const X86Subtarget *Subtarget,
9180 SelectionDAG &DAG) {
9181 // TODO: If minimizing size and one of the inputs is a zero vector and the
9182 // the zero vector has only one use, we could use a VPERM2X128 to save the
9183 // instruction bytes needed to explicitly generate the zero vector.
9185 // Blends are faster and handle all the non-lane-crossing cases.
9186 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9190 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9191 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9193 // If either input operand is a zero vector, use VPERM2X128 because its mask
9194 // allows us to replace the zero input with an implicit zero.
9195 if (!IsV1Zero && !IsV2Zero) {
9196 // Check for patterns which can be matched with a single insert of a 128-bit
9198 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9199 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9200 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9201 VT.getVectorNumElements() / 2);
9202 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9203 DAG.getIntPtrConstant(0, DL));
9204 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9205 OnlyUsesV1 ? V1 : V2,
9206 DAG.getIntPtrConstant(0, DL));
9207 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9211 // Otherwise form a 128-bit permutation. After accounting for undefs,
9212 // convert the 64-bit shuffle mask selection values into 128-bit
9213 // selection bits by dividing the indexes by 2 and shifting into positions
9214 // defined by a vperm2*128 instruction's immediate control byte.
9216 // The immediate permute control byte looks like this:
9217 // [1:0] - select 128 bits from sources for low half of destination
9219 // [3] - zero low half of destination
9220 // [5:4] - select 128 bits from sources for high half of destination
9222 // [7] - zero high half of destination
9224 int MaskLO = Mask[0];
9225 if (MaskLO == SM_SentinelUndef)
9226 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9228 int MaskHI = Mask[2];
9229 if (MaskHI == SM_SentinelUndef)
9230 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9232 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9234 // If either input is a zero vector, replace it with an undef input.
9235 // Shuffle mask values < 4 are selecting elements of V1.
9236 // Shuffle mask values >= 4 are selecting elements of V2.
9237 // Adjust each half of the permute mask by clearing the half that was
9238 // selecting the zero vector and setting the zero mask bit.
9240 V1 = DAG.getUNDEF(VT);
9242 PermMask = (PermMask & 0xf0) | 0x08;
9244 PermMask = (PermMask & 0x0f) | 0x80;
9247 V2 = DAG.getUNDEF(VT);
9249 PermMask = (PermMask & 0xf0) | 0x08;
9251 PermMask = (PermMask & 0x0f) | 0x80;
9254 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9255 DAG.getConstant(PermMask, DL, MVT::i8));
9258 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9259 /// shuffling each lane.
9261 /// This will only succeed when the result of fixing the 128-bit lanes results
9262 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9263 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9264 /// the lane crosses early and then use simpler shuffles within each lane.
9266 /// FIXME: It might be worthwhile at some point to support this without
9267 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9268 /// in x86 only floating point has interesting non-repeating shuffles, and even
9269 /// those are still *marginally* more expensive.
9270 static SDValue lowerVectorShuffleByMerging128BitLanes(
9271 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9272 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9273 assert(!isSingleInputShuffleMask(Mask) &&
9274 "This is only useful with multiple inputs.");
9276 int Size = Mask.size();
9277 int LaneSize = 128 / VT.getScalarSizeInBits();
9278 int NumLanes = Size / LaneSize;
9279 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9281 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9282 // check whether the in-128-bit lane shuffles share a repeating pattern.
9283 SmallVector<int, 4> Lanes;
9284 Lanes.resize(NumLanes, -1);
9285 SmallVector<int, 4> InLaneMask;
9286 InLaneMask.resize(LaneSize, -1);
9287 for (int i = 0; i < Size; ++i) {
9291 int j = i / LaneSize;
9294 // First entry we've seen for this lane.
9295 Lanes[j] = Mask[i] / LaneSize;
9296 } else if (Lanes[j] != Mask[i] / LaneSize) {
9297 // This doesn't match the lane selected previously!
9301 // Check that within each lane we have a consistent shuffle mask.
9302 int k = i % LaneSize;
9303 if (InLaneMask[k] < 0) {
9304 InLaneMask[k] = Mask[i] % LaneSize;
9305 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9306 // This doesn't fit a repeating in-lane mask.
9311 // First shuffle the lanes into place.
9312 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9313 VT.getSizeInBits() / 64);
9314 SmallVector<int, 8> LaneMask;
9315 LaneMask.resize(NumLanes * 2, -1);
9316 for (int i = 0; i < NumLanes; ++i)
9317 if (Lanes[i] >= 0) {
9318 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9319 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9322 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9323 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9324 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9326 // Cast it back to the type we actually want.
9327 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9329 // Now do a simple shuffle that isn't lane crossing.
9330 SmallVector<int, 8> NewMask;
9331 NewMask.resize(Size, -1);
9332 for (int i = 0; i < Size; ++i)
9334 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9335 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9336 "Must not introduce lane crosses at this point!");
9338 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9341 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9344 /// This returns true if the elements from a particular input are already in the
9345 /// slot required by the given mask and require no permutation.
9346 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9347 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9348 int Size = Mask.size();
9349 for (int i = 0; i < Size; ++i)
9350 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9356 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9358 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9359 /// isn't available.
9360 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9361 const X86Subtarget *Subtarget,
9362 SelectionDAG &DAG) {
9364 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9365 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9366 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9367 ArrayRef<int> Mask = SVOp->getMask();
9368 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9370 SmallVector<int, 4> WidenedMask;
9371 if (canWidenShuffleElements(Mask, WidenedMask))
9372 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9375 if (isSingleInputShuffleMask(Mask)) {
9376 // Check for being able to broadcast a single element.
9377 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9378 Mask, Subtarget, DAG))
9381 // Use low duplicate instructions for masks that match their pattern.
9382 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9383 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9385 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9386 // Non-half-crossing single input shuffles can be lowerid with an
9387 // interleaved permutation.
9388 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9389 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9390 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9391 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9394 // With AVX2 we have direct support for this permutation.
9395 if (Subtarget->hasAVX2())
9396 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9397 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9399 // Otherwise, fall back.
9400 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9404 // X86 has dedicated unpack instructions that can handle specific blend
9405 // operations: UNPCKH and UNPCKL.
9406 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9407 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9408 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9409 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9410 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9411 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9412 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9413 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9415 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9419 // Check if the blend happens to exactly fit that of SHUFPD.
9420 if ((Mask[0] == -1 || Mask[0] < 2) &&
9421 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9422 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9423 (Mask[3] == -1 || Mask[3] >= 6)) {
9424 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9425 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9426 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9427 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9429 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9430 (Mask[1] == -1 || Mask[1] < 2) &&
9431 (Mask[2] == -1 || Mask[2] >= 6) &&
9432 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9433 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9434 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9435 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9436 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9439 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9440 // shuffle. However, if we have AVX2 and either inputs are already in place,
9441 // we will be able to shuffle even across lanes the other input in a single
9442 // instruction so skip this pattern.
9443 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9444 isShuffleMaskInputInPlace(1, Mask))))
9445 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9446 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9449 // If we have AVX2 then we always want to lower with a blend because an v4 we
9450 // can fully permute the elements.
9451 if (Subtarget->hasAVX2())
9452 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9455 // Otherwise fall back on generic lowering.
9456 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9459 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9461 /// This routine is only called when we have AVX2 and thus a reasonable
9462 /// instruction set for v4i64 shuffling..
9463 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9464 const X86Subtarget *Subtarget,
9465 SelectionDAG &DAG) {
9467 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9468 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9469 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9470 ArrayRef<int> Mask = SVOp->getMask();
9471 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9472 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9474 SmallVector<int, 4> WidenedMask;
9475 if (canWidenShuffleElements(Mask, WidenedMask))
9476 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9479 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9483 // Check for being able to broadcast a single element.
9484 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9485 Mask, Subtarget, DAG))
9488 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9489 // use lower latency instructions that will operate on both 128-bit lanes.
9490 SmallVector<int, 2> RepeatedMask;
9491 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9492 if (isSingleInputShuffleMask(Mask)) {
9493 int PSHUFDMask[] = {-1, -1, -1, -1};
9494 for (int i = 0; i < 2; ++i)
9495 if (RepeatedMask[i] >= 0) {
9496 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9497 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9500 ISD::BITCAST, DL, MVT::v4i64,
9501 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9502 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9503 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9507 // AVX2 provides a direct instruction for permuting a single input across
9509 if (isSingleInputShuffleMask(Mask))
9510 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9511 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9513 // Try to use shift instructions.
9515 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9518 // Use dedicated unpack instructions for masks that match their pattern.
9519 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9520 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9521 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9522 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9523 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9524 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9525 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9526 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9528 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9529 // shuffle. However, if we have AVX2 and either inputs are already in place,
9530 // we will be able to shuffle even across lanes the other input in a single
9531 // instruction so skip this pattern.
9532 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9533 isShuffleMaskInputInPlace(1, Mask))))
9534 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9535 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9538 // Otherwise fall back on generic blend lowering.
9539 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9543 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9545 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9546 /// isn't available.
9547 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9548 const X86Subtarget *Subtarget,
9549 SelectionDAG &DAG) {
9551 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9552 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9553 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9554 ArrayRef<int> Mask = SVOp->getMask();
9555 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9557 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9561 // Check for being able to broadcast a single element.
9562 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9563 Mask, Subtarget, DAG))
9566 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9567 // options to efficiently lower the shuffle.
9568 SmallVector<int, 4> RepeatedMask;
9569 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9570 assert(RepeatedMask.size() == 4 &&
9571 "Repeated masks must be half the mask width!");
9573 // Use even/odd duplicate instructions for masks that match their pattern.
9574 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9575 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9576 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9577 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9579 if (isSingleInputShuffleMask(Mask))
9580 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9581 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9583 // Use dedicated unpack instructions for masks that match their pattern.
9584 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9585 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9586 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9587 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9588 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9589 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9590 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9591 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9593 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9594 // have already handled any direct blends. We also need to squash the
9595 // repeated mask into a simulated v4f32 mask.
9596 for (int i = 0; i < 4; ++i)
9597 if (RepeatedMask[i] >= 8)
9598 RepeatedMask[i] -= 4;
9599 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9602 // If we have a single input shuffle with different shuffle patterns in the
9603 // two 128-bit lanes use the variable mask to VPERMILPS.
9604 if (isSingleInputShuffleMask(Mask)) {
9605 SDValue VPermMask[8];
9606 for (int i = 0; i < 8; ++i)
9607 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9608 : DAG.getConstant(Mask[i], DL, MVT::i32);
9609 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9611 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9612 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9614 if (Subtarget->hasAVX2())
9615 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9616 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9617 DAG.getNode(ISD::BUILD_VECTOR, DL,
9618 MVT::v8i32, VPermMask)),
9621 // Otherwise, fall back.
9622 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9626 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9628 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9629 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9632 // If we have AVX2 then we always want to lower with a blend because at v8 we
9633 // can fully permute the elements.
9634 if (Subtarget->hasAVX2())
9635 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9638 // Otherwise fall back on generic lowering.
9639 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9642 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9644 /// This routine is only called when we have AVX2 and thus a reasonable
9645 /// instruction set for v8i32 shuffling..
9646 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9647 const X86Subtarget *Subtarget,
9648 SelectionDAG &DAG) {
9650 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9651 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9652 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9653 ArrayRef<int> Mask = SVOp->getMask();
9654 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9655 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9657 // Whenever we can lower this as a zext, that instruction is strictly faster
9658 // than any alternative. It also allows us to fold memory operands into the
9659 // shuffle in many cases.
9660 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9661 Mask, Subtarget, DAG))
9664 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9668 // Check for being able to broadcast a single element.
9669 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9670 Mask, Subtarget, DAG))
9673 // If the shuffle mask is repeated in each 128-bit lane we can use more
9674 // efficient instructions that mirror the shuffles across the two 128-bit
9676 SmallVector<int, 4> RepeatedMask;
9677 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9678 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9679 if (isSingleInputShuffleMask(Mask))
9680 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9681 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9683 // Use dedicated unpack instructions for masks that match their pattern.
9684 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9685 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9686 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9687 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9688 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9689 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9690 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9691 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9694 // Try to use shift instructions.
9696 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9699 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9700 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9703 // If the shuffle patterns aren't repeated but it is a single input, directly
9704 // generate a cross-lane VPERMD instruction.
9705 if (isSingleInputShuffleMask(Mask)) {
9706 SDValue VPermMask[8];
9707 for (int i = 0; i < 8; ++i)
9708 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9709 : DAG.getConstant(Mask[i], DL, MVT::i32);
9711 X86ISD::VPERMV, DL, MVT::v8i32,
9712 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9715 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9717 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9718 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9721 // Otherwise fall back on generic blend lowering.
9722 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9726 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9728 /// This routine is only called when we have AVX2 and thus a reasonable
9729 /// instruction set for v16i16 shuffling..
9730 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9731 const X86Subtarget *Subtarget,
9732 SelectionDAG &DAG) {
9734 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9735 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9736 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9737 ArrayRef<int> Mask = SVOp->getMask();
9738 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9739 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9741 // Whenever we can lower this as a zext, that instruction is strictly faster
9742 // than any alternative. It also allows us to fold memory operands into the
9743 // shuffle in many cases.
9744 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9745 Mask, Subtarget, DAG))
9748 // Check for being able to broadcast a single element.
9749 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9750 Mask, Subtarget, DAG))
9753 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9757 // Use dedicated unpack instructions for masks that match their pattern.
9758 if (isShuffleEquivalent(V1, V2, Mask,
9759 {// First 128-bit lane:
9760 0, 16, 1, 17, 2, 18, 3, 19,
9761 // Second 128-bit lane:
9762 8, 24, 9, 25, 10, 26, 11, 27}))
9763 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9764 if (isShuffleEquivalent(V1, V2, Mask,
9765 {// First 128-bit lane:
9766 4, 20, 5, 21, 6, 22, 7, 23,
9767 // Second 128-bit lane:
9768 12, 28, 13, 29, 14, 30, 15, 31}))
9769 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9771 // Try to use shift instructions.
9773 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9776 // Try to use byte rotation instructions.
9777 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9778 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9781 if (isSingleInputShuffleMask(Mask)) {
9782 // There are no generalized cross-lane shuffle operations available on i16
9784 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9785 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9788 SmallVector<int, 8> RepeatedMask;
9789 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
9790 // As this is a single-input shuffle, the repeated mask should be
9791 // a strictly valid v8i16 mask that we can pass through to the v8i16
9792 // lowering to handle even the v16 case.
9793 return lowerV8I16GeneralSingleInputVectorShuffle(
9794 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
9797 SDValue PSHUFBMask[32];
9798 for (int i = 0; i < 16; ++i) {
9799 if (Mask[i] == -1) {
9800 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9804 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9805 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9806 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
9807 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
9810 ISD::BITCAST, DL, MVT::v16i16,
9812 X86ISD::PSHUFB, DL, MVT::v32i8,
9813 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9814 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9817 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9819 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9820 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9823 // Otherwise fall back on generic lowering.
9824 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9827 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9829 /// This routine is only called when we have AVX2 and thus a reasonable
9830 /// instruction set for v32i8 shuffling..
9831 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9832 const X86Subtarget *Subtarget,
9833 SelectionDAG &DAG) {
9835 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9836 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9837 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9838 ArrayRef<int> Mask = SVOp->getMask();
9839 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9840 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9842 // Whenever we can lower this as a zext, that instruction is strictly faster
9843 // than any alternative. It also allows us to fold memory operands into the
9844 // shuffle in many cases.
9845 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9846 Mask, Subtarget, DAG))
9849 // Check for being able to broadcast a single element.
9850 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9851 Mask, Subtarget, DAG))
9854 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9858 // Use dedicated unpack instructions for masks that match their pattern.
9859 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9861 if (isShuffleEquivalent(
9863 {// First 128-bit lane:
9864 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9865 // Second 128-bit lane:
9866 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9867 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9868 if (isShuffleEquivalent(
9870 {// First 128-bit lane:
9871 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9872 // Second 128-bit lane:
9873 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9874 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9876 // Try to use shift instructions.
9878 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9881 // Try to use byte rotation instructions.
9882 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9883 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9886 if (isSingleInputShuffleMask(Mask)) {
9887 // There are no generalized cross-lane shuffle operations available on i8
9889 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9890 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9893 SDValue PSHUFBMask[32];
9894 for (int i = 0; i < 32; ++i)
9897 ? DAG.getUNDEF(MVT::i8)
9898 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
9902 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9903 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9906 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9908 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9909 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9912 // Otherwise fall back on generic lowering.
9913 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9916 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9918 /// This routine either breaks down the specific type of a 256-bit x86 vector
9919 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9920 /// together based on the available instructions.
9921 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9922 MVT VT, const X86Subtarget *Subtarget,
9923 SelectionDAG &DAG) {
9925 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9926 ArrayRef<int> Mask = SVOp->getMask();
9928 // If we have a single input to the zero element, insert that into V1 if we
9929 // can do so cheaply.
9930 int NumElts = VT.getVectorNumElements();
9931 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
9932 return M >= NumElts;
9935 if (NumV2Elements == 1 && Mask[0] >= NumElts)
9936 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9937 DL, VT, V1, V2, Mask, Subtarget, DAG))
9940 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9941 // check for those subtargets here and avoid much of the subtarget querying in
9942 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9943 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9944 // floating point types there eventually, just immediately cast everything to
9945 // a float and operate entirely in that domain.
9946 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9947 int ElementBits = VT.getScalarSizeInBits();
9948 if (ElementBits < 32)
9949 // No floating point type available, decompose into 128-bit vectors.
9950 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9952 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9953 VT.getVectorNumElements());
9954 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9955 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9956 return DAG.getNode(ISD::BITCAST, DL, VT,
9957 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9960 switch (VT.SimpleTy) {
9962 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9964 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9966 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9968 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9970 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9972 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9975 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9979 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
9980 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9981 const X86Subtarget *Subtarget,
9982 SelectionDAG &DAG) {
9984 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9985 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9986 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9987 ArrayRef<int> Mask = SVOp->getMask();
9988 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9990 // X86 has dedicated unpack instructions that can handle specific blend
9991 // operations: UNPCKH and UNPCKL.
9992 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9993 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
9994 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9995 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
9997 // FIXME: Implement direct support for this type!
9998 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10001 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10002 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10003 const X86Subtarget *Subtarget,
10004 SelectionDAG &DAG) {
10006 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10007 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10008 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10009 ArrayRef<int> Mask = SVOp->getMask();
10010 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10012 // Use dedicated unpack instructions for masks that match their pattern.
10013 if (isShuffleEquivalent(V1, V2, Mask,
10014 {// First 128-bit lane.
10015 0, 16, 1, 17, 4, 20, 5, 21,
10016 // Second 128-bit lane.
10017 8, 24, 9, 25, 12, 28, 13, 29}))
10018 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10019 if (isShuffleEquivalent(V1, V2, Mask,
10020 {// First 128-bit lane.
10021 2, 18, 3, 19, 6, 22, 7, 23,
10022 // Second 128-bit lane.
10023 10, 26, 11, 27, 14, 30, 15, 31}))
10024 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10026 // FIXME: Implement direct support for this type!
10027 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10030 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10031 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10032 const X86Subtarget *Subtarget,
10033 SelectionDAG &DAG) {
10035 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10036 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10037 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10038 ArrayRef<int> Mask = SVOp->getMask();
10039 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10041 // X86 has dedicated unpack instructions that can handle specific blend
10042 // operations: UNPCKH and UNPCKL.
10043 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10044 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10045 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10046 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10048 // FIXME: Implement direct support for this type!
10049 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10052 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10053 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10054 const X86Subtarget *Subtarget,
10055 SelectionDAG &DAG) {
10057 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10058 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10059 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10060 ArrayRef<int> Mask = SVOp->getMask();
10061 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10063 // Use dedicated unpack instructions for masks that match their pattern.
10064 if (isShuffleEquivalent(V1, V2, Mask,
10065 {// First 128-bit lane.
10066 0, 16, 1, 17, 4, 20, 5, 21,
10067 // Second 128-bit lane.
10068 8, 24, 9, 25, 12, 28, 13, 29}))
10069 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10070 if (isShuffleEquivalent(V1, V2, Mask,
10071 {// First 128-bit lane.
10072 2, 18, 3, 19, 6, 22, 7, 23,
10073 // Second 128-bit lane.
10074 10, 26, 11, 27, 14, 30, 15, 31}))
10075 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10077 // FIXME: Implement direct support for this type!
10078 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10081 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10082 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10083 const X86Subtarget *Subtarget,
10084 SelectionDAG &DAG) {
10086 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10087 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10088 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10089 ArrayRef<int> Mask = SVOp->getMask();
10090 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10091 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10093 // FIXME: Implement direct support for this type!
10094 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10097 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10098 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10099 const X86Subtarget *Subtarget,
10100 SelectionDAG &DAG) {
10102 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10103 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10104 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10105 ArrayRef<int> Mask = SVOp->getMask();
10106 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10107 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10109 // FIXME: Implement direct support for this type!
10110 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10113 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10115 /// This routine either breaks down the specific type of a 512-bit x86 vector
10116 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10117 /// together based on the available instructions.
10118 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10119 MVT VT, const X86Subtarget *Subtarget,
10120 SelectionDAG &DAG) {
10122 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10123 ArrayRef<int> Mask = SVOp->getMask();
10124 assert(Subtarget->hasAVX512() &&
10125 "Cannot lower 512-bit vectors w/ basic ISA!");
10127 // Check for being able to broadcast a single element.
10128 if (SDValue Broadcast =
10129 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10132 // Dispatch to each element type for lowering. If we don't have supprot for
10133 // specific element type shuffles at 512 bits, immediately split them and
10134 // lower them. Each lowering routine of a given type is allowed to assume that
10135 // the requisite ISA extensions for that element type are available.
10136 switch (VT.SimpleTy) {
10138 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10140 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10142 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10144 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10146 if (Subtarget->hasBWI())
10147 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10150 if (Subtarget->hasBWI())
10151 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10155 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10158 // Otherwise fall back on splitting.
10159 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10162 /// \brief Top-level lowering for x86 vector shuffles.
10164 /// This handles decomposition, canonicalization, and lowering of all x86
10165 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10166 /// above in helper routines. The canonicalization attempts to widen shuffles
10167 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10168 /// s.t. only one of the two inputs needs to be tested, etc.
10169 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10170 SelectionDAG &DAG) {
10171 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10172 ArrayRef<int> Mask = SVOp->getMask();
10173 SDValue V1 = Op.getOperand(0);
10174 SDValue V2 = Op.getOperand(1);
10175 MVT VT = Op.getSimpleValueType();
10176 int NumElements = VT.getVectorNumElements();
10179 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10181 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10182 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10183 if (V1IsUndef && V2IsUndef)
10184 return DAG.getUNDEF(VT);
10186 // When we create a shuffle node we put the UNDEF node to second operand,
10187 // but in some cases the first operand may be transformed to UNDEF.
10188 // In this case we should just commute the node.
10190 return DAG.getCommutedVectorShuffle(*SVOp);
10192 // Check for non-undef masks pointing at an undef vector and make the masks
10193 // undef as well. This makes it easier to match the shuffle based solely on
10197 if (M >= NumElements) {
10198 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10199 for (int &M : NewMask)
10200 if (M >= NumElements)
10202 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10205 // We actually see shuffles that are entirely re-arrangements of a set of
10206 // zero inputs. This mostly happens while decomposing complex shuffles into
10207 // simple ones. Directly lower these as a buildvector of zeros.
10208 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10209 if (Zeroable.all())
10210 return getZeroVector(VT, Subtarget, DAG, dl);
10212 // Try to collapse shuffles into using a vector type with fewer elements but
10213 // wider element types. We cap this to not form integers or floating point
10214 // elements wider than 64 bits, but it might be interesting to form i128
10215 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10216 SmallVector<int, 16> WidenedMask;
10217 if (VT.getScalarSizeInBits() < 64 &&
10218 canWidenShuffleElements(Mask, WidenedMask)) {
10219 MVT NewEltVT = VT.isFloatingPoint()
10220 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10221 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10222 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10223 // Make sure that the new vector type is legal. For example, v2f64 isn't
10225 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10226 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
10227 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
10228 return DAG.getNode(ISD::BITCAST, dl, VT,
10229 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10233 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10234 for (int M : SVOp->getMask())
10236 ++NumUndefElements;
10237 else if (M < NumElements)
10242 // Commute the shuffle as needed such that more elements come from V1 than
10243 // V2. This allows us to match the shuffle pattern strictly on how many
10244 // elements come from V1 without handling the symmetric cases.
10245 if (NumV2Elements > NumV1Elements)
10246 return DAG.getCommutedVectorShuffle(*SVOp);
10248 // When the number of V1 and V2 elements are the same, try to minimize the
10249 // number of uses of V2 in the low half of the vector. When that is tied,
10250 // ensure that the sum of indices for V1 is equal to or lower than the sum
10251 // indices for V2. When those are equal, try to ensure that the number of odd
10252 // indices for V1 is lower than the number of odd indices for V2.
10253 if (NumV1Elements == NumV2Elements) {
10254 int LowV1Elements = 0, LowV2Elements = 0;
10255 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10256 if (M >= NumElements)
10260 if (LowV2Elements > LowV1Elements) {
10261 return DAG.getCommutedVectorShuffle(*SVOp);
10262 } else if (LowV2Elements == LowV1Elements) {
10263 int SumV1Indices = 0, SumV2Indices = 0;
10264 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10265 if (SVOp->getMask()[i] >= NumElements)
10267 else if (SVOp->getMask()[i] >= 0)
10269 if (SumV2Indices < SumV1Indices) {
10270 return DAG.getCommutedVectorShuffle(*SVOp);
10271 } else if (SumV2Indices == SumV1Indices) {
10272 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10273 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10274 if (SVOp->getMask()[i] >= NumElements)
10275 NumV2OddIndices += i % 2;
10276 else if (SVOp->getMask()[i] >= 0)
10277 NumV1OddIndices += i % 2;
10278 if (NumV2OddIndices < NumV1OddIndices)
10279 return DAG.getCommutedVectorShuffle(*SVOp);
10284 // For each vector width, delegate to a specialized lowering routine.
10285 if (VT.getSizeInBits() == 128)
10286 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10288 if (VT.getSizeInBits() == 256)
10289 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10291 // Force AVX-512 vectors to be scalarized for now.
10292 // FIXME: Implement AVX-512 support!
10293 if (VT.getSizeInBits() == 512)
10294 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10296 llvm_unreachable("Unimplemented!");
10299 // This function assumes its argument is a BUILD_VECTOR of constants or
10300 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10302 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10303 unsigned &MaskValue) {
10305 unsigned NumElems = BuildVector->getNumOperands();
10306 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10307 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10308 unsigned NumElemsInLane = NumElems / NumLanes;
10310 // Blend for v16i16 should be symetric for the both lanes.
10311 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10312 SDValue EltCond = BuildVector->getOperand(i);
10313 SDValue SndLaneEltCond =
10314 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10316 int Lane1Cond = -1, Lane2Cond = -1;
10317 if (isa<ConstantSDNode>(EltCond))
10318 Lane1Cond = !isZero(EltCond);
10319 if (isa<ConstantSDNode>(SndLaneEltCond))
10320 Lane2Cond = !isZero(SndLaneEltCond);
10322 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10323 // Lane1Cond != 0, means we want the first argument.
10324 // Lane1Cond == 0, means we want the second argument.
10325 // The encoding of this argument is 0 for the first argument, 1
10326 // for the second. Therefore, invert the condition.
10327 MaskValue |= !Lane1Cond << i;
10328 else if (Lane1Cond < 0)
10329 MaskValue |= !Lane2Cond << i;
10336 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10337 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10338 const X86Subtarget *Subtarget,
10339 SelectionDAG &DAG) {
10340 SDValue Cond = Op.getOperand(0);
10341 SDValue LHS = Op.getOperand(1);
10342 SDValue RHS = Op.getOperand(2);
10344 MVT VT = Op.getSimpleValueType();
10346 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10348 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10350 // Only non-legal VSELECTs reach this lowering, convert those into generic
10351 // shuffles and re-use the shuffle lowering path for blends.
10352 SmallVector<int, 32> Mask;
10353 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10354 SDValue CondElt = CondBV->getOperand(i);
10356 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10358 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10361 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10362 // A vselect where all conditions and data are constants can be optimized into
10363 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10364 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10365 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10366 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10369 // Try to lower this to a blend-style vector shuffle. This can handle all
10370 // constant condition cases.
10371 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10374 // Variable blends are only legal from SSE4.1 onward.
10375 if (!Subtarget->hasSSE41())
10378 // Only some types will be legal on some subtargets. If we can emit a legal
10379 // VSELECT-matching blend, return Op, and but if we need to expand, return
10381 switch (Op.getSimpleValueType().SimpleTy) {
10383 // Most of the vector types have blends past SSE4.1.
10387 // The byte blends for AVX vectors were introduced only in AVX2.
10388 if (Subtarget->hasAVX2())
10395 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10396 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10399 // FIXME: We should custom lower this by fixing the condition and using i8
10405 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10406 MVT VT = Op.getSimpleValueType();
10409 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10412 if (VT.getSizeInBits() == 8) {
10413 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10414 Op.getOperand(0), Op.getOperand(1));
10415 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10416 DAG.getValueType(VT));
10417 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10420 if (VT.getSizeInBits() == 16) {
10421 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10422 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10424 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10425 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10426 DAG.getNode(ISD::BITCAST, dl,
10429 Op.getOperand(1)));
10430 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10431 Op.getOperand(0), Op.getOperand(1));
10432 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10433 DAG.getValueType(VT));
10434 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10437 if (VT == MVT::f32) {
10438 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10439 // the result back to FR32 register. It's only worth matching if the
10440 // result has a single use which is a store or a bitcast to i32. And in
10441 // the case of a store, it's not worth it if the index is a constant 0,
10442 // because a MOVSSmr can be used instead, which is smaller and faster.
10443 if (!Op.hasOneUse())
10445 SDNode *User = *Op.getNode()->use_begin();
10446 if ((User->getOpcode() != ISD::STORE ||
10447 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10448 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10449 (User->getOpcode() != ISD::BITCAST ||
10450 User->getValueType(0) != MVT::i32))
10452 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10453 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10456 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10459 if (VT == MVT::i32 || VT == MVT::i64) {
10460 // ExtractPS/pextrq works with constant index.
10461 if (isa<ConstantSDNode>(Op.getOperand(1)))
10467 /// Extract one bit from mask vector, like v16i1 or v8i1.
10468 /// AVX-512 feature.
10470 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10471 SDValue Vec = Op.getOperand(0);
10473 MVT VecVT = Vec.getSimpleValueType();
10474 SDValue Idx = Op.getOperand(1);
10475 MVT EltVT = Op.getSimpleValueType();
10477 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10478 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10479 "Unexpected vector type in ExtractBitFromMaskVector");
10481 // variable index can't be handled in mask registers,
10482 // extend vector to VR512
10483 if (!isa<ConstantSDNode>(Idx)) {
10484 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10485 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10486 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10487 ExtVT.getVectorElementType(), Ext, Idx);
10488 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10491 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10492 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10493 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10494 rc = getRegClassFor(MVT::v16i1);
10495 unsigned MaxSift = rc->getSize()*8 - 1;
10496 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10497 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10498 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10499 DAG.getConstant(MaxSift, dl, MVT::i8));
10500 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10501 DAG.getIntPtrConstant(0, dl));
10505 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10506 SelectionDAG &DAG) const {
10508 SDValue Vec = Op.getOperand(0);
10509 MVT VecVT = Vec.getSimpleValueType();
10510 SDValue Idx = Op.getOperand(1);
10512 if (Op.getSimpleValueType() == MVT::i1)
10513 return ExtractBitFromMaskVector(Op, DAG);
10515 if (!isa<ConstantSDNode>(Idx)) {
10516 if (VecVT.is512BitVector() ||
10517 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10518 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10521 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10522 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10523 MaskEltVT.getSizeInBits());
10525 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10526 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10527 getZeroVector(MaskVT, Subtarget, DAG, dl),
10528 Idx, DAG.getConstant(0, dl, getPointerTy()));
10529 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10530 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10531 Perm, DAG.getConstant(0, dl, getPointerTy()));
10536 // If this is a 256-bit vector result, first extract the 128-bit vector and
10537 // then extract the element from the 128-bit vector.
10538 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10540 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10541 // Get the 128-bit vector.
10542 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10543 MVT EltVT = VecVT.getVectorElementType();
10545 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10547 //if (IdxVal >= NumElems/2)
10548 // IdxVal -= NumElems/2;
10549 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10550 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10551 DAG.getConstant(IdxVal, dl, MVT::i32));
10554 assert(VecVT.is128BitVector() && "Unexpected vector length");
10556 if (Subtarget->hasSSE41()) {
10557 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10562 MVT VT = Op.getSimpleValueType();
10563 // TODO: handle v16i8.
10564 if (VT.getSizeInBits() == 16) {
10565 SDValue Vec = Op.getOperand(0);
10566 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10568 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10569 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10570 DAG.getNode(ISD::BITCAST, dl,
10572 Op.getOperand(1)));
10573 // Transform it so it match pextrw which produces a 32-bit result.
10574 MVT EltVT = MVT::i32;
10575 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10576 Op.getOperand(0), Op.getOperand(1));
10577 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10578 DAG.getValueType(VT));
10579 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10582 if (VT.getSizeInBits() == 32) {
10583 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10587 // SHUFPS the element to the lowest double word, then movss.
10588 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10589 MVT VVT = Op.getOperand(0).getSimpleValueType();
10590 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10591 DAG.getUNDEF(VVT), Mask);
10592 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10593 DAG.getIntPtrConstant(0, dl));
10596 if (VT.getSizeInBits() == 64) {
10597 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10598 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10599 // to match extract_elt for f64.
10600 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10604 // UNPCKHPD the element to the lowest double word, then movsd.
10605 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10606 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10607 int Mask[2] = { 1, -1 };
10608 MVT VVT = Op.getOperand(0).getSimpleValueType();
10609 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10610 DAG.getUNDEF(VVT), Mask);
10611 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10612 DAG.getIntPtrConstant(0, dl));
10618 /// Insert one bit to mask vector, like v16i1 or v8i1.
10619 /// AVX-512 feature.
10621 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10623 SDValue Vec = Op.getOperand(0);
10624 SDValue Elt = Op.getOperand(1);
10625 SDValue Idx = Op.getOperand(2);
10626 MVT VecVT = Vec.getSimpleValueType();
10628 if (!isa<ConstantSDNode>(Idx)) {
10629 // Non constant index. Extend source and destination,
10630 // insert element and then truncate the result.
10631 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10632 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10633 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10634 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10635 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10636 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10639 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10640 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10641 if (Vec.getOpcode() == ISD::UNDEF)
10642 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10643 DAG.getConstant(IdxVal, dl, MVT::i8));
10644 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10645 unsigned MaxSift = rc->getSize()*8 - 1;
10646 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10647 DAG.getConstant(MaxSift, dl, MVT::i8));
10648 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10649 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10650 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10653 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10654 SelectionDAG &DAG) const {
10655 MVT VT = Op.getSimpleValueType();
10656 MVT EltVT = VT.getVectorElementType();
10658 if (EltVT == MVT::i1)
10659 return InsertBitToMaskVector(Op, DAG);
10662 SDValue N0 = Op.getOperand(0);
10663 SDValue N1 = Op.getOperand(1);
10664 SDValue N2 = Op.getOperand(2);
10665 if (!isa<ConstantSDNode>(N2))
10667 auto *N2C = cast<ConstantSDNode>(N2);
10668 unsigned IdxVal = N2C->getZExtValue();
10670 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10671 // into that, and then insert the subvector back into the result.
10672 if (VT.is256BitVector() || VT.is512BitVector()) {
10673 // With a 256-bit vector, we can insert into the zero element efficiently
10674 // using a blend if we have AVX or AVX2 and the right data type.
10675 if (VT.is256BitVector() && IdxVal == 0) {
10676 // TODO: It is worthwhile to cast integer to floating point and back
10677 // and incur a domain crossing penalty if that's what we'll end up
10678 // doing anyway after extracting to a 128-bit vector.
10679 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
10680 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
10681 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
10682 N2 = DAG.getIntPtrConstant(1, dl);
10683 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
10687 // Get the desired 128-bit vector chunk.
10688 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10690 // Insert the element into the desired chunk.
10691 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10692 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10694 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10695 DAG.getConstant(IdxIn128, dl, MVT::i32));
10697 // Insert the changed part back into the bigger vector
10698 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10700 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10702 if (Subtarget->hasSSE41()) {
10703 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10705 if (VT == MVT::v8i16) {
10706 Opc = X86ISD::PINSRW;
10708 assert(VT == MVT::v16i8);
10709 Opc = X86ISD::PINSRB;
10712 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10714 if (N1.getValueType() != MVT::i32)
10715 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10716 if (N2.getValueType() != MVT::i32)
10717 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10718 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10721 if (EltVT == MVT::f32) {
10722 // Bits [7:6] of the constant are the source select. This will always be
10723 // zero here. The DAG Combiner may combine an extract_elt index into
10724 // these bits. For example (insert (extract, 3), 2) could be matched by
10725 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
10726 // Bits [5:4] of the constant are the destination select. This is the
10727 // value of the incoming immediate.
10728 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10729 // combine either bitwise AND or insert of float 0.0 to set these bits.
10731 const Function *F = DAG.getMachineFunction().getFunction();
10732 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
10733 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
10734 // If this is an insertion of 32-bits into the low 32-bits of
10735 // a vector, we prefer to generate a blend with immediate rather
10736 // than an insertps. Blends are simpler operations in hardware and so
10737 // will always have equal or better performance than insertps.
10738 // But if optimizing for size and there's a load folding opportunity,
10739 // generate insertps because blendps does not have a 32-bit memory
10741 N2 = DAG.getIntPtrConstant(1, dl);
10742 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10743 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
10745 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
10746 // Create this as a scalar to vector..
10747 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10748 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10751 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10752 // PINSR* works with constant index.
10757 if (EltVT == MVT::i8)
10760 if (EltVT.getSizeInBits() == 16) {
10761 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10762 // as its second argument.
10763 if (N1.getValueType() != MVT::i32)
10764 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10765 if (N2.getValueType() != MVT::i32)
10766 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10767 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10772 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10774 MVT OpVT = Op.getSimpleValueType();
10776 // If this is a 256-bit vector result, first insert into a 128-bit
10777 // vector and then insert into the 256-bit vector.
10778 if (!OpVT.is128BitVector()) {
10779 // Insert into a 128-bit vector.
10780 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10781 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10782 OpVT.getVectorNumElements() / SizeFactor);
10784 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10786 // Insert the 128-bit vector.
10787 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10790 if (OpVT == MVT::v1i64 &&
10791 Op.getOperand(0).getValueType() == MVT::i64)
10792 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10794 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10795 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10796 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10797 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10800 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10801 // a simple subregister reference or explicit instructions to grab
10802 // upper bits of a vector.
10803 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10804 SelectionDAG &DAG) {
10806 SDValue In = Op.getOperand(0);
10807 SDValue Idx = Op.getOperand(1);
10808 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10809 MVT ResVT = Op.getSimpleValueType();
10810 MVT InVT = In.getSimpleValueType();
10812 if (Subtarget->hasFp256()) {
10813 if (ResVT.is128BitVector() &&
10814 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10815 isa<ConstantSDNode>(Idx)) {
10816 return Extract128BitVector(In, IdxVal, DAG, dl);
10818 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10819 isa<ConstantSDNode>(Idx)) {
10820 return Extract256BitVector(In, IdxVal, DAG, dl);
10826 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10827 // simple superregister reference or explicit instructions to insert
10828 // the upper bits of a vector.
10829 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10830 SelectionDAG &DAG) {
10831 if (!Subtarget->hasAVX())
10835 SDValue Vec = Op.getOperand(0);
10836 SDValue SubVec = Op.getOperand(1);
10837 SDValue Idx = Op.getOperand(2);
10839 if (!isa<ConstantSDNode>(Idx))
10842 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10843 MVT OpVT = Op.getSimpleValueType();
10844 MVT SubVecVT = SubVec.getSimpleValueType();
10846 // Fold two 16-byte subvector loads into one 32-byte load:
10847 // (insert_subvector (insert_subvector undef, (load addr), 0),
10848 // (load addr + 16), Elts/2)
10850 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10851 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10852 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10853 !Subtarget->isUnalignedMem32Slow()) {
10854 SDValue SubVec2 = Vec.getOperand(1);
10855 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10856 if (Idx2->getZExtValue() == 0) {
10857 SDValue Ops[] = { SubVec2, SubVec };
10858 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10865 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10866 SubVecVT.is128BitVector())
10867 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10869 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10870 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10872 if (OpVT.getVectorElementType() == MVT::i1) {
10873 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
10875 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
10876 SDValue Undef = DAG.getUNDEF(OpVT);
10877 unsigned NumElems = OpVT.getVectorNumElements();
10878 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
10880 if (IdxVal == OpVT.getVectorNumElements() / 2) {
10881 // Zero upper bits of the Vec
10882 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10883 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10885 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10887 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10888 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10891 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10893 // Zero upper bits of the Vec2
10894 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10895 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
10896 // Zero lower bits of the Vec
10897 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10898 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10899 // Merge them together
10900 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10906 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10907 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10908 // one of the above mentioned nodes. It has to be wrapped because otherwise
10909 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10910 // be used to form addressing mode. These wrapped nodes will be selected
10913 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10914 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10916 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10917 // global base reg.
10918 unsigned char OpFlag = 0;
10919 unsigned WrapperKind = X86ISD::Wrapper;
10920 CodeModel::Model M = DAG.getTarget().getCodeModel();
10922 if (Subtarget->isPICStyleRIPRel() &&
10923 (M == CodeModel::Small || M == CodeModel::Kernel))
10924 WrapperKind = X86ISD::WrapperRIP;
10925 else if (Subtarget->isPICStyleGOT())
10926 OpFlag = X86II::MO_GOTOFF;
10927 else if (Subtarget->isPICStyleStubPIC())
10928 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10930 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10931 CP->getAlignment(),
10932 CP->getOffset(), OpFlag);
10934 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10935 // With PIC, the address is actually $g + Offset.
10937 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10938 DAG.getNode(X86ISD::GlobalBaseReg,
10939 SDLoc(), getPointerTy()),
10946 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10947 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10949 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10950 // global base reg.
10951 unsigned char OpFlag = 0;
10952 unsigned WrapperKind = X86ISD::Wrapper;
10953 CodeModel::Model M = DAG.getTarget().getCodeModel();
10955 if (Subtarget->isPICStyleRIPRel() &&
10956 (M == CodeModel::Small || M == CodeModel::Kernel))
10957 WrapperKind = X86ISD::WrapperRIP;
10958 else if (Subtarget->isPICStyleGOT())
10959 OpFlag = X86II::MO_GOTOFF;
10960 else if (Subtarget->isPICStyleStubPIC())
10961 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10963 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10966 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10968 // With PIC, the address is actually $g + Offset.
10970 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10971 DAG.getNode(X86ISD::GlobalBaseReg,
10972 SDLoc(), getPointerTy()),
10979 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10980 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10982 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10983 // global base reg.
10984 unsigned char OpFlag = 0;
10985 unsigned WrapperKind = X86ISD::Wrapper;
10986 CodeModel::Model M = DAG.getTarget().getCodeModel();
10988 if (Subtarget->isPICStyleRIPRel() &&
10989 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10990 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10991 OpFlag = X86II::MO_GOTPCREL;
10992 WrapperKind = X86ISD::WrapperRIP;
10993 } else if (Subtarget->isPICStyleGOT()) {
10994 OpFlag = X86II::MO_GOT;
10995 } else if (Subtarget->isPICStyleStubPIC()) {
10996 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10997 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10998 OpFlag = X86II::MO_DARWIN_NONLAZY;
11001 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
11004 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11006 // With PIC, the address is actually $g + Offset.
11007 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11008 !Subtarget->is64Bit()) {
11009 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11010 DAG.getNode(X86ISD::GlobalBaseReg,
11011 SDLoc(), getPointerTy()),
11015 // For symbols that require a load from a stub to get the address, emit the
11017 if (isGlobalStubReference(OpFlag))
11018 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
11019 MachinePointerInfo::getGOT(), false, false, false, 0);
11025 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11026 // Create the TargetBlockAddressAddress node.
11027 unsigned char OpFlags =
11028 Subtarget->ClassifyBlockAddressReference();
11029 CodeModel::Model M = DAG.getTarget().getCodeModel();
11030 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11031 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11033 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
11036 if (Subtarget->isPICStyleRIPRel() &&
11037 (M == CodeModel::Small || M == CodeModel::Kernel))
11038 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11040 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11042 // With PIC, the address is actually $g + Offset.
11043 if (isGlobalRelativeToPICBase(OpFlags)) {
11044 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11045 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11053 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11054 int64_t Offset, SelectionDAG &DAG) const {
11055 // Create the TargetGlobalAddress node, folding in the constant
11056 // offset if it is legal.
11057 unsigned char OpFlags =
11058 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11059 CodeModel::Model M = DAG.getTarget().getCodeModel();
11061 if (OpFlags == X86II::MO_NO_FLAG &&
11062 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11063 // A direct static reference to a global.
11064 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
11067 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
11070 if (Subtarget->isPICStyleRIPRel() &&
11071 (M == CodeModel::Small || M == CodeModel::Kernel))
11072 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11074 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11076 // With PIC, the address is actually $g + Offset.
11077 if (isGlobalRelativeToPICBase(OpFlags)) {
11078 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11079 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11083 // For globals that require a load from a stub to get the address, emit the
11085 if (isGlobalStubReference(OpFlags))
11086 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
11087 MachinePointerInfo::getGOT(), false, false, false, 0);
11089 // If there was a non-zero offset that we didn't fold, create an explicit
11090 // addition for it.
11092 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
11093 DAG.getConstant(Offset, dl, getPointerTy()));
11099 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11100 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11101 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11102 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11106 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11107 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11108 unsigned char OperandFlags, bool LocalDynamic = false) {
11109 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11110 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11112 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11113 GA->getValueType(0),
11117 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11121 SDValue Ops[] = { Chain, TGA, *InFlag };
11122 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11124 SDValue Ops[] = { Chain, TGA };
11125 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11128 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11129 MFI->setAdjustsStack(true);
11130 MFI->setHasCalls(true);
11132 SDValue Flag = Chain.getValue(1);
11133 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11136 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11138 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11141 SDLoc dl(GA); // ? function entry point might be better
11142 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11143 DAG.getNode(X86ISD::GlobalBaseReg,
11144 SDLoc(), PtrVT), InFlag);
11145 InFlag = Chain.getValue(1);
11147 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11150 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11152 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11154 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11155 X86::RAX, X86II::MO_TLSGD);
11158 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11164 // Get the start address of the TLS block for this module.
11165 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11166 .getInfo<X86MachineFunctionInfo>();
11167 MFI->incNumLocalDynamicTLSAccesses();
11171 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11172 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11175 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11176 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11177 InFlag = Chain.getValue(1);
11178 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11179 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11182 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11186 unsigned char OperandFlags = X86II::MO_DTPOFF;
11187 unsigned WrapperKind = X86ISD::Wrapper;
11188 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11189 GA->getValueType(0),
11190 GA->getOffset(), OperandFlags);
11191 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11193 // Add x@dtpoff with the base.
11194 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11197 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11198 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11199 const EVT PtrVT, TLSModel::Model model,
11200 bool is64Bit, bool isPIC) {
11203 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11204 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11205 is64Bit ? 257 : 256));
11207 SDValue ThreadPointer =
11208 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11209 MachinePointerInfo(Ptr), false, false, false, 0);
11211 unsigned char OperandFlags = 0;
11212 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11214 unsigned WrapperKind = X86ISD::Wrapper;
11215 if (model == TLSModel::LocalExec) {
11216 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11217 } else if (model == TLSModel::InitialExec) {
11219 OperandFlags = X86II::MO_GOTTPOFF;
11220 WrapperKind = X86ISD::WrapperRIP;
11222 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11225 llvm_unreachable("Unexpected model");
11228 // emit "addl x@ntpoff,%eax" (local exec)
11229 // or "addl x@indntpoff,%eax" (initial exec)
11230 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11232 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11233 GA->getOffset(), OperandFlags);
11234 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11236 if (model == TLSModel::InitialExec) {
11237 if (isPIC && !is64Bit) {
11238 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11239 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11243 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11244 MachinePointerInfo::getGOT(), false, false, false, 0);
11247 // The address of the thread local variable is the add of the thread
11248 // pointer with the offset of the variable.
