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::SIGN_EXTEND, MVT::v16i32, Custom);
1316 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1317 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1318 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1319 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1320 if (Subtarget->hasDQI()) {
1321 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1322 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1324 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1325 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1326 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1327 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1328 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1329 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1330 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1331 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1332 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1333 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1335 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1336 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1337 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1338 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1339 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1341 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1342 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1344 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1346 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1347 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1348 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1349 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1350 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1351 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1352 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1353 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1354 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1356 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1357 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1359 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1360 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1362 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1364 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1365 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1367 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1368 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1370 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1371 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1373 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1374 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1375 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1376 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1377 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1378 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1380 if (Subtarget->hasCDI()) {
1381 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1382 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1384 if (Subtarget->hasDQI()) {
1385 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1386 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1387 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1389 // Custom lower several nodes.
1390 for (MVT VT : MVT::vector_valuetypes()) {
1391 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1392 // Extract subvector is special because the value type
1393 // (result) is 256/128-bit but the source is 512-bit wide.
1394 if (VT.is128BitVector() || VT.is256BitVector()) {
1395 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1397 if (VT.getVectorElementType() == MVT::i1)
1398 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1400 // Do not attempt to custom lower other non-512-bit vectors
1401 if (!VT.is512BitVector())
1404 if ( EltSize >= 32) {
1405 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1406 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1407 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1408 setOperationAction(ISD::VSELECT, VT, Legal);
1409 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1410 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1411 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1412 setOperationAction(ISD::MLOAD, VT, Legal);
1413 setOperationAction(ISD::MSTORE, VT, Legal);
1416 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1417 MVT VT = (MVT::SimpleValueType)i;
1419 // Do not attempt to promote non-512-bit vectors.
1420 if (!VT.is512BitVector())
1423 setOperationAction(ISD::SELECT, VT, Promote);
1424 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1428 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1429 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1430 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1432 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1433 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1435 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1436 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1437 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1438 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1439 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1440 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1441 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1442 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1443 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1444 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1445 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1446 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1447 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1449 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1450 const MVT VT = (MVT::SimpleValueType)i;
1452 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1454 // Do not attempt to promote non-512-bit vectors.
1455 if (!VT.is512BitVector())
1459 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1460 setOperationAction(ISD::VSELECT, VT, Legal);
1465 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1466 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1467 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1469 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1470 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1473 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1474 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1476 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1477 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1478 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1479 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1480 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1481 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1484 // We want to custom lower some of our intrinsics.
1485 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1486 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1487 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1488 if (!Subtarget->is64Bit())
1489 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1491 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1492 // handle type legalization for these operations here.
1494 // FIXME: We really should do custom legalization for addition and
1495 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1496 // than generic legalization for 64-bit multiplication-with-overflow, though.
1497 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1498 // Add/Sub/Mul with overflow operations are custom lowered.
1500 setOperationAction(ISD::SADDO, VT, Custom);
1501 setOperationAction(ISD::UADDO, VT, Custom);
1502 setOperationAction(ISD::SSUBO, VT, Custom);
1503 setOperationAction(ISD::USUBO, VT, Custom);
1504 setOperationAction(ISD::SMULO, VT, Custom);
1505 setOperationAction(ISD::UMULO, VT, Custom);
1509 if (!Subtarget->is64Bit()) {
1510 // These libcalls are not available in 32-bit.
1511 setLibcallName(RTLIB::SHL_I128, nullptr);
1512 setLibcallName(RTLIB::SRL_I128, nullptr);
1513 setLibcallName(RTLIB::SRA_I128, nullptr);
1516 // Combine sin / cos into one node or libcall if possible.