11249 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11253 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11255 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11256 const GlobalValue *GV = GA->getGlobal();
11258 if (Subtarget->isTargetELF()) {
11259 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11262 case TLSModel::GeneralDynamic:
11263 if (Subtarget->is64Bit())
11264 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
11265 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
11266 case TLSModel::LocalDynamic:
11267 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
11268 Subtarget->is64Bit());
11269 case TLSModel::InitialExec:
11270 case TLSModel::LocalExec:
11271 return LowerToTLSExecModel(
11272 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
11273 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
11275 llvm_unreachable("Unknown TLS model.");
11278 if (Subtarget->isTargetDarwin()) {
11279 // Darwin only has one model of TLS. Lower to that.
11280 unsigned char OpFlag = 0;
11281 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11282 X86ISD::WrapperRIP : X86ISD::Wrapper;
11284 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11285 // global base reg.
11286 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11287 !Subtarget->is64Bit();
11289 OpFlag = X86II::MO_TLVP_PIC_BASE;
11291 OpFlag = X86II::MO_TLVP;
11293 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11294 GA->getValueType(0),
11295 GA->getOffset(), OpFlag);
11296 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11298 // With PIC32, the address is actually $g + Offset.
11300 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11301 DAG.getNode(X86ISD::GlobalBaseReg,
11302 SDLoc(), getPointerTy()),
11305 // Lowering the machine isd will make sure everything is in the right
11307 SDValue Chain = DAG.getEntryNode();
11308 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11309 SDValue Args[] = { Chain, Offset };
11310 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11312 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11313 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11314 MFI->setAdjustsStack(true);
11316 // And our return value (tls address) is in the standard call return value
11318 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11319 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11320 Chain.getValue(1));
11323 if (Subtarget->isTargetKnownWindowsMSVC() ||
11324 Subtarget->isTargetWindowsGNU()) {
11325 // Just use the implicit TLS architecture
11326 // Need to generate someting similar to:
11327 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11329 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11330 // mov rcx, qword [rdx+rcx*8]
11331 // mov eax, .tls$:tlsvar
11332 // [rax+rcx] contains the address
11333 // Windows 64bit: gs:0x58
11334 // Windows 32bit: fs:__tls_array
11337 SDValue Chain = DAG.getEntryNode();
11339 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11340 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11341 // use its literal value of 0x2C.
11342 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11343 ? Type::getInt8PtrTy(*DAG.getContext(),
11345 : Type::getInt32PtrTy(*DAG.getContext(),
11349 Subtarget->is64Bit()
11350 ? DAG.getIntPtrConstant(0x58, dl)
11351 : (Subtarget->isTargetWindowsGNU()
11352 ? DAG.getIntPtrConstant(0x2C, dl)
11353 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11355 SDValue ThreadPointer =
11356 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11357 MachinePointerInfo(Ptr), false, false, false, 0);
11359 // Load the _tls_index variable
11360 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11361 if (Subtarget->is64Bit())
11362 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
11363 IDX, MachinePointerInfo(), MVT::i32,
11364 false, false, false, 0);
11366 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11367 false, false, false, 0);
11369 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()), dl,
11371 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11373 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11374 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11375 false, false, false, 0);
11377 // Get the offset of start of .tls section
11378 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11379 GA->getValueType(0),
11380 GA->getOffset(), X86II::MO_SECREL);
11381 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11383 // The address of the thread local variable is the add of the thread
11384 // pointer with the offset of the variable.
11385 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11388 llvm_unreachable("TLS not implemented for this target.");
11391 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11392 /// and take a 2 x i32 value to shift plus a shift amount.
11393 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11394 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11395 MVT VT = Op.getSimpleValueType();
11396 unsigned VTBits = VT.getSizeInBits();
11398 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11399 SDValue ShOpLo = Op.getOperand(0);
11400 SDValue ShOpHi = Op.getOperand(1);
11401 SDValue ShAmt = Op.getOperand(2);
11402 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11403 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11405 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11406 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11407 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11408 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11409 : DAG.getConstant(0, dl, VT);
11411 SDValue Tmp2, Tmp3;
11412 if (Op.getOpcode() == ISD::SHL_PARTS) {
11413 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11414 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11416 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11417 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11420 // If the shift amount is larger or equal than the width of a part we can't
11421 // rely on the results of shld/shrd. Insert a test and select the appropriate
11422 // values for large shift amounts.
11423 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11424 DAG.getConstant(VTBits, dl, MVT::i8));
11425 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11426 AndNode, DAG.getConstant(0, dl, MVT::i8));
11429 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11430 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11431 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11433 if (Op.getOpcode() == ISD::SHL_PARTS) {
11434 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11435 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11437 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11438 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11441 SDValue Ops[2] = { Lo, Hi };
11442 return DAG.getMergeValues(Ops, dl);
11445 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11446 SelectionDAG &DAG) const {
11447 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11450 if (SrcVT.isVector()) {
11451 if (SrcVT.getVectorElementType() == MVT::i1) {
11452 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11453 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11454 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11455 Op.getOperand(0)));
11460 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11461 "Unknown SINT_TO_FP to lower!");
11463 // These are really Legal; return the operand so the caller accepts it as
11465 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11467 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11468 Subtarget->is64Bit()) {
11472 unsigned Size = SrcVT.getSizeInBits()/8;
11473 MachineFunction &MF = DAG.getMachineFunction();
11474 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11475 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11476 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11478 MachinePointerInfo::getFixedStack(SSFI),
11480 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11483 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11485 SelectionDAG &DAG) const {
11489 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11491 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11493 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11495 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11497 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11498 MachineMemOperand *MMO;
11500 int SSFI = FI->getIndex();
11502 DAG.getMachineFunction()
11503 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11504 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11506 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11507 StackSlot = StackSlot.getOperand(1);
11509 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11510 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11512 Tys, Ops, SrcVT, MMO);
11515 Chain = Result.getValue(1);
11516 SDValue InFlag = Result.getValue(2);
11518 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11519 // shouldn't be necessary except that RFP cannot be live across
11520 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11521 MachineFunction &MF = DAG.getMachineFunction();
11522 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11523 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11524 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11525 Tys = DAG.getVTList(MVT::Other);
11527 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11529 MachineMemOperand *MMO =
11530 DAG.getMachineFunction()
11531 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11532 MachineMemOperand::MOStore, SSFISize, SSFISize);
11534 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11535 Ops, Op.getValueType(), MMO);
11536 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11537 MachinePointerInfo::getFixedStack(SSFI),
11538 false, false, false, 0);
11544 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11545 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11546 SelectionDAG &DAG) const {
11547 // This algorithm is not obvious. Here it is what we're trying to output:
11550 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11551 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11553 haddpd %xmm0, %xmm0
11555 pshufd $0x4e, %xmm0, %xmm1
11561 LLVMContext *Context = DAG.getContext();
11563 // Build some magic constants.
11564 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11565 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11566 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11568 SmallVector<Constant*,2> CV1;
11570 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11571 APInt(64, 0x4330000000000000ULL))));
11573 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11574 APInt(64, 0x4530000000000000ULL))));
11575 Constant *C1 = ConstantVector::get(CV1);
11576 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11578 // Load the 64-bit value into an XMM register.
11579 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11581 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11582 MachinePointerInfo::getConstantPool(),
11583 false, false, false, 16);
11584 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11585 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11588 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11589 MachinePointerInfo::getConstantPool(),
11590 false, false, false, 16);
11591 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11592 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11595 if (Subtarget->hasSSE3()) {
11596 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11597 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11599 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11600 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11602 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11603 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11607 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11608 DAG.getIntPtrConstant(0, dl));
11611 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11612 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11613 SelectionDAG &DAG) const {
11615 // FP constant to bias correct the final result.
11616 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
11619 // Load the 32-bit value into an XMM register.
11620 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11623 // Zero out the upper parts of the register.
11624 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11626 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11627 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11628 DAG.getIntPtrConstant(0, dl));
11630 // Or the load with the bias.
11631 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11632 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11633 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11634 MVT::v2f64, Load)),
11635 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11636 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11637 MVT::v2f64, Bias)));
11638 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11639 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11640 DAG.getIntPtrConstant(0, dl));
11642 // Subtract the bias.
11643 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11645 // Handle final rounding.
11646 EVT DestVT = Op.getValueType();
11648 if (DestVT.bitsLT(MVT::f64))
11649 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11650 DAG.getIntPtrConstant(0, dl));
11651 if (DestVT.bitsGT(MVT::f64))
11652 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11654 // Handle final rounding.
11658 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11659 const X86Subtarget &Subtarget) {
11660 // The algorithm is the following:
11661 // #ifdef __SSE4_1__
11662 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11663 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11664 // (uint4) 0x53000000, 0xaa);
11666 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11667 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11669 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11670 // return (float4) lo + fhi;
11673 SDValue V = Op->getOperand(0);
11674 EVT VecIntVT = V.getValueType();
11675 bool Is128 = VecIntVT == MVT::v4i32;
11676 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11677 // If we convert to something else than the supported type, e.g., to v4f64,
11679 if (VecFloatVT != Op->getValueType(0))
11682 unsigned NumElts = VecIntVT.getVectorNumElements();
11683 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11684 "Unsupported custom type");
11685 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11687 // In the #idef/#else code, we have in common:
11688 // - The vector of constants:
11694 // Create the splat vector for 0x4b000000.
11695 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
11696 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11697 CstLow, CstLow, CstLow, CstLow};
11698 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11699 makeArrayRef(&CstLowArray[0], NumElts));
11700 // Create the splat vector for 0x53000000.
11701 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
11702 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11703 CstHigh, CstHigh, CstHigh, CstHigh};
11704 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11705 makeArrayRef(&CstHighArray[0], NumElts));
11707 // Create the right shift.
11708 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
11709 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11710 CstShift, CstShift, CstShift, CstShift};
11711 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11712 makeArrayRef(&CstShiftArray[0], NumElts));
11713 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11716 if (Subtarget.hasSSE41()) {
11717 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11718 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11719 SDValue VecCstLowBitcast =
11720 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11721 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11722 // Low will be bitcasted right away, so do not bother bitcasting back to its
11724 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11725 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11726 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11727 // (uint4) 0x53000000, 0xaa);
11728 SDValue VecCstHighBitcast =
11729 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11730 SDValue VecShiftBitcast =
11731 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11732 // High will be bitcasted right away, so do not bother bitcasting back to
11733 // its original type.
11734 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11735 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11737 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
11738 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11739 CstMask, CstMask, CstMask);
11740 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11741 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11742 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11744 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11745 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11748 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11749 SDValue CstFAdd = DAG.getConstantFP(
11750 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
11751 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11752 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11753 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11754 makeArrayRef(&CstFAddArray[0], NumElts));
11756 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11757 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11759 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11760 // return (float4) lo + fhi;
11761 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11762 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11765 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11766 SelectionDAG &DAG) const {
11767 SDValue N0 = Op.getOperand(0);
11768 MVT SVT = N0.getSimpleValueType();
11771 switch (SVT.SimpleTy) {
11773 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11778 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11779 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11780 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11784 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11786 llvm_unreachable(nullptr);
11789 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11790 SelectionDAG &DAG) const {
11791 SDValue N0 = Op.getOperand(0);
11794 if (Op.getValueType().isVector())
11795 return lowerUINT_TO_FP_vec(Op, DAG);
11797 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11798 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11799 // the optimization here.
11800 if (DAG.SignBitIsZero(N0))
11801 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11803 MVT SrcVT = N0.getSimpleValueType();
11804 MVT DstVT = Op.getSimpleValueType();
11805 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11806 return LowerUINT_TO_FP_i64(Op, DAG);
11807 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11808 return LowerUINT_TO_FP_i32(Op, DAG);
11809 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11812 // Make a 64-bit buffer, and use it to build an FILD.
11813 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11814 if (SrcVT == MVT::i32) {
11815 SDValue WordOff = DAG.getConstant(4, dl, getPointerTy());
11816 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11817 getPointerTy(), StackSlot, WordOff);
11818 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11819 StackSlot, MachinePointerInfo(),
11821 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
11822 OffsetSlot, MachinePointerInfo(),
11824 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11828 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11829 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11830 StackSlot, MachinePointerInfo(),
11832 // For i64 source, we need to add the appropriate power of 2 if the input
11833 // was negative. This is the same as the optimization in
11834 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11835 // we must be careful to do the computation in x87 extended precision, not
11836 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11837 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11838 MachineMemOperand *MMO =
11839 DAG.getMachineFunction()
11840 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11841 MachineMemOperand::MOLoad, 8, 8);
11843 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11844 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11845 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11848 APInt FF(32, 0x5F800000ULL);
11850 // Check whether the sign bit is set.
11851 SDValue SignSet = DAG.getSetCC(dl,
11852 getSetCCResultType(*DAG.getContext(), MVT::i64),
11854 DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
11856 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11857 SDValue FudgePtr = DAG.getConstantPool(
11858 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11861 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11862 SDValue Zero = DAG.getIntPtrConstant(0, dl);
11863 SDValue Four = DAG.getIntPtrConstant(4, dl);
11864 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11866 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11868 // Load the value out, extending it from f32 to f80.
11869 // FIXME: Avoid the extend by constructing the right constant pool?
11870 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11871 FudgePtr, MachinePointerInfo::getConstantPool(),
11872 MVT::f32, false, false, false, 4);
11873 // Extend everything to 80 bits to force it to be done on x87.
11874 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11875 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
11876 DAG.getIntPtrConstant(0, dl));
11879 std::pair<SDValue,SDValue>
11880 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11881 bool IsSigned, bool IsReplace) const {
11884 EVT DstTy = Op.getValueType();
11886 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11887 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11891 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11892 DstTy.getSimpleVT() >= MVT::i16 &&
11893 "Unknown FP_TO_INT to lower!");
11895 // These are really Legal.
11896 if (DstTy == MVT::i32 &&
11897 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11898 return std::make_pair(SDValue(), SDValue());
11899 if (Subtarget->is64Bit() &&
11900 DstTy == MVT::i64 &&
11901 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11902 return std::make_pair(SDValue(), SDValue());
11904 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11905 // stack slot, or into the FTOL runtime function.
11906 MachineFunction &MF = DAG.getMachineFunction();
11907 unsigned MemSize = DstTy.getSizeInBits()/8;
11908 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11909 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11912 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11913 Opc = X86ISD::WIN_FTOL;
11915 switch (DstTy.getSimpleVT().SimpleTy) {
11916 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11917 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11918 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11919 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11922 SDValue Chain = DAG.getEntryNode();
11923 SDValue Value = Op.getOperand(0);
11924 EVT TheVT = Op.getOperand(0).getValueType();
11925 // FIXME This causes a redundant load/store if the SSE-class value is already
11926 // in memory, such as if it is on the callstack.
11927 if (isScalarFPTypeInSSEReg(TheVT)) {
11928 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11929 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11930 MachinePointerInfo::getFixedStack(SSFI),
11932 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11934 Chain, StackSlot, DAG.getValueType(TheVT)
11937 MachineMemOperand *MMO =
11938 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11939 MachineMemOperand::MOLoad, MemSize, MemSize);
11940 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11941 Chain = Value.getValue(1);
11942 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11943 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11946 MachineMemOperand *MMO =
11947 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11948 MachineMemOperand::MOStore, MemSize, MemSize);
11950 if (Opc != X86ISD::WIN_FTOL) {
11951 // Build the FP_TO_INT*_IN_MEM
11952 SDValue Ops[] = { Chain, Value, StackSlot };
11953 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11955 return std::make_pair(FIST, StackSlot);
11957 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11958 DAG.getVTList(MVT::Other, MVT::Glue),
11960 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11961 MVT::i32, ftol.getValue(1));
11962 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11963 MVT::i32, eax.getValue(2));
11964 SDValue Ops[] = { eax, edx };
11965 SDValue pair = IsReplace
11966 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11967 : DAG.getMergeValues(Ops, DL);
11968 return std::make_pair(pair, SDValue());
11972 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11973 const X86Subtarget *Subtarget) {
11974 MVT VT = Op->getSimpleValueType(0);
11975 SDValue In = Op->getOperand(0);
11976 MVT InVT = In.getSimpleValueType();
11979 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
11980 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
11982 // Optimize vectors in AVX mode:
11985 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11986 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11987 // Concat upper and lower parts.
11990 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11991 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11992 // Concat upper and lower parts.
11995 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11996 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11997 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12000 if (Subtarget->hasInt256())
12001 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12003 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12004 SDValue Undef = DAG.getUNDEF(InVT);
12005 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12006 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12007 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12009 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12010 VT.getVectorNumElements()/2);
12012 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
12013 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
12015 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12018 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12019 SelectionDAG &DAG) {
12020 MVT VT = Op->getSimpleValueType(0);
12021 SDValue In = Op->getOperand(0);
12022 MVT InVT = In.getSimpleValueType();
12024 unsigned int NumElts = VT.getVectorNumElements();
12025 if (NumElts != 8 && NumElts != 16)
12028 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12029 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12031 assert(InVT.getVectorElementType() == MVT::i1);
12032 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12034 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12036 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12038 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12039 if (VT.is512BitVector())
12041 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12044 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12045 SelectionDAG &DAG) {
12046 if (Subtarget->hasFp256()) {
12047 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12055 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12056 SelectionDAG &DAG) {
12058 MVT VT = Op.getSimpleValueType();
12059 SDValue In = Op.getOperand(0);
12060 MVT SVT = In.getSimpleValueType();
12062 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12063 return LowerZERO_EXTEND_AVX512(Op, DAG);
12065 if (Subtarget->hasFp256()) {
12066 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12071 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12072 VT.getVectorNumElements() != SVT.getVectorNumElements());
12076 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12078 MVT VT = Op.getSimpleValueType();
12079 SDValue In = Op.getOperand(0);
12080 MVT InVT = In.getSimpleValueType();
12082 if (VT == MVT::i1) {
12083 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12084 "Invalid scalar TRUNCATE operation");
12085 if (InVT.getSizeInBits() >= 32)
12087 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12088 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12090 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12091 "Invalid TRUNCATE operation");
12093 // move vector to mask - truncate solution for SKX
12094 if (VT.getVectorElementType() == MVT::i1) {
12095 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12096 Subtarget->hasBWI())
12097 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12098 if ((InVT.is256BitVector() || InVT.is128BitVector())
12099 && InVT.getScalarSizeInBits() <= 16 &&
12100 Subtarget->hasBWI() && Subtarget->hasVLX())
12101 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12102 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12103 Subtarget->hasDQI())
12104 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12105 if ((InVT.is256BitVector() || InVT.is128BitVector())
12106 && InVT.getScalarSizeInBits() >= 32 &&
12107 Subtarget->hasDQI() && Subtarget->hasVLX())
12108 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12110 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
12111 if (VT.getVectorElementType().getSizeInBits() >=8)
12112 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12114 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12115 unsigned NumElts = InVT.getVectorNumElements();
12116 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12117 if (InVT.getSizeInBits() < 512) {
12118 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12119 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12124 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12125 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12126 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12129 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12130 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12131 if (Subtarget->hasInt256()) {
12132 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12133 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
12134 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12136 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12137 DAG.getIntPtrConstant(0, DL));
12140 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12141 DAG.getIntPtrConstant(0, DL));
12142 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12143 DAG.getIntPtrConstant(2, DL));
12144 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12145 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12146 static const int ShufMask[] = {0, 2, 4, 6};
12147 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12150 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12151 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12152 if (Subtarget->hasInt256()) {
12153 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
12155 SmallVector<SDValue,32> pshufbMask;
12156 for (unsigned i = 0; i < 2; ++i) {
12157 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12158 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12159 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12160 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12161 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12162 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12163 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12164 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12165 for (unsigned j = 0; j < 8; ++j)
12166 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12168 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12169 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12170 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
12172 static const int ShufMask[] = {0, 2, -1, -1};
12173 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12175 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12176 DAG.getIntPtrConstant(0, DL));
12177 return DAG.getNode(ISD::BITCAST, DL, VT, In);
12180 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12181 DAG.getIntPtrConstant(0, DL));
12183 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12184 DAG.getIntPtrConstant(4, DL));
12186 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
12187 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
12189 // The PSHUFB mask:
12190 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12191 -1, -1, -1, -1, -1, -1, -1, -1};
12193 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12194 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12195 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12197 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12198 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12200 // The MOVLHPS Mask:
12201 static const int ShufMask2[] = {0, 1, 4, 5};
12202 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12203 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
12206 // Handle truncation of V256 to V128 using shuffles.
12207 if (!VT.is128BitVector() || !InVT.is256BitVector())
12210 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12212 unsigned NumElems = VT.getVectorNumElements();
12213 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12215 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12216 // Prepare truncation shuffle mask
12217 for (unsigned i = 0; i != NumElems; ++i)
12218 MaskVec[i] = i * 2;
12219 SDValue V = DAG.getVectorShuffle(NVT, DL,
12220 DAG.getNode(ISD::BITCAST, DL, NVT, In),
12221 DAG.getUNDEF(NVT), &MaskVec[0]);
12222 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12223 DAG.getIntPtrConstant(0, DL));
12226 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12227 SelectionDAG &DAG) const {
12228 assert(!Op.getSimpleValueType().isVector());
12230 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12231 /*IsSigned=*/ true, /*IsReplace=*/ false);
12232 SDValue FIST = Vals.first, StackSlot = Vals.second;
12233 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12234 if (!FIST.getNode()) return Op;
12236 if (StackSlot.getNode())
12237 // Load the result.
12238 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12239 FIST, StackSlot, MachinePointerInfo(),
12240 false, false, false, 0);
12242 // The node is the result.
12246 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12247 SelectionDAG &DAG) const {
12248 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12249 /*IsSigned=*/ false, /*IsReplace=*/ false);
12250 SDValue FIST = Vals.first, StackSlot = Vals.second;
12251 assert(FIST.getNode() && "Unexpected failure");
12253 if (StackSlot.getNode())
12254 // Load the result.
12255 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12256 FIST, StackSlot, MachinePointerInfo(),
12257 false, false, false, 0);
12259 // The node is the result.
12263 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12265 MVT VT = Op.getSimpleValueType();
12266 SDValue In = Op.getOperand(0);
12267 MVT SVT = In.getSimpleValueType();
12269 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12271 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12272 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12273 In, DAG.getUNDEF(SVT)));
12276 /// The only differences between FABS and FNEG are the mask and the logic op.
12277 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12278 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12279 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12280 "Wrong opcode for lowering FABS or FNEG.");
12282 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12284 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12285 // into an FNABS. We'll lower the FABS after that if it is still in use.
12287 for (SDNode *User : Op->uses())
12288 if (User->getOpcode() == ISD::FNEG)
12291 SDValue Op0 = Op.getOperand(0);
12292 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12295 MVT VT = Op.getSimpleValueType();
12296 // Assume scalar op for initialization; update for vector if needed.
12297 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
12298 // generate a 16-byte vector constant and logic op even for the scalar case.
12299 // Using a 16-byte mask allows folding the load of the mask with
12300 // the logic op, so it can save (~4 bytes) on code size.
12302 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
12303 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12304 // decide if we should generate a 16-byte constant mask when we only need 4 or
12305 // 8 bytes for the scalar case.
12306 if (VT.isVector()) {
12307 EltVT = VT.getVectorElementType();
12308 NumElts = VT.getVectorNumElements();
12311 unsigned EltBits = EltVT.getSizeInBits();
12312 LLVMContext *Context = DAG.getContext();
12313 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12315 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12316 Constant *C = ConstantInt::get(*Context, MaskElt);
12317 C = ConstantVector::getSplat(NumElts, C);
12318 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12319 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
12320 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12321 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12322 MachinePointerInfo::getConstantPool(),
12323 false, false, false, Alignment);
12325 if (VT.isVector()) {
12326 // For a vector, cast operands to a vector type, perform the logic op,
12327 // and cast the result back to the original value type.
12328 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12329 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12330 SDValue Operand = IsFNABS ?
12331 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12332 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12333 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12334 return DAG.getNode(ISD::BITCAST, dl, VT,
12335 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12338 // If not vector, then scalar.
12339 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12340 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12341 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12344 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12345 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12346 LLVMContext *Context = DAG.getContext();
12347 SDValue Op0 = Op.getOperand(0);
12348 SDValue Op1 = Op.getOperand(1);
12350 MVT VT = Op.getSimpleValueType();
12351 MVT SrcVT = Op1.getSimpleValueType();
12353 // If second operand is smaller, extend it first.
12354 if (SrcVT.bitsLT(VT)) {
12355 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12358 // And if it is bigger, shrink it first.
12359 if (SrcVT.bitsGT(VT)) {
12360 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12364 // At this point the operands and the result should have the same
12365 // type, and that won't be f80 since that is not custom lowered.
12367 const fltSemantics &Sem =
12368 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12369 const unsigned SizeInBits = VT.getSizeInBits();
12371 SmallVector<Constant *, 4> CV(
12372 VT == MVT::f64 ? 2 : 4,
12373 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12375 // First, clear all bits but the sign bit from the second operand (sign).
12376 CV[0] = ConstantFP::get(*Context,
12377 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12378 Constant *C = ConstantVector::get(CV);
12379 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12380 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12381 MachinePointerInfo::getConstantPool(),
12382 false, false, false, 16);
12383 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12385 // Next, clear the sign bit from the first operand (magnitude).
12386 // If it's a constant, we can clear it here.
12387 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12388 APFloat APF = Op0CN->getValueAPF();
12389 // If the magnitude is a positive zero, the sign bit alone is enough.
12390 if (APF.isPosZero())
12393 CV[0] = ConstantFP::get(*Context, APF);
12395 CV[0] = ConstantFP::get(
12397 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12399 C = ConstantVector::get(CV);
12400 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12401 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12402 MachinePointerInfo::getConstantPool(),
12403 false, false, false, 16);
12404 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12405 if (!isa<ConstantFPSDNode>(Op0))
12406 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12408 // OR the magnitude value with the sign bit.
12409 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12412 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12413 SDValue N0 = Op.getOperand(0);
12415 MVT VT = Op.getSimpleValueType();
12417 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12418 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12419 DAG.getConstant(1, dl, VT));
12420 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12423 // Check whether an OR'd tree is PTEST-able.
12424 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12425 SelectionDAG &DAG) {
12426 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12428 if (!Subtarget->hasSSE41())
12431 if (!Op->hasOneUse())
12434 SDNode *N = Op.getNode();
12437 SmallVector<SDValue, 8> Opnds;
12438 DenseMap<SDValue, unsigned> VecInMap;
12439 SmallVector<SDValue, 8> VecIns;
12440 EVT VT = MVT::Other;
12442 // Recognize a special case where a vector is casted into wide integer to
12444 Opnds.push_back(N->getOperand(0));
12445 Opnds.push_back(N->getOperand(1));
12447 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12448 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12449 // BFS traverse all OR'd operands.
12450 if (I->getOpcode() == ISD::OR) {
12451 Opnds.push_back(I->getOperand(0));
12452 Opnds.push_back(I->getOperand(1));
12453 // Re-evaluate the number of nodes to be traversed.
12454 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12458 // Quit if a non-EXTRACT_VECTOR_ELT
12459 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12462 // Quit if without a constant index.
12463 SDValue Idx = I->getOperand(1);
12464 if (!isa<ConstantSDNode>(Idx))
12467 SDValue ExtractedFromVec = I->getOperand(0);
12468 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12469 if (M == VecInMap.end()) {
12470 VT = ExtractedFromVec.getValueType();
12471 // Quit if not 128/256-bit vector.
12472 if (!VT.is128BitVector() && !VT.is256BitVector())
12474 // Quit if not the same type.
12475 if (VecInMap.begin() != VecInMap.end() &&
12476 VT != VecInMap.begin()->first.getValueType())
12478 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12479 VecIns.push_back(ExtractedFromVec);
12481 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12484 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12485 "Not extracted from 128-/256-bit vector.");
12487 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12489 for (DenseMap<SDValue, unsigned>::const_iterator
12490 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12491 // Quit if not all elements are used.
12492 if (I->second != FullMask)
12496 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12498 // Cast all vectors into TestVT for PTEST.
12499 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12500 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12502 // If more than one full vectors are evaluated, OR them first before PTEST.
12503 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12504 // Each iteration will OR 2 nodes and append the result until there is only
12505 // 1 node left, i.e. the final OR'd value of all vectors.
12506 SDValue LHS = VecIns[Slot];
12507 SDValue RHS = VecIns[Slot + 1];
12508 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12511 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12512 VecIns.back(), VecIns.back());
12515 /// \brief return true if \c Op has a use that doesn't just read flags.
12516 static bool hasNonFlagsUse(SDValue Op) {
12517 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12519 SDNode *User = *UI;
12520 unsigned UOpNo = UI.getOperandNo();
12521 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12522 // Look pass truncate.
12523 UOpNo = User->use_begin().getOperandNo();
12524 User = *User->use_begin();
12527 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12528 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12534 /// Emit nodes that will be selected as "test Op0,Op0", or something
12536 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12537 SelectionDAG &DAG) const {
12538 if (Op.getValueType() == MVT::i1) {
12539 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12540 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12541 DAG.getConstant(0, dl, MVT::i8));
12543 // CF and OF aren't always set the way we want. Determine which
12544 // of these we need.
12545 bool NeedCF = false;
12546 bool NeedOF = false;
12549 case X86::COND_A: case X86::COND_AE:
12550 case X86::COND_B: case X86::COND_BE:
12553 case X86::COND_G: case X86::COND_GE:
12554 case X86::COND_L: case X86::COND_LE:
12555 case X86::COND_O: case X86::COND_NO: {
12556 // Check if we really need to set the
12557 // Overflow flag. If NoSignedWrap is present
12558 // that is not actually needed.
12559 switch (Op->getOpcode()) {
12564 const BinaryWithFlagsSDNode *BinNode =
12565 cast<BinaryWithFlagsSDNode>(Op.getNode());
12566 if (BinNode->Flags.hasNoSignedWrap())
12576 // See if we can use the EFLAGS value from the operand instead of
12577 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12578 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12579 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12580 // Emit a CMP with 0, which is the TEST pattern.
12581 //if (Op.getValueType() == MVT::i1)
12582 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12583 // DAG.getConstant(0, MVT::i1));
12584 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12585 DAG.getConstant(0, dl, Op.getValueType()));
12587 unsigned Opcode = 0;
12588 unsigned NumOperands = 0;
12590 // Truncate operations may prevent the merge of the SETCC instruction
12591 // and the arithmetic instruction before it. Attempt to truncate the operands
12592 // of the arithmetic instruction and use a reduced bit-width instruction.
12593 bool NeedTruncation = false;
12594 SDValue ArithOp = Op;
12595 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12596 SDValue Arith = Op->getOperand(0);
12597 // Both the trunc and the arithmetic op need to have one user each.
12598 if (Arith->hasOneUse())
12599 switch (Arith.getOpcode()) {
12606 NeedTruncation = true;
12612 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12613 // which may be the result of a CAST. We use the variable 'Op', which is the
12614 // non-casted variable when we check for possible users.
12615 switch (ArithOp.getOpcode()) {
12617 // Due to an isel shortcoming, be conservative if this add is likely to be
12618 // selected as part of a load-modify-store instruction. When the root node
12619 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12620 // uses of other nodes in the match, such as the ADD in this case. This
12621 // leads to the ADD being left around and reselected, with the result being
12622 // two adds in the output. Alas, even if none our users are stores, that
12623 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12624 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12625 // climbing the DAG back to the root, and it doesn't seem to be worth the
12627 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12628 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12629 if (UI->getOpcode() != ISD::CopyToReg &&
12630 UI->getOpcode() != ISD::SETCC &&
12631 UI->getOpcode() != ISD::STORE)
12634 if (ConstantSDNode *C =
12635 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12636 // An add of one will be selected as an INC.
12637 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12638 Opcode = X86ISD::INC;
12643 // An add of negative one (subtract of one) will be selected as a DEC.
12644 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12645 Opcode = X86ISD::DEC;
12651 // Otherwise use a regular EFLAGS-setting add.
12652 Opcode = X86ISD::ADD;
12657 // If we have a constant logical shift that's only used in a comparison
12658 // against zero turn it into an equivalent AND. This allows turning it into
12659 // a TEST instruction later.
12660 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12661 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12662 EVT VT = Op.getValueType();
12663 unsigned BitWidth = VT.getSizeInBits();
12664 unsigned ShAmt = Op->getConstantOperandVal(1);
12665 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12667 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12668 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12669 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12670 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12672 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12673 DAG.getConstant(Mask, dl, VT));
12674 DAG.ReplaceAllUsesWith(Op, New);
12680 // If the primary and result isn't used, don't bother using X86ISD::AND,
12681 // because a TEST instruction will be better.
12682 if (!hasNonFlagsUse(Op))
12688 // Due to the ISEL shortcoming noted above, be conservative if this op is
12689 // likely to be selected as part of a load-modify-store instruction.
12690 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12691 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12692 if (UI->getOpcode() == ISD::STORE)
12695 // Otherwise use a regular EFLAGS-setting instruction.
12696 switch (ArithOp.getOpcode()) {
12697 default: llvm_unreachable("unexpected operator!");
12698 case ISD::SUB: Opcode = X86ISD::SUB; break;
12699 case ISD::XOR: Opcode = X86ISD::XOR; break;
12700 case ISD::AND: Opcode = X86ISD::AND; break;
12702 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12703 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12704 if (EFLAGS.getNode())
12707 Opcode = X86ISD::OR;
12721 return SDValue(Op.getNode(), 1);
12727 // If we found that truncation is beneficial, perform the truncation and
12729 if (NeedTruncation) {
12730 EVT VT = Op.getValueType();
12731 SDValue WideVal = Op->getOperand(0);
12732 EVT WideVT = WideVal.getValueType();
12733 unsigned ConvertedOp = 0;
12734 // Use a target machine opcode to prevent further DAGCombine
12735 // optimizations that may separate the arithmetic operations
12736 // from the setcc node.
12737 switch (WideVal.getOpcode()) {
12739 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12740 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12741 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12742 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12743 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12747 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12748 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12749 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12750 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12751 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12757 // Emit a CMP with 0, which is the TEST pattern.
12758 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12759 DAG.getConstant(0, dl, Op.getValueType()));
12761 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12762 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12764 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12765 DAG.ReplaceAllUsesWith(Op, New);
12766 return SDValue(New.getNode(), 1);
12769 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12771 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12772 SDLoc dl, SelectionDAG &DAG) const {
12773 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12774 if (C->getAPIntValue() == 0)
12775 return EmitTest(Op0, X86CC, dl, DAG);
12777 if (Op0.getValueType() == MVT::i1)
12778 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12781 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12782 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12783 // Do the comparison at i32 if it's smaller, besides the Atom case.
12784 // This avoids subregister aliasing issues. Keep the smaller reference
12785 // if we're optimizing for size, however, as that'll allow better folding
12786 // of memory operations.
12787 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12788 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12789 Attribute::MinSize) &&
12790 !Subtarget->isAtom()) {
12791 unsigned ExtendOp =
12792 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12793 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12794 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12796 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12797 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12798 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12800 return SDValue(Sub.getNode(), 1);
12802 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12805 /// Convert a comparison if required by the subtarget.
12806 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12807 SelectionDAG &DAG) const {
12808 // If the subtarget does not support the FUCOMI instruction, floating-point
12809 // comparisons have to be converted.
12810 if (Subtarget->hasCMov() ||
12811 Cmp.getOpcode() != X86ISD::CMP ||
12812 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12813 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12816 // The instruction selector will select an FUCOM instruction instead of
12817 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12818 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12819 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12821 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12822 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12823 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12824 DAG.getConstant(8, dl, MVT::i8));
12825 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12826 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12829 /// The minimum architected relative accuracy is 2^-12. We need one
12830 /// Newton-Raphson step to have a good float result (24 bits of precision).
12831 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12832 DAGCombinerInfo &DCI,
12833 unsigned &RefinementSteps,
12834 bool &UseOneConstNR) const {
12835 // FIXME: We should use instruction latency models to calculate the cost of
12836 // each potential sequence, but this is very hard to do reliably because
12837 // at least Intel's Core* chips have variable timing based on the number of
12838 // significant digits in the divisor and/or sqrt operand.
12839 if (!Subtarget->useSqrtEst())
12842 EVT VT = Op.getValueType();
12844 // SSE1 has rsqrtss and rsqrtps.
12845 // TODO: Add support for AVX512 (v16f32).
12846 // It is likely not profitable to do this for f64 because a double-precision
12847 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12848 // instructions: convert to single, rsqrtss, convert back to double, refine
12849 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12850 // along with FMA, this could be a throughput win.
12851 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12852 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12853 RefinementSteps = 1;
12854 UseOneConstNR = false;
12855 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12860 /// The minimum architected relative accuracy is 2^-12. We need one
12861 /// Newton-Raphson step to have a good float result (24 bits of precision).
12862 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12863 DAGCombinerInfo &DCI,
12864 unsigned &RefinementSteps) const {
12865 // FIXME: We should use instruction latency models to calculate the cost of
12866 // each potential sequence, but this is very hard to do reliably because
12867 // at least Intel's Core* chips have variable timing based on the number of
12868 // significant digits in the divisor.
12869 if (!Subtarget->useReciprocalEst())
12872 EVT VT = Op.getValueType();
12874 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12875 // TODO: Add support for AVX512 (v16f32).
12876 // It is likely not profitable to do this for f64 because a double-precision
12877 // reciprocal estimate with refinement on x86 prior to FMA requires
12878 // 15 instructions: convert to single, rcpss, convert back to double, refine
12879 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12880 // along with FMA, this could be a throughput win.
12881 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12882 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12883 RefinementSteps = ReciprocalEstimateRefinementSteps;
12884 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12889 /// If we have at least two divisions that use the same divisor, convert to
12890 /// multplication by a reciprocal. This may need to be adjusted for a given
12891 /// CPU if a division's cost is not at least twice the cost of a multiplication.
12892 /// This is because we still need one division to calculate the reciprocal and
12893 /// then we need two multiplies by that reciprocal as replacements for the
12894 /// original divisions.
12895 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
12896 return NumUsers > 1;
12899 static bool isAllOnes(SDValue V) {
12900 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12901 return C && C->isAllOnesValue();
12904 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12905 /// if it's possible.
12906 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12907 SDLoc dl, SelectionDAG &DAG) const {
12908 SDValue Op0 = And.getOperand(0);
12909 SDValue Op1 = And.getOperand(1);
12910 if (Op0.getOpcode() == ISD::TRUNCATE)
12911 Op0 = Op0.getOperand(0);
12912 if (Op1.getOpcode() == ISD::TRUNCATE)
12913 Op1 = Op1.getOperand(0);
12916 if (Op1.getOpcode() == ISD::SHL)
12917 std::swap(Op0, Op1);
12918 if (Op0.getOpcode() == ISD::SHL) {
12919 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12920 if (And00C->getZExtValue() == 1) {
12921 // If we looked past a truncate, check that it's only truncating away
12923 unsigned BitWidth = Op0.getValueSizeInBits();
12924 unsigned AndBitWidth = And.getValueSizeInBits();
12925 if (BitWidth > AndBitWidth) {
12927 DAG.computeKnownBits(Op0, Zeros, Ones);
12928 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12932 RHS = Op0.getOperand(1);
12934 } else if (Op1.getOpcode() == ISD::Constant) {
12935 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
12936 uint64_t AndRHSVal = AndRHS->getZExtValue();
12937 SDValue AndLHS = Op0;
12939 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
12940 LHS = AndLHS.getOperand(0);
12941 RHS = AndLHS.getOperand(1);
12944 // Use BT if the immediate can't be encoded in a TEST instruction.
12945 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
12947 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
12951 if (LHS.getNode()) {
12952 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12953 // instruction. Since the shift amount is in-range-or-undefined, we know
12954 // that doing a bittest on the i32 value is ok. We extend to i32 because
12955 // the encoding for the i16 version is larger than the i32 version.
12956 // Also promote i16 to i32 for performance / code size reason.
12957 if (LHS.getValueType() == MVT::i8 ||
12958 LHS.getValueType() == MVT::i16)
12959 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12961 // If the operand types disagree, extend the shift amount to match. Since
12962 // BT ignores high bits (like shifts) we can use anyextend.
12963 if (LHS.getValueType() != RHS.getValueType())
12964 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12966 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12967 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12968 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12969 DAG.getConstant(Cond, dl, MVT::i8), BT);
12975 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12977 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12982 // SSE Condition code mapping:
12991 switch (SetCCOpcode) {
12992 default: llvm_unreachable("Unexpected SETCC condition");
12994 case ISD::SETEQ: SSECC = 0; break;
12996 case ISD::SETGT: Swap = true; // Fallthrough
12998 case ISD::SETOLT: SSECC = 1; break;
13000 case ISD::SETGE: Swap = true; // Fallthrough
13002 case ISD::SETOLE: SSECC = 2; break;
13003 case ISD::SETUO: SSECC = 3; break;
13005 case ISD::SETNE: SSECC = 4; break;
13006 case ISD::SETULE: Swap = true; // Fallthrough
13007 case ISD::SETUGE: SSECC = 5; break;
13008 case ISD::SETULT: Swap = true; // Fallthrough
13009 case ISD::SETUGT: SSECC = 6; break;
13010 case ISD::SETO: SSECC = 7; break;
13012 case ISD::SETONE: SSECC = 8; break;
13015 std::swap(Op0, Op1);
13020 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13021 // ones, and then concatenate the result back.
13022 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13023 MVT VT = Op.getSimpleValueType();
13025 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13026 "Unsupported value type for operation");
13028 unsigned NumElems = VT.getVectorNumElements();
13030 SDValue CC = Op.getOperand(2);
13032 // Extract the LHS vectors
13033 SDValue LHS = Op.getOperand(0);
13034 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13035 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13037 // Extract the RHS vectors
13038 SDValue RHS = Op.getOperand(1);
13039 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13040 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13042 // Issue the operation on the smaller types and concatenate the result back
13043 MVT EltVT = VT.getVectorElementType();
13044 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13045 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13046 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13047 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13050 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13051 SDValue Op0 = Op.getOperand(0);
13052 SDValue Op1 = Op.getOperand(1);
13053 SDValue CC = Op.getOperand(2);
13054 MVT VT = Op.getSimpleValueType();
13057 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13058 "Unexpected type for boolean compare operation");
13059 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13060 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13061 DAG.getConstant(-1, dl, VT));
13062 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13063 DAG.getConstant(-1, dl, VT));
13064 switch (SetCCOpcode) {
13065 default: llvm_unreachable("Unexpected SETCC condition");
13067 // (x != y) -> ~(x ^ y)
13068 return DAG.getNode(ISD::XOR, dl, VT,
13069 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13070 DAG.getConstant(-1, dl, VT));
13072 // (x == y) -> (x ^ y)
13073 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13076 // (x > y) -> (x & ~y)
13077 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13080 // (x < y) -> (~x & y)
13081 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13084 // (x <= y) -> (~x | y)
13085 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13088 // (x >=y) -> (x | ~y)
13089 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13093 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13094 const X86Subtarget *Subtarget) {
13095 SDValue Op0 = Op.getOperand(0);
13096 SDValue Op1 = Op.getOperand(1);
13097 SDValue CC = Op.getOperand(2);
13098 MVT VT = Op.getSimpleValueType();
13101 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13102 Op.getValueType().getScalarType() == MVT::i1 &&
13103 "Cannot set masked compare for this operation");
13105 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13107 bool Unsigned = false;
13110 switch (SetCCOpcode) {
13111 default: llvm_unreachable("Unexpected SETCC condition");
13112 case ISD::SETNE: SSECC = 4; break;
13113 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13114 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13115 case ISD::SETLT: Swap = true; //fall-through
13116 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13117 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13118 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13119 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13120 case ISD::SETULE: Unsigned = true; //fall-through
13121 case ISD::SETLE: SSECC = 2; break;
13125 std::swap(Op0, Op1);
13127 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13128 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13129 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13130 DAG.getConstant(SSECC, dl, MVT::i8));
13133 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13134 /// operand \p Op1. If non-trivial (for example because it's not constant)
13135 /// return an empty value.
13136 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13138 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13142 MVT VT = Op1.getSimpleValueType();
13143 MVT EVT = VT.getVectorElementType();
13144 unsigned n = VT.getVectorNumElements();
13145 SmallVector<SDValue, 8> ULTOp1;
13147 for (unsigned i = 0; i < n; ++i) {
13148 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13149 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13152 // Avoid underflow.
13153 APInt Val = Elt->getAPIntValue();
13157 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13160 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13163 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13164 SelectionDAG &DAG) {
13165 SDValue Op0 = Op.getOperand(0);
13166 SDValue Op1 = Op.getOperand(1);
13167 SDValue CC = Op.getOperand(2);
13168 MVT VT = Op.getSimpleValueType();
13169 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13170 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13175 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13176 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13179 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13180 unsigned Opc = X86ISD::CMPP;
13181 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13182 assert(VT.getVectorNumElements() <= 16);
13183 Opc = X86ISD::CMPM;
13185 // In the two special cases we can't handle, emit two comparisons.
13188 unsigned CombineOpc;
13189 if (SetCCOpcode == ISD::SETUEQ) {
13190 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13192 assert(SetCCOpcode == ISD::SETONE);
13193 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13196 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13197 DAG.getConstant(CC0, dl, MVT::i8));
13198 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13199 DAG.getConstant(CC1, dl, MVT::i8));
13200 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13202 // Handle all other FP comparisons here.
13203 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13204 DAG.getConstant(SSECC, dl, MVT::i8));
13207 // Break 256-bit integer vector compare into smaller ones.
13208 if (VT.is256BitVector() && !Subtarget->hasInt256())
13209 return Lower256IntVSETCC(Op, DAG);
13211 EVT OpVT = Op1.getValueType();
13212 if (OpVT.getVectorElementType() == MVT::i1)
13213 return LowerBoolVSETCC_AVX512(Op, DAG);
13215 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13216 if (Subtarget->hasAVX512()) {
13217 if (Op1.getValueType().is512BitVector() ||
13218 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13219 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13220 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13222 // In AVX-512 architecture setcc returns mask with i1 elements,
13223 // But there is no compare instruction for i8 and i16 elements in KNL.
13224 // We are not talking about 512-bit operands in this case, these
13225 // types are illegal.
13227 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13228 OpVT.getVectorElementType().getSizeInBits() >= 8))
13229 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13230 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13233 // We are handling one of the integer comparisons here. Since SSE only has
13234 // GT and EQ comparisons for integer, swapping operands and multiple
13235 // operations may be required for some comparisons.
13237 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13238 bool Subus = false;
13240 switch (SetCCOpcode) {
13241 default: llvm_unreachable("Unexpected SETCC condition");
13242 case ISD::SETNE: Invert = true;
13243 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13244 case ISD::SETLT: Swap = true;
13245 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13246 case ISD::SETGE: Swap = true;
13247 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13248 Invert = true; break;
13249 case ISD::SETULT: Swap = true;
13250 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13251 FlipSigns = true; break;
13252 case ISD::SETUGE: Swap = true;
13253 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13254 FlipSigns = true; Invert = true; break;
13257 // Special case: Use min/max operations for SETULE/SETUGE
13258 MVT VET = VT.getVectorElementType();
13260 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13261 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13264 switch (SetCCOpcode) {
13266 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
13267 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
13270 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13273 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13274 if (!MinMax && hasSubus) {
13275 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13277 // t = psubus Op0, Op1
13278 // pcmpeq t, <0..0>
13279 switch (SetCCOpcode) {
13281 case ISD::SETULT: {
13282 // If the comparison is against a constant we can turn this into a
13283 // setule. With psubus, setule does not require a swap. This is
13284 // beneficial because the constant in the register is no longer
13285 // destructed as the destination so it can be hoisted out of a loop.
13286 // Only do this pre-AVX since vpcmp* is no longer destructive.
13287 if (Subtarget->hasAVX())
13289 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13290 if (ULEOp1.getNode()) {
13292 Subus = true; Invert = false; Swap = false;
13296 // Psubus is better than flip-sign because it requires no inversion.
13297 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13298 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13302 Opc = X86ISD::SUBUS;
13308 std::swap(Op0, Op1);
13310 // Check that the operation in question is available (most are plain SSE2,
13311 // but PCMPGTQ and PCMPEQQ have different requirements).
13312 if (VT == MVT::v2i64) {
13313 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13314 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13316 // First cast everything to the right type.
13317 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13318 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13320 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13321 // bits of the inputs before performing those operations. The lower
13322 // compare is always unsigned.
13325 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13327 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13328 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13329 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13330 Sign, Zero, Sign, Zero);
13332 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13333 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13335 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13336 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13337 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13339 // Create masks for only the low parts/high parts of the 64 bit integers.
13340 static const int MaskHi[] = { 1, 1, 3, 3 };
13341 static const int MaskLo[] = { 0, 0, 2, 2 };
13342 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13343 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13344 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13346 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13347 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13350 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13352 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13355 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13356 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13357 // pcmpeqd + pshufd + pand.
13358 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13360 // First cast everything to the right type.
13361 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13362 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13365 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13367 // Make sure the lower and upper halves are both all-ones.
13368 static const int Mask[] = { 1, 0, 3, 2 };
13369 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13370 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13373 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13375 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13379 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13380 // bits of the inputs before performing those operations.
13382 EVT EltVT = VT.getVectorElementType();
13383 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13385 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13386 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13389 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13391 // If the logical-not of the result is required, perform that now.
13393 Result = DAG.getNOT(dl, Result, VT);
13396 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13399 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13400 getZeroVector(VT, Subtarget, DAG, dl));
13405 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13407 MVT VT = Op.getSimpleValueType();
13409 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13411 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13412 && "SetCC type must be 8-bit or 1-bit integer");
13413 SDValue Op0 = Op.getOperand(0);
13414 SDValue Op1 = Op.getOperand(1);
13416 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13418 // Optimize to BT if possible.
13419 // Lower (X & (1 << N)) == 0 to BT(X, N).
13420 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13421 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13422 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13423 Op1.getOpcode() == ISD::Constant &&
13424 cast<ConstantSDNode>(Op1)->isNullValue() &&
13425 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13426 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13427 if (NewSetCC.getNode()) {
13429 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13434 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13436 if (Op1.getOpcode() == ISD::Constant &&
13437 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13438 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13439 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13441 // If the input is a setcc, then reuse the input setcc or use a new one with
13442 // the inverted condition.
13443 if (Op0.getOpcode() == X86ISD::SETCC) {
13444 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13445 bool Invert = (CC == ISD::SETNE) ^
13446 cast<ConstantSDNode>(Op1)->isNullValue();
13450 CCode = X86::GetOppositeBranchCondition(CCode);
13451 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13452 DAG.getConstant(CCode, dl, MVT::i8),
13453 Op0.getOperand(1));
13455 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13459 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13460 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13461 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13463 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13464 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13467 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13468 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13469 if (X86CC == X86::COND_INVALID)
13472 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13473 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13474 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13475 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13477 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13481 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13482 static bool isX86LogicalCmp(SDValue Op) {
13483 unsigned Opc = Op.getNode()->getOpcode();
13484 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13485 Opc == X86ISD::SAHF)
13487 if (Op.getResNo() == 1 &&
13488 (Opc == X86ISD::ADD ||
13489 Opc == X86ISD::SUB ||
13490 Opc == X86ISD::ADC ||
13491 Opc == X86ISD::SBB ||
13492 Opc == X86ISD::SMUL ||
13493 Opc == X86ISD::UMUL ||
13494 Opc == X86ISD::INC ||
13495 Opc == X86ISD::DEC ||
13496 Opc == X86ISD::OR ||
13497 Opc == X86ISD::XOR ||
13498 Opc == X86ISD::AND))
13501 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13507 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13508 if (V.getOpcode() != ISD::TRUNCATE)
13511 SDValue VOp0 = V.getOperand(0);
13512 unsigned InBits = VOp0.getValueSizeInBits();
13513 unsigned Bits = V.getValueSizeInBits();
13514 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13517 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13518 bool addTest = true;
13519 SDValue Cond = Op.getOperand(0);
13520 SDValue Op1 = Op.getOperand(1);
13521 SDValue Op2 = Op.getOperand(2);
13523 EVT VT = Op1.getValueType();
13526 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13527 // are available or VBLENDV if AVX is available.
13528 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13529 if (Cond.getOpcode() == ISD::SETCC &&
13530 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13531 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13532 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13533 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13534 int SSECC = translateX86FSETCC(
13535 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13538 if (Subtarget->hasAVX512()) {
13539 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13540 DAG.getConstant(SSECC, DL, MVT::i8));
13541 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13544 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13545 DAG.getConstant(SSECC, DL, MVT::i8));
13547 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13548 // of 3 logic instructions for size savings and potentially speed.
13549 // Unfortunately, there is no scalar form of VBLENDV.
13551 // If either operand is a constant, don't try this. We can expect to
13552 // optimize away at least one of the logic instructions later in that
13553 // case, so that sequence would be faster than a variable blend.
13555 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13556 // uses XMM0 as the selection register. That may need just as many
13557 // instructions as the AND/ANDN/OR sequence due to register moves, so
13560 if (Subtarget->hasAVX() &&
13561 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13563 // Convert to vectors, do a VSELECT, and convert back to scalar.
13564 // All of the conversions should be optimized away.
13566 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13567 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13568 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13569 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13571 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13572 VCmp = DAG.getNode(ISD::BITCAST, DL, VCmpVT, VCmp);
13574 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13576 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13577 VSel, DAG.getIntPtrConstant(0, DL));
13579 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13580 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13581 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13585 if (Cond.getOpcode() == ISD::SETCC) {
13586 SDValue NewCond = LowerSETCC(Cond, DAG);
13587 if (NewCond.getNode())
13591 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13592 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13593 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13594 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13595 if (Cond.getOpcode() == X86ISD::SETCC &&
13596 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13597 isZero(Cond.getOperand(1).getOperand(1))) {
13598 SDValue Cmp = Cond.getOperand(1);
13600 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13602 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13603 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13604 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13606 SDValue CmpOp0 = Cmp.getOperand(0);
13607 // Apply further optimizations for special cases
13608 // (select (x != 0), -1, 0) -> neg & sbb
13609 // (select (x == 0), 0, -1) -> neg & sbb
13610 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13611 if (YC->isNullValue() &&
13612 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13613 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13614 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13615 DAG.getConstant(0, DL,
13616 CmpOp0.getValueType()),
13618 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13619 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13620 SDValue(Neg.getNode(), 1));
13624 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13625 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
13626 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13628 SDValue Res = // Res = 0 or -1.
13629 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13630 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
13632 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13633 Res = DAG.getNOT(DL, Res, Res.getValueType());
13635 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13636 if (!N2C || !N2C->isNullValue())
13637 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13642 // Look past (and (setcc_carry (cmp ...)), 1).
13643 if (Cond.getOpcode() == ISD::AND &&
13644 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13645 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13646 if (C && C->getAPIntValue() == 1)
13647 Cond = Cond.getOperand(0);
13650 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13651 // setting operand in place of the X86ISD::SETCC.
13652 unsigned CondOpcode = Cond.getOpcode();
13653 if (CondOpcode == X86ISD::SETCC ||
13654 CondOpcode == X86ISD::SETCC_CARRY) {
13655 CC = Cond.getOperand(0);
13657 SDValue Cmp = Cond.getOperand(1);
13658 unsigned Opc = Cmp.getOpcode();
13659 MVT VT = Op.getSimpleValueType();
13661 bool IllegalFPCMov = false;
13662 if (VT.isFloatingPoint() && !VT.isVector() &&
13663 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13664 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13666 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13667 Opc == X86ISD::BT) { // FIXME
13671 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13672 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13673 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13674 Cond.getOperand(0).getValueType() != MVT::i8)) {
13675 SDValue LHS = Cond.getOperand(0);
13676 SDValue RHS = Cond.getOperand(1);
13677 unsigned X86Opcode;
13680 switch (CondOpcode) {
13681 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13682 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13683 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13684 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13685 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13686 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13687 default: llvm_unreachable("unexpected overflowing operator");
13689 if (CondOpcode == ISD::UMULO)
13690 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13693 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13695 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13697 if (CondOpcode == ISD::UMULO)
13698 Cond = X86Op.getValue(2);
13700 Cond = X86Op.getValue(1);
13702 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
13707 // Look pass the truncate if the high bits are known zero.
13708 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13709 Cond = Cond.getOperand(0);
13711 // We know the result of AND is compared against zero. Try to match
13713 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13714 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13715 if (NewSetCC.getNode()) {
13716 CC = NewSetCC.getOperand(0);
13717 Cond = NewSetCC.getOperand(1);
13724 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
13725 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13728 // a < b ? -1 : 0 -> RES = ~setcc_carry
13729 // a < b ? 0 : -1 -> RES = setcc_carry
13730 // a >= b ? -1 : 0 -> RES = setcc_carry
13731 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13732 if (Cond.getOpcode() == X86ISD::SUB) {
13733 Cond = ConvertCmpIfNecessary(Cond, DAG);
13734 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13736 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13737 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13738 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13739 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13741 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13742 return DAG.getNOT(DL, Res, Res.getValueType());
13747 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13748 // widen the cmov and push the truncate through. This avoids introducing a new
13749 // branch during isel and doesn't add any extensions.
13750 if (Op.getValueType() == MVT::i8 &&
13751 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13752 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13753 if (T1.getValueType() == T2.getValueType() &&
13754 // Blacklist CopyFromReg to avoid partial register stalls.
13755 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13756 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13757 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13758 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13762 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13763 // condition is true.
13764 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13765 SDValue Ops[] = { Op2, Op1, CC, Cond };
13766 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13769 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13770 SelectionDAG &DAG) {
13771 MVT VT = Op->getSimpleValueType(0);
13772 SDValue In = Op->getOperand(0);
13773 MVT InVT = In.getSimpleValueType();
13774 MVT VTElt = VT.getVectorElementType();
13775 MVT InVTElt = InVT.getVectorElementType();
13779 if ((InVTElt == MVT::i1) &&
13780 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13781 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13783 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13784 VTElt.getSizeInBits() <= 16)) ||
13786 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13787 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13789 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13790 VTElt.getSizeInBits() >= 32))))
13791 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13793 unsigned int NumElts = VT.getVectorNumElements();
13795 if (NumElts != 8 && NumElts != 16)
13798 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13799 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13800 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13801 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13804 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13805 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13807 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
13810 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
13812 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
13813 if (VT.is512BitVector())
13815 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
13818 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13819 SelectionDAG &DAG) {
13820 MVT VT = Op->getSimpleValueType(0);
13821 SDValue In = Op->getOperand(0);
13822 MVT InVT = In.getSimpleValueType();
13825 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13826 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13828 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13829 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13830 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13833 if (Subtarget->hasInt256())
13834 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13836 // Optimize vectors in AVX mode
13837 // Sign extend v8i16 to v8i32 and
13840 // Divide input vector into two parts
13841 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13842 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13843 // concat the vectors to original VT
13845 unsigned NumElems = InVT.getVectorNumElements();
13846 SDValue Undef = DAG.getUNDEF(InVT);
13848 SmallVector<int,8> ShufMask1(NumElems, -1);
13849 for (unsigned i = 0; i != NumElems/2; ++i)
13852 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13854 SmallVector<int,8> ShufMask2(NumElems, -1);
13855 for (unsigned i = 0; i != NumElems/2; ++i)
13856 ShufMask2[i] = i + NumElems/2;
13858 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13860 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13861 VT.getVectorNumElements()/2);
13863 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13864 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13866 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13869 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13870 // may emit an illegal shuffle but the expansion is still better than scalar
13871 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13872 // we'll emit a shuffle and a arithmetic shift.
13873 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13874 // TODO: It is possible to support ZExt by zeroing the undef values during
13875 // the shuffle phase or after the shuffle.
13876 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13877 SelectionDAG &DAG) {
13878 MVT RegVT = Op.getSimpleValueType();
13879 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13880 assert(RegVT.isInteger() &&
13881 "We only custom lower integer vector sext loads.");
13883 // Nothing useful we can do without SSE2 shuffles.
13884 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13886 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13888 EVT MemVT = Ld->getMemoryVT();
13889 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13890 unsigned RegSz = RegVT.getSizeInBits();
13892 ISD::LoadExtType Ext = Ld->getExtensionType();
13894 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13895 && "Only anyext and sext are currently implemented.");
13896 assert(MemVT != RegVT && "Cannot extend to the same type");
13897 assert(MemVT.isVector() && "Must load a vector from memory");
13899 unsigned NumElems = RegVT.getVectorNumElements();
13900 unsigned MemSz = MemVT.getSizeInBits();
13901 assert(RegSz > MemSz && "Register size must be greater than the mem size");
13903 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
13904 // The only way in which we have a legal 256-bit vector result but not the
13905 // integer 256-bit operations needed to directly lower a sextload is if we
13906 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
13907 // a 128-bit vector and a normal sign_extend to 256-bits that should get
13908 // correctly legalized. We do this late to allow the canonical form of
13909 // sextload to persist throughout the rest of the DAG combiner -- it wants
13910 // to fold together any extensions it can, and so will fuse a sign_extend
13911 // of an sextload into a sextload targeting a wider value.
13913 if (MemSz == 128) {
13914 // Just switch this to a normal load.
13915 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
13916 "it must be a legal 128-bit vector "
13918 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
13919 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
13920 Ld->isInvariant(), Ld->getAlignment());
13922 assert(MemSz < 128 &&
13923 "Can't extend a type wider than 128 bits to a 256 bit vector!");
13924 // Do an sext load to a 128-bit vector type. We want to use the same
13925 // number of elements, but elements half as wide. This will end up being
13926 // recursively lowered by this routine, but will succeed as we definitely
13927 // have all the necessary features if we're using AVX1.
13929 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
13930 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
13932 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
13933 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
13934 Ld->isNonTemporal(), Ld->isInvariant(),
13935 Ld->getAlignment());
13938 // Replace chain users with the new chain.
13939 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
13940 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
13942 // Finally, do a normal sign-extend to the desired register.
13943 return DAG.getSExtOrTrunc(Load, dl, RegVT);
13946 // All sizes must be a power of two.
13947 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
13948 "Non-power-of-two elements are not custom lowered!");
13950 // Attempt to load the original value using scalar loads.
13951 // Find the largest scalar type that divides the total loaded size.
13952 MVT SclrLoadTy = MVT::i8;
13953 for (MVT Tp : MVT::integer_valuetypes()) {
13954 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
13959 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
13960 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
13962 SclrLoadTy = MVT::f64;
13964 // Calculate the number of scalar loads that we need to perform
13965 // in order to load our vector from memory.
13966 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
13968 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
13969 "Can only lower sext loads with a single scalar load!");
13971 unsigned loadRegZize = RegSz;
13972 if (Ext == ISD::SEXTLOAD && RegSz == 256)
13975 // Represent our vector as a sequence of elements which are the
13976 // largest scalar that we can load.
13977 EVT LoadUnitVecVT = EVT::getVectorVT(
13978 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
13980 // Represent the data using the same element type that is stored in
13981 // memory. In practice, we ''widen'' MemVT.
13983 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
13984 loadRegZize / MemVT.getScalarType().getSizeInBits());
13986 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
13987 "Invalid vector type");
13989 // We can't shuffle using an illegal type.
13990 assert(TLI.isTypeLegal(WideVecVT) &&
13991 "We only lower types that form legal widened vector types");
13993 SmallVector<SDValue, 8> Chains;
13994 SDValue Ptr = Ld->getBasePtr();
13995 SDValue Increment =
13996 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl, TLI.getPointerTy());
13997 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
13999 for (unsigned i = 0; i < NumLoads; ++i) {
14000 // Perform a single load.
14001 SDValue ScalarLoad =
14002 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14003 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14004 Ld->getAlignment());
14005 Chains.push_back(ScalarLoad.getValue(1));
14006 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14007 // another round of DAGCombining.
14009 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14011 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14012 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14014 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14017 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14019 // Bitcast the loaded value to a vector of the original element type, in
14020 // the size of the target vector type.
14021 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
14022 unsigned SizeRatio = RegSz / MemSz;
14024 if (Ext == ISD::SEXTLOAD) {
14025 // If we have SSE4.1, we can directly emit a VSEXT node.
14026 if (Subtarget->hasSSE41()) {
14027 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14028 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14032 // Otherwise we'll shuffle the small elements in the high bits of the
14033 // larger type and perform an arithmetic shift. If the shift is not legal
14034 // it's better to scalarize.
14035 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14036 "We can't implement a sext load without an arithmetic right shift!");
14038 // Redistribute the loaded elements into the different locations.
14039 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14040 for (unsigned i = 0; i != NumElems; ++i)
14041 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14043 SDValue Shuff = DAG.getVectorShuffle(
14044 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14046 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14048 // Build the arithmetic shift.
14049 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14050 MemVT.getVectorElementType().getSizeInBits();
14052 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14053 DAG.getConstant(Amt, dl, RegVT));
14055 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14059 // Redistribute the loaded elements into the different locations.
14060 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14061 for (unsigned i = 0; i != NumElems; ++i)
14062 ShuffleVec[i * SizeRatio] = i;
14064 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14065 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14067 // Bitcast to the requested type.
14068 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14069 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14073 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14074 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14075 // from the AND / OR.
14076 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14077 Opc = Op.getOpcode();
14078 if (Opc != ISD::OR && Opc != ISD::AND)
14080 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14081 Op.getOperand(0).hasOneUse() &&
14082 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14083 Op.getOperand(1).hasOneUse());
14086 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14087 // 1 and that the SETCC node has a single use.
14088 static bool isXor1OfSetCC(SDValue Op) {
14089 if (Op.getOpcode() != ISD::XOR)
14091 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14092 if (N1C && N1C->getAPIntValue() == 1) {
14093 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14094 Op.getOperand(0).hasOneUse();
14099 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14100 bool addTest = true;
14101 SDValue Chain = Op.getOperand(0);
14102 SDValue Cond = Op.getOperand(1);
14103 SDValue Dest = Op.getOperand(2);
14106 bool Inverted = false;
14108 if (Cond.getOpcode() == ISD::SETCC) {
14109 // Check for setcc([su]{add,sub,mul}o == 0).
14110 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14111 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14112 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14113 Cond.getOperand(0).getResNo() == 1 &&
14114 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14115 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14116 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14117 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14118 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14119 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14121 Cond = Cond.getOperand(0);
14123 SDValue NewCond = LowerSETCC(Cond, DAG);
14124 if (NewCond.getNode())
14129 // FIXME: LowerXALUO doesn't handle these!!
14130 else if (Cond.getOpcode() == X86ISD::ADD ||
14131 Cond.getOpcode() == X86ISD::SUB ||
14132 Cond.getOpcode() == X86ISD::SMUL ||
14133 Cond.getOpcode() == X86ISD::UMUL)
14134 Cond = LowerXALUO(Cond, DAG);
14137 // Look pass (and (setcc_carry (cmp ...)), 1).
14138 if (Cond.getOpcode() == ISD::AND &&
14139 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14140 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14141 if (C && C->getAPIntValue() == 1)
14142 Cond = Cond.getOperand(0);
14145 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14146 // setting operand in place of the X86ISD::SETCC.
14147 unsigned CondOpcode = Cond.getOpcode();
14148 if (CondOpcode == X86ISD::SETCC ||
14149 CondOpcode == X86ISD::SETCC_CARRY) {
14150 CC = Cond.getOperand(0);
14152 SDValue Cmp = Cond.getOperand(1);
14153 unsigned Opc = Cmp.getOpcode();
14154 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14155 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14159 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14163 // These can only come from an arithmetic instruction with overflow,
14164 // e.g. SADDO, UADDO.
14165 Cond = Cond.getNode()->getOperand(1);
14171 CondOpcode = Cond.getOpcode();
14172 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14173 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14174 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14175 Cond.getOperand(0).getValueType() != MVT::i8)) {
14176 SDValue LHS = Cond.getOperand(0);
14177 SDValue RHS = Cond.getOperand(1);
14178 unsigned X86Opcode;
14181 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14182 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14184 switch (CondOpcode) {
14185 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14187 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14189 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14192 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14193 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14195 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14197 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14200 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14201 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14202 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14203 default: llvm_unreachable("unexpected overflowing operator");
14206 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14207 if (CondOpcode == ISD::UMULO)
14208 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14211 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14213 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14215 if (CondOpcode == ISD::UMULO)
14216 Cond = X86Op.getValue(2);
14218 Cond = X86Op.getValue(1);
14220 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14224 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14225 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14226 if (CondOpc == ISD::OR) {
14227 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14228 // two branches instead of an explicit OR instruction with a
14230 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14231 isX86LogicalCmp(Cmp)) {
14232 CC = Cond.getOperand(0).getOperand(0);
14233 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14234 Chain, Dest, CC, Cmp);
14235 CC = Cond.getOperand(1).getOperand(0);
14239 } else { // ISD::AND
14240 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14241 // two branches instead of an explicit AND instruction with a
14242 // separate test. However, we only do this if this block doesn't
14243 // have a fall-through edge, because this requires an explicit
14244 // jmp when the condition is false.
14245 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14246 isX86LogicalCmp(Cmp) &&
14247 Op.getNode()->hasOneUse()) {
14248 X86::CondCode CCode =
14249 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14250 CCode = X86::GetOppositeBranchCondition(CCode);
14251 CC = DAG.getConstant(CCode, dl, MVT::i8);
14252 SDNode *User = *Op.getNode()->use_begin();
14253 // Look for an unconditional branch following this conditional branch.
14254 // We need this because we need to reverse the successors in order
14255 // to implement FCMP_OEQ.
14256 if (User->getOpcode() == ISD::BR) {
14257 SDValue FalseBB = User->getOperand(1);
14259 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14260 assert(NewBR == User);
14264 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14265 Chain, Dest, CC, Cmp);
14266 X86::CondCode CCode =
14267 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14268 CCode = X86::GetOppositeBranchCondition(CCode);
14269 CC = DAG.getConstant(CCode, dl, MVT::i8);
14275 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14276 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14277 // It should be transformed during dag combiner except when the condition
14278 // is set by a arithmetics with overflow node.
14279 X86::CondCode CCode =
14280 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14281 CCode = X86::GetOppositeBranchCondition(CCode);
14282 CC = DAG.getConstant(CCode, dl, MVT::i8);
14283 Cond = Cond.getOperand(0).getOperand(1);
14285 } else if (Cond.getOpcode() == ISD::SETCC &&
14286 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14287 // For FCMP_OEQ, we can emit
14288 // two branches instead of an explicit AND instruction with a
14289 // separate test. However, we only do this if this block doesn't
14290 // have a fall-through edge, because this requires an explicit
14291 // jmp when the condition is false.
14292 if (Op.getNode()->hasOneUse()) {
14293 SDNode *User = *Op.getNode()->use_begin();
14294 // Look for an unconditional branch following this conditional branch.
14295 // We need this because we need to reverse the successors in order
14296 // to implement FCMP_OEQ.
14297 if (User->getOpcode() == ISD::BR) {
14298 SDValue FalseBB = User->getOperand(1);
14300 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14301 assert(NewBR == User);
14305 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14306 Cond.getOperand(0), Cond.getOperand(1));
14307 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14308 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14309 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14310 Chain, Dest, CC, Cmp);
14311 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14316 } else if (Cond.getOpcode() == ISD::SETCC &&
14317 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14318 // For FCMP_UNE, we can emit
14319 // two branches instead of an explicit AND instruction with a
14320 // separate test. However, we only do this if this block doesn't
14321 // have a fall-through edge, because this requires an explicit
14322 // jmp when the condition is false.
14323 if (Op.getNode()->hasOneUse()) {
14324 SDNode *User = *Op.getNode()->use_begin();
14325 // Look for an unconditional branch following this conditional branch.
14326 // We need this because we need to reverse the successors in order
14327 // to implement FCMP_UNE.
14328 if (User->getOpcode() == ISD::BR) {
14329 SDValue FalseBB = User->getOperand(1);
14331 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14332 assert(NewBR == User);
14335 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14336 Cond.getOperand(0), Cond.getOperand(1));
14337 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14338 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14339 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14340 Chain, Dest, CC, Cmp);
14341 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14351 // Look pass the truncate if the high bits are known zero.
14352 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14353 Cond = Cond.getOperand(0);
14355 // We know the result of AND is compared against zero. Try to match
14357 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14358 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14359 if (NewSetCC.getNode()) {
14360 CC = NewSetCC.getOperand(0);
14361 Cond = NewSetCC.getOperand(1);
14368 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14369 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14370 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14372 Cond = ConvertCmpIfNecessary(Cond, DAG);
14373 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14374 Chain, Dest, CC, Cond);
14377 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14378 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14379 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14380 // that the guard pages used by the OS virtual memory manager are allocated in
14381 // correct sequence.
14383 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14384 SelectionDAG &DAG) const {
14385 MachineFunction &MF = DAG.getMachineFunction();
14386 bool SplitStack = MF.shouldSplitStack();
14387 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14392 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14393 SDNode* Node = Op.getNode();
14395 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14396 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14397 " not tell us which reg is the stack pointer!");
14398 EVT VT = Node->getValueType(0);
14399 SDValue Tmp1 = SDValue(Node, 0);
14400 SDValue Tmp2 = SDValue(Node, 1);
14401 SDValue Tmp3 = Node->getOperand(2);
14402 SDValue Chain = Tmp1.getOperand(0);
14404 // Chain the dynamic stack allocation so that it doesn't modify the stack
14405 // pointer when other instructions are using the stack.
14406 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14409 SDValue Size = Tmp2.getOperand(1);
14410 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14411 Chain = SP.getValue(1);
14412 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14413 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14414 unsigned StackAlign = TFI.getStackAlignment();
14415 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14416 if (Align > StackAlign)
14417 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14418 DAG.getConstant(-(uint64_t)Align, dl, VT));
14419 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14421 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14422 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14425 SDValue Ops[2] = { Tmp1, Tmp2 };
14426 return DAG.getMergeValues(Ops, dl);
14430 SDValue Chain = Op.getOperand(0);
14431 SDValue Size = Op.getOperand(1);
14432 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14433 EVT VT = Op.getNode()->getValueType(0);
14435 bool Is64Bit = Subtarget->is64Bit();
14436 EVT SPTy = getPointerTy();
14439 MachineRegisterInfo &MRI = MF.getRegInfo();
14442 // The 64 bit implementation of segmented stacks needs to clobber both r10
14443 // r11. This makes it impossible to use it along with nested parameters.
14444 const Function *F = MF.getFunction();
14446 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14448 if (I->hasNestAttr())
14449 report_fatal_error("Cannot use segmented stacks with functions that "
14450 "have nested arguments.");
14453 const TargetRegisterClass *AddrRegClass =
14454 getRegClassFor(getPointerTy());
14455 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14456 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14457 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14458 DAG.getRegister(Vreg, SPTy));
14459 SDValue Ops1[2] = { Value, Chain };
14460 return DAG.getMergeValues(Ops1, dl);
14463 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14465 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14466 Flag = Chain.getValue(1);
14467 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14469 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14471 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14472 unsigned SPReg = RegInfo->getStackRegister();
14473 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14474 Chain = SP.getValue(1);
14477 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14478 DAG.getConstant(-(uint64_t)Align, dl, VT));
14479 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14482 SDValue Ops1[2] = { SP, Chain };
14483 return DAG.getMergeValues(Ops1, dl);
14487 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14488 MachineFunction &MF = DAG.getMachineFunction();
14489 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14491 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14494 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14495 // vastart just stores the address of the VarArgsFrameIndex slot into the
14496 // memory location argument.
14497 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14499 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14500 MachinePointerInfo(SV), false, false, 0);
14504 // gp_offset (0 - 6 * 8)
14505 // fp_offset (48 - 48 + 8 * 16)
14506 // overflow_arg_area (point to parameters coming in memory).
14508 SmallVector<SDValue, 8> MemOps;
14509 SDValue FIN = Op.getOperand(1);
14511 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14512 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14514 FIN, MachinePointerInfo(SV), false, false, 0);
14515 MemOps.push_back(Store);
14518 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14519 FIN, DAG.getIntPtrConstant(4, DL));
14520 Store = DAG.getStore(Op.getOperand(0), DL,
14521 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
14523 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14524 MemOps.push_back(Store);
14526 // Store ptr to overflow_arg_area
14527 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14528 FIN, DAG.getIntPtrConstant(4, DL));
14529 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14531 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14532 MachinePointerInfo(SV, 8),
14534 MemOps.push_back(Store);
14536 // Store ptr to reg_save_area.
14537 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14538 FIN, DAG.getIntPtrConstant(8, DL));
14539 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14541 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14542 MachinePointerInfo(SV, 16), false, false, 0);
14543 MemOps.push_back(Store);
14544 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14547 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14548 assert(Subtarget->is64Bit() &&
14549 "LowerVAARG only handles 64-bit va_arg!");
14550 assert((Subtarget->isTargetLinux() ||
14551 Subtarget->isTargetDarwin()) &&
14552 "Unhandled target in LowerVAARG");
14553 assert(Op.getNode()->getNumOperands() == 4);
14554 SDValue Chain = Op.getOperand(0);
14555 SDValue SrcPtr = Op.getOperand(1);
14556 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14557 unsigned Align = Op.getConstantOperandVal(3);
14560 EVT ArgVT = Op.getNode()->getValueType(0);
14561 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14562 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14565 // Decide which area this value should be read from.
14566 // TODO: Implement the AMD64 ABI in its entirety. This simple
14567 // selection mechanism works only for the basic types.
14568 if (ArgVT == MVT::f80) {
14569 llvm_unreachable("va_arg for f80 not yet implemented");
14570 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14571 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14572 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14573 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14575 llvm_unreachable("Unhandled argument type in LowerVAARG");
14578 if (ArgMode == 2) {
14579 // Sanity Check: Make sure using fp_offset makes sense.
14580 assert(!DAG.getTarget().Options.UseSoftFloat &&
14581 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14582 Attribute::NoImplicitFloat)) &&
14583 Subtarget->hasSSE1());
14586 // Insert VAARG_64 node into the DAG
14587 // VAARG_64 returns two values: Variable Argument Address, Chain
14588 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
14589 DAG.getConstant(ArgMode, dl, MVT::i8),
14590 DAG.getConstant(Align, dl, MVT::i32)};
14591 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14592 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14593 VTs, InstOps, MVT::i64,
14594 MachinePointerInfo(SV),
14596 /*Volatile=*/false,
14598 /*WriteMem=*/true);
14599 Chain = VAARG.getValue(1);
14601 // Load the next argument and return it
14602 return DAG.getLoad(ArgVT, dl,
14605 MachinePointerInfo(),
14606 false, false, false, 0);
14609 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14610 SelectionDAG &DAG) {
14611 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14612 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14613 SDValue Chain = Op.getOperand(0);
14614 SDValue DstPtr = Op.getOperand(1);
14615 SDValue SrcPtr = Op.getOperand(2);
14616 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14617 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14620 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14621 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
14623 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14626 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14627 // amount is a constant. Takes immediate version of shift as input.
14628 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14629 SDValue SrcOp, uint64_t ShiftAmt,
14630 SelectionDAG &DAG) {
14631 MVT ElementType = VT.getVectorElementType();
14633 // Fold this packed shift into its first operand if ShiftAmt is 0.
14637 // Check for ShiftAmt >= element width
14638 if (ShiftAmt >= ElementType.getSizeInBits()) {
14639 if (Opc == X86ISD::VSRAI)
14640 ShiftAmt = ElementType.getSizeInBits() - 1;
14642 return DAG.getConstant(0, dl, VT);
14645 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14646 && "Unknown target vector shift-by-constant node");
14648 // Fold this packed vector shift into a build vector if SrcOp is a
14649 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14650 if (VT == SrcOp.getSimpleValueType() &&
14651 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14652 SmallVector<SDValue, 8> Elts;
14653 unsigned NumElts = SrcOp->getNumOperands();
14654 ConstantSDNode *ND;
14657 default: llvm_unreachable(nullptr);
14658 case X86ISD::VSHLI:
14659 for (unsigned i=0; i!=NumElts; ++i) {
14660 SDValue CurrentOp = SrcOp->getOperand(i);
14661 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14662 Elts.push_back(CurrentOp);
14665 ND = cast<ConstantSDNode>(CurrentOp);
14666 const APInt &C = ND->getAPIntValue();
14667 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
14670 case X86ISD::VSRLI:
14671 for (unsigned i=0; i!=NumElts; ++i) {
14672 SDValue CurrentOp = SrcOp->getOperand(i);
14673 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14674 Elts.push_back(CurrentOp);
14677 ND = cast<ConstantSDNode>(CurrentOp);
14678 const APInt &C = ND->getAPIntValue();
14679 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
14682 case X86ISD::VSRAI:
14683 for (unsigned i=0; i!=NumElts; ++i) {
14684 SDValue CurrentOp = SrcOp->getOperand(i);
14685 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14686 Elts.push_back(CurrentOp);
14689 ND = cast<ConstantSDNode>(CurrentOp);
14690 const APInt &C = ND->getAPIntValue();
14691 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
14696 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14699 return DAG.getNode(Opc, dl, VT, SrcOp,
14700 DAG.getConstant(ShiftAmt, dl, MVT::i8));
14703 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14704 // may or may not be a constant. Takes immediate version of shift as input.
14705 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14706 SDValue SrcOp, SDValue ShAmt,
14707 SelectionDAG &DAG) {
14708 MVT SVT = ShAmt.getSimpleValueType();
14709 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14711 // Catch shift-by-constant.
14712 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14713 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14714 CShAmt->getZExtValue(), DAG);
14716 // Change opcode to non-immediate version
14718 default: llvm_unreachable("Unknown target vector shift node");
14719 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14720 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14721 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14724 const X86Subtarget &Subtarget =
14725 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14726 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14727 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14728 // Let the shuffle legalizer expand this shift amount node.
14729 SDValue Op0 = ShAmt.getOperand(0);
14730 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14731 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14733 // Need to build a vector containing shift amount.
14734 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14735 SmallVector<SDValue, 4> ShOps;
14736 ShOps.push_back(ShAmt);
14737 if (SVT == MVT::i32) {
14738 ShOps.push_back(DAG.getConstant(0, dl, SVT));
14739 ShOps.push_back(DAG.getUNDEF(SVT));
14741 ShOps.push_back(DAG.getUNDEF(SVT));
14743 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14744 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14747 // The return type has to be a 128-bit type with the same element
14748 // type as the input type.
14749 MVT EltVT = VT.getVectorElementType();
14750 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14752 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14753 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14756 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14757 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14758 /// necessary casting for \p Mask when lowering masking intrinsics.
14759 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14760 SDValue PreservedSrc,
14761 const X86Subtarget *Subtarget,
14762 SelectionDAG &DAG) {
14763 EVT VT = Op.getValueType();
14764 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14765 MVT::i1, VT.getVectorNumElements());
14766 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14767 Mask.getValueType().getSizeInBits());
14770 assert(MaskVT.isSimple() && "invalid mask type");
14772 if (isAllOnes(Mask))
14775 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14776 // are extracted by EXTRACT_SUBVECTOR.
14777 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14778 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14779 DAG.getIntPtrConstant(0, dl));
14781 switch (Op.getOpcode()) {
14783 case X86ISD::PCMPEQM:
14784 case X86ISD::PCMPGTM:
14786 case X86ISD::CMPMU:
14787 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14789 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14790 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14791 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14794 /// \brief Creates an SDNode for a predicated scalar operation.
14795 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14796 /// The mask is comming as MVT::i8 and it should be truncated
14797 /// to MVT::i1 while lowering masking intrinsics.
14798 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14799 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14800 /// a scalar instruction.
14801 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14802 SDValue PreservedSrc,
14803 const X86Subtarget *Subtarget,
14804 SelectionDAG &DAG) {
14805 if (isAllOnes(Mask))
14808 EVT VT = Op.getValueType();
14810 // The mask should be of type MVT::i1
14811 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14813 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14814 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14815 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14818 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14819 SelectionDAG &DAG) {
14821 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14822 EVT VT = Op.getValueType();
14823 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14825 switch(IntrData->Type) {
14826 case INTR_TYPE_1OP:
14827 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14828 case INTR_TYPE_2OP:
14829 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14831 case INTR_TYPE_3OP:
14832 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14833 Op.getOperand(2), Op.getOperand(3));
14834 case INTR_TYPE_1OP_MASK_RM: {
14835 SDValue Src = Op.getOperand(1);
14836 SDValue Src0 = Op.getOperand(2);
14837 SDValue Mask = Op.getOperand(3);
14838 SDValue RoundingMode = Op.getOperand(4);
14839 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14841 Mask, Src0, Subtarget, DAG);
14843 case INTR_TYPE_SCALAR_MASK_RM: {
14844 SDValue Src1 = Op.getOperand(1);
14845 SDValue Src2 = Op.getOperand(2);
14846 SDValue Src0 = Op.getOperand(3);
14847 SDValue Mask = Op.getOperand(4);
14848 // There are 2 kinds of intrinsics in this group:
14849 // (1) With supress-all-exceptions (sae) - 6 operands
14850 // (2) With rounding mode and sae - 7 operands.
14851 if (Op.getNumOperands() == 6) {
14852 SDValue Sae = Op.getOperand(5);
14853 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14855 Mask, Src0, Subtarget, DAG);
14857 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
14858 SDValue RoundingMode = Op.getOperand(5);
14859 SDValue Sae = Op.getOperand(6);
14860 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14861 RoundingMode, Sae),
14862 Mask, Src0, Subtarget, DAG);
14864 case INTR_TYPE_2OP_MASK: {
14865 SDValue Src1 = Op.getOperand(1);
14866 SDValue Src2 = Op.getOperand(2);
14867 SDValue PassThru = Op.getOperand(3);
14868 SDValue Mask = Op.getOperand(4);
14869 // We specify 2 possible opcodes for intrinsics with rounding modes.
14870 // First, we check if the intrinsic may have non-default rounding mode,
14871 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14872 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14873 if (IntrWithRoundingModeOpcode != 0) {
14874 SDValue Rnd = Op.getOperand(5);
14875 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14876 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14877 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14878 dl, Op.getValueType(),
14880 Mask, PassThru, Subtarget, DAG);
14883 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14885 Mask, PassThru, Subtarget, DAG);
14887 case FMA_OP_MASK: {
14888 SDValue Src1 = Op.getOperand(1);
14889 SDValue Src2 = Op.getOperand(2);
14890 SDValue Src3 = Op.getOperand(3);
14891 SDValue Mask = Op.getOperand(4);
14892 // We specify 2 possible opcodes for intrinsics with rounding modes.
14893 // First, we check if the intrinsic may have non-default rounding mode,
14894 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14895 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14896 if (IntrWithRoundingModeOpcode != 0) {
14897 SDValue Rnd = Op.getOperand(5);
14898 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14899 X86::STATIC_ROUNDING::CUR_DIRECTION)
14900 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14901 dl, Op.getValueType(),
14902 Src1, Src2, Src3, Rnd),
14903 Mask, Src1, Subtarget, DAG);
14905 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
14906 dl, Op.getValueType(),
14908 Mask, Src1, Subtarget, DAG);
14911 case CMP_MASK_CC: {
14912 // Comparison intrinsics with masks.
14913 // Example of transformation:
14914 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
14915 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
14917 // (v8i1 (insert_subvector undef,
14918 // (v2i1 (and (PCMPEQM %a, %b),
14919 // (extract_subvector
14920 // (v8i1 (bitcast %mask)), 0))), 0))))
14921 EVT VT = Op.getOperand(1).getValueType();
14922 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14923 VT.getVectorNumElements());
14924 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
14925 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14926 Mask.getValueType().getSizeInBits());
14928 if (IntrData->Type == CMP_MASK_CC) {
14929 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14930 Op.getOperand(2), Op.getOperand(3));
14932 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
14933 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14936 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
14937 DAG.getTargetConstant(0, dl,
14940 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
14941 DAG.getUNDEF(BitcastVT), CmpMask,
14942 DAG.getIntPtrConstant(0, dl));
14943 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
14945 case COMI: { // Comparison intrinsics
14946 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
14947 SDValue LHS = Op.getOperand(1);
14948 SDValue RHS = Op.getOperand(2);
14949 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
14950 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
14951 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
14952 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14953 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
14954 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14957 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
14958 Op.getOperand(1), Op.getOperand(2), DAG);
14960 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
14961 Op.getSimpleValueType(),
14963 Op.getOperand(2), DAG),
14964 Op.getOperand(4), Op.getOperand(3), Subtarget,
14966 case COMPRESS_EXPAND_IN_REG: {
14967 SDValue Mask = Op.getOperand(3);
14968 SDValue DataToCompress = Op.getOperand(1);
14969 SDValue PassThru = Op.getOperand(2);
14970 if (isAllOnes(Mask)) // return data as is
14971 return Op.getOperand(1);
14972 EVT VT = Op.getValueType();
14973 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14974 VT.getVectorNumElements());
14975 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14976 Mask.getValueType().getSizeInBits());
14978 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14979 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14980 DAG.getIntPtrConstant(0, dl));
14982 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
14986 SDValue Mask = Op.getOperand(3);
14987 EVT VT = Op.getValueType();
14988 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14989 VT.getVectorNumElements());
14990 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14991 Mask.getValueType().getSizeInBits());
14993 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14994 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14995 DAG.getIntPtrConstant(0, dl));
14996 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15005 default: return SDValue(); // Don't custom lower most intrinsics.
15007 case Intrinsic::x86_avx2_permd:
15008 case Intrinsic::x86_avx2_permps:
15009 // Operands intentionally swapped. Mask is last operand to intrinsic,
15010 // but second operand for node/instruction.
15011 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15012 Op.getOperand(2), Op.getOperand(1));
15014 case Intrinsic::x86_avx512_mask_valign_q_512:
15015 case Intrinsic::x86_avx512_mask_valign_d_512:
15016 // Vector source operands are swapped.
15017 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
15018 Op.getValueType(), Op.getOperand(2),
15021 Op.getOperand(5), Op.getOperand(4),
15024 // ptest and testp intrinsics. The intrinsic these come from are designed to
15025 // return an integer value, not just an instruction so lower it to the ptest
15026 // or testp pattern and a setcc for the result.
15027 case Intrinsic::x86_sse41_ptestz:
15028 case Intrinsic::x86_sse41_ptestc:
15029 case Intrinsic::x86_sse41_ptestnzc:
15030 case Intrinsic::x86_avx_ptestz_256:
15031 case Intrinsic::x86_avx_ptestc_256:
15032 case Intrinsic::x86_avx_ptestnzc_256:
15033 case Intrinsic::x86_avx_vtestz_ps:
15034 case Intrinsic::x86_avx_vtestc_ps:
15035 case Intrinsic::x86_avx_vtestnzc_ps:
15036 case Intrinsic::x86_avx_vtestz_pd:
15037 case Intrinsic::x86_avx_vtestc_pd:
15038 case Intrinsic::x86_avx_vtestnzc_pd:
15039 case Intrinsic::x86_avx_vtestz_ps_256:
15040 case Intrinsic::x86_avx_vtestc_ps_256:
15041 case Intrinsic::x86_avx_vtestnzc_ps_256:
15042 case Intrinsic::x86_avx_vtestz_pd_256:
15043 case Intrinsic::x86_avx_vtestc_pd_256:
15044 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15045 bool IsTestPacked = false;
15048 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15049 case Intrinsic::x86_avx_vtestz_ps:
15050 case Intrinsic::x86_avx_vtestz_pd:
15051 case Intrinsic::x86_avx_vtestz_ps_256:
15052 case Intrinsic::x86_avx_vtestz_pd_256:
15053 IsTestPacked = true; // Fallthrough
15054 case Intrinsic::x86_sse41_ptestz:
15055 case Intrinsic::x86_avx_ptestz_256:
15057 X86CC = X86::COND_E;
15059 case Intrinsic::x86_avx_vtestc_ps:
15060 case Intrinsic::x86_avx_vtestc_pd:
15061 case Intrinsic::x86_avx_vtestc_ps_256:
15062 case Intrinsic::x86_avx_vtestc_pd_256:
15063 IsTestPacked = true; // Fallthrough
15064 case Intrinsic::x86_sse41_ptestc:
15065 case Intrinsic::x86_avx_ptestc_256:
15067 X86CC = X86::COND_B;
15069 case Intrinsic::x86_avx_vtestnzc_ps:
15070 case Intrinsic::x86_avx_vtestnzc_pd:
15071 case Intrinsic::x86_avx_vtestnzc_ps_256:
15072 case Intrinsic::x86_avx_vtestnzc_pd_256:
15073 IsTestPacked = true; // Fallthrough
15074 case Intrinsic::x86_sse41_ptestnzc:
15075 case Intrinsic::x86_avx_ptestnzc_256:
15077 X86CC = X86::COND_A;
15081 SDValue LHS = Op.getOperand(1);
15082 SDValue RHS = Op.getOperand(2);
15083 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15084 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15085 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15086 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15087 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15089 case Intrinsic::x86_avx512_kortestz_w:
15090 case Intrinsic::x86_avx512_kortestc_w: {
15091 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15092 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
15093 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
15094 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15095 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15096 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15097 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15100 case Intrinsic::x86_sse42_pcmpistria128:
15101 case Intrinsic::x86_sse42_pcmpestria128:
15102 case Intrinsic::x86_sse42_pcmpistric128:
15103 case Intrinsic::x86_sse42_pcmpestric128:
15104 case Intrinsic::x86_sse42_pcmpistrio128:
15105 case Intrinsic::x86_sse42_pcmpestrio128:
15106 case Intrinsic::x86_sse42_pcmpistris128:
15107 case Intrinsic::x86_sse42_pcmpestris128:
15108 case Intrinsic::x86_sse42_pcmpistriz128:
15109 case Intrinsic::x86_sse42_pcmpestriz128: {
15113 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15114 case Intrinsic::x86_sse42_pcmpistria128:
15115 Opcode = X86ISD::PCMPISTRI;
15116 X86CC = X86::COND_A;
15118 case Intrinsic::x86_sse42_pcmpestria128:
15119 Opcode = X86ISD::PCMPESTRI;
15120 X86CC = X86::COND_A;
15122 case Intrinsic::x86_sse42_pcmpistric128:
15123 Opcode = X86ISD::PCMPISTRI;
15124 X86CC = X86::COND_B;
15126 case Intrinsic::x86_sse42_pcmpestric128:
15127 Opcode = X86ISD::PCMPESTRI;
15128 X86CC = X86::COND_B;
15130 case Intrinsic::x86_sse42_pcmpistrio128:
15131 Opcode = X86ISD::PCMPISTRI;
15132 X86CC = X86::COND_O;
15134 case Intrinsic::x86_sse42_pcmpestrio128:
15135 Opcode = X86ISD::PCMPESTRI;
15136 X86CC = X86::COND_O;
15138 case Intrinsic::x86_sse42_pcmpistris128:
15139 Opcode = X86ISD::PCMPISTRI;
15140 X86CC = X86::COND_S;
15142 case Intrinsic::x86_sse42_pcmpestris128:
15143 Opcode = X86ISD::PCMPESTRI;
15144 X86CC = X86::COND_S;
15146 case Intrinsic::x86_sse42_pcmpistriz128:
15147 Opcode = X86ISD::PCMPISTRI;
15148 X86CC = X86::COND_E;
15150 case Intrinsic::x86_sse42_pcmpestriz128:
15151 Opcode = X86ISD::PCMPESTRI;
15152 X86CC = X86::COND_E;
15155 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15156 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15157 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15158 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15159 DAG.getConstant(X86CC, dl, MVT::i8),
15160 SDValue(PCMP.getNode(), 1));
15161 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15164 case Intrinsic::x86_sse42_pcmpistri128:
15165 case Intrinsic::x86_sse42_pcmpestri128: {
15167 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15168 Opcode = X86ISD::PCMPISTRI;
15170 Opcode = X86ISD::PCMPESTRI;
15172 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15173 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15174 return DAG.getNode(Opcode, dl, VTs, NewOps);
15179 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15180 SDValue Src, SDValue Mask, SDValue Base,
15181 SDValue Index, SDValue ScaleOp, SDValue Chain,
15182 const X86Subtarget * Subtarget) {
15184 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15185 assert(C && "Invalid scale type");
15186 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15187 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15188 Index.getSimpleValueType().getVectorNumElements());
15190 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15192 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15194 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15195 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15196 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15197 SDValue Segment = DAG.getRegister(0, MVT::i32);
15198 if (Src.getOpcode() == ISD::UNDEF)
15199 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15200 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15201 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15202 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15203 return DAG.getMergeValues(RetOps, dl);
15206 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15207 SDValue Src, SDValue Mask, SDValue Base,
15208 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15210 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15211 assert(C && "Invalid scale type");
15212 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15213 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15214 SDValue Segment = DAG.getRegister(0, MVT::i32);
15215 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15216 Index.getSimpleValueType().getVectorNumElements());
15218 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15220 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15222 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15223 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15224 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15225 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15226 return SDValue(Res, 1);
15229 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15230 SDValue Mask, SDValue Base, SDValue Index,
15231 SDValue ScaleOp, SDValue Chain) {
15233 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15234 assert(C && "Invalid scale type");
15235 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15236 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15237 SDValue Segment = DAG.getRegister(0, MVT::i32);
15239 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15241 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15243 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15245 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15246 //SDVTList VTs = DAG.getVTList(MVT::Other);
15247 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15248 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15249 return SDValue(Res, 0);
15252 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15253 // read performance monitor counters (x86_rdpmc).