1517 if (Subtarget->hasSinCos()) {
1518 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1519 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1520 if (Subtarget->isTargetDarwin()) {
1521 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1522 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1523 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1524 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1528 if (Subtarget->isTargetWin64()) {
1529 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1530 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1531 setOperationAction(ISD::SREM, MVT::i128, Custom);
1532 setOperationAction(ISD::UREM, MVT::i128, Custom);
1533 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1534 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1537 // We have target-specific dag combine patterns for the following nodes:
1538 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1539 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1540 setTargetDAGCombine(ISD::BITCAST);
1541 setTargetDAGCombine(ISD::VSELECT);
1542 setTargetDAGCombine(ISD::SELECT);
1543 setTargetDAGCombine(ISD::SHL);
1544 setTargetDAGCombine(ISD::SRA);
1545 setTargetDAGCombine(ISD::SRL);
1546 setTargetDAGCombine(ISD::OR);
1547 setTargetDAGCombine(ISD::AND);
1548 setTargetDAGCombine(ISD::ADD);
1549 setTargetDAGCombine(ISD::FADD);
1550 setTargetDAGCombine(ISD::FSUB);
1551 setTargetDAGCombine(ISD::FMA);
1552 setTargetDAGCombine(ISD::SUB);
1553 setTargetDAGCombine(ISD::LOAD);
1554 setTargetDAGCombine(ISD::MLOAD);
1555 setTargetDAGCombine(ISD::STORE);
1556 setTargetDAGCombine(ISD::MSTORE);
1557 setTargetDAGCombine(ISD::ZERO_EXTEND);
1558 setTargetDAGCombine(ISD::ANY_EXTEND);
1559 setTargetDAGCombine(ISD::SIGN_EXTEND);
1560 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1561 setTargetDAGCombine(ISD::TRUNCATE);
1562 setTargetDAGCombine(ISD::SINT_TO_FP);
1563 setTargetDAGCombine(ISD::SETCC);
1564 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1565 setTargetDAGCombine(ISD::BUILD_VECTOR);
1566 setTargetDAGCombine(ISD::MUL);
1567 setTargetDAGCombine(ISD::XOR);
1569 computeRegisterProperties(Subtarget->getRegisterInfo());
1571 // On Darwin, -Os means optimize for size without hurting performance,
1572 // do not reduce the limit.
1573 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1574 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1575 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1576 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1577 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1578 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1579 setPrefLoopAlignment(4); // 2^4 bytes.
1581 // Predictable cmov don't hurt on atom because it's in-order.
1582 PredictableSelectIsExpensive = !Subtarget->isAtom();
1583 EnableExtLdPromotion = true;
1584 setPrefFunctionAlignment(4); // 2^4 bytes.
1586 verifyIntrinsicTables();
1589 // This has so far only been implemented for 64-bit MachO.
1590 bool X86TargetLowering::useLoadStackGuardNode() const {
1591 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1594 TargetLoweringBase::LegalizeTypeAction
1595 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1596 if (ExperimentalVectorWideningLegalization &&
1597 VT.getVectorNumElements() != 1 &&
1598 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1599 return TypeWidenVector;
1601 return TargetLoweringBase::getPreferredVectorAction(VT);
1604 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1606 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1608 const unsigned NumElts = VT.getVectorNumElements();
1609 const EVT EltVT = VT.getVectorElementType();
1610 if (VT.is512BitVector()) {
1611 if (Subtarget->hasAVX512())
1612 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1613 EltVT == MVT::f32 || EltVT == MVT::f64)
1615 case 8: return MVT::v8i1;
1616 case 16: return MVT::v16i1;
1618 if (Subtarget->hasBWI())
1619 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1621 case 32: return MVT::v32i1;
1622 case 64: return MVT::v64i1;
1626 if (VT.is256BitVector() || VT.is128BitVector()) {
1627 if (Subtarget->hasVLX())
1628 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1629 EltVT == MVT::f32 || EltVT == MVT::f64)
1631 case 2: return MVT::v2i1;
1632 case 4: return MVT::v4i1;
1633 case 8: return MVT::v8i1;
1635 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1636 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1638 case 8: return MVT::v8i1;
1639 case 16: return MVT::v16i1;
1640 case 32: return MVT::v32i1;
1644 return VT.changeVectorElementTypeToInteger();
1647 /// Helper for getByValTypeAlignment to determine
1648 /// the desired ByVal argument alignment.
1649 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1652 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1653 if (VTy->getBitWidth() == 128)
1655 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1656 unsigned EltAlign = 0;
1657 getMaxByValAlign(ATy->getElementType(), EltAlign);
1658 if (EltAlign > MaxAlign)
1659 MaxAlign = EltAlign;
1660 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1661 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1662 unsigned EltAlign = 0;
1663 getMaxByValAlign(STy->getElementType(i), EltAlign);
1664 if (EltAlign > MaxAlign)
1665 MaxAlign = EltAlign;
1672 /// Return the desired alignment for ByVal aggregate
1673 /// function arguments in the caller parameter area. For X86, aggregates
1674 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1675 /// are at 4-byte boundaries.
1676 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1677 if (Subtarget->is64Bit()) {
1678 // Max of 8 and alignment of type.
1679 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1686 if (Subtarget->hasSSE1())
1687 getMaxByValAlign(Ty, Align);
1691 /// Returns the target specific optimal type for load
1692 /// and store operations as a result of memset, memcpy, and memmove
1693 /// lowering. If DstAlign is zero that means it's safe to destination
1694 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1695 /// means there isn't a need to check it against alignment requirement,
1696 /// probably because the source does not need to be loaded. If 'IsMemset' is
1697 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1698 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1699 /// source is constant so it does not need to be loaded.