15254 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15255 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15256 SmallVectorImpl<SDValue> &Results) {
15257 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15258 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15261 // The ECX register is used to select the index of the performance counter
15263 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
15265 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
15267 // Reads the content of a 64-bit performance counter and returns it in the
15268 // registers EDX:EAX.
15269 if (Subtarget->is64Bit()) {
15270 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15271 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15274 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15275 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15278 Chain = HI.getValue(1);
15280 if (Subtarget->is64Bit()) {
15281 // The EAX register is loaded with the low-order 32 bits. The EDX register
15282 // is loaded with the supported high-order bits of the counter.
15283 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15284 DAG.getConstant(32, DL, MVT::i8));
15285 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15286 Results.push_back(Chain);
15290 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15291 SDValue Ops[] = { LO, HI };
15292 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15293 Results.push_back(Pair);
15294 Results.push_back(Chain);
15297 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
15298 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
15299 // also used to custom lower READCYCLECOUNTER nodes.
15300 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
15301 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15302 SmallVectorImpl<SDValue> &Results) {
15303 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15304 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
15307 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
15308 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
15309 // and the EAX register is loaded with the low-order 32 bits.
15310 if (Subtarget->is64Bit()) {
15311 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15312 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15315 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15316 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15319 SDValue Chain = HI.getValue(1);
15321 if (Opcode == X86ISD::RDTSCP_DAG) {
15322 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15324 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
15325 // the ECX register. Add 'ecx' explicitly to the chain.
15326 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
15328 // Explicitly store the content of ECX at the location passed in input
15329 // to the 'rdtscp' intrinsic.
15330 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
15331 MachinePointerInfo(), false, false, 0);
15334 if (Subtarget->is64Bit()) {
15335 // The EDX register is loaded with the high-order 32 bits of the MSR, and
15336 // the EAX register is loaded with the low-order 32 bits.
15337 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15338 DAG.getConstant(32, DL, MVT::i8));
15339 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15340 Results.push_back(Chain);
15344 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15345 SDValue Ops[] = { LO, HI };
15346 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15347 Results.push_back(Pair);
15348 Results.push_back(Chain);
15351 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
15352 SelectionDAG &DAG) {
15353 SmallVector<SDValue, 2> Results;
15355 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
15357 return DAG.getMergeValues(Results, DL);
15361 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15362 SelectionDAG &DAG) {
15363 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
15365 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
15370 switch(IntrData->Type) {
15372 llvm_unreachable("Unknown Intrinsic Type");
15376 // Emit the node with the right value type.
15377 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
15378 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15380 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
15381 // Otherwise return the value from Rand, which is always 0, casted to i32.
15382 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
15383 DAG.getConstant(1, dl, Op->getValueType(1)),
15384 DAG.getConstant(X86::COND_B, dl, MVT::i32),
15385 SDValue(Result.getNode(), 1) };
15386 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
15387 DAG.getVTList(Op->getValueType(1), MVT::Glue),
15390 // Return { result, isValid, chain }.
15391 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
15392 SDValue(Result.getNode(), 2));
15395 //gather(v1, mask, index, base, scale);
15396 SDValue Chain = Op.getOperand(0);
15397 SDValue Src = Op.getOperand(2);
15398 SDValue Base = Op.getOperand(3);
15399 SDValue Index = Op.getOperand(4);
15400 SDValue Mask = Op.getOperand(5);
15401 SDValue Scale = Op.getOperand(6);
15402 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
15406 //scatter(base, mask, index, v1, scale);
15407 SDValue Chain = Op.getOperand(0);
15408 SDValue Base = Op.getOperand(2);
15409 SDValue Mask = Op.getOperand(3);
15410 SDValue Index = Op.getOperand(4);
15411 SDValue Src = Op.getOperand(5);
15412 SDValue Scale = Op.getOperand(6);
15413 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
15417 SDValue Hint = Op.getOperand(6);
15418 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
15419 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
15420 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
15421 SDValue Chain = Op.getOperand(0);
15422 SDValue Mask = Op.getOperand(2);
15423 SDValue Index = Op.getOperand(3);
15424 SDValue Base = Op.getOperand(4);
15425 SDValue Scale = Op.getOperand(5);
15426 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15428 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15430 SmallVector<SDValue, 2> Results;
15431 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
15433 return DAG.getMergeValues(Results, dl);
15435 // Read Performance Monitoring Counters.
15437 SmallVector<SDValue, 2> Results;
15438 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15439 return DAG.getMergeValues(Results, dl);
15441 // XTEST intrinsics.
15443 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15444 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15445 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15446 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
15448 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15449 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15450 Ret, SDValue(InTrans.getNode(), 1));
15454 SmallVector<SDValue, 2> Results;
15455 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15456 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15457 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15458 DAG.getConstant(-1, dl, MVT::i8));
15459 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15460 Op.getOperand(4), GenCF.getValue(1));
15461 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15462 Op.getOperand(5), MachinePointerInfo(),
15464 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15465 DAG.getConstant(X86::COND_B, dl, MVT::i8),
15467 Results.push_back(SetCC);
15468 Results.push_back(Store);
15469 return DAG.getMergeValues(Results, dl);
15471 case COMPRESS_TO_MEM: {
15473 SDValue Mask = Op.getOperand(4);
15474 SDValue DataToCompress = Op.getOperand(3);
15475 SDValue Addr = Op.getOperand(2);
15476 SDValue Chain = Op.getOperand(0);
15478 if (isAllOnes(Mask)) // return just a store
15479 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15480 MachinePointerInfo(), false, false, 0);
15482 EVT VT = DataToCompress.getValueType();
15483 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15484 VT.getVectorNumElements());
15485 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15486 Mask.getValueType().getSizeInBits());
15487 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15488 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15489 DAG.getIntPtrConstant(0, dl));
15491 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15492 DataToCompress, DAG.getUNDEF(VT));
15493 return DAG.getStore(Chain, dl, Compressed, Addr,
15494 MachinePointerInfo(), false, false, 0);
15496 case EXPAND_FROM_MEM: {
15498 SDValue Mask = Op.getOperand(4);
15499 SDValue PathThru = Op.getOperand(3);
15500 SDValue Addr = Op.getOperand(2);
15501 SDValue Chain = Op.getOperand(0);
15502 EVT VT = Op.getValueType();
15504 if (isAllOnes(Mask)) // return just a load
15505 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15507 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15508 VT.getVectorNumElements());
15509 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15510 Mask.getValueType().getSizeInBits());
15511 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15512 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15513 DAG.getIntPtrConstant(0, dl));
15515 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15516 false, false, false, 0);
15518 SDValue Results[] = {
15519 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15521 return DAG.getMergeValues(Results, dl);
15526 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15527 SelectionDAG &DAG) const {
15528 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15529 MFI->setReturnAddressIsTaken(true);
15531 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15534 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15536 EVT PtrVT = getPointerTy();
15539 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15540 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15541 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
15542 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15543 DAG.getNode(ISD::ADD, dl, PtrVT,
15544 FrameAddr, Offset),
15545 MachinePointerInfo(), false, false, false, 0);
15548 // Just load the return address.
15549 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15550 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15551 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15554 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15555 MachineFunction &MF = DAG.getMachineFunction();
15556 MachineFrameInfo *MFI = MF.getFrameInfo();
15557 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15558 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15559 EVT VT = Op.getValueType();
15561 MFI->setFrameAddressIsTaken(true);
15563 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15564 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15565 // is not possible to crawl up the stack without looking at the unwind codes
15567 int FrameAddrIndex = FuncInfo->getFAIndex();
15568 if (!FrameAddrIndex) {
15569 // Set up a frame object for the return address.
15570 unsigned SlotSize = RegInfo->getSlotSize();
15571 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15572 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
15573 FuncInfo->setFAIndex(FrameAddrIndex);
15575 return DAG.getFrameIndex(FrameAddrIndex, VT);
15578 unsigned FrameReg =
15579 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15580 SDLoc dl(Op); // FIXME probably not meaningful
15581 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15582 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15583 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15584 "Invalid Frame Register!");
15585 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15587 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15588 MachinePointerInfo(),
15589 false, false, false, 0);
15593 // FIXME? Maybe this could be a TableGen attribute on some registers and
15594 // this table could be generated automatically from RegInfo.
15595 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15597 unsigned Reg = StringSwitch<unsigned>(RegName)
15598 .Case("esp", X86::ESP)
15599 .Case("rsp", X86::RSP)
15603 report_fatal_error("Invalid register name global variable");
15606 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15607 SelectionDAG &DAG) const {
15608 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15609 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
15612 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15613 SDValue Chain = Op.getOperand(0);
15614 SDValue Offset = Op.getOperand(1);
15615 SDValue Handler = Op.getOperand(2);
15618 EVT PtrVT = getPointerTy();
15619 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15620 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15621 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15622 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15623 "Invalid Frame Register!");
15624 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15625 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15627 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15628 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
15630 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15631 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15633 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15635 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15636 DAG.getRegister(StoreAddrReg, PtrVT));
15639 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15640 SelectionDAG &DAG) const {
15642 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15643 DAG.getVTList(MVT::i32, MVT::Other),
15644 Op.getOperand(0), Op.getOperand(1));
15647 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15648 SelectionDAG &DAG) const {
15650 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15651 Op.getOperand(0), Op.getOperand(1));
15654 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15655 return Op.getOperand(0);
15658 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15659 SelectionDAG &DAG) const {
15660 SDValue Root = Op.getOperand(0);
15661 SDValue Trmp = Op.getOperand(1); // trampoline
15662 SDValue FPtr = Op.getOperand(2); // nested function
15663 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15666 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15667 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15669 if (Subtarget->is64Bit()) {
15670 SDValue OutChains[6];
15672 // Large code-model.
15673 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15674 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15676 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15677 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15679 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15681 // Load the pointer to the nested function into R11.
15682 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15683 SDValue Addr = Trmp;
15684 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15685 Addr, MachinePointerInfo(TrmpAddr),
15688 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15689 DAG.getConstant(2, dl, MVT::i64));
15690 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15691 MachinePointerInfo(TrmpAddr, 2),
15694 // Load the 'nest' parameter value into R10.
15695 // R10 is specified in X86CallingConv.td
15696 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15697 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15698 DAG.getConstant(10, dl, MVT::i64));
15699 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15700 Addr, MachinePointerInfo(TrmpAddr, 10),
15703 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15704 DAG.getConstant(12, dl, MVT::i64));
15705 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15706 MachinePointerInfo(TrmpAddr, 12),
15709 // Jump to the nested function.
15710 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15711 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15712 DAG.getConstant(20, dl, MVT::i64));
15713 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15714 Addr, MachinePointerInfo(TrmpAddr, 20),
15717 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15718 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15719 DAG.getConstant(22, dl, MVT::i64));
15720 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
15721 Addr, MachinePointerInfo(TrmpAddr, 22),
15724 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15726 const Function *Func =
15727 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15728 CallingConv::ID CC = Func->getCallingConv();
15733 llvm_unreachable("Unsupported calling convention");
15734 case CallingConv::C:
15735 case CallingConv::X86_StdCall: {
15736 // Pass 'nest' parameter in ECX.
15737 // Must be kept in sync with X86CallingConv.td
15738 NestReg = X86::ECX;
15740 // Check that ECX wasn't needed by an 'inreg' parameter.
15741 FunctionType *FTy = Func->getFunctionType();
15742 const AttributeSet &Attrs = Func->getAttributes();
15744 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15745 unsigned InRegCount = 0;
15748 for (FunctionType::param_iterator I = FTy->param_begin(),
15749 E = FTy->param_end(); I != E; ++I, ++Idx)
15750 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15751 // FIXME: should only count parameters that are lowered to integers.
15752 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15754 if (InRegCount > 2) {
15755 report_fatal_error("Nest register in use - reduce number of inreg"
15761 case CallingConv::X86_FastCall:
15762 case CallingConv::X86_ThisCall:
15763 case CallingConv::Fast:
15764 // Pass 'nest' parameter in EAX.
15765 // Must be kept in sync with X86CallingConv.td
15766 NestReg = X86::EAX;
15770 SDValue OutChains[4];
15771 SDValue Addr, Disp;
15773 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15774 DAG.getConstant(10, dl, MVT::i32));
15775 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15777 // This is storing the opcode for MOV32ri.
15778 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15779 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15780 OutChains[0] = DAG.getStore(Root, dl,
15781 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
15782 Trmp, MachinePointerInfo(TrmpAddr),
15785 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15786 DAG.getConstant(1, dl, MVT::i32));
15787 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15788 MachinePointerInfo(TrmpAddr, 1),
15791 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15792 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15793 DAG.getConstant(5, dl, MVT::i32));
15794 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
15795 Addr, MachinePointerInfo(TrmpAddr, 5),
15798 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15799 DAG.getConstant(6, dl, MVT::i32));
15800 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15801 MachinePointerInfo(TrmpAddr, 6),
15804 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15808 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15809 SelectionDAG &DAG) const {
15811 The rounding mode is in bits 11:10 of FPSR, and has the following
15813 00 Round to nearest
15818 FLT_ROUNDS, on the other hand, expects the following:
15825 To perform the conversion, we do:
15826 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15829 MachineFunction &MF = DAG.getMachineFunction();
15830 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15831 unsigned StackAlignment = TFI.getStackAlignment();
15832 MVT VT = Op.getSimpleValueType();
15835 // Save FP Control Word to stack slot
15836 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15837 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15839 MachineMemOperand *MMO =
15840 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15841 MachineMemOperand::MOStore, 2, 2);
15843 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15844 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15845 DAG.getVTList(MVT::Other),
15846 Ops, MVT::i16, MMO);
15848 // Load FP Control Word from stack slot
15849 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15850 MachinePointerInfo(), false, false, false, 0);
15852 // Transform as necessary
15854 DAG.getNode(ISD::SRL, DL, MVT::i16,
15855 DAG.getNode(ISD::AND, DL, MVT::i16,
15856 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
15857 DAG.getConstant(11, DL, MVT::i8));
15859 DAG.getNode(ISD::SRL, DL, MVT::i16,
15860 DAG.getNode(ISD::AND, DL, MVT::i16,
15861 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
15862 DAG.getConstant(9, DL, MVT::i8));
15865 DAG.getNode(ISD::AND, DL, MVT::i16,
15866 DAG.getNode(ISD::ADD, DL, MVT::i16,
15867 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
15868 DAG.getConstant(1, DL, MVT::i16)),
15869 DAG.getConstant(3, DL, MVT::i16));
15871 return DAG.getNode((VT.getSizeInBits() < 16 ?
15872 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
15875 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
15876 MVT VT = Op.getSimpleValueType();
15878 unsigned NumBits = VT.getSizeInBits();
15881 Op = Op.getOperand(0);
15882 if (VT == MVT::i8) {
15883 // Zero extend to i32 since there is not an i8 bsr.
15885 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15888 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
15889 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15890 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15892 // If src is zero (i.e. bsr sets ZF), returns NumBits.
15895 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
15896 DAG.getConstant(X86::COND_E, dl, MVT::i8),
15899 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
15901 // Finally xor with NumBits-1.
15902 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
15903 DAG.getConstant(NumBits - 1, dl, OpVT));
15906 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15910 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
15911 MVT VT = Op.getSimpleValueType();
15913 unsigned NumBits = VT.getSizeInBits();
15916 Op = Op.getOperand(0);
15917 if (VT == MVT::i8) {
15918 // Zero extend to i32 since there is not an i8 bsr.
15920 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15923 // Issue a bsr (scan bits in reverse).
15924 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15925 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15927 // And xor with NumBits-1.
15928 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
15929 DAG.getConstant(NumBits - 1, dl, OpVT));
15932 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15936 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
15937 MVT VT = Op.getSimpleValueType();
15938 unsigned NumBits = VT.getSizeInBits();
15940 Op = Op.getOperand(0);
15942 // Issue a bsf (scan bits forward) which also sets EFLAGS.
15943 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
15944 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
15946 // If src is zero (i.e. bsf sets ZF), returns NumBits.
15949 DAG.getConstant(NumBits, dl, VT),
15950 DAG.getConstant(X86::COND_E, dl, MVT::i8),
15953 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
15956 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
15957 // ones, and then concatenate the result back.
15958 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
15959 MVT VT = Op.getSimpleValueType();
15961 assert(VT.is256BitVector() && VT.isInteger() &&
15962 "Unsupported value type for operation");
15964 unsigned NumElems = VT.getVectorNumElements();
15967 // Extract the LHS vectors
15968 SDValue LHS = Op.getOperand(0);
15969 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15970 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15972 // Extract the RHS vectors
15973 SDValue RHS = Op.getOperand(1);
15974 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15975 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15977 MVT EltVT = VT.getVectorElementType();
15978 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15980 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15981 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15982 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15985 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15986 assert(Op.getSimpleValueType().is256BitVector() &&
15987 Op.getSimpleValueType().isInteger() &&
15988 "Only handle AVX 256-bit vector integer operation");
15989 return Lower256IntArith(Op, DAG);
15992 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15993 assert(Op.getSimpleValueType().is256BitVector() &&
15994 Op.getSimpleValueType().isInteger() &&
15995 "Only handle AVX 256-bit vector integer operation");
15996 return Lower256IntArith(Op, DAG);
15999 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16000 SelectionDAG &DAG) {
16002 MVT VT = Op.getSimpleValueType();
16004 // Decompose 256-bit ops into smaller 128-bit ops.
16005 if (VT.is256BitVector() && !Subtarget->hasInt256())
16006 return Lower256IntArith(Op, DAG);
16008 SDValue A = Op.getOperand(0);
16009 SDValue B = Op.getOperand(1);
16011 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16012 // pairs, multiply and truncate.
16013 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16014 if (Subtarget->hasInt256()) {
16015 if (VT == MVT::v32i8) {
16016 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16017 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16018 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16019 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16020 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16021 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16022 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16023 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16024 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16025 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16028 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16029 return DAG.getNode(
16030 ISD::TRUNCATE, dl, VT,
16031 DAG.getNode(ISD::MUL, dl, ExVT,
16032 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16033 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16036 assert(VT == MVT::v16i8 &&
16037 "Pre-AVX2 support only supports v16i8 multiplication");
16038 MVT ExVT = MVT::v8i16;
16040 // Extract the lo parts and sign extend to i16
16042 if (Subtarget->hasSSE41()) {
16043 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16044 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16046 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16047 -1, 4, -1, 5, -1, 6, -1, 7};
16048 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16049 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16050 ALo = DAG.getNode(ISD::BITCAST, dl, ExVT, ALo);
16051 BLo = DAG.getNode(ISD::BITCAST, dl, ExVT, BLo);
16052 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16053 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16056 // Extract the hi parts and sign extend to i16
16058 if (Subtarget->hasSSE41()) {
16059 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16060 -1, -1, -1, -1, -1, -1, -1, -1};
16061 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16062 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16063 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16064 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16066 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16067 -1, 12, -1, 13, -1, 14, -1, 15};
16068 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16069 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16070 AHi = DAG.getNode(ISD::BITCAST, dl, ExVT, AHi);
16071 BHi = DAG.getNode(ISD::BITCAST, dl, ExVT, BHi);
16072 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16073 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16076 // Multiply, mask the lower 8bits of the lo/hi results and pack
16077 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16078 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16079 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16080 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16081 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16084 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16085 if (VT == MVT::v4i32) {
16086 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16087 "Should not custom lower when pmuldq is available!");
16089 // Extract the odd parts.
16090 static const int UnpackMask[] = { 1, -1, 3, -1 };
16091 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16092 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16094 // Multiply the even parts.
16095 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16096 // Now multiply odd parts.
16097 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16099 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
16100 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
16102 // Merge the two vectors back together with a shuffle. This expands into 2
16104 static const int ShufMask[] = { 0, 4, 2, 6 };
16105 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16108 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16109 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16111 // Ahi = psrlqi(a, 32);
16112 // Bhi = psrlqi(b, 32);
16114 // AloBlo = pmuludq(a, b);
16115 // AloBhi = pmuludq(a, Bhi);
16116 // AhiBlo = pmuludq(Ahi, b);
16118 // AloBhi = psllqi(AloBhi, 32);
16119 // AhiBlo = psllqi(AhiBlo, 32);
16120 // return AloBlo + AloBhi + AhiBlo;
16122 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16123 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16125 // Bit cast to 32-bit vectors for MULUDQ
16126 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16127 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16128 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
16129 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
16130 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
16131 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
16133 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16134 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16135 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16137 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16138 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16140 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16141 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16144 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16145 assert(Subtarget->isTargetWin64() && "Unexpected target");
16146 EVT VT = Op.getValueType();
16147 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16148 "Unexpected return type for lowering");
16152 switch (Op->getOpcode()) {
16153 default: llvm_unreachable("Unexpected request for libcall!");
16154 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16155 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16156 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16157 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16158 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16159 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16163 SDValue InChain = DAG.getEntryNode();
16165 TargetLowering::ArgListTy Args;
16166 TargetLowering::ArgListEntry Entry;
16167 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16168 EVT ArgVT = Op->getOperand(i).getValueType();
16169 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16170 "Unexpected argument type for lowering");
16171 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16172 Entry.Node = StackPtr;
16173 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16175 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16176 Entry.Ty = PointerType::get(ArgTy,0);
16177 Entry.isSExt = false;
16178 Entry.isZExt = false;
16179 Args.push_back(Entry);
16182 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16185 TargetLowering::CallLoweringInfo CLI(DAG);
16186 CLI.setDebugLoc(dl).setChain(InChain)
16187 .setCallee(getLibcallCallingConv(LC),
16188 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16189 Callee, std::move(Args), 0)
16190 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16192 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16193 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
16196 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16197 SelectionDAG &DAG) {
16198 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16199 EVT VT = Op0.getValueType();
16202 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16203 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16205 // PMULxD operations multiply each even value (starting at 0) of LHS with
16206 // the related value of RHS and produce a widen result.
16207 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16208 // => <2 x i64> <ae|cg>
16210 // In other word, to have all the results, we need to perform two PMULxD:
16211 // 1. one with the even values.
16212 // 2. one with the odd values.
16213 // To achieve #2, with need to place the odd values at an even position.
16215 // Place the odd value at an even position (basically, shift all values 1
16216 // step to the left):
16217 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16218 // <a|b|c|d> => <b|undef|d|undef>
16219 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16220 // <e|f|g|h> => <f|undef|h|undef>
16221 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16223 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16225 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16226 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16228 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16229 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16230 // => <2 x i64> <ae|cg>
16231 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
16232 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
16233 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
16234 // => <2 x i64> <bf|dh>
16235 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
16236 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
16238 // Shuffle it back into the right order.
16239 SDValue Highs, Lows;
16240 if (VT == MVT::v8i32) {
16241 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
16242 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16243 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
16244 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16246 const int HighMask[] = {1, 5, 3, 7};
16247 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16248 const int LowMask[] = {0, 4, 2, 6};
16249 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16252 // If we have a signed multiply but no PMULDQ fix up the high parts of a
16253 // unsigned multiply.
16254 if (IsSigned && !Subtarget->hasSSE41()) {
16256 DAG.getConstant(31, dl,
16257 DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
16258 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
16259 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
16260 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
16261 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
16263 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
16264 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
16267 // The first result of MUL_LOHI is actually the low value, followed by the
16269 SDValue Ops[] = {Lows, Highs};
16270 return DAG.getMergeValues(Ops, dl);
16273 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
16274 const X86Subtarget *Subtarget) {
16275 MVT VT = Op.getSimpleValueType();
16277 SDValue R = Op.getOperand(0);
16278 SDValue Amt = Op.getOperand(1);
16280 // Optimize shl/srl/sra with constant shift amount.
16281 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
16282 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
16283 uint64_t ShiftAmt = ShiftConst->getZExtValue();
16285 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
16286 (Subtarget->hasInt256() &&
16287 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
16288 (Subtarget->hasAVX512() &&
16289 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
16290 if (Op.getOpcode() == ISD::SHL)
16291 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
16293 if (Op.getOpcode() == ISD::SRL)
16294 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
16296 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
16297 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
16301 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
16302 unsigned NumElts = VT.getVectorNumElements();
16303 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
16305 if (Op.getOpcode() == ISD::SHL) {
16306 // Make a large shift.
16307 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
16309 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
16310 // Zero out the rightmost bits.
16311 SmallVector<SDValue, 32> V(
16312 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
16313 return DAG.getNode(ISD::AND, dl, VT, SHL,
16314 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16316 if (Op.getOpcode() == ISD::SRL) {
16317 // Make a large shift.
16318 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
16320 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
16321 // Zero out the leftmost bits.
16322 SmallVector<SDValue, 32> V(
16323 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
16324 return DAG.getNode(ISD::AND, dl, VT, SRL,
16325 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16327 if (Op.getOpcode() == ISD::SRA) {
16328 if (ShiftAmt == 7) {
16329 // R s>> 7 === R s< 0
16330 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
16331 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
16334 // R s>> a === ((R u>> a) ^ m) - m
16335 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
16336 SmallVector<SDValue, 32> V(NumElts,
16337 DAG.getConstant(128 >> ShiftAmt, dl,
16339 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
16340 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
16341 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
16344 llvm_unreachable("Unknown shift opcode.");
16349 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16350 if (!Subtarget->is64Bit() &&
16351 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
16352 Amt.getOpcode() == ISD::BITCAST &&
16353 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16354 Amt = Amt.getOperand(0);
16355 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16356 VT.getVectorNumElements();
16357 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
16358 uint64_t ShiftAmt = 0;
16359 for (unsigned i = 0; i != Ratio; ++i) {
16360 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
16364 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
16366 // Check remaining shift amounts.
16367 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16368 uint64_t ShAmt = 0;
16369 for (unsigned j = 0; j != Ratio; ++j) {
16370 ConstantSDNode *C =
16371 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
16375 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
16377 if (ShAmt != ShiftAmt)
16380 switch (Op.getOpcode()) {
16382 llvm_unreachable("Unknown shift opcode!");
16384 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
16387 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
16390 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
16398 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
16399 const X86Subtarget* Subtarget) {
16400 MVT VT = Op.getSimpleValueType();
16402 SDValue R = Op.getOperand(0);
16403 SDValue Amt = Op.getOperand(1);
16405 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
16406 VT == MVT::v4i32 || VT == MVT::v8i16 ||
16407 (Subtarget->hasInt256() &&
16408 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
16409 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
16410 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
16412 EVT EltVT = VT.getVectorElementType();
16414 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
16415 // Check if this build_vector node is doing a splat.
16416 // If so, then set BaseShAmt equal to the splat value.
16417 BaseShAmt = BV->getSplatValue();
16418 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
16419 BaseShAmt = SDValue();
16421 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
16422 Amt = Amt.getOperand(0);
16424 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
16425 if (SVN && SVN->isSplat()) {
16426 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
16427 SDValue InVec = Amt.getOperand(0);
16428 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
16429 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
16430 "Unexpected shuffle index found!");
16431 BaseShAmt = InVec.getOperand(SplatIdx);
16432 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
16433 if (ConstantSDNode *C =
16434 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
16435 if (C->getZExtValue() == SplatIdx)
16436 BaseShAmt = InVec.getOperand(1);
16441 // Avoid introducing an extract element from a shuffle.
16442 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
16443 DAG.getIntPtrConstant(SplatIdx, dl));
16447 if (BaseShAmt.getNode()) {
16448 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
16449 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
16450 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
16451 else if (EltVT.bitsLT(MVT::i32))
16452 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
16454 switch (Op.getOpcode()) {
16456 llvm_unreachable("Unknown shift opcode!");
16458 switch (VT.SimpleTy) {
16459 default: return SDValue();
16468 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
16471 switch (VT.SimpleTy) {
16472 default: return SDValue();
16479 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
16482 switch (VT.SimpleTy) {
16483 default: return SDValue();
16492 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
16498 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16499 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
16500 Amt.getOpcode() == ISD::BITCAST &&
16501 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16502 Amt = Amt.getOperand(0);
16503 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16504 VT.getVectorNumElements();
16505 std::vector<SDValue> Vals(Ratio);
16506 for (unsigned i = 0; i != Ratio; ++i)
16507 Vals[i] = Amt.getOperand(i);
16508 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16509 for (unsigned j = 0; j != Ratio; ++j)
16510 if (Vals[j] != Amt.getOperand(i + j))
16513 switch (Op.getOpcode()) {
16515 llvm_unreachable("Unknown shift opcode!");
16517 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
16519 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
16521 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
16528 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16529 SelectionDAG &DAG) {
16530 MVT VT = Op.getSimpleValueType();
16532 SDValue R = Op.getOperand(0);
16533 SDValue Amt = Op.getOperand(1);
16535 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16536 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16538 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
16541 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
16544 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
16547 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
16548 if (Subtarget->hasInt256()) {
16549 if (Op.getOpcode() == ISD::SRL &&
16550 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16551 VT == MVT::v4i64 || VT == MVT::v8i32))
16553 if (Op.getOpcode() == ISD::SHL &&
16554 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16555 VT == MVT::v4i64 || VT == MVT::v8i32))
16557 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
16561 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
16562 // shifts per-lane and then shuffle the partial results back together.
16563 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
16564 // Splat the shift amounts so the scalar shifts above will catch it.
16565 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
16566 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
16567 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
16568 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
16569 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
16572 // If possible, lower this packed shift into a vector multiply instead of
16573 // expanding it into a sequence of scalar shifts.
16574 // Do this only if the vector shift count is a constant build_vector.
16575 if (Op.getOpcode() == ISD::SHL &&
16576 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16577 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16578 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16579 SmallVector<SDValue, 8> Elts;
16580 EVT SVT = VT.getScalarType();
16581 unsigned SVTBits = SVT.getSizeInBits();
16582 const APInt &One = APInt(SVTBits, 1);
16583 unsigned NumElems = VT.getVectorNumElements();
16585 for (unsigned i=0; i !=NumElems; ++i) {
16586 SDValue Op = Amt->getOperand(i);
16587 if (Op->getOpcode() == ISD::UNDEF) {
16588 Elts.push_back(Op);
16592 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16593 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16594 uint64_t ShAmt = C.getZExtValue();
16595 if (ShAmt >= SVTBits) {
16596 Elts.push_back(DAG.getUNDEF(SVT));
16599 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
16601 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16602 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16605 // Lower SHL with variable shift amount.
16606 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16607 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
16609 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
16610 DAG.getConstant(0x3f800000U, dl, VT));
16611 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16612 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16613 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16616 // If possible, lower this shift as a sequence of two shifts by
16617 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16619 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16621 // Could be rewritten as:
16622 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16624 // The advantage is that the two shifts from the example would be
16625 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16626 // the vector shift into four scalar shifts plus four pairs of vector
16628 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16629 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16630 unsigned TargetOpcode = X86ISD::MOVSS;
16631 bool CanBeSimplified;
16632 // The splat value for the first packed shift (the 'X' from the example).
16633 SDValue Amt1 = Amt->getOperand(0);
16634 // The splat value for the second packed shift (the 'Y' from the example).
16635 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16636 Amt->getOperand(2);
16638 // See if it is possible to replace this node with a sequence of
16639 // two shifts followed by a MOVSS/MOVSD
16640 if (VT == MVT::v4i32) {
16641 // Check if it is legal to use a MOVSS.
16642 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16643 Amt2 == Amt->getOperand(3);
16644 if (!CanBeSimplified) {
16645 // Otherwise, check if we can still simplify this node using a MOVSD.
16646 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16647 Amt->getOperand(2) == Amt->getOperand(3);
16648 TargetOpcode = X86ISD::MOVSD;
16649 Amt2 = Amt->getOperand(2);
16652 // Do similar checks for the case where the machine value type
16654 CanBeSimplified = Amt1 == Amt->getOperand(1);
16655 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16656 CanBeSimplified = Amt2 == Amt->getOperand(i);
16658 if (!CanBeSimplified) {
16659 TargetOpcode = X86ISD::MOVSD;
16660 CanBeSimplified = true;
16661 Amt2 = Amt->getOperand(4);
16662 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16663 CanBeSimplified = Amt1 == Amt->getOperand(i);
16664 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16665 CanBeSimplified = Amt2 == Amt->getOperand(j);
16669 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16670 isa<ConstantSDNode>(Amt2)) {
16671 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16672 EVT CastVT = MVT::v4i32;
16674 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
16675 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16677 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
16678 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16679 if (TargetOpcode == X86ISD::MOVSD)
16680 CastVT = MVT::v2i64;
16681 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16682 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16683 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16685 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16689 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16690 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
16691 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, dl, VT));
16693 SDValue VSelM = DAG.getConstant(0x80, dl, VT);
16694 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16695 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16697 // r = VSELECT(r, shl(r, 4), a);
16698 SDValue M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(4, dl, VT));
16699 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16702 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16703 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16704 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16706 // r = VSELECT(r, shl(r, 2), a);
16707 M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(2, dl, VT));
16708 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16711 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16712 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16713 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16715 // return VSELECT(r, r+r, a);
16716 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16717 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16721 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16722 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16723 // solution better.
16724 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16725 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16727 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16728 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16729 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16730 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16731 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16734 // Decompose 256-bit shifts into smaller 128-bit shifts.
16735 if (VT.is256BitVector()) {
16736 unsigned NumElems = VT.getVectorNumElements();
16737 MVT EltVT = VT.getVectorElementType();
16738 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16740 // Extract the two vectors
16741 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16742 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16744 // Recreate the shift amount vectors
16745 SDValue Amt1, Amt2;
16746 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16747 // Constant shift amount
16748 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
16749 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
16750 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
16752 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16753 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16755 // Variable shift amount
16756 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16757 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16760 // Issue new vector shifts for the smaller types
16761 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16762 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16764 // Concatenate the result back
16765 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16771 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16772 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16773 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16774 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16775 // has only one use.
16776 SDNode *N = Op.getNode();
16777 SDValue LHS = N->getOperand(0);
16778 SDValue RHS = N->getOperand(1);
16779 unsigned BaseOp = 0;
16782 switch (Op.getOpcode()) {
16783 default: llvm_unreachable("Unknown ovf instruction!");
16785 // A subtract of one will be selected as a INC. Note that INC doesn't
16786 // set CF, so we can't do this for UADDO.
16787 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16789 BaseOp = X86ISD::INC;
16790 Cond = X86::COND_O;
16793 BaseOp = X86ISD::ADD;
16794 Cond = X86::COND_O;
16797 BaseOp = X86ISD::ADD;
16798 Cond = X86::COND_B;
16801 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16802 // set CF, so we can't do this for USUBO.
16803 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16805 BaseOp = X86ISD::DEC;
16806 Cond = X86::COND_O;
16809 BaseOp = X86ISD::SUB;
16810 Cond = X86::COND_O;
16813 BaseOp = X86ISD::SUB;
16814 Cond = X86::COND_B;
16817 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16818 Cond = X86::COND_O;
16820 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16821 if (N->getValueType(0) == MVT::i8) {
16822 BaseOp = X86ISD::UMUL8;
16823 Cond = X86::COND_O;
16826 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16828 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16831 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16832 DAG.getConstant(X86::COND_O, DL, MVT::i32),
16833 SDValue(Sum.getNode(), 2));
16835 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16839 // Also sets EFLAGS.
16840 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16841 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16844 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16845 DAG.getConstant(Cond, DL, MVT::i32),
16846 SDValue(Sum.getNode(), 1));
16848 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16851 /// Returns true if the operand type is exactly twice the native width, and
16852 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16853 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16854 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16855 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16856 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16859 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16860 else if (OpWidth == 128)
16861 return Subtarget->hasCmpxchg16b();
16866 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16867 return needsCmpXchgNb(SI->getValueOperand()->getType());
16870 // Note: this turns large loads into lock cmpxchg8b/16b.
16871 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16872 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16873 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16874 return needsCmpXchgNb(PTy->getElementType());
16877 TargetLoweringBase::AtomicRMWExpansionKind
16878 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16879 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16880 const Type *MemType = AI->getType();
16882 // If the operand is too big, we must see if cmpxchg8/16b is available
16883 // and default to library calls otherwise.
16884 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
16885 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
16886 : AtomicRMWExpansionKind::None;
16889 AtomicRMWInst::BinOp Op = AI->getOperation();
16892 llvm_unreachable("Unknown atomic operation");
16893 case AtomicRMWInst::Xchg:
16894 case AtomicRMWInst::Add:
16895 case AtomicRMWInst::Sub:
16896 // It's better to use xadd, xsub or xchg for these in all cases.
16897 return AtomicRMWExpansionKind::None;
16898 case AtomicRMWInst::Or:
16899 case AtomicRMWInst::And:
16900 case AtomicRMWInst::Xor:
16901 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16902 // prefix to a normal instruction for these operations.
16903 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
16904 : AtomicRMWExpansionKind::None;
16905 case AtomicRMWInst::Nand:
16906 case AtomicRMWInst::Max:
16907 case AtomicRMWInst::Min:
16908 case AtomicRMWInst::UMax:
16909 case AtomicRMWInst::UMin:
16910 // These always require a non-trivial set of data operations on x86. We must
16911 // use a cmpxchg loop.
16912 return AtomicRMWExpansionKind::CmpXChg;
16916 static bool hasMFENCE(const X86Subtarget& Subtarget) {
16917 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
16918 // no-sse2). There isn't any reason to disable it if the target processor
16920 return Subtarget.hasSSE2() || Subtarget.is64Bit();
16924 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
16925 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16926 const Type *MemType = AI->getType();
16927 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
16928 // there is no benefit in turning such RMWs into loads, and it is actually
16929 // harmful as it introduces a mfence.
16930 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16933 auto Builder = IRBuilder<>(AI);
16934 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16935 auto SynchScope = AI->getSynchScope();
16936 // We must restrict the ordering to avoid generating loads with Release or
16937 // ReleaseAcquire orderings.
16938 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
16939 auto Ptr = AI->getPointerOperand();
16941 // Before the load we need a fence. Here is an example lifted from
16942 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
16945 // x.store(1, relaxed);
16946 // r1 = y.fetch_add(0, release);
16948 // y.fetch_add(42, acquire);
16949 // r2 = x.load(relaxed);
16950 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
16951 // lowered to just a load without a fence. A mfence flushes the store buffer,
16952 // making the optimization clearly correct.
16953 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
16954 // otherwise, we might be able to be more agressive on relaxed idempotent
16955 // rmw. In practice, they do not look useful, so we don't try to be
16956 // especially clever.
16957 if (SynchScope == SingleThread) {
16958 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
16959 // the IR level, so we must wrap it in an intrinsic.
16961 } else if (hasMFENCE(*Subtarget)) {
16962 Function *MFence = llvm::Intrinsic::getDeclaration(M,
16963 Intrinsic::x86_sse2_mfence);
16964 Builder.CreateCall(MFence);
16966 // FIXME: it might make sense to use a locked operation here but on a
16967 // different cache-line to prevent cache-line bouncing. In practice it
16968 // is probably a small win, and x86 processors without mfence are rare
16969 // enough that we do not bother.
16973 // Finally we can emit the atomic load.
16974 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
16975 AI->getType()->getPrimitiveSizeInBits());
16976 Loaded->setAtomic(Order, SynchScope);
16977 AI->replaceAllUsesWith(Loaded);
16978 AI->eraseFromParent();
16982 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
16983 SelectionDAG &DAG) {
16985 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
16986 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
16987 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
16988 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
16990 // The only fence that needs an instruction is a sequentially-consistent
16991 // cross-thread fence.
16992 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
16993 if (hasMFENCE(*Subtarget))
16994 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
16996 SDValue Chain = Op.getOperand(0);
16997 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
16999 DAG.getRegister(X86::ESP, MVT::i32), // Base
17000 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
17001 DAG.getRegister(0, MVT::i32), // Index
17002 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
17003 DAG.getRegister(0, MVT::i32), // Segment.
17007 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
17008 return SDValue(Res, 0);
17011 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
17012 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17015 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17016 SelectionDAG &DAG) {
17017 MVT T = Op.getSimpleValueType();
17021 switch(T.SimpleTy) {
17022 default: llvm_unreachable("Invalid value type!");
17023 case MVT::i8: Reg = X86::AL; size = 1; break;
17024 case MVT::i16: Reg = X86::AX; size = 2; break;
17025 case MVT::i32: Reg = X86::EAX; size = 4; break;
17027 assert(Subtarget->is64Bit() && "Node not type legal!");
17028 Reg = X86::RAX; size = 8;
17031 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17032 Op.getOperand(2), SDValue());
17033 SDValue Ops[] = { cpIn.getValue(0),
17036 DAG.getTargetConstant(size, DL, MVT::i8),
17037 cpIn.getValue(1) };
17038 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17039 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17040 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17044 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17045 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17046 MVT::i32, cpOut.getValue(2));
17047 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17048 DAG.getConstant(X86::COND_E, DL, MVT::i8),
17051 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17052 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17053 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17057 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17058 SelectionDAG &DAG) {
17059 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17060 MVT DstVT = Op.getSimpleValueType();
17062 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17063 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17064 if (DstVT != MVT::f64)
17065 // This conversion needs to be expanded.
17068 SDValue InVec = Op->getOperand(0);
17070 unsigned NumElts = SrcVT.getVectorNumElements();
17071 EVT SVT = SrcVT.getVectorElementType();
17073 // Widen the vector in input in the case of MVT::v2i32.
17074 // Example: from MVT::v2i32 to MVT::v4i32.
17075 SmallVector<SDValue, 16> Elts;
17076 for (unsigned i = 0, e = NumElts; i != e; ++i)
17077 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17078 DAG.getIntPtrConstant(i, dl)));
17080 // Explicitly mark the extra elements as Undef.
17081 Elts.append(NumElts, DAG.getUNDEF(SVT));
17083 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17084 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17085 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
17086 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17087 DAG.getIntPtrConstant(0, dl));
17090 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17091 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17092 assert((DstVT == MVT::i64 ||
17093 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17094 "Unexpected custom BITCAST");
17095 // i64 <=> MMX conversions are Legal.
17096 if (SrcVT==MVT::i64 && DstVT.isVector())
17098 if (DstVT==MVT::i64 && SrcVT.isVector())
17100 // MMX <=> MMX conversions are Legal.
17101 if (SrcVT.isVector() && DstVT.isVector())
17103 // All other conversions need to be expanded.
17107 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17108 SelectionDAG &DAG) {
17109 SDNode *Node = Op.getNode();
17112 Op = Op.getOperand(0);
17113 EVT VT = Op.getValueType();
17114 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17115 "CTPOP lowering only implemented for 128/256-bit wide vector types");
17117 unsigned NumElts = VT.getVectorNumElements();
17118 EVT EltVT = VT.getVectorElementType();
17119 unsigned Len = EltVT.getSizeInBits();
17121 // This is the vectorized version of the "best" algorithm from
17122 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
17123 // with a minor tweak to use a series of adds + shifts instead of vector
17124 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
17126 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
17127 // v8i32 => Always profitable
17129 // FIXME: There a couple of possible improvements:
17131 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
17132 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
17134 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
17135 "CTPOP not implemented for this vector element type.");
17137 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
17138 // extra legalization.
17139 bool NeedsBitcast = EltVT == MVT::i32;
17140 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
17142 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl,
17144 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl,
17146 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl,
17149 // v = v - ((v >> 1) & 0x55555555...)
17150 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, dl, EltVT));
17151 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
17152 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
17154 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17156 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
17157 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
17159 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
17161 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
17162 if (VT != And.getValueType())
17163 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17164 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
17166 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
17167 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
17168 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
17169 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, dl, EltVT));
17170 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
17172 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
17173 if (NeedsBitcast) {
17174 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17175 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
17176 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
17179 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
17180 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
17181 if (VT != AndRHS.getValueType()) {
17182 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
17183 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
17185 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
17187 // v = (v + (v >> 4)) & 0x0F0F0F0F...
17188 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, dl, EltVT));
17189 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
17190 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
17191 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17193 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
17194 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
17195 if (NeedsBitcast) {
17196 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17197 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
17199 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
17200 if (VT != And.getValueType())
17201 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17203 // The algorithm mentioned above uses:
17204 // v = (v * 0x01010101...) >> (Len - 8)
17206 // Change it to use vector adds + vector shifts which yield faster results on
17207 // Haswell than using vector integer multiplication.
17209 // For i32 elements:
17210 // v = v + (v >> 8)
17211 // v = v + (v >> 16)
17213 // For i64 elements:
17214 // v = v + (v >> 8)
17215 // v = v + (v >> 16)
17216 // v = v + (v >> 32)
17219 SmallVector<SDValue, 8> Csts;
17220 for (unsigned i = 8; i <= Len/2; i *= 2) {
17221 Csts.assign(NumElts, DAG.getConstant(i, dl, EltVT));
17222 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
17223 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
17224 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17228 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
17229 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), dl,
17231 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
17232 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
17233 if (NeedsBitcast) {
17234 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17235 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
17237 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
17238 if (VT != And.getValueType())
17239 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17244 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
17245 SDNode *Node = Op.getNode();
17247 EVT T = Node->getValueType(0);
17248 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
17249 DAG.getConstant(0, dl, T), Node->getOperand(2));
17250 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
17251 cast<AtomicSDNode>(Node)->getMemoryVT(),
17252 Node->getOperand(0),
17253 Node->getOperand(1), negOp,
17254 cast<AtomicSDNode>(Node)->getMemOperand(),
17255 cast<AtomicSDNode>(Node)->getOrdering(),
17256 cast<AtomicSDNode>(Node)->getSynchScope());
17259 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
17260 SDNode *Node = Op.getNode();
17262 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
17264 // Convert seq_cst store -> xchg
17265 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
17266 // FIXME: On 32-bit, store -> fist or movq would be more efficient
17267 // (The only way to get a 16-byte store is cmpxchg16b)
17268 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
17269 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
17270 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17271 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
17272 cast<AtomicSDNode>(Node)->getMemoryVT(),
17273 Node->getOperand(0),
17274 Node->getOperand(1), Node->getOperand(2),
17275 cast<AtomicSDNode>(Node)->getMemOperand(),
17276 cast<AtomicSDNode>(Node)->getOrdering(),
17277 cast<AtomicSDNode>(Node)->getSynchScope());
17278 return Swap.getValue(1);
17280 // Other atomic stores have a simple pattern.
17284 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
17285 EVT VT = Op.getNode()->getSimpleValueType(0);
17287 // Let legalize expand this if it isn't a legal type yet.
17288 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17291 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17294 bool ExtraOp = false;
17295 switch (Op.getOpcode()) {
17296 default: llvm_unreachable("Invalid code");
17297 case ISD::ADDC: Opc = X86ISD::ADD; break;
17298 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
17299 case ISD::SUBC: Opc = X86ISD::SUB; break;
17300 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
17304 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17306 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17307 Op.getOperand(1), Op.getOperand(2));
17310 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
17311 SelectionDAG &DAG) {
17312 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
17314 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
17315 // which returns the values as { float, float } (in XMM0) or
17316 // { double, double } (which is returned in XMM0, XMM1).
17318 SDValue Arg = Op.getOperand(0);
17319 EVT ArgVT = Arg.getValueType();
17320 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17322 TargetLowering::ArgListTy Args;
17323 TargetLowering::ArgListEntry Entry;
17327 Entry.isSExt = false;
17328 Entry.isZExt = false;
17329 Args.push_back(Entry);
17331 bool isF64 = ArgVT == MVT::f64;
17332 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
17333 // the small struct {f32, f32} is returned in (eax, edx). For f64,
17334 // the results are returned via SRet in memory.
17335 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
17336 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17337 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
17339 Type *RetTy = isF64
17340 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
17341 : (Type*)VectorType::get(ArgTy, 4);
17343 TargetLowering::CallLoweringInfo CLI(DAG);
17344 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
17345 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
17347 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
17350 // Returned in xmm0 and xmm1.
17351 return CallResult.first;
17353 // Returned in bits 0:31 and 32:64 xmm0.
17354 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17355 CallResult.first, DAG.getIntPtrConstant(0, dl));
17356 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17357 CallResult.first, DAG.getIntPtrConstant(1, dl));
17358 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
17359 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
17362 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
17363 SelectionDAG &DAG) {
17364 assert(Subtarget->hasAVX512() &&
17365 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17367 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
17368 EVT VT = N->getValue().getValueType();
17369 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
17372 // X86 scatter kills mask register, so its type should be added to
17373 // the list of return values
17374 if (N->getNumValues() == 1) {
17375 SDValue Index = N->getIndex();
17376 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17377 !Index.getValueType().is512BitVector())
17378 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17380 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
17381 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17382 N->getOperand(3), Index };
17384 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
17385 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
17386 return SDValue(NewScatter.getNode(), 0);
17391 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
17392 SelectionDAG &DAG) {
17393 assert(Subtarget->hasAVX512() &&
17394 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17396 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
17397 EVT VT = Op.getValueType();
17398 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
17401 SDValue Index = N->getIndex();
17402 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17403 !Index.getValueType().is512BitVector()) {
17404 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17405 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17406 N->getOperand(3), Index };
17407 DAG.UpdateNodeOperands(N, Ops);
17412 /// LowerOperation - Provide custom lowering hooks for some operations.
17414 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
17415 switch (Op.getOpcode()) {
17416 default: llvm_unreachable("Should not custom lower this!");
17417 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
17418 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
17419 return LowerCMP_SWAP(Op, Subtarget, DAG);
17420 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
17421 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
17422 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
17423 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
17424 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
17425 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
17426 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
17427 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
17428 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
17429 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
17430 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
17431 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
17432 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
17433 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
17434 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
17435 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
17436 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
17437 case ISD::SHL_PARTS:
17438 case ISD::SRA_PARTS:
17439 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
17440 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
17441 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
17442 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
17443 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
17444 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
17445 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
17446 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
17447 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
17448 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
17449 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
17451 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
17452 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
17453 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
17454 case ISD::SETCC: return LowerSETCC(Op, DAG);
17455 case ISD::SELECT: return LowerSELECT(Op, DAG);
17456 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
17457 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
17458 case ISD::VASTART: return LowerVASTART(Op, DAG);
17459 case ISD::VAARG: return LowerVAARG(Op, DAG);
17460 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
17461 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
17462 case ISD::INTRINSIC_VOID:
17463 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
17464 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
17465 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
17466 case ISD::FRAME_TO_ARGS_OFFSET:
17467 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
17468 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
17469 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
17470 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
17471 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
17472 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
17473 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
17474 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
17475 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
17476 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
17477 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
17478 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
17479 case ISD::UMUL_LOHI:
17480 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
17483 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
17489 case ISD::UMULO: return LowerXALUO(Op, DAG);
17490 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
17491 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
17495 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
17496 case ISD::ADD: return LowerADD(Op, DAG);
17497 case ISD::SUB: return LowerSUB(Op, DAG);
17498 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
17499 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
17500 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
17504 /// ReplaceNodeResults - Replace a node with an illegal result type
17505 /// with a new node built out of custom code.
17506 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
17507 SmallVectorImpl<SDValue>&Results,
17508 SelectionDAG &DAG) const {
17510 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17511 switch (N->getOpcode()) {
17513 llvm_unreachable("Do not know how to custom type legalize this operation!");
17514 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17515 case X86ISD::FMINC:
17517 case X86ISD::FMAXC:
17518 case X86ISD::FMAX: {
17519 EVT VT = N->getValueType(0);
17520 if (VT != MVT::v2f32)
17521 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17522 SDValue UNDEF = DAG.getUNDEF(VT);
17523 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17524 N->getOperand(0), UNDEF);
17525 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17526 N->getOperand(1), UNDEF);
17527 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17530 case ISD::SIGN_EXTEND_INREG:
17535 // We don't want to expand or promote these.
17542 case ISD::UDIVREM: {
17543 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17544 Results.push_back(V);
17547 case ISD::FP_TO_SINT:
17548 case ISD::FP_TO_UINT: {
17549 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17551 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17554 std::pair<SDValue,SDValue> Vals =
17555 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17556 SDValue FIST = Vals.first, StackSlot = Vals.second;
17557 if (FIST.getNode()) {
17558 EVT VT = N->getValueType(0);
17559 // Return a load from the stack slot.
17560 if (StackSlot.getNode())
17561 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17562 MachinePointerInfo(),
17563 false, false, false, 0));
17565 Results.push_back(FIST);
17569 case ISD::UINT_TO_FP: {
17570 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17571 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17572 N->getValueType(0) != MVT::v2f32)
17574 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17576 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
17578 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17579 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17580 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17581 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17582 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17583 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17586 case ISD::FP_ROUND: {
17587 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17589 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17590 Results.push_back(V);
17593 case ISD::INTRINSIC_W_CHAIN: {
17594 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17596 default : llvm_unreachable("Do not know how to custom type "
17597 "legalize this intrinsic operation!");
17598 case Intrinsic::x86_rdtsc:
17599 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17601 case Intrinsic::x86_rdtscp:
17602 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17604 case Intrinsic::x86_rdpmc:
17605 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17608 case ISD::READCYCLECOUNTER: {
17609 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17612 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17613 EVT T = N->getValueType(0);
17614 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17615 bool Regs64bit = T == MVT::i128;
17616 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17617 SDValue cpInL, cpInH;
17618 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17619 DAG.getConstant(0, dl, HalfT));
17620 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17621 DAG.getConstant(1, dl, HalfT));
17622 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17623 Regs64bit ? X86::RAX : X86::EAX,
17625 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17626 Regs64bit ? X86::RDX : X86::EDX,
17627 cpInH, cpInL.getValue(1));
17628 SDValue swapInL, swapInH;
17629 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17630 DAG.getConstant(0, dl, HalfT));
17631 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17632 DAG.getConstant(1, dl, HalfT));
17633 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17634 Regs64bit ? X86::RBX : X86::EBX,
17635 swapInL, cpInH.getValue(1));
17636 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17637 Regs64bit ? X86::RCX : X86::ECX,
17638 swapInH, swapInL.getValue(1));
17639 SDValue Ops[] = { swapInH.getValue(0),
17641 swapInH.getValue(1) };
17642 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17643 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17644 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17645 X86ISD::LCMPXCHG8_DAG;
17646 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17647 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17648 Regs64bit ? X86::RAX : X86::EAX,
17649 HalfT, Result.getValue(1));
17650 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17651 Regs64bit ? X86::RDX : X86::EDX,
17652 HalfT, cpOutL.getValue(2));
17653 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17655 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17656 MVT::i32, cpOutH.getValue(2));
17658 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17659 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
17660 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17662 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17663 Results.push_back(Success);
17664 Results.push_back(EFLAGS.getValue(1));
17667 case ISD::ATOMIC_SWAP:
17668 case ISD::ATOMIC_LOAD_ADD:
17669 case ISD::ATOMIC_LOAD_SUB:
17670 case ISD::ATOMIC_LOAD_AND:
17671 case ISD::ATOMIC_LOAD_OR:
17672 case ISD::ATOMIC_LOAD_XOR:
17673 case ISD::ATOMIC_LOAD_NAND:
17674 case ISD::ATOMIC_LOAD_MIN:
17675 case ISD::ATOMIC_LOAD_MAX:
17676 case ISD::ATOMIC_LOAD_UMIN:
17677 case ISD::ATOMIC_LOAD_UMAX:
17678 case ISD::ATOMIC_LOAD: {
17679 // Delegate to generic TypeLegalization. Situations we can really handle
17680 // should have already been dealt with by AtomicExpandPass.cpp.
17683 case ISD::BITCAST: {
17684 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17685 EVT DstVT = N->getValueType(0);
17686 EVT SrcVT = N->getOperand(0)->getValueType(0);
17688 if (SrcVT != MVT::f64 ||
17689 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17692 unsigned NumElts = DstVT.getVectorNumElements();
17693 EVT SVT = DstVT.getVectorElementType();
17694 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17695 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17696 MVT::v2f64, N->getOperand(0));
17697 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17699 if (ExperimentalVectorWideningLegalization) {
17700 // If we are legalizing vectors by widening, we already have the desired
17701 // legal vector type, just return it.
17702 Results.push_back(ToVecInt);
17706 SmallVector<SDValue, 8> Elts;
17707 for (unsigned i = 0, e = NumElts; i != e; ++i)
17708 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17709 ToVecInt, DAG.getIntPtrConstant(i, dl)));
17711 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17716 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17718 default: return nullptr;
17719 case X86ISD::BSF: return "X86ISD::BSF";
17720 case X86ISD::BSR: return "X86ISD::BSR";
17721 case X86ISD::SHLD: return "X86ISD::SHLD";
17722 case X86ISD::SHRD: return "X86ISD::SHRD";
17723 case X86ISD::FAND: return "X86ISD::FAND";
17724 case X86ISD::FANDN: return "X86ISD::FANDN";
17725 case X86ISD::FOR: return "X86ISD::FOR";
17726 case X86ISD::FXOR: return "X86ISD::FXOR";
17727 case X86ISD::FSRL: return "X86ISD::FSRL";
17728 case X86ISD::FILD: return "X86ISD::FILD";
17729 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17730 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17731 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17732 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17733 case X86ISD::FLD: return "X86ISD::FLD";
17734 case X86ISD::FST: return "X86ISD::FST";
17735 case X86ISD::CALL: return "X86ISD::CALL";
17736 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17737 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17738 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17739 case X86ISD::BT: return "X86ISD::BT";
17740 case X86ISD::CMP: return "X86ISD::CMP";
17741 case X86ISD::COMI: return "X86ISD::COMI";
17742 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17743 case X86ISD::CMPM: return "X86ISD::CMPM";
17744 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17745 case X86ISD::SETCC: return "X86ISD::SETCC";
17746 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17747 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17748 case X86ISD::CMOV: return "X86ISD::CMOV";
17749 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17750 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17751 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17752 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17753 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17754 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17755 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17756 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17757 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17758 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17759 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17760 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17761 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17762 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17763 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17764 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17765 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17766 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17767 case X86ISD::HADD: return "X86ISD::HADD";
17768 case X86ISD::HSUB: return "X86ISD::HSUB";
17769 case X86ISD::FHADD: return "X86ISD::FHADD";
17770 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17771 case X86ISD::UMAX: return "X86ISD::UMAX";
17772 case X86ISD::UMIN: return "X86ISD::UMIN";
17773 case X86ISD::SMAX: return "X86ISD::SMAX";
17774 case X86ISD::SMIN: return "X86ISD::SMIN";
17775 case X86ISD::FMAX: return "X86ISD::FMAX";
17776 case X86ISD::FMIN: return "X86ISD::FMIN";
17777 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17778 case X86ISD::FMINC: return "X86ISD::FMINC";
17779 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17780 case X86ISD::FRCP: return "X86ISD::FRCP";
17781 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17782 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17783 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17784 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17785 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17786 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17787 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17788 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17789 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17790 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17791 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17792 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17793 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17794 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17795 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17796 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17797 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17798 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17799 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17800 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17801 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17802 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17803 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17804 case X86ISD::VSHL: return "X86ISD::VSHL";
17805 case X86ISD::VSRL: return "X86ISD::VSRL";
17806 case X86ISD::VSRA: return "X86ISD::VSRA";
17807 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17808 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17809 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17810 case X86ISD::CMPP: return "X86ISD::CMPP";
17811 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17812 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17813 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17814 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17815 case X86ISD::ADD: return "X86ISD::ADD";
17816 case X86ISD::SUB: return "X86ISD::SUB";
17817 case X86ISD::ADC: return "X86ISD::ADC";
17818 case X86ISD::SBB: return "X86ISD::SBB";
17819 case X86ISD::SMUL: return "X86ISD::SMUL";
17820 case X86ISD::UMUL: return "X86ISD::UMUL";
17821 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17822 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17823 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17824 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17825 case X86ISD::INC: return "X86ISD::INC";
17826 case X86ISD::DEC: return "X86ISD::DEC";
17827 case X86ISD::OR: return "X86ISD::OR";
17828 case X86ISD::XOR: return "X86ISD::XOR";
17829 case X86ISD::AND: return "X86ISD::AND";
17830 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17831 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17832 case X86ISD::PTEST: return "X86ISD::PTEST";
17833 case X86ISD::TESTP: return "X86ISD::TESTP";
17834 case X86ISD::TESTM: return "X86ISD::TESTM";
17835 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17836 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17837 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17838 case X86ISD::PACKUS: return "X86ISD::PACKUS";
17839 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
17840 case X86ISD::VALIGN: return "X86ISD::VALIGN";
17841 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
17842 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
17843 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
17844 case X86ISD::SHUFP: return "X86ISD::SHUFP";
17845 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
17846 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
17847 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
17848 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
17849 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
17850 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
17851 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
17852 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
17853 case X86ISD::MOVSD: return "X86ISD::MOVSD";
17854 case X86ISD::MOVSS: return "X86ISD::MOVSS";
17855 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
17856 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
17857 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
17858 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
17859 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
17860 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
17861 case X86ISD::VPERMV: return "X86ISD::VPERMV";
17862 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
17863 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
17864 case X86ISD::VPERMI: return "X86ISD::VPERMI";
17865 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
17866 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
17867 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
17868 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
17869 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
17870 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
17871 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
17872 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
17873 case X86ISD::SAHF: return "X86ISD::SAHF";
17874 case X86ISD::RDRAND: return "X86ISD::RDRAND";
17875 case X86ISD::RDSEED: return "X86ISD::RDSEED";
17876 case X86ISD::FMADD: return "X86ISD::FMADD";
17877 case X86ISD::FMSUB: return "X86ISD::FMSUB";
17878 case X86ISD::FNMADD: return "X86ISD::FNMADD";
17879 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
17880 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
17881 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
17882 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
17883 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
17884 case X86ISD::XTEST: return "X86ISD::XTEST";
17885 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
17886 case X86ISD::EXPAND: return "X86ISD::EXPAND";
17887 case X86ISD::SELECT: return "X86ISD::SELECT";
17888 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
17889 case X86ISD::RCP28: return "X86ISD::RCP28";
17890 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
17891 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
17892 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
17893 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
17894 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
17898 // isLegalAddressingMode - Return true if the addressing mode represented
17899 // by AM is legal for this target, for a load/store of the specified type.
17900 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
17902 // X86 supports extremely general addressing modes.
17903 CodeModel::Model M = getTargetMachine().getCodeModel();
17904 Reloc::Model R = getTargetMachine().getRelocationModel();
17906 // X86 allows a sign-extended 32-bit immediate field as a displacement.
17907 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
17912 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
17914 // If a reference to this global requires an extra load, we can't fold it.
17915 if (isGlobalStubReference(GVFlags))
17918 // If BaseGV requires a register for the PIC base, we cannot also have a
17919 // BaseReg specified.
17920 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
17923 // If lower 4G is not available, then we must use rip-relative addressing.
17924 if ((M != CodeModel::Small || R != Reloc::Static) &&
17925 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
17929 switch (AM.Scale) {
17935 // These scales always work.
17940 // These scales are formed with basereg+scalereg. Only accept if there is
17945 default: // Other stuff never works.
17952 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
17953 unsigned Bits = Ty->getScalarSizeInBits();
17955 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
17956 // particularly cheaper than those without.
17960 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
17961 // variable shifts just as cheap as scalar ones.
17962 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
17965 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
17966 // fully general vector.
17970 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
17971 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17973 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
17974 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
17975 return NumBits1 > NumBits2;
17978 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
17979 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17982 if (!isTypeLegal(EVT::getEVT(Ty1)))
17985 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
17987 // Assuming the caller doesn't have a zeroext or signext return parameter,
17988 // truncation all the way down to i1 is valid.
17992 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
17993 return isInt<32>(Imm);
17996 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
17997 // Can also use sub to handle negated immediates.
17998 return isInt<32>(Imm);
18001 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
18002 if (!VT1.isInteger() || !VT2.isInteger())
18004 unsigned NumBits1 = VT1.getSizeInBits();
18005 unsigned NumBits2 = VT2.getSizeInBits();
18006 return NumBits1 > NumBits2;
18009 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
18010 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18011 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
18014 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
18015 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18016 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
18019 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
18020 EVT VT1 = Val.getValueType();
18021 if (isZExtFree(VT1, VT2))
18024 if (Val.getOpcode() != ISD::LOAD)
18027 if (!VT1.isSimple() || !VT1.isInteger() ||
18028 !VT2.isSimple() || !VT2.isInteger())
18031 switch (VT1.getSimpleVT().SimpleTy) {
18036 // X86 has 8, 16, and 32-bit zero-extending loads.
18043 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
18046 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
18047 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
18050 VT = VT.getScalarType();
18052 if (!VT.isSimple())
18055 switch (VT.getSimpleVT().SimpleTy) {
18066 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
18067 // i16 instructions are longer (0x66 prefix) and potentially slower.
18068 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18071 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18072 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18073 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18074 /// are assumed to be legal.
18076 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18078 if (!VT.isSimple())
18081 // Very little shuffling can be done for 64-bit vectors right now.
18082 if (VT.getSizeInBits() == 64)
18085 // We only care that the types being shuffled are legal. The lowering can
18086 // handle any possible shuffle mask that results.
18087 return isTypeLegal(VT.getSimpleVT());
18091 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18093 // Just delegate to the generic legality, clear masks aren't special.
18094 return isShuffleMaskLegal(Mask, VT);
18097 //===----------------------------------------------------------------------===//
18098 // X86 Scheduler Hooks
18099 //===----------------------------------------------------------------------===//
18101 /// Utility function to emit xbegin specifying the start of an RTM region.
18102 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18103 const TargetInstrInfo *TII) {
18104 DebugLoc DL = MI->getDebugLoc();
18106 const BasicBlock *BB = MBB->getBasicBlock();
18107 MachineFunction::iterator I = MBB;
18110 // For the v = xbegin(), we generate
18121 MachineBasicBlock *thisMBB = MBB;
18122 MachineFunction *MF = MBB->getParent();
18123 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18124 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18125 MF->insert(I, mainMBB);
18126 MF->insert(I, sinkMBB);
18128 // Transfer the remainder of BB and its successor edges to sinkMBB.
18129 sinkMBB->splice(sinkMBB->begin(), MBB,
18130 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18131 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18135 // # fallthrough to mainMBB
18136 // # abortion to sinkMBB
18137 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18138 thisMBB->addSuccessor(mainMBB);
18139 thisMBB->addSuccessor(sinkMBB);
18143 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18144 mainMBB->addSuccessor(sinkMBB);
18147 // EAX is live into the sinkMBB
18148 sinkMBB->addLiveIn(X86::EAX);
18149 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18150 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18153 MI->eraseFromParent();
18157 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18158 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18159 // in the .td file.
18160 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18161 const TargetInstrInfo *TII) {
18163 switch (MI->getOpcode()) {
18164 default: llvm_unreachable("illegal opcode!");
18165 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18166 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18167 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18168 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18169 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18170 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18171 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18172 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18175 DebugLoc dl = MI->getDebugLoc();
18176 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18178 unsigned NumArgs = MI->getNumOperands();
18179 for (unsigned i = 1; i < NumArgs; ++i) {
18180 MachineOperand &Op = MI->getOperand(i);
18181 if (!(Op.isReg() && Op.isImplicit()))
18182 MIB.addOperand(Op);
18184 if (MI->hasOneMemOperand())
18185 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18187 BuildMI(*BB, MI, dl,
18188 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18189 .addReg(X86::XMM0);
18191 MI->eraseFromParent();
18195 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18196 // defs in an instruction pattern
18197 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18198 const TargetInstrInfo *TII) {
18200 switch (MI->getOpcode()) {
18201 default: llvm_unreachable("illegal opcode!");
18202 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18203 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18204 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18205 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18206 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18207 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18208 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18209 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18212 DebugLoc dl = MI->getDebugLoc();
18213 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18215 unsigned NumArgs = MI->getNumOperands(); // remove the results
18216 for (unsigned i = 1; i < NumArgs; ++i) {
18217 MachineOperand &Op = MI->getOperand(i);
18218 if (!(Op.isReg() && Op.isImplicit()))
18219 MIB.addOperand(Op);
18221 if (MI->hasOneMemOperand())
18222 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18224 BuildMI(*BB, MI, dl,
18225 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18228 MI->eraseFromParent();
18232 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18233 const X86Subtarget *Subtarget) {
18234 DebugLoc dl = MI->getDebugLoc();
18235 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18236 // Address into RAX/EAX, other two args into ECX, EDX.
18237 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18238 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18239 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18240 for (int i = 0; i < X86::AddrNumOperands; ++i)
18241 MIB.addOperand(MI->getOperand(i));
18243 unsigned ValOps = X86::AddrNumOperands;
18244 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18245 .addReg(MI->getOperand(ValOps).getReg());
18246 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18247 .addReg(MI->getOperand(ValOps+1).getReg());
18249 // The instruction doesn't actually take any operands though.
18250 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18252 MI->eraseFromParent(); // The pseudo is gone now.
18256 MachineBasicBlock *
18257 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
18258 MachineBasicBlock *MBB) const {
18259 // Emit va_arg instruction on X86-64.
18261 // Operands to this pseudo-instruction:
18262 // 0 ) Output : destination address (reg)
18263 // 1-5) Input : va_list address (addr, i64mem)
18264 // 6 ) ArgSize : Size (in bytes) of vararg type
18265 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18266 // 8 ) Align : Alignment of type
18267 // 9 ) EFLAGS (implicit-def)
18269 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18270 static_assert(X86::AddrNumOperands == 5,
18271 "VAARG_64 assumes 5 address operands");
18273 unsigned DestReg = MI->getOperand(0).getReg();
18274 MachineOperand &Base = MI->getOperand(1);
18275 MachineOperand &Scale = MI->getOperand(2);
18276 MachineOperand &Index = MI->getOperand(3);
18277 MachineOperand &Disp = MI->getOperand(4);
18278 MachineOperand &Segment = MI->getOperand(5);
18279 unsigned ArgSize = MI->getOperand(6).getImm();
18280 unsigned ArgMode = MI->getOperand(7).getImm();
18281 unsigned Align = MI->getOperand(8).getImm();
18283 // Memory Reference
18284 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
18285 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18286 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18288 // Machine Information
18289 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18290 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
18291 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
18292 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
18293 DebugLoc DL = MI->getDebugLoc();
18295 // struct va_list {
18298 // i64 overflow_area (address)
18299 // i64 reg_save_area (address)
18301 // sizeof(va_list) = 24
18302 // alignment(va_list) = 8
18304 unsigned TotalNumIntRegs = 6;
18305 unsigned TotalNumXMMRegs = 8;
18306 bool UseGPOffset = (ArgMode == 1);
18307 bool UseFPOffset = (ArgMode == 2);
18308 unsigned MaxOffset = TotalNumIntRegs * 8 +
18309 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
18311 /* Align ArgSize to a multiple of 8 */
18312 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
18313 bool NeedsAlign = (Align > 8);
18315 MachineBasicBlock *thisMBB = MBB;
18316 MachineBasicBlock *overflowMBB;
18317 MachineBasicBlock *offsetMBB;
18318 MachineBasicBlock *endMBB;
18320 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
18321 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
18322 unsigned OffsetReg = 0;
18324 if (!UseGPOffset && !UseFPOffset) {
18325 // If we only pull from the overflow region, we don't create a branch.
18326 // We don't need to alter control flow.
18327 OffsetDestReg = 0; // unused
18328 OverflowDestReg = DestReg;
18330 offsetMBB = nullptr;
18331 overflowMBB = thisMBB;
18334 // First emit code to check if gp_offset (or fp_offset) is below the bound.
18335 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
18336 // If not, pull from overflow_area. (branch to overflowMBB)
18341 // offsetMBB overflowMBB
18346 // Registers for the PHI in endMBB
18347 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
18348 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
18350 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18351 MachineFunction *MF = MBB->getParent();
18352 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18353 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18354 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18356 MachineFunction::iterator MBBIter = MBB;
18359 // Insert the new basic blocks
18360 MF->insert(MBBIter, offsetMBB);
18361 MF->insert(MBBIter, overflowMBB);
18362 MF->insert(MBBIter, endMBB);
18364 // Transfer the remainder of MBB and its successor edges to endMBB.
18365 endMBB->splice(endMBB->begin(), thisMBB,
18366 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
18367 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
18369 // Make offsetMBB and overflowMBB successors of thisMBB
18370 thisMBB->addSuccessor(offsetMBB);
18371 thisMBB->addSuccessor(overflowMBB);
18373 // endMBB is a successor of both offsetMBB and overflowMBB
18374 offsetMBB->addSuccessor(endMBB);
18375 overflowMBB->addSuccessor(endMBB);
18377 // Load the offset value into a register
18378 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18379 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
18383 .addDisp(Disp, UseFPOffset ? 4 : 0)
18384 .addOperand(Segment)
18385 .setMemRefs(MMOBegin, MMOEnd);
18387 // Check if there is enough room left to pull this argument.
18388 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
18390 .addImm(MaxOffset + 8 - ArgSizeA8);
18392 // Branch to "overflowMBB" if offset >= max
18393 // Fall through to "offsetMBB" otherwise
18394 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
18395 .addMBB(overflowMBB);
18398 // In offsetMBB, emit code to use the reg_save_area.
18400 assert(OffsetReg != 0);
18402 // Read the reg_save_area address.
18403 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
18404 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
18409 .addOperand(Segment)
18410 .setMemRefs(MMOBegin, MMOEnd);
18412 // Zero-extend the offset
18413 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
18414 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
18417 .addImm(X86::sub_32bit);
18419 // Add the offset to the reg_save_area to get the final address.
18420 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
18421 .addReg(OffsetReg64)
18422 .addReg(RegSaveReg);
18424 // Compute the offset for the next argument
18425 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18426 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
18428 .addImm(UseFPOffset ? 16 : 8);
18430 // Store it back into the va_list.
18431 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
18435 .addDisp(Disp, UseFPOffset ? 4 : 0)
18436 .addOperand(Segment)
18437 .addReg(NextOffsetReg)
18438 .setMemRefs(MMOBegin, MMOEnd);
18441 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
18446 // Emit code to use overflow area
18449 // Load the overflow_area address into a register.
18450 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
18451 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
18456 .addOperand(Segment)
18457 .setMemRefs(MMOBegin, MMOEnd);
18459 // If we need to align it, do so. Otherwise, just copy the address
18460 // to OverflowDestReg.
18462 // Align the overflow address
18463 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
18464 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
18466 // aligned_addr = (addr + (align-1)) & ~(align-1)
18467 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
18468 .addReg(OverflowAddrReg)
18471 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
18473 .addImm(~(uint64_t)(Align-1));
18475 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
18476 .addReg(OverflowAddrReg);
18479 // Compute the next overflow address after this argument.
18480 // (the overflow address should be kept 8-byte aligned)
18481 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
18482 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
18483 .addReg(OverflowDestReg)
18484 .addImm(ArgSizeA8);
18486 // Store the new overflow address.
18487 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
18492 .addOperand(Segment)
18493 .addReg(NextAddrReg)
18494 .setMemRefs(MMOBegin, MMOEnd);
18496 // If we branched, emit the PHI to the front of endMBB.
18498 BuildMI(*endMBB, endMBB->begin(), DL,
18499 TII->get(X86::PHI), DestReg)
18500 .addReg(OffsetDestReg).addMBB(offsetMBB)
18501 .addReg(OverflowDestReg).addMBB(overflowMBB);
18504 // Erase the pseudo instruction
18505 MI->eraseFromParent();
18510 MachineBasicBlock *
18511 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18513 MachineBasicBlock *MBB) const {
18514 // Emit code to save XMM registers to the stack. The ABI says that the
18515 // number of registers to save is given in %al, so it's theoretically
18516 // possible to do an indirect jump trick to avoid saving all of them,
18517 // however this code takes a simpler approach and just executes all
18518 // of the stores if %al is non-zero. It's less code, and it's probably
18519 // easier on the hardware branch predictor, and stores aren't all that
18520 // expensive anyway.
18522 // Create the new basic blocks. One block contains all the XMM stores,
18523 // and one block is the final destination regardless of whether any
18524 // stores were performed.
18525 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18526 MachineFunction *F = MBB->getParent();
18527 MachineFunction::iterator MBBIter = MBB;
18529 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
18530 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
18531 F->insert(MBBIter, XMMSaveMBB);
18532 F->insert(MBBIter, EndMBB);
18534 // Transfer the remainder of MBB and its successor edges to EndMBB.
18535 EndMBB->splice(EndMBB->begin(), MBB,
18536 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18537 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
18539 // The original block will now fall through to the XMM save block.
18540 MBB->addSuccessor(XMMSaveMBB);
18541 // The XMMSaveMBB will fall through to the end block.
18542 XMMSaveMBB->addSuccessor(EndMBB);
18544 // Now add the instructions.
18545 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18546 DebugLoc DL = MI->getDebugLoc();
18548 unsigned CountReg = MI->getOperand(0).getReg();
18549 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
18550 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
18552 if (!Subtarget->isTargetWin64()) {
18553 // If %al is 0, branch around the XMM save block.
18554 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
18555 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18556 MBB->addSuccessor(EndMBB);
18559 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18560 // that was just emitted, but clearly shouldn't be "saved".
18561 assert((MI->getNumOperands() <= 3 ||
18562 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18563 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18564 && "Expected last argument to be EFLAGS");
18565 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18566 // In the XMM save block, save all the XMM argument registers.
18567 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18568 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18569 MachineMemOperand *MMO =
18570 F->getMachineMemOperand(
18571 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18572 MachineMemOperand::MOStore,
18573 /*Size=*/16, /*Align=*/16);
18574 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18575 .addFrameIndex(RegSaveFrameIndex)
18576 .addImm(/*Scale=*/1)
18577 .addReg(/*IndexReg=*/0)
18578 .addImm(/*Disp=*/Offset)
18579 .addReg(/*Segment=*/0)
18580 .addReg(MI->getOperand(i).getReg())
18581 .addMemOperand(MMO);
18584 MI->eraseFromParent(); // The pseudo instruction is gone now.
18589 // The EFLAGS operand of SelectItr might be missing a kill marker
18590 // because there were multiple uses of EFLAGS, and ISel didn't know
18591 // which to mark. Figure out whether SelectItr should have had a
18592 // kill marker, and set it if it should. Returns the correct kill
18594 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18595 MachineBasicBlock* BB,
18596 const TargetRegisterInfo* TRI) {
18597 // Scan forward through BB for a use/def of EFLAGS.
18598 MachineBasicBlock::iterator miI(std::next(SelectItr));
18599 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18600 const MachineInstr& mi = *miI;
18601 if (mi.readsRegister(X86::EFLAGS))
18603 if (mi.definesRegister(X86::EFLAGS))
18604 break; // Should have kill-flag - update below.
18607 // If we hit the end of the block, check whether EFLAGS is live into a
18609 if (miI == BB->end()) {
18610 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18611 sEnd = BB->succ_end();
18612 sItr != sEnd; ++sItr) {
18613 MachineBasicBlock* succ = *sItr;
18614 if (succ->isLiveIn(X86::EFLAGS))
18619 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18620 // out. SelectMI should have a kill flag on EFLAGS.
18621 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18625 MachineBasicBlock *
18626 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18627 MachineBasicBlock *BB) const {
18628 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18629 DebugLoc DL = MI->getDebugLoc();
18631 // To "insert" a SELECT_CC instruction, we actually have to insert the
18632 // diamond control-flow pattern. The incoming instruction knows the
18633 // destination vreg to set, the condition code register to branch on, the
18634 // true/false values to select between, and a branch opcode to use.
18635 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18636 MachineFunction::iterator It = BB;
18642 // cmpTY ccX, r1, r2
18644 // fallthrough --> copy0MBB
18645 MachineBasicBlock *thisMBB = BB;
18646 MachineFunction *F = BB->getParent();
18648 // We also lower double CMOVs:
18649 // (CMOV (CMOV F, T, cc1), T, cc2)
18650 // to two successives branches. For that, we look for another CMOV as the
18651 // following instruction.
18653 // Without this, we would add a PHI between the two jumps, which ends up
18654 // creating a few copies all around. For instance, for
18656 // (sitofp (zext (fcmp une)))
18658 // we would generate:
18660 // ucomiss %xmm1, %xmm0
18661 // movss <1.0f>, %xmm0
18662 // movaps %xmm0, %xmm1
18664 // xorps %xmm1, %xmm1
18667 // movaps %xmm1, %xmm0
18671 // because this custom-inserter would have generated:
18683 // A: X = ...; Y = ...
18685 // C: Z = PHI [X, A], [Y, B]
18687 // E: PHI [X, C], [Z, D]
18689 // If we lower both CMOVs in a single step, we can instead generate:
18701 // A: X = ...; Y = ...
18703 // E: PHI [X, A], [X, C], [Y, D]
18705 // Which, in our sitofp/fcmp example, gives us something like:
18707 // ucomiss %xmm1, %xmm0
18708 // movss <1.0f>, %xmm0
18711 // xorps %xmm0, %xmm0
18715 MachineInstr *NextCMOV = nullptr;
18716 MachineBasicBlock::iterator NextMIIt =
18717 std::next(MachineBasicBlock::iterator(MI));
18718 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
18719 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
18720 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
18721 NextCMOV = &*NextMIIt;
18723 MachineBasicBlock *jcc1MBB = nullptr;
18725 // If we have a double CMOV, we lower it to two successive branches to
18726 // the same block. EFLAGS is used by both, so mark it as live in the second.
18728 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
18729 F->insert(It, jcc1MBB);
18730 jcc1MBB->addLiveIn(X86::EFLAGS);
18733 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18734 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18735 F->insert(It, copy0MBB);
18736 F->insert(It, sinkMBB);
18738 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18739 // live into the sink and copy blocks.
18740 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18742 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
18743 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
18744 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
18745 copy0MBB->addLiveIn(X86::EFLAGS);
18746 sinkMBB->addLiveIn(X86::EFLAGS);
18749 // Transfer the remainder of BB and its successor edges to sinkMBB.
18750 sinkMBB->splice(sinkMBB->begin(), BB,
18751 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18752 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18754 // Add the true and fallthrough blocks as its successors.
18756 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
18757 BB->addSuccessor(jcc1MBB);
18759 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
18760 // jump to the sinkMBB.
18761 jcc1MBB->addSuccessor(copy0MBB);
18762 jcc1MBB->addSuccessor(sinkMBB);
18764 BB->addSuccessor(copy0MBB);
18767 // The true block target of the first (or only) branch is always sinkMBB.
18768 BB->addSuccessor(sinkMBB);
18770 // Create the conditional branch instruction.
18772 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18773 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18776 unsigned Opc2 = X86::GetCondBranchFromCond(
18777 (X86::CondCode)NextCMOV->getOperand(3).getImm());
18778 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
18782 // %FalseValue = ...
18783 // # fallthrough to sinkMBB
18784 copy0MBB->addSuccessor(sinkMBB);
18787 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18789 MachineInstrBuilder MIB =
18790 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
18791 MI->getOperand(0).getReg())
18792 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18793 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18795 // If we have a double CMOV, the second Jcc provides the same incoming
18796 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
18798 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
18799 // Copy the PHI result to the register defined by the second CMOV.
18800 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
18801 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
18802 .addReg(MI->getOperand(0).getReg());
18803 NextCMOV->eraseFromParent();
18806 MI->eraseFromParent(); // The pseudo instruction is gone now.
18810 MachineBasicBlock *
18811 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18812 MachineBasicBlock *BB) const {
18813 MachineFunction *MF = BB->getParent();
18814 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18815 DebugLoc DL = MI->getDebugLoc();
18816 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18818 assert(MF->shouldSplitStack());
18820 const bool Is64Bit = Subtarget->is64Bit();
18821 const bool IsLP64 = Subtarget->isTarget64BitLP64();
18823 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
18824 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
18827 // ... [Till the alloca]
18828 // If stacklet is not large enough, jump to mallocMBB
18831 // Allocate by subtracting from RSP
18832 // Jump to continueMBB
18835 // Allocate by call to runtime
18839 // [rest of original BB]
18842 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18843 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18844 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18846 MachineRegisterInfo &MRI = MF->getRegInfo();
18847 const TargetRegisterClass *AddrRegClass =
18848 getRegClassFor(getPointerTy());
18850 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18851 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18852 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
18853 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
18854 sizeVReg = MI->getOperand(1).getReg(),
18855 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
18857 MachineFunction::iterator MBBIter = BB;
18860 MF->insert(MBBIter, bumpMBB);
18861 MF->insert(MBBIter, mallocMBB);
18862 MF->insert(MBBIter, continueMBB);
18864 continueMBB->splice(continueMBB->begin(), BB,
18865 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18866 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
18868 // Add code to the main basic block to check if the stack limit has been hit,
18869 // and if so, jump to mallocMBB otherwise to bumpMBB.
18870 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
18871 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
18872 .addReg(tmpSPVReg).addReg(sizeVReg);
18873 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
18874 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
18875 .addReg(SPLimitVReg);
18876 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
18878 // bumpMBB simply decreases the stack pointer, since we know the current
18879 // stacklet has enough space.
18880 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
18881 .addReg(SPLimitVReg);
18882 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
18883 .addReg(SPLimitVReg);
18884 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18886 // Calls into a routine in libgcc to allocate more space from the heap.
18887 const uint32_t *RegMask =
18888 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
18890 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
18892 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18893 .addExternalSymbol("__morestack_allocate_stack_space")
18894 .addRegMask(RegMask)
18895 .addReg(X86::RDI, RegState::Implicit)
18896 .addReg(X86::RAX, RegState::ImplicitDefine);
18897 } else if (Is64Bit) {
18898 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
18900 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18901 .addExternalSymbol("__morestack_allocate_stack_space")
18902 .addRegMask(RegMask)
18903 .addReg(X86::EDI, RegState::Implicit)
18904 .addReg(X86::EAX, RegState::ImplicitDefine);
18906 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
18908 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
18909 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
18910 .addExternalSymbol("__morestack_allocate_stack_space")
18911 .addRegMask(RegMask)
18912 .addReg(X86::EAX, RegState::ImplicitDefine);
18916 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
18919 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
18920 .addReg(IsLP64 ? X86::RAX : X86::EAX);
18921 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18923 // Set up the CFG correctly.
18924 BB->addSuccessor(bumpMBB);
18925 BB->addSuccessor(mallocMBB);
18926 mallocMBB->addSuccessor(continueMBB);
18927 bumpMBB->addSuccessor(continueMBB);
18929 // Take care of the PHI nodes.
18930 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
18931 MI->getOperand(0).getReg())
18932 .addReg(mallocPtrVReg).addMBB(mallocMBB)
18933 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
18935 // Delete the original pseudo instruction.
18936 MI->eraseFromParent();
18939 return continueMBB;
18942 MachineBasicBlock *
18943 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
18944 MachineBasicBlock *BB) const {
18945 DebugLoc DL = MI->getDebugLoc();
18947 assert(!Subtarget->isTargetMachO());
18949 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
18951 MI->eraseFromParent(); // The pseudo instruction is gone now.
18955 MachineBasicBlock *
18956 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
18957 MachineBasicBlock *BB) const {
18958 // This is pretty easy. We're taking the value that we received from
18959 // our load from the relocation, sticking it in either RDI (x86-64)
18960 // or EAX and doing an indirect call. The return value will then
18961 // be in the normal return register.
18962 MachineFunction *F = BB->getParent();
18963 const X86InstrInfo *TII = Subtarget->getInstrInfo();
18964 DebugLoc DL = MI->getDebugLoc();
18966 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
18967 assert(MI->getOperand(3).isGlobal() && "This should be a global");
18969 // Get a register mask for the lowered call.
18970 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
18971 // proper register mask.
18972 const uint32_t *RegMask =
18973 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
18974 if (Subtarget->is64Bit()) {
18975 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18976 TII->get(X86::MOV64rm), X86::RDI)
18978 .addImm(0).addReg(0)
18979 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18980 MI->getOperand(3).getTargetFlags())
18982 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
18983 addDirectMem(MIB, X86::RDI);
18984 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
18985 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
18986 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18987 TII->get(X86::MOV32rm), X86::EAX)
18989 .addImm(0).addReg(0)
18990 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18991 MI->getOperand(3).getTargetFlags())
18993 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18994 addDirectMem(MIB, X86::EAX);
18995 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18997 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18998 TII->get(X86::MOV32rm), X86::EAX)
18999 .addReg(TII->getGlobalBaseReg(F))
19000 .addImm(0).addReg(0)
19001 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19002 MI->getOperand(3).getTargetFlags())
19004 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19005 addDirectMem(MIB, X86::EAX);
19006 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19009 MI->eraseFromParent(); // The pseudo instruction is gone now.
19013 MachineBasicBlock *
19014 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
19015 MachineBasicBlock *MBB) const {
19016 DebugLoc DL = MI->getDebugLoc();
19017 MachineFunction *MF = MBB->getParent();
19018 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19019 MachineRegisterInfo &MRI = MF->getRegInfo();
19021 const BasicBlock *BB = MBB->getBasicBlock();
19022 MachineFunction::iterator I = MBB;
19025 // Memory Reference
19026 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19027 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19030 unsigned MemOpndSlot = 0;
19032 unsigned CurOp = 0;
19034 DstReg = MI->getOperand(CurOp++).getReg();
19035 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
19036 assert(RC->hasType(MVT::i32) && "Invalid destination!");
19037 unsigned mainDstReg = MRI.createVirtualRegister(RC);
19038 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
19040 MemOpndSlot = CurOp;
19042 MVT PVT = getPointerTy();
19043 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19044 "Invalid Pointer Size!");
19046 // For v = setjmp(buf), we generate
19049 // buf[LabelOffset] = restoreMBB
19050 // SjLjSetup restoreMBB
19056 // v = phi(main, restore)
19059 // if base pointer being used, load it from frame
19062 MachineBasicBlock *thisMBB = MBB;
19063 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19064 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19065 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
19066 MF->insert(I, mainMBB);
19067 MF->insert(I, sinkMBB);
19068 MF->push_back(restoreMBB);
19070 MachineInstrBuilder MIB;
19072 // Transfer the remainder of BB and its successor edges to sinkMBB.
19073 sinkMBB->splice(sinkMBB->begin(), MBB,
19074 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19075 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19078 unsigned PtrStoreOpc = 0;
19079 unsigned LabelReg = 0;
19080 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19081 Reloc::Model RM = MF->getTarget().getRelocationModel();
19082 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19083 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19085 // Prepare IP either in reg or imm.
19086 if (!UseImmLabel) {
19087 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19088 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19089 LabelReg = MRI.createVirtualRegister(PtrRC);
19090 if (Subtarget->is64Bit()) {
19091 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19095 .addMBB(restoreMBB)
19098 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19099 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19100 .addReg(XII->getGlobalBaseReg(MF))
19103 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19107 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19109 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19110 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19111 if (i == X86::AddrDisp)
19112 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19114 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19117 MIB.addReg(LabelReg);
19119 MIB.addMBB(restoreMBB);
19120 MIB.setMemRefs(MMOBegin, MMOEnd);
19122 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19123 .addMBB(restoreMBB);
19125 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19126 MIB.addRegMask(RegInfo->getNoPreservedMask());
19127 thisMBB->addSuccessor(mainMBB);
19128 thisMBB->addSuccessor(restoreMBB);
19132 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19133 mainMBB->addSuccessor(sinkMBB);
19136 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19137 TII->get(X86::PHI), DstReg)
19138 .addReg(mainDstReg).addMBB(mainMBB)
19139 .addReg(restoreDstReg).addMBB(restoreMBB);
19142 if (RegInfo->hasBasePointer(*MF)) {
19143 const bool Uses64BitFramePtr =
19144 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
19145 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
19146 X86FI->setRestoreBasePointer(MF);
19147 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
19148 unsigned BasePtr = RegInfo->getBaseRegister();
19149 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
19150 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
19151 FramePtr, true, X86FI->getRestoreBasePointerOffset())
19152 .setMIFlag(MachineInstr::FrameSetup);
19154 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19155 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
19156 restoreMBB->addSuccessor(sinkMBB);
19158 MI->eraseFromParent();
19162 MachineBasicBlock *
19163 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19164 MachineBasicBlock *MBB) const {
19165 DebugLoc DL = MI->getDebugLoc();
19166 MachineFunction *MF = MBB->getParent();
19167 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19168 MachineRegisterInfo &MRI = MF->getRegInfo();
19170 // Memory Reference
19171 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19172 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19174 MVT PVT = getPointerTy();
19175 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19176 "Invalid Pointer Size!");
19178 const TargetRegisterClass *RC =
19179 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19180 unsigned Tmp = MRI.createVirtualRegister(RC);
19181 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19182 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19183 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19184 unsigned SP = RegInfo->getStackRegister();
19186 MachineInstrBuilder MIB;
19188 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19189 const int64_t SPOffset = 2 * PVT.getStoreSize();
19191 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19192 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19195 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19196 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19197 MIB.addOperand(MI->getOperand(i));
19198 MIB.setMemRefs(MMOBegin, MMOEnd);
19200 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19201 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19202 if (i == X86::AddrDisp)
19203 MIB.addDisp(MI->getOperand(i), LabelOffset);
19205 MIB.addOperand(MI->getOperand(i));
19207 MIB.setMemRefs(MMOBegin, MMOEnd);
19209 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19210 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19211 if (i == X86::AddrDisp)
19212 MIB.addDisp(MI->getOperand(i), SPOffset);
19214 MIB.addOperand(MI->getOperand(i));
19216 MIB.setMemRefs(MMOBegin, MMOEnd);
19218 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19220 MI->eraseFromParent();
19224 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19225 // accumulator loops. Writing back to the accumulator allows the coalescer
19226 // to remove extra copies in the loop.
19227 MachineBasicBlock *
19228 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19229 MachineBasicBlock *MBB) const {
19230 MachineOperand &AddendOp = MI->getOperand(3);
19232 // Bail out early if the addend isn't a register - we can't switch these.
19233 if (!AddendOp.isReg())
19236 MachineFunction &MF = *MBB->getParent();
19237 MachineRegisterInfo &MRI = MF.getRegInfo();
19239 // Check whether the addend is defined by a PHI:
19240 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19241 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19242 if (!AddendDef.isPHI())
19245 // Look for the following pattern:
19247 // %addend = phi [%entry, 0], [%loop, %result]
19249 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19253 // %addend = phi [%entry, 0], [%loop, %result]
19255 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19257 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19258 assert(AddendDef.getOperand(i).isReg());
19259 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19260 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19261 if (&PHISrcInst == MI) {
19262 // Found a matching instruction.