1700 /// It returns EVT::Other if the type should be determined using generic
1701 /// target-independent logic.
1703 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1704 unsigned DstAlign, unsigned SrcAlign,
1705 bool IsMemset, bool ZeroMemset,
1707 MachineFunction &MF) const {
1708 const Function *F = MF.getFunction();
1709 if ((!IsMemset || ZeroMemset) &&
1710 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1712 (Subtarget->isUnalignedMemAccessFast() ||
1713 ((DstAlign == 0 || DstAlign >= 16) &&
1714 (SrcAlign == 0 || SrcAlign >= 16)))) {
1716 if (Subtarget->hasInt256())
1718 if (Subtarget->hasFp256())
1721 if (Subtarget->hasSSE2())
1723 if (Subtarget->hasSSE1())
1725 } else if (!MemcpyStrSrc && Size >= 8 &&
1726 !Subtarget->is64Bit() &&
1727 Subtarget->hasSSE2()) {
1728 // Do not use f64 to lower memcpy if source is string constant. It's
1729 // better to use i32 to avoid the loads.
1733 if (Subtarget->is64Bit() && Size >= 8)
1738 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1740 return X86ScalarSSEf32;
1741 else if (VT == MVT::f64)
1742 return X86ScalarSSEf64;
1747 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1752 *Fast = Subtarget->isUnalignedMemAccessFast();
1756 /// Return the entry encoding for a jump table in the
1757 /// current function. The returned value is a member of the
1758 /// MachineJumpTableInfo::JTEntryKind enum.
1759 unsigned X86TargetLowering::getJumpTableEncoding() const {
1760 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1762 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1763 Subtarget->isPICStyleGOT())
1764 return MachineJumpTableInfo::EK_Custom32;
1766 // Otherwise, use the normal jump table encoding heuristics.
1767 return TargetLowering::getJumpTableEncoding();
1771 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1772 const MachineBasicBlock *MBB,
1773 unsigned uid,MCContext &Ctx) const{
1774 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1775 Subtarget->isPICStyleGOT());
1776 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1778 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1779 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1782 /// Returns relocation base for the given PIC jumptable.
1783 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1784 SelectionDAG &DAG) const {
1785 if (!Subtarget->is64Bit())
1786 // This doesn't have SDLoc associated with it, but is not really the
1787 // same as a Register.
1788 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1792 /// This returns the relocation base for the given PIC jumptable,
1793 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1794 const MCExpr *X86TargetLowering::
1795 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1796 MCContext &Ctx) const {
1797 // X86-64 uses RIP relative addressing based on the jump table label.
1798 if (Subtarget->isPICStyleRIPRel())
1799 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1801 // Otherwise, the reference is relative to the PIC base.
1802 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1805 std::pair<const TargetRegisterClass *, uint8_t>
1806 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1808 const TargetRegisterClass *RRC = nullptr;
1810 switch (VT.SimpleTy) {
1812 return TargetLowering::findRepresentativeClass(TRI, VT);
1813 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1814 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1817 RRC = &X86::VR64RegClass;
1819 case MVT::f32: case MVT::f64:
1820 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1821 case MVT::v4f32: case MVT::v2f64:
1822 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1824 RRC = &X86::VR128RegClass;
1827 return std::make_pair(RRC, Cost);
1830 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1831 unsigned &Offset) const {
1832 if (!Subtarget->isTargetLinux())
1835 if (Subtarget->is64Bit()) {
1836 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1838 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1850 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1851 unsigned DestAS) const {
1852 assert(SrcAS != DestAS && "Expected different address spaces!");
1854 return SrcAS < 256 && DestAS < 256;
1857 //===----------------------------------------------------------------------===//
1858 // Return Value Calling Convention Implementation
1859 //===----------------------------------------------------------------------===//
1861 #include "X86GenCallingConv.inc"
1864 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1865 MachineFunction &MF, bool isVarArg,
1866 const SmallVectorImpl<ISD::OutputArg> &Outs,
1867 LLVMContext &Context) const {
1868 SmallVector<CCValAssign, 16> RVLocs;
1869 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1870 return CCInfo.CheckReturn(Outs, RetCC_X86);
1873 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1874 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1879 X86TargetLowering::LowerReturn(SDValue Chain,
1880 CallingConv::ID CallConv, bool isVarArg,
1881 const SmallVectorImpl<ISD::OutputArg> &Outs,
1882 const SmallVectorImpl<SDValue> &OutVals,
1883 SDLoc dl, SelectionDAG &DAG) const {
1884 MachineFunction &MF = DAG.getMachineFunction();
1885 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1887 SmallVector<CCValAssign, 16> RVLocs;
1888 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1889 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1892 SmallVector<SDValue, 6> RetOps;
1893 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1894 // Operand #1 = Bytes To Pop
1895 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1898 // Copy the result values into the output registers.