19263 unsigned NewFMAOpc = 0;
19264 switch (MI->getOpcode()) {
19265 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19266 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19267 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19268 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19269 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19270 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19271 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19272 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19273 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19274 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19275 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19276 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19277 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19278 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19279 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19280 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19281 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
19282 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
19283 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
19284 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
19286 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19287 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19288 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19289 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19290 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19291 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19292 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19293 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19294 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
19295 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
19296 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
19297 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
19298 default: llvm_unreachable("Unrecognized FMA variant.");
19301 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
19302 MachineInstrBuilder MIB =
19303 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19304 .addOperand(MI->getOperand(0))
19305 .addOperand(MI->getOperand(3))
19306 .addOperand(MI->getOperand(2))
19307 .addOperand(MI->getOperand(1));
19308 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19309 MI->eraseFromParent();
19316 MachineBasicBlock *
19317 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19318 MachineBasicBlock *BB) const {
19319 switch (MI->getOpcode()) {
19320 default: llvm_unreachable("Unexpected instr type to insert");
19321 case X86::TAILJMPd64:
19322 case X86::TAILJMPr64:
19323 case X86::TAILJMPm64:
19324 case X86::TAILJMPd64_REX:
19325 case X86::TAILJMPr64_REX:
19326 case X86::TAILJMPm64_REX:
19327 llvm_unreachable("TAILJMP64 would not be touched here.");
19328 case X86::TCRETURNdi64:
19329 case X86::TCRETURNri64:
19330 case X86::TCRETURNmi64:
19332 case X86::WIN_ALLOCA:
19333 return EmitLoweredWinAlloca(MI, BB);
19334 case X86::SEG_ALLOCA_32:
19335 case X86::SEG_ALLOCA_64:
19336 return EmitLoweredSegAlloca(MI, BB);
19337 case X86::TLSCall_32:
19338 case X86::TLSCall_64:
19339 return EmitLoweredTLSCall(MI, BB);
19340 case X86::CMOV_GR8:
19341 case X86::CMOV_FR32:
19342 case X86::CMOV_FR64:
19343 case X86::CMOV_V4F32:
19344 case X86::CMOV_V2F64:
19345 case X86::CMOV_V2I64:
19346 case X86::CMOV_V8F32:
19347 case X86::CMOV_V4F64:
19348 case X86::CMOV_V4I64:
19349 case X86::CMOV_V16F32:
19350 case X86::CMOV_V8F64:
19351 case X86::CMOV_V8I64:
19352 case X86::CMOV_GR16:
19353 case X86::CMOV_GR32:
19354 case X86::CMOV_RFP32:
19355 case X86::CMOV_RFP64:
19356 case X86::CMOV_RFP80:
19357 return EmitLoweredSelect(MI, BB);
19359 case X86::FP32_TO_INT16_IN_MEM:
19360 case X86::FP32_TO_INT32_IN_MEM:
19361 case X86::FP32_TO_INT64_IN_MEM:
19362 case X86::FP64_TO_INT16_IN_MEM:
19363 case X86::FP64_TO_INT32_IN_MEM:
19364 case X86::FP64_TO_INT64_IN_MEM:
19365 case X86::FP80_TO_INT16_IN_MEM:
19366 case X86::FP80_TO_INT32_IN_MEM:
19367 case X86::FP80_TO_INT64_IN_MEM: {
19368 MachineFunction *F = BB->getParent();
19369 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19370 DebugLoc DL = MI->getDebugLoc();
19372 // Change the floating point control register to use "round towards zero"
19373 // mode when truncating to an integer value.
19374 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19375 addFrameReference(BuildMI(*BB, MI, DL,
19376 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19378 // Load the old value of the high byte of the control word...
19380 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
19381 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
19384 // Set the high part to be round to zero...
19385 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
19388 // Reload the modified control word now...
19389 addFrameReference(BuildMI(*BB, MI, DL,
19390 TII->get(X86::FLDCW16m)), CWFrameIdx);
19392 // Restore the memory image of control word to original value
19393 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
19396 // Get the X86 opcode to use.
19398 switch (MI->getOpcode()) {
19399 default: llvm_unreachable("illegal opcode!");
19400 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
19401 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
19402 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
19403 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
19404 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
19405 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
19406 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
19407 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
19408 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
19412 MachineOperand &Op = MI->getOperand(0);
19414 AM.BaseType = X86AddressMode::RegBase;
19415 AM.Base.Reg = Op.getReg();
19417 AM.BaseType = X86AddressMode::FrameIndexBase;
19418 AM.Base.FrameIndex = Op.getIndex();
19420 Op = MI->getOperand(1);
19422 AM.Scale = Op.getImm();
19423 Op = MI->getOperand(2);
19425 AM.IndexReg = Op.getImm();
19426 Op = MI->getOperand(3);
19427 if (Op.isGlobal()) {
19428 AM.GV = Op.getGlobal();
19430 AM.Disp = Op.getImm();
19432 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
19433 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
19435 // Reload the original control word now.
19436 addFrameReference(BuildMI(*BB, MI, DL,
19437 TII->get(X86::FLDCW16m)), CWFrameIdx);
19439 MI->eraseFromParent(); // The pseudo instruction is gone now.
19442 // String/text processing lowering.
19443 case X86::PCMPISTRM128REG:
19444 case X86::VPCMPISTRM128REG:
19445 case X86::PCMPISTRM128MEM:
19446 case X86::VPCMPISTRM128MEM:
19447 case X86::PCMPESTRM128REG:
19448 case X86::VPCMPESTRM128REG:
19449 case X86::PCMPESTRM128MEM:
19450 case X86::VPCMPESTRM128MEM:
19451 assert(Subtarget->hasSSE42() &&
19452 "Target must have SSE4.2 or AVX features enabled");
19453 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
19455 // String/text processing lowering.
19456 case X86::PCMPISTRIREG:
19457 case X86::VPCMPISTRIREG:
19458 case X86::PCMPISTRIMEM:
19459 case X86::VPCMPISTRIMEM:
19460 case X86::PCMPESTRIREG:
19461 case X86::VPCMPESTRIREG:
19462 case X86::PCMPESTRIMEM:
19463 case X86::VPCMPESTRIMEM:
19464 assert(Subtarget->hasSSE42() &&
19465 "Target must have SSE4.2 or AVX features enabled");
19466 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
19468 // Thread synchronization.
19470 return EmitMonitor(MI, BB, Subtarget);
19474 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
19476 case X86::VASTART_SAVE_XMM_REGS:
19477 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
19479 case X86::VAARG_64:
19480 return EmitVAARG64WithCustomInserter(MI, BB);
19482 case X86::EH_SjLj_SetJmp32:
19483 case X86::EH_SjLj_SetJmp64:
19484 return emitEHSjLjSetJmp(MI, BB);
19486 case X86::EH_SjLj_LongJmp32:
19487 case X86::EH_SjLj_LongJmp64:
19488 return emitEHSjLjLongJmp(MI, BB);
19490 case TargetOpcode::STATEPOINT:
19491 // As an implementation detail, STATEPOINT shares the STACKMAP format at
19492 // this point in the process. We diverge later.
19493 return emitPatchPoint(MI, BB);
19495 case TargetOpcode::STACKMAP:
19496 case TargetOpcode::PATCHPOINT:
19497 return emitPatchPoint(MI, BB);
19499 case X86::VFMADDPDr213r:
19500 case X86::VFMADDPSr213r:
19501 case X86::VFMADDSDr213r:
19502 case X86::VFMADDSSr213r:
19503 case X86::VFMSUBPDr213r:
19504 case X86::VFMSUBPSr213r:
19505 case X86::VFMSUBSDr213r:
19506 case X86::VFMSUBSSr213r:
19507 case X86::VFNMADDPDr213r:
19508 case X86::VFNMADDPSr213r:
19509 case X86::VFNMADDSDr213r:
19510 case X86::VFNMADDSSr213r:
19511 case X86::VFNMSUBPDr213r:
19512 case X86::VFNMSUBPSr213r:
19513 case X86::VFNMSUBSDr213r:
19514 case X86::VFNMSUBSSr213r:
19515 case X86::VFMADDSUBPDr213r:
19516 case X86::VFMADDSUBPSr213r:
19517 case X86::VFMSUBADDPDr213r:
19518 case X86::VFMSUBADDPSr213r:
19519 case X86::VFMADDPDr213rY:
19520 case X86::VFMADDPSr213rY:
19521 case X86::VFMSUBPDr213rY:
19522 case X86::VFMSUBPSr213rY:
19523 case X86::VFNMADDPDr213rY:
19524 case X86::VFNMADDPSr213rY:
19525 case X86::VFNMSUBPDr213rY:
19526 case X86::VFNMSUBPSr213rY:
19527 case X86::VFMADDSUBPDr213rY:
19528 case X86::VFMADDSUBPSr213rY:
19529 case X86::VFMSUBADDPDr213rY:
19530 case X86::VFMSUBADDPSr213rY:
19531 return emitFMA3Instr(MI, BB);
19535 //===----------------------------------------------------------------------===//
19536 // X86 Optimization Hooks
19537 //===----------------------------------------------------------------------===//
19539 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
19542 const SelectionDAG &DAG,
19543 unsigned Depth) const {
19544 unsigned BitWidth = KnownZero.getBitWidth();
19545 unsigned Opc = Op.getOpcode();
19546 assert((Opc >= ISD::BUILTIN_OP_END ||
19547 Opc == ISD::INTRINSIC_WO_CHAIN ||
19548 Opc == ISD::INTRINSIC_W_CHAIN ||
19549 Opc == ISD::INTRINSIC_VOID) &&
19550 "Should use MaskedValueIsZero if you don't know whether Op"
19551 " is a target node!");
19553 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
19567 // These nodes' second result is a boolean.
19568 if (Op.getResNo() == 0)
19571 case X86ISD::SETCC:
19572 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
19574 case ISD::INTRINSIC_WO_CHAIN: {
19575 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
19576 unsigned NumLoBits = 0;
19579 case Intrinsic::x86_sse_movmsk_ps:
19580 case Intrinsic::x86_avx_movmsk_ps_256:
19581 case Intrinsic::x86_sse2_movmsk_pd:
19582 case Intrinsic::x86_avx_movmsk_pd_256:
19583 case Intrinsic::x86_mmx_pmovmskb:
19584 case Intrinsic::x86_sse2_pmovmskb_128:
19585 case Intrinsic::x86_avx2_pmovmskb: {
19586 // High bits of movmskp{s|d}, pmovmskb are known zero.
19588 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
19589 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
19590 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
19591 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
19592 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
19593 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
19594 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
19595 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
19597 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
19606 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
19608 const SelectionDAG &,
19609 unsigned Depth) const {
19610 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
19611 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
19612 return Op.getValueType().getScalarType().getSizeInBits();
19618 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
19619 /// node is a GlobalAddress + offset.
19620 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
19621 const GlobalValue* &GA,
19622 int64_t &Offset) const {
19623 if (N->getOpcode() == X86ISD::Wrapper) {
19624 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
19625 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
19626 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
19630 return TargetLowering::isGAPlusOffset(N, GA, Offset);
19633 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
19634 /// same as extracting the high 128-bit part of 256-bit vector and then
19635 /// inserting the result into the low part of a new 256-bit vector
19636 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
19637 EVT VT = SVOp->getValueType(0);
19638 unsigned NumElems = VT.getVectorNumElements();
19640 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19641 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
19642 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19643 SVOp->getMaskElt(j) >= 0)
19649 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
19650 /// same as extracting the low 128-bit part of 256-bit vector and then
19651 /// inserting the result into the high part of a new 256-bit vector
19652 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
19653 EVT VT = SVOp->getValueType(0);
19654 unsigned NumElems = VT.getVectorNumElements();
19656 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19657 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
19658 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19659 SVOp->getMaskElt(j) >= 0)
19665 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
19666 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
19667 TargetLowering::DAGCombinerInfo &DCI,
19668 const X86Subtarget* Subtarget) {
19670 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19671 SDValue V1 = SVOp->getOperand(0);
19672 SDValue V2 = SVOp->getOperand(1);
19673 EVT VT = SVOp->getValueType(0);
19674 unsigned NumElems = VT.getVectorNumElements();
19676 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19677 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19681 // V UNDEF BUILD_VECTOR UNDEF
19683 // CONCAT_VECTOR CONCAT_VECTOR
19686 // RESULT: V + zero extended
19688 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19689 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19690 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19693 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19696 // To match the shuffle mask, the first half of the mask should
19697 // be exactly the first vector, and all the rest a splat with the
19698 // first element of the second one.
19699 for (unsigned i = 0; i != NumElems/2; ++i)
19700 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19701 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19704 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19705 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19706 if (Ld->hasNUsesOfValue(1, 0)) {
19707 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19708 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19710 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19712 Ld->getPointerInfo(),
19713 Ld->getAlignment(),
19714 false/*isVolatile*/, true/*ReadMem*/,
19715 false/*WriteMem*/);
19717 // Make sure the newly-created LOAD is in the same position as Ld in
19718 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19719 // and update uses of Ld's output chain to use the TokenFactor.
19720 if (Ld->hasAnyUseOfValue(1)) {
19721 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19722 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19723 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19724 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19725 SDValue(ResNode.getNode(), 1));
19728 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19732 // Emit a zeroed vector and insert the desired subvector on its
19734 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19735 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19736 return DCI.CombineTo(N, InsV);
19739 //===--------------------------------------------------------------------===//
19740 // Combine some shuffles into subvector extracts and inserts:
19743 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19744 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19745 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19746 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19747 return DCI.CombineTo(N, InsV);
19750 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19751 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19752 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19753 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19754 return DCI.CombineTo(N, InsV);
19760 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19763 /// This is the leaf of the recursive combinine below. When we have found some
19764 /// chain of single-use x86 shuffle instructions and accumulated the combined
19765 /// shuffle mask represented by them, this will try to pattern match that mask
19766 /// into either a single instruction if there is a special purpose instruction
19767 /// for this operation, or into a PSHUFB instruction which is a fully general
19768 /// instruction but should only be used to replace chains over a certain depth.
19769 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19770 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19771 TargetLowering::DAGCombinerInfo &DCI,
19772 const X86Subtarget *Subtarget) {
19773 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19775 // Find the operand that enters the chain. Note that multiple uses are OK
19776 // here, we're not going to remove the operand we find.
19777 SDValue Input = Op.getOperand(0);
19778 while (Input.getOpcode() == ISD::BITCAST)
19779 Input = Input.getOperand(0);
19781 MVT VT = Input.getSimpleValueType();
19782 MVT RootVT = Root.getSimpleValueType();
19785 // Just remove no-op shuffle masks.
19786 if (Mask.size() == 1) {
19787 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19792 // Use the float domain if the operand type is a floating point type.
19793 bool FloatDomain = VT.isFloatingPoint();
19795 // For floating point shuffles, we don't have free copies in the shuffle
19796 // instructions or the ability to load as part of the instruction, so
19797 // canonicalize their shuffles to UNPCK or MOV variants.
19799 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19800 // vectors because it can have a load folded into it that UNPCK cannot. This
19801 // doesn't preclude something switching to the shorter encoding post-RA.
19803 // FIXME: Should teach these routines about AVX vector widths.
19804 if (FloatDomain && VT.getSizeInBits() == 128) {
19805 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
19806 bool Lo = Mask.equals({0, 0});
19809 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19810 // is no slower than UNPCKLPD but has the option to fold the input operand
19811 // into even an unaligned memory load.
19812 if (Lo && Subtarget->hasSSE3()) {
19813 Shuffle = X86ISD::MOVDDUP;
19814 ShuffleVT = MVT::v2f64;
19816 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
19817 // than the UNPCK variants.
19818 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
19819 ShuffleVT = MVT::v4f32;
19821 if (Depth == 1 && Root->getOpcode() == Shuffle)
19822 return false; // Nothing to do!
19823 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19824 DCI.AddToWorklist(Op.getNode());
19825 if (Shuffle == X86ISD::MOVDDUP)
19826 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19828 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19829 DCI.AddToWorklist(Op.getNode());
19830 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19834 if (Subtarget->hasSSE3() &&
19835 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
19836 bool Lo = Mask.equals({0, 0, 2, 2});
19837 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
19838 MVT ShuffleVT = MVT::v4f32;
19839 if (Depth == 1 && Root->getOpcode() == Shuffle)
19840 return false; // Nothing to do!
19841 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19842 DCI.AddToWorklist(Op.getNode());
19843 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19844 DCI.AddToWorklist(Op.getNode());
19845 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19849 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
19850 bool Lo = Mask.equals({0, 0, 1, 1});
19851 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19852 MVT ShuffleVT = MVT::v4f32;
19853 if (Depth == 1 && Root->getOpcode() == Shuffle)
19854 return false; // Nothing to do!
19855 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19856 DCI.AddToWorklist(Op.getNode());
19857 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19858 DCI.AddToWorklist(Op.getNode());
19859 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19865 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
19866 // variants as none of these have single-instruction variants that are
19867 // superior to the UNPCK formulation.
19868 if (!FloatDomain && VT.getSizeInBits() == 128 &&
19869 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
19870 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
19871 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
19873 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
19874 bool Lo = Mask[0] == 0;
19875 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19876 if (Depth == 1 && Root->getOpcode() == Shuffle)
19877 return false; // Nothing to do!
19879 switch (Mask.size()) {
19881 ShuffleVT = MVT::v8i16;
19884 ShuffleVT = MVT::v16i8;
19887 llvm_unreachable("Impossible mask size!");
19889 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19890 DCI.AddToWorklist(Op.getNode());
19891 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19892 DCI.AddToWorklist(Op.getNode());
19893 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19898 // Don't try to re-form single instruction chains under any circumstances now
19899 // that we've done encoding canonicalization for them.
19903 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
19904 // can replace them with a single PSHUFB instruction profitably. Intel's
19905 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
19906 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
19907 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
19908 SmallVector<SDValue, 16> PSHUFBMask;
19909 int NumBytes = VT.getSizeInBits() / 8;
19910 int Ratio = NumBytes / Mask.size();
19911 for (int i = 0; i < NumBytes; ++i) {
19912 if (Mask[i / Ratio] == SM_SentinelUndef) {
19913 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
19916 int M = Mask[i / Ratio] != SM_SentinelZero
19917 ? Ratio * Mask[i / Ratio] + i % Ratio
19919 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
19921 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
19922 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Input);
19923 DCI.AddToWorklist(Op.getNode());
19924 SDValue PSHUFBMaskOp =
19925 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
19926 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
19927 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
19928 DCI.AddToWorklist(Op.getNode());
19929 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19934 // Failed to find any combines.
19938 /// \brief Fully generic combining of x86 shuffle instructions.
19940 /// This should be the last combine run over the x86 shuffle instructions. Once
19941 /// they have been fully optimized, this will recursively consider all chains
19942 /// of single-use shuffle instructions, build a generic model of the cumulative
19943 /// shuffle operation, and check for simpler instructions which implement this
19944 /// operation. We use this primarily for two purposes:
19946 /// 1) Collapse generic shuffles to specialized single instructions when
19947 /// equivalent. In most cases, this is just an encoding size win, but
19948 /// sometimes we will collapse multiple generic shuffles into a single
19949 /// special-purpose shuffle.
19950 /// 2) Look for sequences of shuffle instructions with 3 or more total
19951 /// instructions, and replace them with the slightly more expensive SSSE3
19952 /// PSHUFB instruction if available. We do this as the last combining step
19953 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
19954 /// a suitable short sequence of other instructions. The PHUFB will either
19955 /// use a register or have to read from memory and so is slightly (but only
19956 /// slightly) more expensive than the other shuffle instructions.
19958 /// Because this is inherently a quadratic operation (for each shuffle in
19959 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
19960 /// This should never be an issue in practice as the shuffle lowering doesn't
19961 /// produce sequences of more than 8 instructions.
19963 /// FIXME: We will currently miss some cases where the redundant shuffling
19964 /// would simplify under the threshold for PSHUFB formation because of
19965 /// combine-ordering. To fix this, we should do the redundant instruction
19966 /// combining in this recursive walk.
19967 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
19968 ArrayRef<int> RootMask,
19969 int Depth, bool HasPSHUFB,
19971 TargetLowering::DAGCombinerInfo &DCI,
19972 const X86Subtarget *Subtarget) {
19973 // Bound the depth of our recursive combine because this is ultimately
19974 // quadratic in nature.
19978 // Directly rip through bitcasts to find the underlying operand.
19979 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
19980 Op = Op.getOperand(0);
19982 MVT VT = Op.getSimpleValueType();
19983 if (!VT.isVector())
19984 return false; // Bail if we hit a non-vector.
19986 assert(Root.getSimpleValueType().isVector() &&
19987 "Shuffles operate on vector types!");
19988 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
19989 "Can only combine shuffles of the same vector register size.");
19991 if (!isTargetShuffle(Op.getOpcode()))
19993 SmallVector<int, 16> OpMask;
19995 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
19996 // We only can combine unary shuffles which we can decode the mask for.
19997 if (!HaveMask || !IsUnary)
20000 assert(VT.getVectorNumElements() == OpMask.size() &&
20001 "Different mask size from vector size!");
20002 assert(((RootMask.size() > OpMask.size() &&
20003 RootMask.size() % OpMask.size() == 0) ||
20004 (OpMask.size() > RootMask.size() &&
20005 OpMask.size() % RootMask.size() == 0) ||
20006 OpMask.size() == RootMask.size()) &&
20007 "The smaller number of elements must divide the larger.");
20008 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
20009 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
20010 assert(((RootRatio == 1 && OpRatio == 1) ||
20011 (RootRatio == 1) != (OpRatio == 1)) &&
20012 "Must not have a ratio for both incoming and op masks!");
20014 SmallVector<int, 16> Mask;
20015 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
20017 // Merge this shuffle operation's mask into our accumulated mask. Note that
20018 // this shuffle's mask will be the first applied to the input, followed by the
20019 // root mask to get us all the way to the root value arrangement. The reason
20020 // for this order is that we are recursing up the operation chain.
20021 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
20022 int RootIdx = i / RootRatio;
20023 if (RootMask[RootIdx] < 0) {
20024 // This is a zero or undef lane, we're done.
20025 Mask.push_back(RootMask[RootIdx]);
20029 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
20030 int OpIdx = RootMaskedIdx / OpRatio;
20031 if (OpMask[OpIdx] < 0) {
20032 // The incoming lanes are zero or undef, it doesn't matter which ones we
20034 Mask.push_back(OpMask[OpIdx]);
20038 // Ok, we have non-zero lanes, map them through.
20039 Mask.push_back(OpMask[OpIdx] * OpRatio +
20040 RootMaskedIdx % OpRatio);
20043 // See if we can recurse into the operand to combine more things.
20044 switch (Op.getOpcode()) {
20045 case X86ISD::PSHUFB:
20047 case X86ISD::PSHUFD:
20048 case X86ISD::PSHUFHW:
20049 case X86ISD::PSHUFLW:
20050 if (Op.getOperand(0).hasOneUse() &&
20051 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20052 HasPSHUFB, DAG, DCI, Subtarget))
20056 case X86ISD::UNPCKL:
20057 case X86ISD::UNPCKH:
20058 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
20059 // We can't check for single use, we have to check that this shuffle is the only user.
20060 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
20061 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20062 HasPSHUFB, DAG, DCI, Subtarget))
20067 // Minor canonicalization of the accumulated shuffle mask to make it easier
20068 // to match below. All this does is detect masks with squential pairs of
20069 // elements, and shrink them to the half-width mask. It does this in a loop
20070 // so it will reduce the size of the mask to the minimal width mask which
20071 // performs an equivalent shuffle.
20072 SmallVector<int, 16> WidenedMask;
20073 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
20074 Mask = std::move(WidenedMask);
20075 WidenedMask.clear();
20078 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20082 /// \brief Get the PSHUF-style mask from PSHUF node.
20084 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20085 /// PSHUF-style masks that can be reused with such instructions.
20086 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20087 MVT VT = N.getSimpleValueType();
20088 SmallVector<int, 4> Mask;
20090 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
20094 // If we have more than 128-bits, only the low 128-bits of shuffle mask
20095 // matter. Check that the upper masks are repeats and remove them.
20096 if (VT.getSizeInBits() > 128) {
20097 int LaneElts = 128 / VT.getScalarSizeInBits();
20099 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
20100 for (int j = 0; j < LaneElts; ++j)
20101 assert(Mask[j] == Mask[i * LaneElts + j] - LaneElts &&
20102 "Mask doesn't repeat in high 128-bit lanes!");
20104 Mask.resize(LaneElts);
20107 switch (N.getOpcode()) {
20108 case X86ISD::PSHUFD:
20110 case X86ISD::PSHUFLW:
20113 case X86ISD::PSHUFHW:
20114 Mask.erase(Mask.begin(), Mask.begin() + 4);
20115 for (int &M : Mask)
20119 llvm_unreachable("No valid shuffle instruction found!");
20123 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20125 /// We walk up the chain and look for a combinable shuffle, skipping over
20126 /// shuffles that we could hoist this shuffle's transformation past without
20127 /// altering anything.
20129 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20131 TargetLowering::DAGCombinerInfo &DCI) {
20132 assert(N.getOpcode() == X86ISD::PSHUFD &&
20133 "Called with something other than an x86 128-bit half shuffle!");
20136 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20137 // of the shuffles in the chain so that we can form a fresh chain to replace
20139 SmallVector<SDValue, 8> Chain;
20140 SDValue V = N.getOperand(0);
20141 for (; V.hasOneUse(); V = V.getOperand(0)) {
20142 switch (V.getOpcode()) {
20144 return SDValue(); // Nothing combined!
20147 // Skip bitcasts as we always know the type for the target specific
20151 case X86ISD::PSHUFD:
20152 // Found another dword shuffle.
20155 case X86ISD::PSHUFLW:
20156 // Check that the low words (being shuffled) are the identity in the
20157 // dword shuffle, and the high words are self-contained.
20158 if (Mask[0] != 0 || Mask[1] != 1 ||
20159 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20162 Chain.push_back(V);
20165 case X86ISD::PSHUFHW:
20166 // Check that the high words (being shuffled) are the identity in the
20167 // dword shuffle, and the low words are self-contained.
20168 if (Mask[2] != 2 || Mask[3] != 3 ||
20169 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20172 Chain.push_back(V);
20175 case X86ISD::UNPCKL:
20176 case X86ISD::UNPCKH:
20177 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20178 // shuffle into a preceding word shuffle.
20179 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
20180 V.getSimpleValueType().getScalarType() != MVT::i16)
20183 // Search for a half-shuffle which we can combine with.
20184 unsigned CombineOp =
20185 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20186 if (V.getOperand(0) != V.getOperand(1) ||
20187 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20189 Chain.push_back(V);
20190 V = V.getOperand(0);
20192 switch (V.getOpcode()) {
20194 return SDValue(); // Nothing to combine.
20196 case X86ISD::PSHUFLW:
20197 case X86ISD::PSHUFHW:
20198 if (V.getOpcode() == CombineOp)
20201 Chain.push_back(V);
20205 V = V.getOperand(0);
20209 } while (V.hasOneUse());
20212 // Break out of the loop if we break out of the switch.
20216 if (!V.hasOneUse())
20217 // We fell out of the loop without finding a viable combining instruction.
20220 // Merge this node's mask and our incoming mask.
20221 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20222 for (int &M : Mask)
20224 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20225 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20227 // Rebuild the chain around this new shuffle.
20228 while (!Chain.empty()) {
20229 SDValue W = Chain.pop_back_val();
20231 if (V.getValueType() != W.getOperand(0).getValueType())
20232 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
20234 switch (W.getOpcode()) {
20236 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20238 case X86ISD::UNPCKL:
20239 case X86ISD::UNPCKH:
20240 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20243 case X86ISD::PSHUFD:
20244 case X86ISD::PSHUFLW:
20245 case X86ISD::PSHUFHW:
20246 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20250 if (V.getValueType() != N.getValueType())
20251 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
20253 // Return the new chain to replace N.
20257 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20259 /// We walk up the chain, skipping shuffles of the other half and looking
20260 /// through shuffles which switch halves trying to find a shuffle of the same
20261 /// pair of dwords.
20262 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20264 TargetLowering::DAGCombinerInfo &DCI) {
20266 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20267 "Called with something other than an x86 128-bit half shuffle!");
20269 unsigned CombineOpcode = N.getOpcode();
20271 // Walk up a single-use chain looking for a combinable shuffle.
20272 SDValue V = N.getOperand(0);
20273 for (; V.hasOneUse(); V = V.getOperand(0)) {
20274 switch (V.getOpcode()) {
20276 return false; // Nothing combined!
20279 // Skip bitcasts as we always know the type for the target specific
20283 case X86ISD::PSHUFLW:
20284 case X86ISD::PSHUFHW:
20285 if (V.getOpcode() == CombineOpcode)
20288 // Other-half shuffles are no-ops.
20291 // Break out of the loop if we break out of the switch.
20295 if (!V.hasOneUse())
20296 // We fell out of the loop without finding a viable combining instruction.
20299 // Combine away the bottom node as its shuffle will be accumulated into
20300 // a preceding shuffle.
20301 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20303 // Record the old value.
20306 // Merge this node's mask and our incoming mask (adjusted to account for all
20307 // the pshufd instructions encountered).
20308 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20309 for (int &M : Mask)
20311 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20312 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20314 // Check that the shuffles didn't cancel each other out. If not, we need to
20315 // combine to the new one.
20317 // Replace the combinable shuffle with the combined one, updating all users
20318 // so that we re-evaluate the chain here.
20319 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20324 /// \brief Try to combine x86 target specific shuffles.
20325 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20326 TargetLowering::DAGCombinerInfo &DCI,
20327 const X86Subtarget *Subtarget) {
20329 MVT VT = N.getSimpleValueType();
20330 SmallVector<int, 4> Mask;
20332 switch (N.getOpcode()) {
20333 case X86ISD::PSHUFD:
20334 case X86ISD::PSHUFLW:
20335 case X86ISD::PSHUFHW:
20336 Mask = getPSHUFShuffleMask(N);
20337 assert(Mask.size() == 4);
20343 // Nuke no-op shuffles that show up after combining.
20344 if (isNoopShuffleMask(Mask))
20345 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20347 // Look for simplifications involving one or two shuffle instructions.
20348 SDValue V = N.getOperand(0);
20349 switch (N.getOpcode()) {
20352 case X86ISD::PSHUFLW:
20353 case X86ISD::PSHUFHW:
20354 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
20356 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20357 return SDValue(); // We combined away this shuffle, so we're done.
20359 // See if this reduces to a PSHUFD which is no more expensive and can
20360 // combine with more operations. Note that it has to at least flip the
20361 // dwords as otherwise it would have been removed as a no-op.
20362 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
20363 int DMask[] = {0, 1, 2, 3};
20364 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20365 DMask[DOffset + 0] = DOffset + 1;
20366 DMask[DOffset + 1] = DOffset + 0;
20367 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
20368 V = DAG.getNode(ISD::BITCAST, DL, DVT, V);
20369 DCI.AddToWorklist(V.getNode());
20370 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
20371 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
20372 DCI.AddToWorklist(V.getNode());
20373 return DAG.getNode(ISD::BITCAST, DL, VT, V);
20376 // Look for shuffle patterns which can be implemented as a single unpack.
20377 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20378 // only works when we have a PSHUFD followed by two half-shuffles.
20379 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20380 (V.getOpcode() == X86ISD::PSHUFLW ||
20381 V.getOpcode() == X86ISD::PSHUFHW) &&
20382 V.getOpcode() != N.getOpcode() &&
20384 SDValue D = V.getOperand(0);
20385 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20386 D = D.getOperand(0);
20387 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20388 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20389 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20390 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20391 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20393 for (int i = 0; i < 4; ++i) {
20394 WordMask[i + NOffset] = Mask[i] + NOffset;
20395 WordMask[i + VOffset] = VMask[i] + VOffset;
20397 // Map the word mask through the DWord mask.
20399 for (int i = 0; i < 8; ++i)
20400 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20401 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20402 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
20403 // We can replace all three shuffles with an unpack.
20404 V = DAG.getNode(ISD::BITCAST, DL, VT, D.getOperand(0));
20405 DCI.AddToWorklist(V.getNode());
20406 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
20415 case X86ISD::PSHUFD:
20416 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
20425 /// \brief Try to combine a shuffle into a target-specific add-sub node.
20427 /// We combine this directly on the abstract vector shuffle nodes so it is
20428 /// easier to generically match. We also insert dummy vector shuffle nodes for
20429 /// the operands which explicitly discard the lanes which are unused by this
20430 /// operation to try to flow through the rest of the combiner the fact that
20431 /// they're unused.
20432 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
20434 EVT VT = N->getValueType(0);
20436 // We only handle target-independent shuffles.
20437 // FIXME: It would be easy and harmless to use the target shuffle mask
20438 // extraction tool to support more.
20439 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
20442 auto *SVN = cast<ShuffleVectorSDNode>(N);
20443 ArrayRef<int> Mask = SVN->getMask();
20444 SDValue V1 = N->getOperand(0);
20445 SDValue V2 = N->getOperand(1);
20447 // We require the first shuffle operand to be the SUB node, and the second to
20448 // be the ADD node.
20449 // FIXME: We should support the commuted patterns.
20450 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
20453 // If there are other uses of these operations we can't fold them.
20454 if (!V1->hasOneUse() || !V2->hasOneUse())
20457 // Ensure that both operations have the same operands. Note that we can
20458 // commute the FADD operands.
20459 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
20460 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
20461 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
20464 // We're looking for blends between FADD and FSUB nodes. We insist on these
20465 // nodes being lined up in a specific expected pattern.
20466 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
20467 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
20468 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
20471 // Only specific types are legal at this point, assert so we notice if and
20472 // when these change.
20473 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
20474 VT == MVT::v4f64) &&
20475 "Unknown vector type encountered!");
20477 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
20480 /// PerformShuffleCombine - Performs several different shuffle combines.
20481 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
20482 TargetLowering::DAGCombinerInfo &DCI,
20483 const X86Subtarget *Subtarget) {
20485 SDValue N0 = N->getOperand(0);
20486 SDValue N1 = N->getOperand(1);
20487 EVT VT = N->getValueType(0);
20489 // Don't create instructions with illegal types after legalize types has run.
20490 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20491 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
20494 // If we have legalized the vector types, look for blends of FADD and FSUB
20495 // nodes that we can fuse into an ADDSUB node.
20496 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
20497 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
20500 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
20501 if (Subtarget->hasFp256() && VT.is256BitVector() &&
20502 N->getOpcode() == ISD::VECTOR_SHUFFLE)
20503 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
20505 // During Type Legalization, when promoting illegal vector types,
20506 // the backend might introduce new shuffle dag nodes and bitcasts.
20508 // This code performs the following transformation:
20509 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
20510 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
20512 // We do this only if both the bitcast and the BINOP dag nodes have
20513 // one use. Also, perform this transformation only if the new binary
20514 // operation is legal. This is to avoid introducing dag nodes that
20515 // potentially need to be further expanded (or custom lowered) into a
20516 // less optimal sequence of dag nodes.
20517 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
20518 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
20519 N0.getOpcode() == ISD::BITCAST) {
20520 SDValue BC0 = N0.getOperand(0);
20521 EVT SVT = BC0.getValueType();
20522 unsigned Opcode = BC0.getOpcode();
20523 unsigned NumElts = VT.getVectorNumElements();
20525 if (BC0.hasOneUse() && SVT.isVector() &&
20526 SVT.getVectorNumElements() * 2 == NumElts &&
20527 TLI.isOperationLegal(Opcode, VT)) {
20528 bool CanFold = false;
20540 unsigned SVTNumElts = SVT.getVectorNumElements();
20541 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20542 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
20543 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
20544 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
20545 CanFold = SVOp->getMaskElt(i) < 0;
20548 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
20549 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
20550 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
20551 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
20556 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
20557 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
20558 // consecutive, non-overlapping, and in the right order.
20559 SmallVector<SDValue, 16> Elts;
20560 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
20561 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
20563 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
20567 if (isTargetShuffle(N->getOpcode())) {
20569 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
20570 if (Shuffle.getNode())
20573 // Try recursively combining arbitrary sequences of x86 shuffle
20574 // instructions into higher-order shuffles. We do this after combining
20575 // specific PSHUF instruction sequences into their minimal form so that we
20576 // can evaluate how many specialized shuffle instructions are involved in
20577 // a particular chain.
20578 SmallVector<int, 1> NonceMask; // Just a placeholder.
20579 NonceMask.push_back(0);
20580 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
20581 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
20583 return SDValue(); // This routine will use CombineTo to replace N.
20589 /// PerformTruncateCombine - Converts truncate operation to
20590 /// a sequence of vector shuffle operations.
20591 /// It is possible when we truncate 256-bit vector to 128-bit vector
20592 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
20593 TargetLowering::DAGCombinerInfo &DCI,
20594 const X86Subtarget *Subtarget) {
20598 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
20599 /// specific shuffle of a load can be folded into a single element load.
20600 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
20601 /// shuffles have been custom lowered so we need to handle those here.
20602 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
20603 TargetLowering::DAGCombinerInfo &DCI) {
20604 if (DCI.isBeforeLegalizeOps())
20607 SDValue InVec = N->getOperand(0);
20608 SDValue EltNo = N->getOperand(1);
20610 if (!isa<ConstantSDNode>(EltNo))
20613 EVT OriginalVT = InVec.getValueType();
20615 if (InVec.getOpcode() == ISD::BITCAST) {
20616 // Don't duplicate a load with other uses.
20617 if (!InVec.hasOneUse())
20619 EVT BCVT = InVec.getOperand(0).getValueType();
20620 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
20622 InVec = InVec.getOperand(0);
20625 EVT CurrentVT = InVec.getValueType();
20627 if (!isTargetShuffle(InVec.getOpcode()))
20630 // Don't duplicate a load with other uses.
20631 if (!InVec.hasOneUse())
20634 SmallVector<int, 16> ShuffleMask;
20636 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
20637 ShuffleMask, UnaryShuffle))
20640 // Select the input vector, guarding against out of range extract vector.
20641 unsigned NumElems = CurrentVT.getVectorNumElements();
20642 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
20643 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
20644 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
20645 : InVec.getOperand(1);
20647 // If inputs to shuffle are the same for both ops, then allow 2 uses
20648 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
20649 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
20651 if (LdNode.getOpcode() == ISD::BITCAST) {
20652 // Don't duplicate a load with other uses.
20653 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
20656 AllowedUses = 1; // only allow 1 load use if we have a bitcast
20657 LdNode = LdNode.getOperand(0);
20660 if (!ISD::isNormalLoad(LdNode.getNode()))
20663 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
20665 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
20668 EVT EltVT = N->getValueType(0);
20669 // If there's a bitcast before the shuffle, check if the load type and
20670 // alignment is valid.
20671 unsigned Align = LN0->getAlignment();
20672 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20673 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
20674 EltVT.getTypeForEVT(*DAG.getContext()));
20676 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
20679 // All checks match so transform back to vector_shuffle so that DAG combiner
20680 // can finish the job
20683 // Create shuffle node taking into account the case that its a unary shuffle
20684 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20685 : InVec.getOperand(1);
20686 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20687 InVec.getOperand(0), Shuffle,
20689 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20690 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20694 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20695 /// special and don't usually play with other vector types, it's better to
20696 /// handle them early to be sure we emit efficient code by avoiding
20697 /// store-load conversions.
20698 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20699 if (N->getValueType(0) != MVT::x86mmx ||
20700 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20701 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20704 SDValue V = N->getOperand(0);
20705 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20706 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20707 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20708 N->getValueType(0), V.getOperand(0));
20713 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20714 /// generation and convert it from being a bunch of shuffles and extracts
20715 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20716 /// storing the value and loading scalars back, while for x64 we should
20717 /// use 64-bit extracts and shifts.
20718 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20719 TargetLowering::DAGCombinerInfo &DCI) {
20720 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20721 if (NewOp.getNode())
20724 SDValue InputVector = N->getOperand(0);
20726 // Detect mmx to i32 conversion through a v2i32 elt extract.
20727 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20728 N->getValueType(0) == MVT::i32 &&
20729 InputVector.getValueType() == MVT::v2i32) {
20731 // The bitcast source is a direct mmx result.
20732 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20733 if (MMXSrc.getValueType() == MVT::x86mmx)
20734 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20735 N->getValueType(0),
20736 InputVector.getNode()->getOperand(0));
20738 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20739 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20740 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20741 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20742 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20743 MMXSrcOp.getValueType() == MVT::v1i64 &&
20744 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20745 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20746 N->getValueType(0),
20747 MMXSrcOp.getOperand(0));
20750 // Only operate on vectors of 4 elements, where the alternative shuffling
20751 // gets to be more expensive.
20752 if (InputVector.getValueType() != MVT::v4i32)
20755 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20756 // single use which is a sign-extend or zero-extend, and all elements are
20758 SmallVector<SDNode *, 4> Uses;
20759 unsigned ExtractedElements = 0;
20760 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20761 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20762 if (UI.getUse().getResNo() != InputVector.getResNo())
20765 SDNode *Extract = *UI;
20766 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20769 if (Extract->getValueType(0) != MVT::i32)
20771 if (!Extract->hasOneUse())
20773 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20774 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20776 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20779 // Record which element was extracted.
20780 ExtractedElements |=
20781 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20783 Uses.push_back(Extract);
20786 // If not all the elements were used, this may not be worthwhile.
20787 if (ExtractedElements != 15)
20790 // Ok, we've now decided to do the transformation.
20791 // If 64-bit shifts are legal, use the extract-shift sequence,
20792 // otherwise bounce the vector off the cache.
20793 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20795 SDLoc dl(InputVector);
20797 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20798 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20799 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20800 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20801 DAG.getConstant(0, dl, VecIdxTy));
20802 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20803 DAG.getConstant(1, dl, VecIdxTy));
20805 SDValue ShAmt = DAG.getConstant(32, dl,
20806 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20807 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20808 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20809 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20810 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20811 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20812 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
20814 // Store the value to a temporary stack slot.
20815 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
20816 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
20817 MachinePointerInfo(), false, false, 0);
20819 EVT ElementType = InputVector.getValueType().getVectorElementType();
20820 unsigned EltSize = ElementType.getSizeInBits() / 8;
20822 // Replace each use (extract) with a load of the appropriate element.
20823 for (unsigned i = 0; i < 4; ++i) {
20824 uint64_t Offset = EltSize * i;
20825 SDValue OffsetVal = DAG.getConstant(Offset, dl, TLI.getPointerTy());
20827 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
20828 StackPtr, OffsetVal);
20830 // Load the scalar.
20831 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
20832 ScalarAddr, MachinePointerInfo(),
20833 false, false, false, 0);
20838 // Replace the extracts
20839 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
20840 UE = Uses.end(); UI != UE; ++UI) {
20841 SDNode *Extract = *UI;
20843 SDValue Idx = Extract->getOperand(1);
20844 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
20845 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
20848 // The replacement was made in place; don't return anything.
20852 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
20853 static std::pair<unsigned, bool>
20854 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
20855 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
20856 if (!VT.isVector())
20857 return std::make_pair(0, false);
20859 bool NeedSplit = false;
20860 switch (VT.getSimpleVT().SimpleTy) {
20861 default: return std::make_pair(0, false);
20864 if (!Subtarget->hasVLX())
20865 return std::make_pair(0, false);
20869 if (!Subtarget->hasBWI())
20870 return std::make_pair(0, false);
20874 if (!Subtarget->hasAVX512())
20875 return std::make_pair(0, false);
20880 if (!Subtarget->hasAVX2())
20882 if (!Subtarget->hasAVX())
20883 return std::make_pair(0, false);
20888 if (!Subtarget->hasSSE2())
20889 return std::make_pair(0, false);
20892 // SSE2 has only a small subset of the operations.
20893 bool hasUnsigned = Subtarget->hasSSE41() ||
20894 (Subtarget->hasSSE2() && VT == MVT::v16i8);
20895 bool hasSigned = Subtarget->hasSSE41() ||
20896 (Subtarget->hasSSE2() && VT == MVT::v8i16);
20898 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20901 // Check for x CC y ? x : y.
20902 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20903 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20908 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20911 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20914 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20917 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20919 // Check for x CC y ? y : x -- a min/max with reversed arms.
20920 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20921 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20926 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20929 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20932 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20935 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20939 return std::make_pair(Opc, NeedSplit);
20943 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
20944 const X86Subtarget *Subtarget) {
20946 SDValue Cond = N->getOperand(0);
20947 SDValue LHS = N->getOperand(1);
20948 SDValue RHS = N->getOperand(2);
20950 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
20951 SDValue CondSrc = Cond->getOperand(0);
20952 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
20953 Cond = CondSrc->getOperand(0);
20956 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
20959 // A vselect where all conditions and data are constants can be optimized into
20960 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
20961 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
20962 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
20965 unsigned MaskValue = 0;
20966 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
20969 MVT VT = N->getSimpleValueType(0);
20970 unsigned NumElems = VT.getVectorNumElements();
20971 SmallVector<int, 8> ShuffleMask(NumElems, -1);
20972 for (unsigned i = 0; i < NumElems; ++i) {
20973 // Be sure we emit undef where we can.
20974 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
20975 ShuffleMask[i] = -1;
20977 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
20980 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20981 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
20983 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
20986 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
20988 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
20989 TargetLowering::DAGCombinerInfo &DCI,
20990 const X86Subtarget *Subtarget) {
20992 SDValue Cond = N->getOperand(0);
20993 // Get the LHS/RHS of the select.
20994 SDValue LHS = N->getOperand(1);
20995 SDValue RHS = N->getOperand(2);
20996 EVT VT = LHS.getValueType();
20997 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20999 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
21000 // instructions match the semantics of the common C idiom x<y?x:y but not
21001 // x<=y?x:y, because of how they handle negative zero (which can be
21002 // ignored in unsafe-math mode).
21003 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
21004 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
21005 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
21006 (Subtarget->hasSSE2() ||
21007 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
21008 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21010 unsigned Opcode = 0;
21011 // Check for x CC y ? x : y.
21012 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21013 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21017 // Converting this to a min would handle NaNs incorrectly, and swapping
21018 // the operands would cause it to handle comparisons between positive
21019 // and negative zero incorrectly.
21020 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21021 if (!DAG.getTarget().Options.UnsafeFPMath &&
21022 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21024 std::swap(LHS, RHS);
21026 Opcode = X86ISD::FMIN;
21029 // Converting this to a min would handle comparisons between positive
21030 // and negative zero incorrectly.
21031 if (!DAG.getTarget().Options.UnsafeFPMath &&
21032 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21034 Opcode = X86ISD::FMIN;
21037 // Converting this to a min would handle both negative zeros and NaNs
21038 // incorrectly, but we can swap the operands to fix both.
21039 std::swap(LHS, RHS);
21043 Opcode = X86ISD::FMIN;
21047 // Converting this to a max would handle comparisons between positive
21048 // and negative zero incorrectly.
21049 if (!DAG.getTarget().Options.UnsafeFPMath &&
21050 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21052 Opcode = X86ISD::FMAX;
21055 // Converting this to a max would handle NaNs incorrectly, and swapping
21056 // the operands would cause it to handle comparisons between positive
21057 // and negative zero incorrectly.
21058 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21059 if (!DAG.getTarget().Options.UnsafeFPMath &&
21060 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21062 std::swap(LHS, RHS);
21064 Opcode = X86ISD::FMAX;
21067 // Converting this to a max would handle both negative zeros and NaNs
21068 // incorrectly, but we can swap the operands to fix both.
21069 std::swap(LHS, RHS);
21073 Opcode = X86ISD::FMAX;
21076 // Check for x CC y ? y : x -- a min/max with reversed arms.
21077 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21078 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21082 // Converting this to a min would handle comparisons between positive
21083 // and negative zero incorrectly, and swapping the operands would
21084 // cause it to handle NaNs incorrectly.
21085 if (!DAG.getTarget().Options.UnsafeFPMath &&
21086 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21087 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21089 std::swap(LHS, RHS);
21091 Opcode = X86ISD::FMIN;
21094 // Converting this to a min would handle NaNs incorrectly.
21095 if (!DAG.getTarget().Options.UnsafeFPMath &&
21096 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21098 Opcode = X86ISD::FMIN;
21101 // Converting this to a min would handle both negative zeros and NaNs
21102 // incorrectly, but we can swap the operands to fix both.
21103 std::swap(LHS, RHS);
21107 Opcode = X86ISD::FMIN;
21111 // Converting this to a max would handle NaNs incorrectly.
21112 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21114 Opcode = X86ISD::FMAX;
21117 // Converting this to a max would handle comparisons between positive
21118 // and negative zero incorrectly, and swapping the operands would
21119 // cause it to handle NaNs incorrectly.
21120 if (!DAG.getTarget().Options.UnsafeFPMath &&
21121 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21122 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21124 std::swap(LHS, RHS);
21126 Opcode = X86ISD::FMAX;
21129 // Converting this to a max would handle both negative zeros and NaNs
21130 // incorrectly, but we can swap the operands to fix both.
21131 std::swap(LHS, RHS);
21135 Opcode = X86ISD::FMAX;
21141 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21144 EVT CondVT = Cond.getValueType();
21145 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21146 CondVT.getVectorElementType() == MVT::i1) {
21147 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21148 // lowering on KNL. In this case we convert it to
21149 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21150 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21151 // Since SKX these selects have a proper lowering.
21152 EVT OpVT = LHS.getValueType();
21153 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21154 (OpVT.getVectorElementType() == MVT::i8 ||
21155 OpVT.getVectorElementType() == MVT::i16) &&
21156 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21157 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21158 DCI.AddToWorklist(Cond.getNode());
21159 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21162 // If this is a select between two integer constants, try to do some
21164 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21165 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21166 // Don't do this for crazy integer types.
21167 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21168 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21169 // so that TrueC (the true value) is larger than FalseC.
21170 bool NeedsCondInvert = false;
21172 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21173 // Efficiently invertible.
21174 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21175 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21176 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21177 NeedsCondInvert = true;
21178 std::swap(TrueC, FalseC);
21181 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21182 if (FalseC->getAPIntValue() == 0 &&
21183 TrueC->getAPIntValue().isPowerOf2()) {
21184 if (NeedsCondInvert) // Invert the condition if needed.
21185 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21186 DAG.getConstant(1, DL, Cond.getValueType()));
21188 // Zero extend the condition if needed.
21189 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21191 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21192 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21193 DAG.getConstant(ShAmt, DL, MVT::i8));
21196 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21197 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21198 if (NeedsCondInvert) // Invert the condition if needed.
21199 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21200 DAG.getConstant(1, DL, Cond.getValueType()));
21202 // Zero extend the condition if needed.
21203 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21204 FalseC->getValueType(0), Cond);
21205 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21206 SDValue(FalseC, 0));
21209 // Optimize cases that will turn into an LEA instruction. This requires
21210 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21211 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21212 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21213 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21215 bool isFastMultiplier = false;
21217 switch ((unsigned char)Diff) {
21219 case 1: // result = add base, cond
21220 case 2: // result = lea base( , cond*2)
21221 case 3: // result = lea base(cond, cond*2)
21222 case 4: // result = lea base( , cond*4)
21223 case 5: // result = lea base(cond, cond*4)
21224 case 8: // result = lea base( , cond*8)
21225 case 9: // result = lea base(cond, cond*8)
21226 isFastMultiplier = true;
21231 if (isFastMultiplier) {
21232 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21233 if (NeedsCondInvert) // Invert the condition if needed.
21234 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21235 DAG.getConstant(1, DL, Cond.getValueType()));
21237 // Zero extend the condition if needed.
21238 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21240 // Scale the condition by the difference.
21242 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21243 DAG.getConstant(Diff, DL,
21244 Cond.getValueType()));
21246 // Add the base if non-zero.
21247 if (FalseC->getAPIntValue() != 0)
21248 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21249 SDValue(FalseC, 0));
21256 // Canonicalize max and min:
21257 // (x > y) ? x : y -> (x >= y) ? x : y
21258 // (x < y) ? x : y -> (x <= y) ? x : y
21259 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21260 // the need for an extra compare
21261 // against zero. e.g.
21262 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21264 // testl %edi, %edi
21266 // cmovgl %edi, %eax
21270 // cmovsl %eax, %edi
21271 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21272 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21273 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21274 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21279 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21280 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21281 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21282 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21287 // Early exit check
21288 if (!TLI.isTypeLegal(VT))
21291 // Match VSELECTs into subs with unsigned saturation.
21292 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21293 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21294 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21295 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21296 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21298 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21299 // left side invert the predicate to simplify logic below.
21301 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21303 CC = ISD::getSetCCInverse(CC, true);
21304 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21308 if (Other.getNode() && Other->getNumOperands() == 2 &&
21309 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21310 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21311 SDValue CondRHS = Cond->getOperand(1);
21313 // Look for a general sub with unsigned saturation first.
21314 // x >= y ? x-y : 0 --> subus x, y
21315 // x > y ? x-y : 0 --> subus x, y
21316 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21317 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21318 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21320 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21321 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21322 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21323 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21324 // If the RHS is a constant we have to reverse the const
21325 // canonicalization.
21326 // x > C-1 ? x+-C : 0 --> subus x, C
21327 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21328 CondRHSConst->getAPIntValue() ==
21329 (-OpRHSConst->getAPIntValue() - 1))
21330 return DAG.getNode(
21331 X86ISD::SUBUS, DL, VT, OpLHS,
21332 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
21334 // Another special case: If C was a sign bit, the sub has been
21335 // canonicalized into a xor.
21336 // FIXME: Would it be better to use computeKnownBits to determine
21337 // whether it's safe to decanonicalize the xor?
21338 // x s< 0 ? x^C : 0 --> subus x, C
21339 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21340 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21341 OpRHSConst->getAPIntValue().isSignBit())
21342 // Note that we have to rebuild the RHS constant here to ensure we
21343 // don't rely on particular values of undef lanes.
21344 return DAG.getNode(
21345 X86ISD::SUBUS, DL, VT, OpLHS,
21346 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
21351 // Try to match a min/max vector operation.
21352 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21353 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21354 unsigned Opc = ret.first;
21355 bool NeedSplit = ret.second;
21357 if (Opc && NeedSplit) {
21358 unsigned NumElems = VT.getVectorNumElements();
21359 // Extract the LHS vectors
21360 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21361 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21363 // Extract the RHS vectors
21364 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21365 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21367 // Create min/max for each subvector
21368 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21369 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21371 // Merge the result
21372 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21374 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21377 // Simplify vector selection if condition value type matches vselect
21379 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
21380 assert(Cond.getValueType().isVector() &&
21381 "vector select expects a vector selector!");
21383 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21384 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21386 // Try invert the condition if true value is not all 1s and false value
21388 if (!TValIsAllOnes && !FValIsAllZeros &&
21389 // Check if the selector will be produced by CMPP*/PCMP*
21390 Cond.getOpcode() == ISD::SETCC &&
21391 // Check if SETCC has already been promoted
21392 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
21393 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21394 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21396 if (TValIsAllZeros || FValIsAllOnes) {
21397 SDValue CC = Cond.getOperand(2);
21398 ISD::CondCode NewCC =
21399 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21400 Cond.getOperand(0).getValueType().isInteger());
21401 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21402 std::swap(LHS, RHS);
21403 TValIsAllOnes = FValIsAllOnes;
21404 FValIsAllZeros = TValIsAllZeros;
21408 if (TValIsAllOnes || FValIsAllZeros) {
21411 if (TValIsAllOnes && FValIsAllZeros)
21413 else if (TValIsAllOnes)
21414 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
21415 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
21416 else if (FValIsAllZeros)
21417 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21418 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
21420 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
21424 // We should generate an X86ISD::BLENDI from a vselect if its argument
21425 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
21426 // constants. This specific pattern gets generated when we split a
21427 // selector for a 512 bit vector in a machine without AVX512 (but with
21428 // 256-bit vectors), during legalization:
21430 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
21432 // Iff we find this pattern and the build_vectors are built from
21433 // constants, we translate the vselect into a shuffle_vector that we
21434 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
21435 if ((N->getOpcode() == ISD::VSELECT ||
21436 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
21437 !DCI.isBeforeLegalize()) {
21438 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
21439 if (Shuffle.getNode())
21443 // If this is a *dynamic* select (non-constant condition) and we can match
21444 // this node with one of the variable blend instructions, restructure the
21445 // condition so that the blends can use the high bit of each element and use
21446 // SimplifyDemandedBits to simplify the condition operand.
21447 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
21448 !DCI.isBeforeLegalize() &&
21449 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
21450 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
21452 // Don't optimize vector selects that map to mask-registers.
21456 // We can only handle the cases where VSELECT is directly legal on the
21457 // subtarget. We custom lower VSELECT nodes with constant conditions and
21458 // this makes it hard to see whether a dynamic VSELECT will correctly
21459 // lower, so we both check the operation's status and explicitly handle the
21460 // cases where a *dynamic* blend will fail even though a constant-condition
21461 // blend could be custom lowered.
21462 // FIXME: We should find a better way to handle this class of problems.
21463 // Potentially, we should combine constant-condition vselect nodes
21464 // pre-legalization into shuffles and not mark as many types as custom
21466 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
21468 // FIXME: We don't support i16-element blends currently. We could and
21469 // should support them by making *all* the bits in the condition be set
21470 // rather than just the high bit and using an i8-element blend.
21471 if (VT.getScalarType() == MVT::i16)
21473 // Dynamic blending was only available from SSE4.1 onward.
21474 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
21476 // Byte blends are only available in AVX2
21477 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
21478 !Subtarget->hasAVX2())
21481 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
21482 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
21484 APInt KnownZero, KnownOne;
21485 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
21486 DCI.isBeforeLegalizeOps());
21487 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
21488 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
21490 // If we changed the computation somewhere in the DAG, this change
21491 // will affect all users of Cond.
21492 // Make sure it is fine and update all the nodes so that we do not
21493 // use the generic VSELECT anymore. Otherwise, we may perform
21494 // wrong optimizations as we messed up with the actual expectation
21495 // for the vector boolean values.
21496 if (Cond != TLO.Old) {
21497 // Check all uses of that condition operand to check whether it will be
21498 // consumed by non-BLEND instructions, which may depend on all bits are
21500 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21502 if (I->getOpcode() != ISD::VSELECT)
21503 // TODO: Add other opcodes eventually lowered into BLEND.
21506 // Update all the users of the condition, before committing the change,
21507 // so that the VSELECT optimizations that expect the correct vector
21508 // boolean value will not be triggered.
21509 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21511 DAG.ReplaceAllUsesOfValueWith(
21513 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
21514 Cond, I->getOperand(1), I->getOperand(2)));
21515 DCI.CommitTargetLoweringOpt(TLO);
21518 // At this point, only Cond is changed. Change the condition
21519 // just for N to keep the opportunity to optimize all other
21520 // users their own way.
21521 DAG.ReplaceAllUsesOfValueWith(
21523 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
21524 TLO.New, N->getOperand(1), N->getOperand(2)));
21532 // Check whether a boolean test is testing a boolean value generated by
21533 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
21536 // Simplify the following patterns:
21537 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
21538 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
21539 // to (Op EFLAGS Cond)
21541 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
21542 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
21543 // to (Op EFLAGS !Cond)
21545 // where Op could be BRCOND or CMOV.
21547 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
21548 // Quit if not CMP and SUB with its value result used.
21549 if (Cmp.getOpcode() != X86ISD::CMP &&
21550 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
21553 // Quit if not used as a boolean value.
21554 if (CC != X86::COND_E && CC != X86::COND_NE)
21557 // Check CMP operands. One of them should be 0 or 1 and the other should be
21558 // an SetCC or extended from it.
21559 SDValue Op1 = Cmp.getOperand(0);
21560 SDValue Op2 = Cmp.getOperand(1);
21563 const ConstantSDNode* C = nullptr;
21564 bool needOppositeCond = (CC == X86::COND_E);
21565 bool checkAgainstTrue = false; // Is it a comparison against 1?
21567 if ((C = dyn_cast<ConstantSDNode>(Op1)))
21569 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
21571 else // Quit if all operands are not constants.
21574 if (C->getZExtValue() == 1) {
21575 needOppositeCond = !needOppositeCond;
21576 checkAgainstTrue = true;
21577 } else if (C->getZExtValue() != 0)
21578 // Quit if the constant is neither 0 or 1.
21581 bool truncatedToBoolWithAnd = false;
21582 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
21583 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
21584 SetCC.getOpcode() == ISD::TRUNCATE ||
21585 SetCC.getOpcode() == ISD::AND) {
21586 if (SetCC.getOpcode() == ISD::AND) {
21588 ConstantSDNode *CS;
21589 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
21590 CS->getZExtValue() == 1)
21592 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
21593 CS->getZExtValue() == 1)
21597 SetCC = SetCC.getOperand(OpIdx);
21598 truncatedToBoolWithAnd = true;
21600 SetCC = SetCC.getOperand(0);
21603 switch (SetCC.getOpcode()) {
21604 case X86ISD::SETCC_CARRY:
21605 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
21606 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
21607 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
21608 // truncated to i1 using 'and'.
21609 if (checkAgainstTrue && !truncatedToBoolWithAnd)
21611 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
21612 "Invalid use of SETCC_CARRY!");
21614 case X86ISD::SETCC:
21615 // Set the condition code or opposite one if necessary.
21616 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
21617 if (needOppositeCond)
21618 CC = X86::GetOppositeBranchCondition(CC);
21619 return SetCC.getOperand(1);
21620 case X86ISD::CMOV: {
21621 // Check whether false/true value has canonical one, i.e. 0 or 1.
21622 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
21623 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
21624 // Quit if true value is not a constant.
21627 // Quit if false value is not a constant.
21629 SDValue Op = SetCC.getOperand(0);
21630 // Skip 'zext' or 'trunc' node.
21631 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
21632 Op.getOpcode() == ISD::TRUNCATE)
21633 Op = Op.getOperand(0);
21634 // A special case for rdrand/rdseed, where 0 is set if false cond is
21636 if ((Op.getOpcode() != X86ISD::RDRAND &&
21637 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
21640 // Quit if false value is not the constant 0 or 1.
21641 bool FValIsFalse = true;
21642 if (FVal && FVal->getZExtValue() != 0) {
21643 if (FVal->getZExtValue() != 1)
21645 // If FVal is 1, opposite cond is needed.
21646 needOppositeCond = !needOppositeCond;
21647 FValIsFalse = false;
21649 // Quit if TVal is not the constant opposite of FVal.
21650 if (FValIsFalse && TVal->getZExtValue() != 1)
21652 if (!FValIsFalse && TVal->getZExtValue() != 0)
21654 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
21655 if (needOppositeCond)
21656 CC = X86::GetOppositeBranchCondition(CC);
21657 return SetCC.getOperand(3);
21664 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
21666 /// (X86or (X86setcc) (X86setcc))
21667 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
21668 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
21669 X86::CondCode &CC1, SDValue &Flags,
21671 if (Cond->getOpcode() == X86ISD::CMP) {
21672 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
21673 if (!CondOp1C || !CondOp1C->isNullValue())
21676 Cond = Cond->getOperand(0);
21681 SDValue SetCC0, SetCC1;
21682 switch (Cond->getOpcode()) {
21683 default: return false;
21690 SetCC0 = Cond->getOperand(0);
21691 SetCC1 = Cond->getOperand(1);
21695 // Make sure we have SETCC nodes, using the same flags value.
21696 if (SetCC0.getOpcode() != X86ISD::SETCC ||
21697 SetCC1.getOpcode() != X86ISD::SETCC ||
21698 SetCC0->getOperand(1) != SetCC1->getOperand(1))
21701 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
21702 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
21703 Flags = SetCC0->getOperand(1);
21707 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
21708 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
21709 TargetLowering::DAGCombinerInfo &DCI,
21710 const X86Subtarget *Subtarget) {
21713 // If the flag operand isn't dead, don't touch this CMOV.
21714 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
21717 SDValue FalseOp = N->getOperand(0);
21718 SDValue TrueOp = N->getOperand(1);
21719 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
21720 SDValue Cond = N->getOperand(3);
21722 if (CC == X86::COND_E || CC == X86::COND_NE) {
21723 switch (Cond.getOpcode()) {
21727 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
21728 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
21729 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21735 Flags = checkBoolTestSetCCCombine(Cond, CC);
21736 if (Flags.getNode() &&
21737 // Extra check as FCMOV only supports a subset of X86 cond.
21738 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21739 SDValue Ops[] = { FalseOp, TrueOp,
21740 DAG.getConstant(CC, DL, MVT::i8), Flags };
21741 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21744 // If this is a select between two integer constants, try to do some
21745 // optimizations. Note that the operands are ordered the opposite of SELECT
21747 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21748 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21749 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21750 // larger than FalseC (the false value).
21751 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21752 CC = X86::GetOppositeBranchCondition(CC);
21753 std::swap(TrueC, FalseC);
21754 std::swap(TrueOp, FalseOp);
21757 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21758 // This is efficient for any integer data type (including i8/i16) and
21760 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21761 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21762 DAG.getConstant(CC, DL, MVT::i8), Cond);
21764 // Zero extend the condition if needed.
21765 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21767 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21768 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21769 DAG.getConstant(ShAmt, DL, MVT::i8));
21770 if (N->getNumValues() == 2) // Dead flag value?
21771 return DCI.CombineTo(N, Cond, SDValue());
21775 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21776 // for any integer data type, including i8/i16.
21777 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21778 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21779 DAG.getConstant(CC, DL, MVT::i8), Cond);
21781 // Zero extend the condition if needed.
21782 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21783 FalseC->getValueType(0), Cond);
21784 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21785 SDValue(FalseC, 0));
21787 if (N->getNumValues() == 2) // Dead flag value?
21788 return DCI.CombineTo(N, Cond, SDValue());
21792 // Optimize cases that will turn into an LEA instruction. This requires
21793 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21794 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21795 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21796 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21798 bool isFastMultiplier = false;
21800 switch ((unsigned char)Diff) {
21802 case 1: // result = add base, cond
21803 case 2: // result = lea base( , cond*2)
21804 case 3: // result = lea base(cond, cond*2)
21805 case 4: // result = lea base( , cond*4)
21806 case 5: // result = lea base(cond, cond*4)
21807 case 8: // result = lea base( , cond*8)
21808 case 9: // result = lea base(cond, cond*8)
21809 isFastMultiplier = true;
21814 if (isFastMultiplier) {
21815 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21816 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21817 DAG.getConstant(CC, DL, MVT::i8), Cond);
21818 // Zero extend the condition if needed.
21819 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21821 // Scale the condition by the difference.
21823 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21824 DAG.getConstant(Diff, DL, Cond.getValueType()));
21826 // Add the base if non-zero.
21827 if (FalseC->getAPIntValue() != 0)
21828 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21829 SDValue(FalseC, 0));
21830 if (N->getNumValues() == 2) // Dead flag value?
21831 return DCI.CombineTo(N, Cond, SDValue());
21838 // Handle these cases:
21839 // (select (x != c), e, c) -> select (x != c), e, x),
21840 // (select (x == c), c, e) -> select (x == c), x, e)
21841 // where the c is an integer constant, and the "select" is the combination
21842 // of CMOV and CMP.
21844 // The rationale for this change is that the conditional-move from a constant
21845 // needs two instructions, however, conditional-move from a register needs
21846 // only one instruction.
21848 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
21849 // some instruction-combining opportunities. This opt needs to be
21850 // postponed as late as possible.
21852 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
21853 // the DCI.xxxx conditions are provided to postpone the optimization as
21854 // late as possible.
21856 ConstantSDNode *CmpAgainst = nullptr;
21857 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
21858 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
21859 !isa<ConstantSDNode>(Cond.getOperand(0))) {
21861 if (CC == X86::COND_NE &&
21862 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
21863 CC = X86::GetOppositeBranchCondition(CC);
21864 std::swap(TrueOp, FalseOp);
21867 if (CC == X86::COND_E &&
21868 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
21869 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
21870 DAG.getConstant(CC, DL, MVT::i8), Cond };
21871 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
21876 // Fold and/or of setcc's to double CMOV:
21877 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
21878 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
21880 // This combine lets us generate:
21881 // cmovcc1 (jcc1 if we don't have CMOV)
21887 // cmovne (jne if we don't have CMOV)
21888 // When we can't use the CMOV instruction, it might increase branch
21890 // When we can use CMOV, or when there is no mispredict, this improves
21891 // throughput and reduces register pressure.
21893 if (CC == X86::COND_NE) {
21895 X86::CondCode CC0, CC1;
21897 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
21899 std::swap(FalseOp, TrueOp);
21900 CC0 = X86::GetOppositeBranchCondition(CC0);
21901 CC1 = X86::GetOppositeBranchCondition(CC1);
21904 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
21906 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
21907 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
21908 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21909 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
21917 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
21918 const X86Subtarget *Subtarget) {
21919 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
21921 default: return SDValue();
21922 // SSE/AVX/AVX2 blend intrinsics.
21923 case Intrinsic::x86_avx2_pblendvb:
21924 // Don't try to simplify this intrinsic if we don't have AVX2.
21925 if (!Subtarget->hasAVX2())
21928 case Intrinsic::x86_avx_blendv_pd_256:
21929 case Intrinsic::x86_avx_blendv_ps_256:
21930 // Don't try to simplify this intrinsic if we don't have AVX.
21931 if (!Subtarget->hasAVX())
21934 case Intrinsic::x86_sse41_blendvps:
21935 case Intrinsic::x86_sse41_blendvpd:
21936 case Intrinsic::x86_sse41_pblendvb: {
21937 SDValue Op0 = N->getOperand(1);
21938 SDValue Op1 = N->getOperand(2);
21939 SDValue Mask = N->getOperand(3);
21941 // Don't try to simplify this intrinsic if we don't have SSE4.1.
21942 if (!Subtarget->hasSSE41())
21945 // fold (blend A, A, Mask) -> A
21948 // fold (blend A, B, allZeros) -> A
21949 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
21951 // fold (blend A, B, allOnes) -> B
21952 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
21955 // Simplify the case where the mask is a constant i32 value.
21956 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
21957 if (C->isNullValue())
21959 if (C->isAllOnesValue())
21966 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
21967 case Intrinsic::x86_sse2_psrai_w:
21968 case Intrinsic::x86_sse2_psrai_d:
21969 case Intrinsic::x86_avx2_psrai_w:
21970 case Intrinsic::x86_avx2_psrai_d:
21971 case Intrinsic::x86_sse2_psra_w:
21972 case Intrinsic::x86_sse2_psra_d:
21973 case Intrinsic::x86_avx2_psra_w:
21974 case Intrinsic::x86_avx2_psra_d: {
21975 SDValue Op0 = N->getOperand(1);
21976 SDValue Op1 = N->getOperand(2);
21977 EVT VT = Op0.getValueType();
21978 assert(VT.isVector() && "Expected a vector type!");
21980 if (isa<BuildVectorSDNode>(Op1))
21981 Op1 = Op1.getOperand(0);
21983 if (!isa<ConstantSDNode>(Op1))
21986 EVT SVT = VT.getVectorElementType();
21987 unsigned SVTBits = SVT.getSizeInBits();
21989 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
21990 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
21991 uint64_t ShAmt = C.getZExtValue();
21993 // Don't try to convert this shift into a ISD::SRA if the shift
21994 // count is bigger than or equal to the element size.
21995 if (ShAmt >= SVTBits)
21998 // Trivial case: if the shift count is zero, then fold this
21999 // into the first operand.
22003 // Replace this packed shift intrinsic with a target independent
22006 SDValue Splat = DAG.getConstant(C, DL, VT);
22007 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
22012 /// PerformMulCombine - Optimize a single multiply with constant into two
22013 /// in order to implement it with two cheaper instructions, e.g.
22014 /// LEA + SHL, LEA + LEA.
22015 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
22016 TargetLowering::DAGCombinerInfo &DCI) {
22017 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
22020 EVT VT = N->getValueType(0);
22021 if (VT != MVT::i64 && VT != MVT::i32)
22024 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
22027 uint64_t MulAmt = C->getZExtValue();
22028 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
22031 uint64_t MulAmt1 = 0;
22032 uint64_t MulAmt2 = 0;
22033 if ((MulAmt % 9) == 0) {
22035 MulAmt2 = MulAmt / 9;
22036 } else if ((MulAmt % 5) == 0) {
22038 MulAmt2 = MulAmt / 5;
22039 } else if ((MulAmt % 3) == 0) {
22041 MulAmt2 = MulAmt / 3;
22044 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
22047 if (isPowerOf2_64(MulAmt2) &&
22048 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
22049 // If second multiplifer is pow2, issue it first. We want the multiply by
22050 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
22052 std::swap(MulAmt1, MulAmt2);
22055 if (isPowerOf2_64(MulAmt1))
22056 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
22057 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
22059 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
22060 DAG.getConstant(MulAmt1, DL, VT));
22062 if (isPowerOf2_64(MulAmt2))
22063 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
22064 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
22066 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
22067 DAG.getConstant(MulAmt2, DL, VT));
22069 // Do not add new nodes to DAG combiner worklist.
22070 DCI.CombineTo(N, NewMul, false);
22075 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22076 SDValue N0 = N->getOperand(0);
22077 SDValue N1 = N->getOperand(1);
22078 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22079 EVT VT = N0.getValueType();
22081 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22082 // since the result of setcc_c is all zero's or all ones.
22083 if (VT.isInteger() && !VT.isVector() &&
22084 N1C && N0.getOpcode() == ISD::AND &&
22085 N0.getOperand(1).getOpcode() == ISD::Constant) {
22086 SDValue N00 = N0.getOperand(0);
22087 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22088 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22089 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22090 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22091 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22092 APInt ShAmt = N1C->getAPIntValue();
22093 Mask = Mask.shl(ShAmt);
22096 return DAG.getNode(ISD::AND, DL, VT,
22097 N00, DAG.getConstant(Mask, DL, VT));
22102 // Hardware support for vector shifts is sparse which makes us scalarize the
22103 // vector operations in many cases. Also, on sandybridge ADD is faster than
22105 // (shl V, 1) -> add V,V
22106 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22107 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22108 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22109 // We shift all of the values by one. In many cases we do not have
22110 // hardware support for this operation. This is better expressed as an ADD
22112 if (N1SplatC->getZExtValue() == 1)
22113 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22119 /// \brief Returns a vector of 0s if the node in input is a vector logical
22120 /// shift by a constant amount which is known to be bigger than or equal
22121 /// to the vector element size in bits.
22122 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22123 const X86Subtarget *Subtarget) {
22124 EVT VT = N->getValueType(0);
22126 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22127 (!Subtarget->hasInt256() ||
22128 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22131 SDValue Amt = N->getOperand(1);
22133 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22134 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22135 APInt ShiftAmt = AmtSplat->getAPIntValue();
22136 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22138 // SSE2/AVX2 logical shifts always return a vector of 0s
22139 // if the shift amount is bigger than or equal to
22140 // the element size. The constant shift amount will be
22141 // encoded as a 8-bit immediate.
22142 if (ShiftAmt.trunc(8).uge(MaxAmount))
22143 return getZeroVector(VT, Subtarget, DAG, DL);
22149 /// PerformShiftCombine - Combine shifts.
22150 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22151 TargetLowering::DAGCombinerInfo &DCI,
22152 const X86Subtarget *Subtarget) {
22153 if (N->getOpcode() == ISD::SHL) {
22154 SDValue V = PerformSHLCombine(N, DAG);
22155 if (V.getNode()) return V;
22158 if (N->getOpcode() != ISD::SRA) {
22159 // Try to fold this logical shift into a zero vector.
22160 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22161 if (V.getNode()) return V;
22167 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22168 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22169 // and friends. Likewise for OR -> CMPNEQSS.
22170 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22171 TargetLowering::DAGCombinerInfo &DCI,
22172 const X86Subtarget *Subtarget) {
22175 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22176 // we're requiring SSE2 for both.
22177 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22178 SDValue N0 = N->getOperand(0);
22179 SDValue N1 = N->getOperand(1);
22180 SDValue CMP0 = N0->getOperand(1);
22181 SDValue CMP1 = N1->getOperand(1);
22184 // The SETCCs should both refer to the same CMP.
22185 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22188 SDValue CMP00 = CMP0->getOperand(0);
22189 SDValue CMP01 = CMP0->getOperand(1);
22190 EVT VT = CMP00.getValueType();
22192 if (VT == MVT::f32 || VT == MVT::f64) {
22193 bool ExpectingFlags = false;
22194 // Check for any users that want flags:
22195 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22196 !ExpectingFlags && UI != UE; ++UI)
22197 switch (UI->getOpcode()) {
22202 ExpectingFlags = true;
22204 case ISD::CopyToReg:
22205 case ISD::SIGN_EXTEND:
22206 case ISD::ZERO_EXTEND:
22207 case ISD::ANY_EXTEND:
22211 if (!ExpectingFlags) {
22212 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22213 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22215 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22216 X86::CondCode tmp = cc0;
22221 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22222 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22223 // FIXME: need symbolic constants for these magic numbers.
22224 // See X86ATTInstPrinter.cpp:printSSECC().
22225 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22226 if (Subtarget->hasAVX512()) {
22227 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22229 DAG.getConstant(x86cc, DL, MVT::i8));
22230 if (N->getValueType(0) != MVT::i1)
22231 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22235 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22236 CMP00.getValueType(), CMP00, CMP01,
22237 DAG.getConstant(x86cc, DL,
22240 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22241 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22243 if (is64BitFP && !Subtarget->is64Bit()) {
22244 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22245 // 64-bit integer, since that's not a legal type. Since
22246 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22247 // bits, but can do this little dance to extract the lowest 32 bits
22248 // and work with those going forward.
22249 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22251 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
22253 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22254 Vector32, DAG.getIntPtrConstant(0, DL));
22258 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
22260 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22261 DAG.getConstant(1, DL, IntVT));
22262 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
22264 return OneBitOfTruth;
22272 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22273 /// so it can be folded inside ANDNP.
22274 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22275 EVT VT = N->getValueType(0);
22277 // Match direct AllOnes for 128 and 256-bit vectors
22278 if (ISD::isBuildVectorAllOnes(N))
22281 // Look through a bit convert.
22282 if (N->getOpcode() == ISD::BITCAST)
22283 N = N->getOperand(0).getNode();
22285 // Sometimes the operand may come from a insert_subvector building a 256-bit
22287 if (VT.is256BitVector() &&
22288 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22289 SDValue V1 = N->getOperand(0);
22290 SDValue V2 = N->getOperand(1);
22292 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22293 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22294 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22295 ISD::isBuildVectorAllOnes(V2.getNode()))
22302 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22303 // register. In most cases we actually compare or select YMM-sized registers
22304 // and mixing the two types creates horrible code. This method optimizes
22305 // some of the transition sequences.
22306 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22307 TargetLowering::DAGCombinerInfo &DCI,
22308 const X86Subtarget *Subtarget) {
22309 EVT VT = N->getValueType(0);
22310 if (!VT.is256BitVector())
22313 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22314 N->getOpcode() == ISD::ZERO_EXTEND ||
22315 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22317 SDValue Narrow = N->getOperand(0);
22318 EVT NarrowVT = Narrow->getValueType(0);
22319 if (!NarrowVT.is128BitVector())
22322 if (Narrow->getOpcode() != ISD::XOR &&
22323 Narrow->getOpcode() != ISD::AND &&
22324 Narrow->getOpcode() != ISD::OR)
22327 SDValue N0 = Narrow->getOperand(0);
22328 SDValue N1 = Narrow->getOperand(1);
22331 // The Left side has to be a trunc.
22332 if (N0.getOpcode() != ISD::TRUNCATE)
22335 // The type of the truncated inputs.
22336 EVT WideVT = N0->getOperand(0)->getValueType(0);
22340 // The right side has to be a 'trunc' or a constant vector.
22341 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22342 ConstantSDNode *RHSConstSplat = nullptr;
22343 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22344 RHSConstSplat = RHSBV->getConstantSplatNode();
22345 if (!RHSTrunc && !RHSConstSplat)
22348 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22350 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22353 // Set N0 and N1 to hold the inputs to the new wide operation.
22354 N0 = N0->getOperand(0);
22355 if (RHSConstSplat) {
22356 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22357 SDValue(RHSConstSplat, 0));
22358 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22359 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22360 } else if (RHSTrunc) {
22361 N1 = N1->getOperand(0);
22364 // Generate the wide operation.
22365 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22366 unsigned Opcode = N->getOpcode();
22368 case ISD::ANY_EXTEND:
22370 case ISD::ZERO_EXTEND: {
22371 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22372 APInt Mask = APInt::getAllOnesValue(InBits);
22373 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22374 return DAG.getNode(ISD::AND, DL, VT,
22375 Op, DAG.getConstant(Mask, DL, VT));
22377 case ISD::SIGN_EXTEND:
22378 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22379 Op, DAG.getValueType(NarrowVT));
22381 llvm_unreachable("Unexpected opcode");
22385 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
22386 TargetLowering::DAGCombinerInfo &DCI,
22387 const X86Subtarget *Subtarget) {
22388 SDValue N0 = N->getOperand(0);
22389 SDValue N1 = N->getOperand(1);
22392 // A vector zext_in_reg may be represented as a shuffle,
22393 // feeding into a bitcast (this represents anyext) feeding into
22394 // an and with a mask.
22395 // We'd like to try to combine that into a shuffle with zero
22396 // plus a bitcast, removing the and.
22397 if (N0.getOpcode() != ISD::BITCAST ||
22398 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
22401 // The other side of the AND should be a splat of 2^C, where C
22402 // is the number of bits in the source type.
22403 if (N1.getOpcode() == ISD::BITCAST)
22404 N1 = N1.getOperand(0);
22405 if (N1.getOpcode() != ISD::BUILD_VECTOR)
22407 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
22409 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
22410 EVT SrcType = Shuffle->getValueType(0);
22412 // We expect a single-source shuffle
22413 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
22416 unsigned SrcSize = SrcType.getScalarSizeInBits();
22418 APInt SplatValue, SplatUndef;
22419 unsigned SplatBitSize;
22421 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
22422 SplatBitSize, HasAnyUndefs))
22425 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
22426 // Make sure the splat matches the mask we expect
22427 if (SplatBitSize > ResSize ||
22428 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
22431 // Make sure the input and output size make sense
22432 if (SrcSize >= ResSize || ResSize % SrcSize)
22435 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
22436 // The number of u's between each two values depends on the ratio between
22437 // the source and dest type.
22438 unsigned ZextRatio = ResSize / SrcSize;
22439 bool IsZext = true;
22440 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
22441 if (i % ZextRatio) {
22442 if (Shuffle->getMaskElt(i) > 0) {
22448 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
22449 // Expected element number
22459 // Ok, perform the transformation - replace the shuffle with
22460 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
22461 // (instead of undef) where the k elements come from the zero vector.
22462 SmallVector<int, 8> Mask;
22463 unsigned NumElems = SrcType.getVectorNumElements();
22464 for (unsigned i = 0; i < NumElems; ++i)
22466 Mask.push_back(NumElems);
22468 Mask.push_back(i / ZextRatio);
22470 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
22471 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
22472 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
22475 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22476 TargetLowering::DAGCombinerInfo &DCI,
22477 const X86Subtarget *Subtarget) {
22478 if (DCI.isBeforeLegalizeOps())
22481 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
22484 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
22487 EVT VT = N->getValueType(0);
22488 SDValue N0 = N->getOperand(0);
22489 SDValue N1 = N->getOperand(1);
22492 // Create BEXTR instructions
22493 // BEXTR is ((X >> imm) & (2**size-1))
22494 if (VT == MVT::i32 || VT == MVT::i64) {
22495 // Check for BEXTR.
22496 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22497 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22498 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22499 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22500 if (MaskNode && ShiftNode) {
22501 uint64_t Mask = MaskNode->getZExtValue();
22502 uint64_t Shift = ShiftNode->getZExtValue();
22503 if (isMask_64(Mask)) {
22504 uint64_t MaskSize = countPopulation(Mask);
22505 if (Shift + MaskSize <= VT.getSizeInBits())
22506 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22507 DAG.getConstant(Shift | (MaskSize << 8), DL,
22516 // Want to form ANDNP nodes:
22517 // 1) In the hopes of then easily combining them with OR and AND nodes
22518 // to form PBLEND/PSIGN.
22519 // 2) To match ANDN packed intrinsics
22520 if (VT != MVT::v2i64 && VT != MVT::v4i64)
22523 // Check LHS for vnot
22524 if (N0.getOpcode() == ISD::XOR &&
22525 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
22526 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
22527 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
22529 // Check RHS for vnot
22530 if (N1.getOpcode() == ISD::XOR &&
22531 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
22532 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
22533 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
22538 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
22539 TargetLowering::DAGCombinerInfo &DCI,
22540 const X86Subtarget *Subtarget) {
22541 if (DCI.isBeforeLegalizeOps())
22544 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22548 SDValue N0 = N->getOperand(0);
22549 SDValue N1 = N->getOperand(1);
22550 EVT VT = N->getValueType(0);
22552 // look for psign/blend
22553 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
22554 if (!Subtarget->hasSSSE3() ||
22555 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
22558 // Canonicalize pandn to RHS
22559 if (N0.getOpcode() == X86ISD::ANDNP)
22561 // or (and (m, y), (pandn m, x))
22562 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
22563 SDValue Mask = N1.getOperand(0);
22564 SDValue X = N1.getOperand(1);
22566 if (N0.getOperand(0) == Mask)
22567 Y = N0.getOperand(1);
22568 if (N0.getOperand(1) == Mask)
22569 Y = N0.getOperand(0);
22571 // Check to see if the mask appeared in both the AND and ANDNP and
22575 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
22576 // Look through mask bitcast.
22577 if (Mask.getOpcode() == ISD::BITCAST)
22578 Mask = Mask.getOperand(0);
22579 if (X.getOpcode() == ISD::BITCAST)
22580 X = X.getOperand(0);
22581 if (Y.getOpcode() == ISD::BITCAST)
22582 Y = Y.getOperand(0);
22584 EVT MaskVT = Mask.getValueType();
22586 // Validate that the Mask operand is a vector sra node.
22587 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
22588 // there is no psrai.b
22589 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
22590 unsigned SraAmt = ~0;
22591 if (Mask.getOpcode() == ISD::SRA) {
22592 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
22593 if (auto *AmtConst = AmtBV->getConstantSplatNode())
22594 SraAmt = AmtConst->getZExtValue();
22595 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
22596 SDValue SraC = Mask.getOperand(1);
22597 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
22599 if ((SraAmt + 1) != EltBits)
22604 // Now we know we at least have a plendvb with the mask val. See if
22605 // we can form a psignb/w/d.
22606 // psign = x.type == y.type == mask.type && y = sub(0, x);
22607 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
22608 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
22609 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
22610 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
22611 "Unsupported VT for PSIGN");
22612 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
22613 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22615 // PBLENDVB only available on SSE 4.1
22616 if (!Subtarget->hasSSE41())
22619 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
22621 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
22622 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
22623 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
22624 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
22625 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22629 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
22632 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
22633 MachineFunction &MF = DAG.getMachineFunction();
22635 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
22637 // SHLD/SHRD instructions have lower register pressure, but on some
22638 // platforms they have higher latency than the equivalent
22639 // series of shifts/or that would otherwise be generated.
22640 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
22641 // have higher latencies and we are not optimizing for size.
22642 if (!OptForSize && Subtarget->isSHLDSlow())
22645 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
22647 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
22649 if (!N0.hasOneUse() || !N1.hasOneUse())
22652 SDValue ShAmt0 = N0.getOperand(1);
22653 if (ShAmt0.getValueType() != MVT::i8)
22655 SDValue ShAmt1 = N1.getOperand(1);
22656 if (ShAmt1.getValueType() != MVT::i8)
22658 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
22659 ShAmt0 = ShAmt0.getOperand(0);
22660 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
22661 ShAmt1 = ShAmt1.getOperand(0);
22664 unsigned Opc = X86ISD::SHLD;
22665 SDValue Op0 = N0.getOperand(0);
22666 SDValue Op1 = N1.getOperand(0);
22667 if (ShAmt0.getOpcode() == ISD::SUB) {
22668 Opc = X86ISD::SHRD;
22669 std::swap(Op0, Op1);
22670 std::swap(ShAmt0, ShAmt1);
22673 unsigned Bits = VT.getSizeInBits();
22674 if (ShAmt1.getOpcode() == ISD::SUB) {
22675 SDValue Sum = ShAmt1.getOperand(0);
22676 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
22677 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
22678 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
22679 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
22680 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
22681 return DAG.getNode(Opc, DL, VT,
22683 DAG.getNode(ISD::TRUNCATE, DL,
22686 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
22687 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
22689 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
22690 return DAG.getNode(Opc, DL, VT,
22691 N0.getOperand(0), N1.getOperand(0),
22692 DAG.getNode(ISD::TRUNCATE, DL,
22699 // Generate NEG and CMOV for integer abs.
22700 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
22701 EVT VT = N->getValueType(0);
22703 // Since X86 does not have CMOV for 8-bit integer, we don't convert
22704 // 8-bit integer abs to NEG and CMOV.
22705 if (VT.isInteger() && VT.getSizeInBits() == 8)
22708 SDValue N0 = N->getOperand(0);
22709 SDValue N1 = N->getOperand(1);
22712 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
22713 // and change it to SUB and CMOV.
22714 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
22715 N0.getOpcode() == ISD::ADD &&
22716 N0.getOperand(1) == N1 &&
22717 N1.getOpcode() == ISD::SRA &&
22718 N1.getOperand(0) == N0.getOperand(0))
22719 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
22720 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
22721 // Generate SUB & CMOV.
22722 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
22723 DAG.getConstant(0, DL, VT), N0.getOperand(0));
22725 SDValue Ops[] = { N0.getOperand(0), Neg,
22726 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
22727 SDValue(Neg.getNode(), 1) };
22728 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
22733 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
22734 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
22735 TargetLowering::DAGCombinerInfo &DCI,
22736 const X86Subtarget *Subtarget) {
22737 if (DCI.isBeforeLegalizeOps())
22740 if (Subtarget->hasCMov()) {
22741 SDValue RV = performIntegerAbsCombine(N, DAG);
22749 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
22750 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
22751 TargetLowering::DAGCombinerInfo &DCI,
22752 const X86Subtarget *Subtarget) {
22753 LoadSDNode *Ld = cast<LoadSDNode>(N);
22754 EVT RegVT = Ld->getValueType(0);
22755 EVT MemVT = Ld->getMemoryVT();
22757 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22759 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
22760 // into two 16-byte operations.
22761 ISD::LoadExtType Ext = Ld->getExtensionType();
22762 unsigned Alignment = Ld->getAlignment();
22763 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
22764 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22765 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
22766 unsigned NumElems = RegVT.getVectorNumElements();
22770 SDValue Ptr = Ld->getBasePtr();
22771 SDValue Increment = DAG.getConstant(16, dl, TLI.getPointerTy());
22773 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
22775 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22776 Ld->getPointerInfo(), Ld->isVolatile(),
22777 Ld->isNonTemporal(), Ld->isInvariant(),
22779 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22780 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22781 Ld->getPointerInfo(), Ld->isVolatile(),
22782 Ld->isNonTemporal(), Ld->isInvariant(),
22783 std::min(16U, Alignment));
22784 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22786 Load2.getValue(1));
22788 SDValue NewVec = DAG.getUNDEF(RegVT);
22789 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22790 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22791 return DCI.CombineTo(N, NewVec, TF, true);
22797 /// PerformMLOADCombine - Resolve extending loads
22798 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22799 TargetLowering::DAGCombinerInfo &DCI,
22800 const X86Subtarget *Subtarget) {
22801 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22802 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22805 EVT VT = Mld->getValueType(0);
22806 unsigned NumElems = VT.getVectorNumElements();
22807 EVT LdVT = Mld->getMemoryVT();
22810 assert(LdVT != VT && "Cannot extend to the same type");
22811 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
22812 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
22813 // From, To sizes and ElemCount must be pow of two
22814 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22815 "Unexpected size for extending masked load");
22817 unsigned SizeRatio = ToSz / FromSz;
22818 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
22820 // Create a type on which we perform the shuffle
22821 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22822 LdVT.getScalarType(), NumElems*SizeRatio);
22823 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22825 // Convert Src0 value
22826 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
22827 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
22828 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22829 for (unsigned i = 0; i != NumElems; ++i)
22830 ShuffleVec[i] = i * SizeRatio;
22832 // Can't shuffle using an illegal type.
22833 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22834 && "WideVecVT should be legal");
22835 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
22836 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
22838 // Prepare the new mask
22840 SDValue Mask = Mld->getMask();
22841 if (Mask.getValueType() == VT) {
22842 // Mask and original value have the same type
22843 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22844 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22845 for (unsigned i = 0; i != NumElems; ++i)
22846 ShuffleVec[i] = i * SizeRatio;
22847 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22848 ShuffleVec[i] = NumElems*SizeRatio;
22849 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22850 DAG.getConstant(0, dl, WideVecVT),
22854 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22855 unsigned WidenNumElts = NumElems*SizeRatio;
22856 unsigned MaskNumElts = VT.getVectorNumElements();
22857 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22860 unsigned NumConcat = WidenNumElts / MaskNumElts;
22861 SmallVector<SDValue, 16> Ops(NumConcat);
22862 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
22864 for (unsigned i = 1; i != NumConcat; ++i)
22867 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22870 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
22871 Mld->getBasePtr(), NewMask, WideSrc0,
22872 Mld->getMemoryVT(), Mld->getMemOperand(),
22874 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
22875 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
22878 /// PerformMSTORECombine - Resolve truncating stores
22879 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
22880 const X86Subtarget *Subtarget) {
22881 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
22882 if (!Mst->isTruncatingStore())
22885 EVT VT = Mst->getValue().getValueType();
22886 unsigned NumElems = VT.getVectorNumElements();
22887 EVT StVT = Mst->getMemoryVT();
22890 assert(StVT != VT && "Cannot truncate to the same type");
22891 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22892 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22894 // From, To sizes and ElemCount must be pow of two
22895 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22896 "Unexpected size for truncating masked store");
22897 // We are going to use the original vector elt for storing.
22898 // Accumulated smaller vector elements must be a multiple of the store size.
22899 assert (((NumElems * FromSz) % ToSz) == 0 &&
22900 "Unexpected ratio for truncating masked store");
22902 unsigned SizeRatio = FromSz / ToSz;
22903 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22905 // Create a type on which we perform the shuffle
22906 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22907 StVT.getScalarType(), NumElems*SizeRatio);
22909 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22911 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
22912 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22913 for (unsigned i = 0; i != NumElems; ++i)
22914 ShuffleVec[i] = i * SizeRatio;
22916 // Can't shuffle using an illegal type.
22917 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22918 && "WideVecVT should be legal");
22920 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22921 DAG.getUNDEF(WideVecVT),
22925 SDValue Mask = Mst->getMask();
22926 if (Mask.getValueType() == VT) {
22927 // Mask and original value have the same type
22928 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22929 for (unsigned i = 0; i != NumElems; ++i)
22930 ShuffleVec[i] = i * SizeRatio;
22931 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22932 ShuffleVec[i] = NumElems*SizeRatio;
22933 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22934 DAG.getConstant(0, dl, WideVecVT),
22938 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22939 unsigned WidenNumElts = NumElems*SizeRatio;
22940 unsigned MaskNumElts = VT.getVectorNumElements();
22941 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22944 unsigned NumConcat = WidenNumElts / MaskNumElts;
22945 SmallVector<SDValue, 16> Ops(NumConcat);
22946 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
22948 for (unsigned i = 1; i != NumConcat; ++i)
22951 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22954 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
22955 NewMask, StVT, Mst->getMemOperand(), false);
22957 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
22958 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
22959 const X86Subtarget *Subtarget) {
22960 StoreSDNode *St = cast<StoreSDNode>(N);
22961 EVT VT = St->getValue().getValueType();
22962 EVT StVT = St->getMemoryVT();
22964 SDValue StoredVal = St->getOperand(1);
22965 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22967 // If we are saving a concatenation of two XMM registers and 32-byte stores
22968 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
22969 unsigned Alignment = St->getAlignment();
22970 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
22971 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22972 StVT == VT && !IsAligned) {
22973 unsigned NumElems = VT.getVectorNumElements();
22977 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
22978 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
22980 SDValue Stride = DAG.getConstant(16, dl, TLI.getPointerTy());
22981 SDValue Ptr0 = St->getBasePtr();
22982 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
22984 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
22985 St->getPointerInfo(), St->isVolatile(),
22986 St->isNonTemporal(), Alignment);
22987 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
22988 St->getPointerInfo(), St->isVolatile(),
22989 St->isNonTemporal(),
22990 std::min(16U, Alignment));
22991 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
22994 // Optimize trunc store (of multiple scalars) to shuffle and store.
22995 // First, pack all of the elements in one place. Next, store to memory
22996 // in fewer chunks.
22997 if (St->isTruncatingStore() && VT.isVector()) {
22998 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22999 unsigned NumElems = VT.getVectorNumElements();
23000 assert(StVT != VT && "Cannot truncate to the same type");
23001 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23002 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23004 // From, To sizes and ElemCount must be pow of two
23005 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
23006 // We are going to use the original vector elt for storing.
23007 // Accumulated smaller vector elements must be a multiple of the store size.
23008 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
23010 unsigned SizeRatio = FromSz / ToSz;
23012 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23014 // Create a type on which we perform the shuffle
23015 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23016 StVT.getScalarType(), NumElems*SizeRatio);
23018 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23020 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
23021 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
23022 for (unsigned i = 0; i != NumElems; ++i)
23023 ShuffleVec[i] = i * SizeRatio;
23025 // Can't shuffle using an illegal type.
23026 if (!TLI.isTypeLegal(WideVecVT))
23029 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23030 DAG.getUNDEF(WideVecVT),
23032 // At this point all of the data is stored at the bottom of the
23033 // register. We now need to save it to mem.
23035 // Find the largest store unit
23036 MVT StoreType = MVT::i8;
23037 for (MVT Tp : MVT::integer_valuetypes()) {
23038 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
23042 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
23043 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
23044 (64 <= NumElems * ToSz))
23045 StoreType = MVT::f64;
23047 // Bitcast the original vector into a vector of store-size units
23048 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
23049 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
23050 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
23051 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
23052 SmallVector<SDValue, 8> Chains;
23053 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, dl,
23054 TLI.getPointerTy());
23055 SDValue Ptr = St->getBasePtr();
23057 // Perform one or more big stores into memory.
23058 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
23059 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
23060 StoreType, ShuffWide,
23061 DAG.getIntPtrConstant(i, dl));
23062 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
23063 St->getPointerInfo(), St->isVolatile(),
23064 St->isNonTemporal(), St->getAlignment());
23065 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23066 Chains.push_back(Ch);
23069 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
23072 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
23073 // the FP state in cases where an emms may be missing.
23074 // A preferable solution to the general problem is to figure out the right
23075 // places to insert EMMS. This qualifies as a quick hack.
23077 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
23078 if (VT.getSizeInBits() != 64)
23081 const Function *F = DAG.getMachineFunction().getFunction();
23082 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
23083 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
23084 && Subtarget->hasSSE2();
23085 if ((VT.isVector() ||
23086 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23087 isa<LoadSDNode>(St->getValue()) &&
23088 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23089 St->getChain().hasOneUse() && !St->isVolatile()) {
23090 SDNode* LdVal = St->getValue().getNode();
23091 LoadSDNode *Ld = nullptr;
23092 int TokenFactorIndex = -1;
23093 SmallVector<SDValue, 8> Ops;
23094 SDNode* ChainVal = St->getChain().getNode();
23095 // Must be a store of a load. We currently handle two cases: the load
23096 // is a direct child, and it's under an intervening TokenFactor. It is
23097 // possible to dig deeper under nested TokenFactors.
23098 if (ChainVal == LdVal)
23099 Ld = cast<LoadSDNode>(St->getChain());
23100 else if (St->getValue().hasOneUse() &&
23101 ChainVal->getOpcode() == ISD::TokenFactor) {
23102 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23103 if (ChainVal->getOperand(i).getNode() == LdVal) {
23104 TokenFactorIndex = i;
23105 Ld = cast<LoadSDNode>(St->getValue());
23107 Ops.push_back(ChainVal->getOperand(i));
23111 if (!Ld || !ISD::isNormalLoad(Ld))
23114 // If this is not the MMX case, i.e. we are just turning i64 load/store
23115 // into f64 load/store, avoid the transformation if there are multiple
23116 // uses of the loaded value.
23117 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23122 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23123 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23125 if (Subtarget->is64Bit() || F64IsLegal) {
23126 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23127 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23128 Ld->getPointerInfo(), Ld->isVolatile(),
23129 Ld->isNonTemporal(), Ld->isInvariant(),
23130 Ld->getAlignment());
23131 SDValue NewChain = NewLd.getValue(1);
23132 if (TokenFactorIndex != -1) {
23133 Ops.push_back(NewChain);
23134 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23136 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23137 St->getPointerInfo(),
23138 St->isVolatile(), St->isNonTemporal(),
23139 St->getAlignment());
23142 // Otherwise, lower to two pairs of 32-bit loads / stores.
23143 SDValue LoAddr = Ld->getBasePtr();
23144 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23145 DAG.getConstant(4, LdDL, MVT::i32));
23147 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23148 Ld->getPointerInfo(),
23149 Ld->isVolatile(), Ld->isNonTemporal(),
23150 Ld->isInvariant(), Ld->getAlignment());
23151 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23152 Ld->getPointerInfo().getWithOffset(4),
23153 Ld->isVolatile(), Ld->isNonTemporal(),
23155 MinAlign(Ld->getAlignment(), 4));
23157 SDValue NewChain = LoLd.getValue(1);
23158 if (TokenFactorIndex != -1) {
23159 Ops.push_back(LoLd);
23160 Ops.push_back(HiLd);
23161 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23164 LoAddr = St->getBasePtr();
23165 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23166 DAG.getConstant(4, StDL, MVT::i32));
23168 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23169 St->getPointerInfo(),
23170 St->isVolatile(), St->isNonTemporal(),
23171 St->getAlignment());
23172 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23173 St->getPointerInfo().getWithOffset(4),
23175 St->isNonTemporal(),
23176 MinAlign(St->getAlignment(), 4));
23177 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23180 // This is similar to the above case, but here we handle a scalar 64-bit
23181 // integer store that is extracted from a vector on a 32-bit target.
23182 // If we have SSE2, then we can treat it like a floating-point double
23183 // to get past legalization. The execution dependencies fixup pass will
23184 // choose the optimal machine instruction for the store if this really is
23185 // an integer or v2f32 rather than an f64.
23186 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
23187 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
23188 SDValue OldExtract = St->getOperand(1);
23189 SDValue ExtOp0 = OldExtract.getOperand(0);
23190 unsigned VecSize = ExtOp0.getValueSizeInBits();
23191 MVT VecVT = MVT::getVectorVT(MVT::f64, VecSize / 64);
23192 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtOp0);
23193 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
23194 BitCast, OldExtract.getOperand(1));
23195 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
23196 St->getPointerInfo(), St->isVolatile(),
23197 St->isNonTemporal(), St->getAlignment());
23203 /// Return 'true' if this vector operation is "horizontal"
23204 /// and return the operands for the horizontal operation in LHS and RHS. A
23205 /// horizontal operation performs the binary operation on successive elements
23206 /// of its first operand, then on successive elements of its second operand,
23207 /// returning the resulting values in a vector. For example, if
23208 /// A = < float a0, float a1, float a2, float a3 >
23210 /// B = < float b0, float b1, float b2, float b3 >
23211 /// then the result of doing a horizontal operation on A and B is
23212 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23213 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23214 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23215 /// set to A, RHS to B, and the routine returns 'true'.
23216 /// Note that the binary operation should have the property that if one of the
23217 /// operands is UNDEF then the result is UNDEF.
23218 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23219 // Look for the following pattern: if
23220 // A = < float a0, float a1, float a2, float a3 >
23221 // B = < float b0, float b1, float b2, float b3 >
23223 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23224 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23225 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23226 // which is A horizontal-op B.
23228 // At least one of the operands should be a vector shuffle.
23229 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23230 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23233 MVT VT = LHS.getSimpleValueType();
23235 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23236 "Unsupported vector type for horizontal add/sub");
23238 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23239 // operate independently on 128-bit lanes.
23240 unsigned NumElts = VT.getVectorNumElements();
23241 unsigned NumLanes = VT.getSizeInBits()/128;
23242 unsigned NumLaneElts = NumElts / NumLanes;
23243 assert((NumLaneElts % 2 == 0) &&
23244 "Vector type should have an even number of elements in each lane");
23245 unsigned HalfLaneElts = NumLaneElts/2;
23247 // View LHS in the form
23248 // LHS = VECTOR_SHUFFLE A, B, LMask
23249 // If LHS is not a shuffle then pretend it is the shuffle
23250 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23251 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23254 SmallVector<int, 16> LMask(NumElts);
23255 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23256 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23257 A = LHS.getOperand(0);
23258 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23259 B = LHS.getOperand(1);
23260 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23261 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23263 if (LHS.getOpcode() != ISD::UNDEF)
23265 for (unsigned i = 0; i != NumElts; ++i)
23269 // Likewise, view RHS in the form
23270 // RHS = VECTOR_SHUFFLE C, D, RMask
23272 SmallVector<int, 16> RMask(NumElts);
23273 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23274 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23275 C = RHS.getOperand(0);
23276 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23277 D = RHS.getOperand(1);
23278 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23279 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23281 if (RHS.getOpcode() != ISD::UNDEF)
23283 for (unsigned i = 0; i != NumElts; ++i)
23287 // Check that the shuffles are both shuffling the same vectors.
23288 if (!(A == C && B == D) && !(A == D && B == C))
23291 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23292 if (!A.getNode() && !B.getNode())
23295 // If A and B occur in reverse order in RHS, then "swap" them (which means
23296 // rewriting the mask).
23298 ShuffleVectorSDNode::commuteMask(RMask);
23300 // At this point LHS and RHS are equivalent to
23301 // LHS = VECTOR_SHUFFLE A, B, LMask
23302 // RHS = VECTOR_SHUFFLE A, B, RMask
23303 // Check that the masks correspond to performing a horizontal operation.
23304 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23305 for (unsigned i = 0; i != NumLaneElts; ++i) {
23306 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23308 // Ignore any UNDEF components.
23309 if (LIdx < 0 || RIdx < 0 ||
23310 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23311 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23314 // Check that successive elements are being operated on. If not, this is
23315 // not a horizontal operation.
23316 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23317 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23318 if (!(LIdx == Index && RIdx == Index + 1) &&
23319 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23324 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23325 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23329 /// Do target-specific dag combines on floating point adds.
23330 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23331 const X86Subtarget *Subtarget) {
23332 EVT VT = N->getValueType(0);
23333 SDValue LHS = N->getOperand(0);
23334 SDValue RHS = N->getOperand(1);
23336 // Try to synthesize horizontal adds from adds of shuffles.
23337 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23338 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23339 isHorizontalBinOp(LHS, RHS, true))
23340 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23344 /// Do target-specific dag combines on floating point subs.
23345 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23346 const X86Subtarget *Subtarget) {
23347 EVT VT = N->getValueType(0);
23348 SDValue LHS = N->getOperand(0);
23349 SDValue RHS = N->getOperand(1);
23351 // Try to synthesize horizontal subs from subs of shuffles.
23352 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23353 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23354 isHorizontalBinOp(LHS, RHS, false))
23355 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23359 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
23360 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23361 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23363 // F[X]OR(0.0, x) -> x
23364 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23365 if (C->getValueAPF().isPosZero())
23366 return N->getOperand(1);
23368 // F[X]OR(x, 0.0) -> x
23369 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23370 if (C->getValueAPF().isPosZero())
23371 return N->getOperand(0);
23375 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
23376 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23377 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23379 // Only perform optimizations if UnsafeMath is used.
23380 if (!DAG.getTarget().Options.UnsafeFPMath)
23383 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23384 // into FMINC and FMAXC, which are Commutative operations.
23385 unsigned NewOp = 0;
23386 switch (N->getOpcode()) {
23387 default: llvm_unreachable("unknown opcode");
23388 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23389 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23392 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23393 N->getOperand(0), N->getOperand(1));
23396 /// Do target-specific dag combines on X86ISD::FAND nodes.
23397 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23398 // FAND(0.0, x) -> 0.0
23399 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23400 if (C->getValueAPF().isPosZero())
23401 return N->getOperand(0);
23403 // FAND(x, 0.0) -> 0.0
23404 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23405 if (C->getValueAPF().isPosZero())
23406 return N->getOperand(1);
23411 /// Do target-specific dag combines on X86ISD::FANDN nodes
23412 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23413 // FANDN(0.0, x) -> x
23414 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23415 if (C->getValueAPF().isPosZero())
23416 return N->getOperand(1);
23418 // FANDN(x, 0.0) -> 0.0
23419 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23420 if (C->getValueAPF().isPosZero())
23421 return N->getOperand(1);
23426 static SDValue PerformBTCombine(SDNode *N,
23428 TargetLowering::DAGCombinerInfo &DCI) {
23429 // BT ignores high bits in the bit index operand.
23430 SDValue Op1 = N->getOperand(1);
23431 if (Op1.hasOneUse()) {
23432 unsigned BitWidth = Op1.getValueSizeInBits();
23433 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23434 APInt KnownZero, KnownOne;
23435 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23436 !DCI.isBeforeLegalizeOps());
23437 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23438 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23439 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23440 DCI.CommitTargetLoweringOpt(TLO);
23445 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23446 SDValue Op = N->getOperand(0);
23447 if (Op.getOpcode() == ISD::BITCAST)
23448 Op = Op.getOperand(0);
23449 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23450 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23451 VT.getVectorElementType().getSizeInBits() ==
23452 OpVT.getVectorElementType().getSizeInBits()) {
23453 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23458 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23459 const X86Subtarget *Subtarget) {
23460 EVT VT = N->getValueType(0);
23461 if (!VT.isVector())
23464 SDValue N0 = N->getOperand(0);
23465 SDValue N1 = N->getOperand(1);
23466 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23469 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23470 // both SSE and AVX2 since there is no sign-extended shift right
23471 // operation on a vector with 64-bit elements.
23472 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23473 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23474 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23475 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23476 SDValue N00 = N0.getOperand(0);
23478 // EXTLOAD has a better solution on AVX2,
23479 // it may be replaced with X86ISD::VSEXT node.
23480 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23481 if (!ISD::isNormalLoad(N00.getNode()))
23484 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23485 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23487 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23493 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23494 TargetLowering::DAGCombinerInfo &DCI,
23495 const X86Subtarget *Subtarget) {
23496 SDValue N0 = N->getOperand(0);
23497 EVT VT = N->getValueType(0);
23499 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
23500 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
23501 // This exposes the sext to the sdivrem lowering, so that it directly extends
23502 // from AH (which we otherwise need to do contortions to access).
23503 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
23504 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
23506 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23507 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
23508 N0.getOperand(0), N0.getOperand(1));
23509 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23510 return R.getValue(1);
23513 if (!DCI.isBeforeLegalizeOps())
23516 if (!Subtarget->hasFp256())
23519 if (VT.isVector() && VT.getSizeInBits() == 256) {
23520 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23528 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
23529 const X86Subtarget* Subtarget) {
23531 EVT VT = N->getValueType(0);
23533 // Let legalize expand this if it isn't a legal type yet.
23534 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
23537 EVT ScalarVT = VT.getScalarType();
23538 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
23539 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
23542 SDValue A = N->getOperand(0);
23543 SDValue B = N->getOperand(1);
23544 SDValue C = N->getOperand(2);
23546 bool NegA = (A.getOpcode() == ISD::FNEG);
23547 bool NegB = (B.getOpcode() == ISD::FNEG);
23548 bool NegC = (C.getOpcode() == ISD::FNEG);
23550 // Negative multiplication when NegA xor NegB
23551 bool NegMul = (NegA != NegB);
23553 A = A.getOperand(0);
23555 B = B.getOperand(0);
23557 C = C.getOperand(0);
23561 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
23563 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
23565 return DAG.getNode(Opcode, dl, VT, A, B, C);
23568 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
23569 TargetLowering::DAGCombinerInfo &DCI,
23570 const X86Subtarget *Subtarget) {
23571 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
23572 // (and (i32 x86isd::setcc_carry), 1)
23573 // This eliminates the zext. This transformation is necessary because
23574 // ISD::SETCC is always legalized to i8.
23576 SDValue N0 = N->getOperand(0);
23577 EVT VT = N->getValueType(0);
23579 if (N0.getOpcode() == ISD::AND &&
23581 N0.getOperand(0).hasOneUse()) {
23582 SDValue N00 = N0.getOperand(0);
23583 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23584 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23585 if (!C || C->getZExtValue() != 1)
23587 return DAG.getNode(ISD::AND, dl, VT,
23588 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23589 N00.getOperand(0), N00.getOperand(1)),
23590 DAG.getConstant(1, dl, VT));
23594 if (N0.getOpcode() == ISD::TRUNCATE &&
23596 N0.getOperand(0).hasOneUse()) {
23597 SDValue N00 = N0.getOperand(0);
23598 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23599 return DAG.getNode(ISD::AND, dl, VT,
23600 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23601 N00.getOperand(0), N00.getOperand(1)),
23602 DAG.getConstant(1, dl, VT));
23605 if (VT.is256BitVector()) {
23606 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23611 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
23612 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
23613 // This exposes the zext to the udivrem lowering, so that it directly extends
23614 // from AH (which we otherwise need to do contortions to access).
23615 if (N0.getOpcode() == ISD::UDIVREM &&
23616 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
23617 (VT == MVT::i32 || VT == MVT::i64)) {
23618 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23619 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
23620 N0.getOperand(0), N0.getOperand(1));
23621 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23622 return R.getValue(1);
23628 // Optimize x == -y --> x+y == 0
23629 // x != -y --> x+y != 0
23630 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
23631 const X86Subtarget* Subtarget) {
23632 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
23633 SDValue LHS = N->getOperand(0);
23634 SDValue RHS = N->getOperand(1);
23635 EVT VT = N->getValueType(0);
23638 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
23639 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
23640 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
23641 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
23642 LHS.getOperand(1));
23643 return DAG.getSetCC(DL, N->getValueType(0), addV,
23644 DAG.getConstant(0, DL, addV.getValueType()), CC);
23646 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
23647 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
23648 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
23649 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
23650 RHS.getOperand(1));
23651 return DAG.getSetCC(DL, N->getValueType(0), addV,
23652 DAG.getConstant(0, DL, addV.getValueType()), CC);
23655 if (VT.getScalarType() == MVT::i1 &&
23656 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
23658 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23659 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23660 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23662 if (!IsSEXT0 || !IsVZero1) {
23663 // Swap the operands and update the condition code.
23664 std::swap(LHS, RHS);
23665 CC = ISD::getSetCCSwappedOperands(CC);
23667 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23668 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23669 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23672 if (IsSEXT0 && IsVZero1) {
23673 assert(VT == LHS.getOperand(0).getValueType() &&
23674 "Uexpected operand type");
23675 if (CC == ISD::SETGT)
23676 return DAG.getConstant(0, DL, VT);
23677 if (CC == ISD::SETLE)
23678 return DAG.getConstant(1, DL, VT);
23679 if (CC == ISD::SETEQ || CC == ISD::SETGE)
23680 return DAG.getNOT(DL, LHS.getOperand(0), VT);
23682 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
23683 "Unexpected condition code!");
23684 return LHS.getOperand(0);
23691 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
23692 SelectionDAG &DAG) {
23694 MVT VT = Load->getSimpleValueType(0);
23695 MVT EVT = VT.getVectorElementType();
23696 SDValue Addr = Load->getOperand(1);
23697 SDValue NewAddr = DAG.getNode(
23698 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
23699 DAG.getConstant(Index * EVT.getStoreSize(), dl,
23700 Addr.getSimpleValueType()));
23703 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
23704 DAG.getMachineFunction().getMachineMemOperand(
23705 Load->getMemOperand(), 0, EVT.getStoreSize()));
23709 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
23710 const X86Subtarget *Subtarget) {
23712 MVT VT = N->getOperand(1)->getSimpleValueType(0);
23713 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
23714 "X86insertps is only defined for v4x32");
23716 SDValue Ld = N->getOperand(1);
23717 if (MayFoldLoad(Ld)) {
23718 // Extract the countS bits from the immediate so we can get the proper
23719 // address when narrowing the vector load to a specific element.
23720 // When the second source op is a memory address, insertps doesn't use
23721 // countS and just gets an f32 from that address.
23722 unsigned DestIndex =
23723 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
23725 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
23727 // Create this as a scalar to vector to match the instruction pattern.
23728 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
23729 // countS bits are ignored when loading from memory on insertps, which
23730 // means we don't need to explicitly set them to 0.
23731 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
23732 LoadScalarToVector, N->getOperand(2));
23737 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
23738 SDValue V0 = N->getOperand(0);
23739 SDValue V1 = N->getOperand(1);
23741 EVT VT = N->getValueType(0);
23743 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23744 // operands and changing the mask to 1. This saves us a bunch of
23745 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23746 // x86InstrInfo knows how to commute this back after instruction selection
23747 // if it would help register allocation.
23749 // TODO: If optimizing for size or a processor that doesn't suffer from
23750 // partial register update stalls, this should be transformed into a MOVSD
23751 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23753 if (VT == MVT::v2f64)
23754 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23755 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23756 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
23757 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23763 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
23764 // as "sbb reg,reg", since it can be extended without zext and produces
23765 // an all-ones bit which is more useful than 0/1 in some cases.
23766 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
23769 return DAG.getNode(ISD::AND, DL, VT,
23770 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23771 DAG.getConstant(X86::COND_B, DL, MVT::i8),
23773 DAG.getConstant(1, DL, VT));
23774 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
23775 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
23776 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23777 DAG.getConstant(X86::COND_B, DL, MVT::i8),
23781 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
23782 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
23783 TargetLowering::DAGCombinerInfo &DCI,
23784 const X86Subtarget *Subtarget) {
23786 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
23787 SDValue EFLAGS = N->getOperand(1);
23789 if (CC == X86::COND_A) {
23790 // Try to convert COND_A into COND_B in an attempt to facilitate
23791 // materializing "setb reg".
23793 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
23794 // cannot take an immediate as its first operand.
23796 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
23797 EFLAGS.getValueType().isInteger() &&
23798 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
23799 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23800 EFLAGS.getNode()->getVTList(),
23801 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23802 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23803 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
23807 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
23808 // a zext and produces an all-ones bit which is more useful than 0/1 in some
23810 if (CC == X86::COND_B)
23811 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23815 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23816 if (Flags.getNode()) {
23817 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
23818 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23824 // Optimize branch condition evaluation.
23826 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23827 TargetLowering::DAGCombinerInfo &DCI,
23828 const X86Subtarget *Subtarget) {
23830 SDValue Chain = N->getOperand(0);
23831 SDValue Dest = N->getOperand(1);
23832 SDValue EFLAGS = N->getOperand(3);
23833 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23837 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23838 if (Flags.getNode()) {
23839 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
23840 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23847 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23848 SelectionDAG &DAG) {
23849 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23850 // optimize away operation when it's from a constant.
23852 // The general transformation is:
23853 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23854 // AND(VECTOR_CMP(x,y), constant2)
23855 // constant2 = UNARYOP(constant)
23857 // Early exit if this isn't a vector operation, the operand of the
23858 // unary operation isn't a bitwise AND, or if the sizes of the operations
23859 // aren't the same.
23860 EVT VT = N->getValueType(0);
23861 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23862 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23863 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23866 // Now check that the other operand of the AND is a constant. We could
23867 // make the transformation for non-constant splats as well, but it's unclear
23868 // that would be a benefit as it would not eliminate any operations, just
23869 // perform one more step in scalar code before moving to the vector unit.
23870 if (BuildVectorSDNode *BV =
23871 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23872 // Bail out if the vector isn't a constant.
23873 if (!BV->isConstant())
23876 // Everything checks out. Build up the new and improved node.
23878 EVT IntVT = BV->getValueType(0);
23879 // Create a new constant of the appropriate type for the transformed
23881 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23882 // The AND node needs bitcasts to/from an integer vector type around it.
23883 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
23884 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
23885 N->getOperand(0)->getOperand(0), MaskConst);
23886 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
23893 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
23894 const X86Subtarget *Subtarget) {
23895 // First try to optimize away the conversion entirely when it's
23896 // conditionally from a constant. Vectors only.
23897 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
23898 if (Res != SDValue())
23901 // Now move on to more general possibilities.
23902 SDValue Op0 = N->getOperand(0);
23903 EVT InVT = Op0->getValueType(0);
23905 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
23906 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
23908 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
23909 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
23910 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
23913 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
23914 // a 32-bit target where SSE doesn't support i64->FP operations.
23915 if (Op0.getOpcode() == ISD::LOAD) {
23916 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
23917 EVT VT = Ld->getValueType(0);
23918 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
23919 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
23920 !Subtarget->is64Bit() && VT == MVT::i64) {
23921 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
23922 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
23923 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
23930 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
23931 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
23932 X86TargetLowering::DAGCombinerInfo &DCI) {
23933 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
23934 // the result is either zero or one (depending on the input carry bit).
23935 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
23936 if (X86::isZeroNode(N->getOperand(0)) &&
23937 X86::isZeroNode(N->getOperand(1)) &&
23938 // We don't have a good way to replace an EFLAGS use, so only do this when
23940 SDValue(N, 1).use_empty()) {
23942 EVT VT = N->getValueType(0);
23943 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
23944 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
23945 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
23946 DAG.getConstant(X86::COND_B, DL,
23949 DAG.getConstant(1, DL, VT));
23950 return DCI.CombineTo(N, Res1, CarryOut);
23956 // fold (add Y, (sete X, 0)) -> adc 0, Y
23957 // (add Y, (setne X, 0)) -> sbb -1, Y
23958 // (sub (sete X, 0), Y) -> sbb 0, Y
23959 // (sub (setne X, 0), Y) -> adc -1, Y
23960 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
23963 // Look through ZExts.
23964 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
23965 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
23968 SDValue SetCC = Ext.getOperand(0);
23969 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
23972 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
23973 if (CC != X86::COND_E && CC != X86::COND_NE)
23976 SDValue Cmp = SetCC.getOperand(1);
23977 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
23978 !X86::isZeroNode(Cmp.getOperand(1)) ||
23979 !Cmp.getOperand(0).getValueType().isInteger())
23982 SDValue CmpOp0 = Cmp.getOperand(0);
23983 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
23984 DAG.getConstant(1, DL, CmpOp0.getValueType()));
23986 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
23987 if (CC == X86::COND_NE)
23988 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
23989 DL, OtherVal.getValueType(), OtherVal,
23990 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
23992 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
23993 DL, OtherVal.getValueType(), OtherVal,
23994 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
23997 /// PerformADDCombine - Do target-specific dag combines on integer adds.
23998 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
23999 const X86Subtarget *Subtarget) {
24000 EVT VT = N->getValueType(0);
24001 SDValue Op0 = N->getOperand(0);
24002 SDValue Op1 = N->getOperand(1);
24004 // Try to synthesize horizontal adds from adds of shuffles.
24005 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24006 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24007 isHorizontalBinOp(Op0, Op1, true))
24008 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
24010 return OptimizeConditionalInDecrement(N, DAG);
24013 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
24014 const X86Subtarget *Subtarget) {
24015 SDValue Op0 = N->getOperand(0);
24016 SDValue Op1 = N->getOperand(1);
24018 // X86 can't encode an immediate LHS of a sub. See if we can push the
24019 // negation into a preceding instruction.
24020 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
24021 // If the RHS of the sub is a XOR with one use and a constant, invert the
24022 // immediate. Then add one to the LHS of the sub so we can turn
24023 // X-Y -> X+~Y+1, saving one register.
24024 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
24025 isa<ConstantSDNode>(Op1.getOperand(1))) {
24026 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
24027 EVT VT = Op0.getValueType();
24028 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
24030 DAG.getConstant(~XorC, SDLoc(Op1), VT));
24031 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
24032 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
24036 // Try to synthesize horizontal adds from adds of shuffles.
24037 EVT VT = N->getValueType(0);
24038 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24039 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24040 isHorizontalBinOp(Op0, Op1, true))
24041 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
24043 return OptimizeConditionalInDecrement(N, DAG);
24046 /// performVZEXTCombine - Performs build vector combines
24047 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
24048 TargetLowering::DAGCombinerInfo &DCI,
24049 const X86Subtarget *Subtarget) {
24051 MVT VT = N->getSimpleValueType(0);
24052 SDValue Op = N->getOperand(0);
24053 MVT OpVT = Op.getSimpleValueType();
24054 MVT OpEltVT = OpVT.getVectorElementType();
24055 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
24057 // (vzext (bitcast (vzext (x)) -> (vzext x)
24059 while (V.getOpcode() == ISD::BITCAST)
24060 V = V.getOperand(0);
24062 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
24063 MVT InnerVT = V.getSimpleValueType();
24064 MVT InnerEltVT = InnerVT.getVectorElementType();
24066 // If the element sizes match exactly, we can just do one larger vzext. This
24067 // is always an exact type match as vzext operates on integer types.
24068 if (OpEltVT == InnerEltVT) {
24069 assert(OpVT == InnerVT && "Types must match for vzext!");
24070 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
24073 // The only other way we can combine them is if only a single element of the
24074 // inner vzext is used in the input to the outer vzext.
24075 if (InnerEltVT.getSizeInBits() < InputBits)
24078 // In this case, the inner vzext is completely dead because we're going to
24079 // only look at bits inside of the low element. Just do the outer vzext on
24080 // a bitcast of the input to the inner.
24081 return DAG.getNode(X86ISD::VZEXT, DL, VT,
24082 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
24085 // Check if we can bypass extracting and re-inserting an element of an input
24086 // vector. Essentialy:
24087 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
24088 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
24089 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
24090 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
24091 SDValue ExtractedV = V.getOperand(0);
24092 SDValue OrigV = ExtractedV.getOperand(0);
24093 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
24094 if (ExtractIdx->getZExtValue() == 0) {
24095 MVT OrigVT = OrigV.getSimpleValueType();
24096 // Extract a subvector if necessary...
24097 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
24098 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
24099 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
24100 OrigVT.getVectorNumElements() / Ratio);
24101 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
24102 DAG.getIntPtrConstant(0, DL));
24104 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
24105 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
24112 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24113 DAGCombinerInfo &DCI) const {
24114 SelectionDAG &DAG = DCI.DAG;
24115 switch (N->getOpcode()) {
24117 case ISD::EXTRACT_VECTOR_ELT:
24118 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24121 case X86ISD::SHRUNKBLEND:
24122 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24123 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
24124 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24125 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24126 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24127 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24128 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24131 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24132 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24133 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24134 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24135 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24136 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
24137 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24138 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
24139 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
24140 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24141 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24143 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24145 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24146 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24147 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24148 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24149 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24150 case ISD::ANY_EXTEND:
24151 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24152 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24153 case ISD::SIGN_EXTEND_INREG:
24154 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24155 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
24156 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24157 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24158 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24159 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24160 case X86ISD::SHUFP: // Handle all target specific shuffles
24161 case X86ISD::PALIGNR:
24162 case X86ISD::UNPCKH:
24163 case X86ISD::UNPCKL:
24164 case X86ISD::MOVHLPS:
24165 case X86ISD::MOVLHPS:
24166 case X86ISD::PSHUFB:
24167 case X86ISD::PSHUFD:
24168 case X86ISD::PSHUFHW:
24169 case X86ISD::PSHUFLW:
24170 case X86ISD::MOVSS:
24171 case X86ISD::MOVSD:
24172 case X86ISD::VPERMILPI:
24173 case X86ISD::VPERM2X128:
24174 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24175 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24176 case ISD::INTRINSIC_WO_CHAIN:
24177 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24178 case X86ISD::INSERTPS: {
24179 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
24180 return PerformINSERTPSCombine(N, DAG, Subtarget);
24183 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
24189 /// isTypeDesirableForOp - Return true if the target has native support for
24190 /// the specified value type and it is 'desirable' to use the type for the
24191 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24192 /// instruction encodings are longer and some i16 instructions are slow.
24193 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24194 if (!isTypeLegal(VT))
24196 if (VT != MVT::i16)
24203 case ISD::SIGN_EXTEND:
24204 case ISD::ZERO_EXTEND:
24205 case ISD::ANY_EXTEND:
24218 /// IsDesirableToPromoteOp - This method query the target whether it is
24219 /// beneficial for dag combiner to promote the specified node. If true, it
24220 /// should return the desired promotion type by reference.
24221 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24222 EVT VT = Op.getValueType();
24223 if (VT != MVT::i16)
24226 bool Promote = false;
24227 bool Commute = false;
24228 switch (Op.getOpcode()) {
24231 LoadSDNode *LD = cast<LoadSDNode>(Op);
24232 // If the non-extending load has a single use and it's not live out, then it
24233 // might be folded.
24234 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24235 Op.hasOneUse()*/) {
24236 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24237 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24238 // The only case where we'd want to promote LOAD (rather then it being
24239 // promoted as an operand is when it's only use is liveout.
24240 if (UI->getOpcode() != ISD::CopyToReg)
24247 case ISD::SIGN_EXTEND:
24248 case ISD::ZERO_EXTEND:
24249 case ISD::ANY_EXTEND:
24254 SDValue N0 = Op.getOperand(0);
24255 // Look out for (store (shl (load), x)).
24256 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24269 SDValue N0 = Op.getOperand(0);
24270 SDValue N1 = Op.getOperand(1);
24271 if (!Commute && MayFoldLoad(N1))
24273 // Avoid disabling potential load folding opportunities.
24274 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24276 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24286 //===----------------------------------------------------------------------===//
24287 // X86 Inline Assembly Support
24288 //===----------------------------------------------------------------------===//
24290 // Helper to match a string separated by whitespace.
24291 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
24292 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
24294 for (StringRef Piece : Pieces) {
24295 if (!S.startswith(Piece)) // Check if the piece matches.
24298 S = S.substr(Piece.size());
24299 StringRef::size_type Pos = S.find_first_not_of(" \t");
24300 if (Pos == 0) // We matched a prefix.
24309 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24311 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24312 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24313 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24314 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24316 if (AsmPieces.size() == 3)
24318 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24325 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24326 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24328 std::string AsmStr = IA->getAsmString();
24330 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24331 if (!Ty || Ty->getBitWidth() % 16 != 0)
24334 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24335 SmallVector<StringRef, 4> AsmPieces;
24336 SplitString(AsmStr, AsmPieces, ";\n");
24338 switch (AsmPieces.size()) {
24339 default: return false;
24341 // FIXME: this should verify that we are targeting a 486 or better. If not,
24342 // we will turn this bswap into something that will be lowered to logical
24343 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24344 // lower so don't worry about this.
24346 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
24347 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
24348 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
24349 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
24350 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
24351 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
24352 // No need to check constraints, nothing other than the equivalent of
24353 // "=r,0" would be valid here.
24354 return IntrinsicLowering::LowerToByteSwap(CI);
24357 // rorw $$8, ${0:w} --> llvm.bswap.i16
24358 if (CI->getType()->isIntegerTy(16) &&
24359 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24360 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
24361 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
24363 const std::string &ConstraintsStr = IA->getConstraintString();
24364 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24365 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24366 if (clobbersFlagRegisters(AsmPieces))
24367 return IntrinsicLowering::LowerToByteSwap(CI);
24371 if (CI->getType()->isIntegerTy(32) &&
24372 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24373 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
24374 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
24375 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
24377 const std::string &ConstraintsStr = IA->getConstraintString();
24378 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24379 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24380 if (clobbersFlagRegisters(AsmPieces))
24381 return IntrinsicLowering::LowerToByteSwap(CI);
24384 if (CI->getType()->isIntegerTy(64)) {
24385 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24386 if (Constraints.size() >= 2 &&
24387 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24388 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24389 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24390 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
24391 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
24392 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
24393 return IntrinsicLowering::LowerToByteSwap(CI);
24401 /// getConstraintType - Given a constraint letter, return the type of
24402 /// constraint it is for this target.
24403 X86TargetLowering::ConstraintType
24404 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24405 if (Constraint.size() == 1) {
24406 switch (Constraint[0]) {
24417 return C_RegisterClass;
24441 return TargetLowering::getConstraintType(Constraint);
24444 /// Examine constraint type and operand type and determine a weight value.
24445 /// This object must already have been set up with the operand type
24446 /// and the current alternative constraint selected.
24447 TargetLowering::ConstraintWeight
24448 X86TargetLowering::getSingleConstraintMatchWeight(
24449 AsmOperandInfo &info, const char *constraint) const {
24450 ConstraintWeight weight = CW_Invalid;
24451 Value *CallOperandVal = info.CallOperandVal;
24452 // If we don't have a value, we can't do a match,
24453 // but allow it at the lowest weight.
24454 if (!CallOperandVal)
24456 Type *type = CallOperandVal->getType();
24457 // Look at the constraint type.
24458 switch (*constraint) {
24460 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
24471 if (CallOperandVal->getType()->isIntegerTy())
24472 weight = CW_SpecificReg;
24477 if (type->isFloatingPointTy())
24478 weight = CW_SpecificReg;
24481 if (type->isX86_MMXTy() && Subtarget->hasMMX())
24482 weight = CW_SpecificReg;
24486 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
24487 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
24488 weight = CW_Register;
24491 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
24492 if (C->getZExtValue() <= 31)
24493 weight = CW_Constant;
24497 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24498 if (C->getZExtValue() <= 63)
24499 weight = CW_Constant;
24503 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24504 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
24505 weight = CW_Constant;
24509 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24510 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
24511 weight = CW_Constant;
24515 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24516 if (C->getZExtValue() <= 3)
24517 weight = CW_Constant;
24521 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24522 if (C->getZExtValue() <= 0xff)
24523 weight = CW_Constant;
24528 if (isa<ConstantFP>(CallOperandVal)) {
24529 weight = CW_Constant;
24533 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24534 if ((C->getSExtValue() >= -0x80000000LL) &&
24535 (C->getSExtValue() <= 0x7fffffffLL))
24536 weight = CW_Constant;
24540 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24541 if (C->getZExtValue() <= 0xffffffff)
24542 weight = CW_Constant;
24549 /// LowerXConstraint - try to replace an X constraint, which matches anything,
24550 /// with another that has more specific requirements based on the type of the
24551 /// corresponding operand.
24552 const char *X86TargetLowering::
24553 LowerXConstraint(EVT ConstraintVT) const {
24554 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
24555 // 'f' like normal targets.
24556 if (ConstraintVT.isFloatingPoint()) {
24557 if (Subtarget->hasSSE2())
24559 if (Subtarget->hasSSE1())
24563 return TargetLowering::LowerXConstraint(ConstraintVT);
24566 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
24567 /// vector. If it is invalid, don't add anything to Ops.
24568 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
24569 std::string &Constraint,
24570 std::vector<SDValue>&Ops,
24571 SelectionDAG &DAG) const {
24574 // Only support length 1 constraints for now.
24575 if (Constraint.length() > 1) return;
24577 char ConstraintLetter = Constraint[0];
24578 switch (ConstraintLetter) {
24581 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24582 if (C->getZExtValue() <= 31) {
24583 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24584 Op.getValueType());
24590 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24591 if (C->getZExtValue() <= 63) {
24592 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24593 Op.getValueType());
24599 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24600 if (isInt<8>(C->getSExtValue())) {
24601 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24602 Op.getValueType());
24608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24609 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
24610 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
24611 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
24612 Op.getValueType());
24618 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24619 if (C->getZExtValue() <= 3) {
24620 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24621 Op.getValueType());
24627 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24628 if (C->getZExtValue() <= 255) {
24629 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24630 Op.getValueType());
24636 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24637 if (C->getZExtValue() <= 127) {
24638 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24639 Op.getValueType());
24645 // 32-bit signed value
24646 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24647 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24648 C->getSExtValue())) {
24649 // Widen to 64 bits here to get it sign extended.
24650 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
24653 // FIXME gcc accepts some relocatable values here too, but only in certain
24654 // memory models; it's complicated.
24659 // 32-bit unsigned value
24660 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24661 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24662 C->getZExtValue())) {
24663 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24664 Op.getValueType());
24668 // FIXME gcc accepts some relocatable values here too, but only in certain
24669 // memory models; it's complicated.
24673 // Literal immediates are always ok.
24674 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
24675 // Widen to 64 bits here to get it sign extended.
24676 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
24680 // In any sort of PIC mode addresses need to be computed at runtime by
24681 // adding in a register or some sort of table lookup. These can't
24682 // be used as immediates.
24683 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
24686 // If we are in non-pic codegen mode, we allow the address of a global (with
24687 // an optional displacement) to be used with 'i'.
24688 GlobalAddressSDNode *GA = nullptr;
24689 int64_t Offset = 0;
24691 // Match either (GA), (GA+C), (GA+C1+C2), etc.
24693 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
24694 Offset += GA->getOffset();
24696 } else if (Op.getOpcode() == ISD::ADD) {
24697 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24698 Offset += C->getZExtValue();
24699 Op = Op.getOperand(0);
24702 } else if (Op.getOpcode() == ISD::SUB) {
24703 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24704 Offset += -C->getZExtValue();
24705 Op = Op.getOperand(0);
24710 // Otherwise, this isn't something we can handle, reject it.
24714 const GlobalValue *GV = GA->getGlobal();
24715 // If we require an extra load to get this address, as in PIC mode, we
24716 // can't accept it.
24717 if (isGlobalStubReference(
24718 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
24721 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
24722 GA->getValueType(0), Offset);
24727 if (Result.getNode()) {
24728 Ops.push_back(Result);
24731 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
24734 std::pair<unsigned, const TargetRegisterClass *>
24735 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
24736 const std::string &Constraint,
24738 // First, see if this is a constraint that directly corresponds to an LLVM
24740 if (Constraint.size() == 1) {
24741 // GCC Constraint Letters
24742 switch (Constraint[0]) {
24744 // TODO: Slight differences here in allocation order and leaving
24745 // RIP in the class. Do they matter any more here than they do
24746 // in the normal allocation?
24747 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
24748 if (Subtarget->is64Bit()) {
24749 if (VT == MVT::i32 || VT == MVT::f32)
24750 return std::make_pair(0U, &X86::GR32RegClass);
24751 if (VT == MVT::i16)
24752 return std::make_pair(0U, &X86::GR16RegClass);
24753 if (VT == MVT::i8 || VT == MVT::i1)
24754 return std::make_pair(0U, &X86::GR8RegClass);
24755 if (VT == MVT::i64 || VT == MVT::f64)
24756 return std::make_pair(0U, &X86::GR64RegClass);
24759 // 32-bit fallthrough
24760 case 'Q': // Q_REGS
24761 if (VT == MVT::i32 || VT == MVT::f32)
24762 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
24763 if (VT == MVT::i16)
24764 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
24765 if (VT == MVT::i8 || VT == MVT::i1)
24766 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
24767 if (VT == MVT::i64)
24768 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
24770 case 'r': // GENERAL_REGS
24771 case 'l': // INDEX_REGS
24772 if (VT == MVT::i8 || VT == MVT::i1)
24773 return std::make_pair(0U, &X86::GR8RegClass);
24774 if (VT == MVT::i16)
24775 return std::make_pair(0U, &X86::GR16RegClass);
24776 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
24777 return std::make_pair(0U, &X86::GR32RegClass);
24778 return std::make_pair(0U, &X86::GR64RegClass);
24779 case 'R': // LEGACY_REGS
24780 if (VT == MVT::i8 || VT == MVT::i1)
24781 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
24782 if (VT == MVT::i16)
24783 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
24784 if (VT == MVT::i32 || !Subtarget->is64Bit())
24785 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
24786 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
24787 case 'f': // FP Stack registers.
24788 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
24789 // value to the correct fpstack register class.
24790 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
24791 return std::make_pair(0U, &X86::RFP32RegClass);
24792 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
24793 return std::make_pair(0U, &X86::RFP64RegClass);
24794 return std::make_pair(0U, &X86::RFP80RegClass);
24795 case 'y': // MMX_REGS if MMX allowed.
24796 if (!Subtarget->hasMMX()) break;
24797 return std::make_pair(0U, &X86::VR64RegClass);
24798 case 'Y': // SSE_REGS if SSE2 allowed
24799 if (!Subtarget->hasSSE2()) break;
24801 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
24802 if (!Subtarget->hasSSE1()) break;
24804 switch (VT.SimpleTy) {
24806 // Scalar SSE types.
24809 return std::make_pair(0U, &X86::FR32RegClass);
24812 return std::make_pair(0U, &X86::FR64RegClass);
24820 return std::make_pair(0U, &X86::VR128RegClass);
24828 return std::make_pair(0U, &X86::VR256RegClass);
24833 return std::make_pair(0U, &X86::VR512RegClass);
24839 // Use the default implementation in TargetLowering to convert the register
24840 // constraint into a member of a register class.
24841 std::pair<unsigned, const TargetRegisterClass*> Res;
24842 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
24844 // Not found as a standard register?
24846 // Map st(0) -> st(7) -> ST0
24847 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24848 tolower(Constraint[1]) == 's' &&
24849 tolower(Constraint[2]) == 't' &&
24850 Constraint[3] == '(' &&
24851 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24852 Constraint[5] == ')' &&
24853 Constraint[6] == '}') {
24855 Res.first = X86::FP0+Constraint[4]-'0';
24856 Res.second = &X86::RFP80RegClass;
24860 // GCC allows "st(0)" to be called just plain "st".
24861 if (StringRef("{st}").equals_lower(Constraint)) {
24862 Res.first = X86::FP0;
24863 Res.second = &X86::RFP80RegClass;
24868 if (StringRef("{flags}").equals_lower(Constraint)) {
24869 Res.first = X86::EFLAGS;
24870 Res.second = &X86::CCRRegClass;
24874 // 'A' means EAX + EDX.
24875 if (Constraint == "A") {
24876 Res.first = X86::EAX;
24877 Res.second = &X86::GR32_ADRegClass;
24883 // Otherwise, check to see if this is a register class of the wrong value
24884 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24885 // turn into {ax},{dx}.
24886 if (Res.second->hasType(VT))
24887 return Res; // Correct type already, nothing to do.
24889 // All of the single-register GCC register classes map their values onto
24890 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24891 // really want an 8-bit or 32-bit register, map to the appropriate register
24892 // class and return the appropriate register.
24893 if (Res.second == &X86::GR16RegClass) {
24894 if (VT == MVT::i8 || VT == MVT::i1) {
24895 unsigned DestReg = 0;
24896 switch (Res.first) {
24898 case X86::AX: DestReg = X86::AL; break;
24899 case X86::DX: DestReg = X86::DL; break;
24900 case X86::CX: DestReg = X86::CL; break;
24901 case X86::BX: DestReg = X86::BL; break;
24904 Res.first = DestReg;
24905 Res.second = &X86::GR8RegClass;
24907 } else if (VT == MVT::i32 || VT == MVT::f32) {
24908 unsigned DestReg = 0;
24909 switch (Res.first) {
24911 case X86::AX: DestReg = X86::EAX; break;
24912 case X86::DX: DestReg = X86::EDX; break;
24913 case X86::CX: DestReg = X86::ECX; break;
24914 case X86::BX: DestReg = X86::EBX; break;
24915 case X86::SI: DestReg = X86::ESI; break;
24916 case X86::DI: DestReg = X86::EDI; break;
24917 case X86::BP: DestReg = X86::EBP; break;
24918 case X86::SP: DestReg = X86::ESP; break;
24921 Res.first = DestReg;
24922 Res.second = &X86::GR32RegClass;
24924 } else if (VT == MVT::i64 || VT == MVT::f64) {
24925 unsigned DestReg = 0;
24926 switch (Res.first) {
24928 case X86::AX: DestReg = X86::RAX; break;
24929 case X86::DX: DestReg = X86::RDX; break;
24930 case X86::CX: DestReg = X86::RCX; break;
24931 case X86::BX: DestReg = X86::RBX; break;
24932 case X86::SI: DestReg = X86::RSI; break;
24933 case X86::DI: DestReg = X86::RDI; break;
24934 case X86::BP: DestReg = X86::RBP; break;
24935 case X86::SP: DestReg = X86::RSP; break;
24938 Res.first = DestReg;
24939 Res.second = &X86::GR64RegClass;
24942 } else if (Res.second == &X86::FR32RegClass ||
24943 Res.second == &X86::FR64RegClass ||
24944 Res.second == &X86::VR128RegClass ||
24945 Res.second == &X86::VR256RegClass ||
24946 Res.second == &X86::FR32XRegClass ||
24947 Res.second == &X86::FR64XRegClass ||
24948 Res.second == &X86::VR128XRegClass ||
24949 Res.second == &X86::VR256XRegClass ||
24950 Res.second == &X86::VR512RegClass) {
24951 // Handle references to XMM physical registers that got mapped into the
24952 // wrong class. This can happen with constraints like {xmm0} where the
24953 // target independent register mapper will just pick the first match it can
24954 // find, ignoring the required type.
24956 if (VT == MVT::f32 || VT == MVT::i32)
24957 Res.second = &X86::FR32RegClass;
24958 else if (VT == MVT::f64 || VT == MVT::i64)
24959 Res.second = &X86::FR64RegClass;
24960 else if (X86::VR128RegClass.hasType(VT))
24961 Res.second = &X86::VR128RegClass;
24962 else if (X86::VR256RegClass.hasType(VT))
24963 Res.second = &X86::VR256RegClass;
24964 else if (X86::VR512RegClass.hasType(VT))
24965 Res.second = &X86::VR512RegClass;
24971 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
24973 // Scaling factors are not free at all.
24974 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
24975 // will take 2 allocations in the out of order engine instead of 1
24976 // for plain addressing mode, i.e. inst (reg1).
24978 // vaddps (%rsi,%drx), %ymm0, %ymm1
24979 // Requires two allocations (one for the load, one for the computation)
24981 // vaddps (%rsi), %ymm0, %ymm1
24982 // Requires just 1 allocation, i.e., freeing allocations for other operations
24983 // and having less micro operations to execute.
24985 // For some X86 architectures, this is even worse because for instance for
24986 // stores, the complex addressing mode forces the instruction to use the
24987 // "load" ports instead of the dedicated "store" port.
24988 // E.g., on Haswell:
24989 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
24990 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
24991 if (isLegalAddressingMode(AM, Ty))
24992 // Scale represents reg2 * scale, thus account for 1
24993 // as soon as we use a second register.
24994 return AM.Scale != 0;
24998 bool X86TargetLowering::isTargetFTOL() const {
24999 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();