1899 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1900 CCValAssign &VA = RVLocs[i];
1901 assert(VA.isRegLoc() && "Can only return in registers!");
1902 SDValue ValToCopy = OutVals[i];
1903 EVT ValVT = ValToCopy.getValueType();
1905 // Promote values to the appropriate types.
1906 if (VA.getLocInfo() == CCValAssign::SExt)
1907 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1908 else if (VA.getLocInfo() == CCValAssign::ZExt)
1909 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1910 else if (VA.getLocInfo() == CCValAssign::AExt) {
1911 if (ValVT.getScalarType() == MVT::i1)
1912 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1914 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1916 else if (VA.getLocInfo() == CCValAssign::BCvt)
1917 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1919 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1920 "Unexpected FP-extend for return value.");
1922 // If this is x86-64, and we disabled SSE, we can't return FP values,
1923 // or SSE or MMX vectors.
1924 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1925 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1926 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1927 report_fatal_error("SSE register return with SSE disabled");
1929 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1930 // llvm-gcc has never done it right and no one has noticed, so this
1931 // should be OK for now.
1932 if (ValVT == MVT::f64 &&
1933 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1934 report_fatal_error("SSE2 register return with SSE2 disabled");
1936 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1937 // the RET instruction and handled by the FP Stackifier.
1938 if (VA.getLocReg() == X86::FP0 ||
1939 VA.getLocReg() == X86::FP1) {
1940 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1941 // change the value to the FP stack register class.
1942 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1943 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1944 RetOps.push_back(ValToCopy);
1945 // Don't emit a copytoreg.
1949 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1950 // which is returned in RAX / RDX.
1951 if (Subtarget->is64Bit()) {
1952 if (ValVT == MVT::x86mmx) {
1953 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1954 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1955 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1957 // If we don't have SSE2 available, convert to v4f32 so the generated
1958 // register is legal.
1959 if (!Subtarget->hasSSE2())
1960 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1965 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1966 Flag = Chain.getValue(1);
1967 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1970 // The x86-64 ABIs require that for returning structs by value we copy
1971 // the sret argument into %rax/%eax (depending on ABI) for the return.
1972 // Win32 requires us to put the sret argument to %eax as well.
1973 // We saved the argument into a virtual register in the entry block,
1974 // so now we copy the value out and into %rax/%eax.
1976 // Checking Function.hasStructRetAttr() here is insufficient because the IR
1977 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
1978 // false, then an sret argument may be implicitly inserted in the SelDAG. In
1979 // either case FuncInfo->setSRetReturnReg() will have been called.
1980 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
1981 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
1982 "No need for an sret register");
1983 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
1986 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1987 X86::RAX : X86::EAX;
1988 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1989 Flag = Chain.getValue(1);
1991 // RAX/EAX now acts like a return value.
1992 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
1995 RetOps[0] = Chain; // Update chain.
1997 // Add the flag if we have it.
1999 RetOps.push_back(Flag);
2001 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2004 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2005 if (N->getNumValues() != 1)
2007 if (!N->hasNUsesOfValue(1, 0))
2010 SDValue TCChain = Chain;
2011 SDNode *Copy = *N->use_begin();
2012 if (Copy->getOpcode() == ISD::CopyToReg) {
2013 // If the copy has a glue operand, we conservatively assume it isn't safe to
2014 // perform a tail call.
2015 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2017 TCChain = Copy->getOperand(0);
2018 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2021 bool HasRet = false;
2022 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2024 if (UI->getOpcode() != X86ISD::RET_FLAG)
2026 // If we are returning more than one value, we can definitely
2027 // not make a tail call see PR19530
2028 if (UI->getNumOperands() > 4)
2030 if (UI->getNumOperands() == 4 &&
2031 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2044 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2045 ISD::NodeType ExtendKind) const {
2047 // TODO: Is this also valid on 32-bit?
2048 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2049 ReturnMVT = MVT::i8;
2051 ReturnMVT = MVT::i32;
2053 EVT MinVT = getRegisterType(Context, ReturnMVT);
2054 return VT.bitsLT(MinVT) ? MinVT : VT;
2057 /// Lower the result values of a call into the
2058 /// appropriate copies out of appropriate physical registers.
2061 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2062 CallingConv::ID CallConv, bool isVarArg,
2063 const SmallVectorImpl<ISD::InputArg> &Ins,
2064 SDLoc dl, SelectionDAG &DAG,
2065 SmallVectorImpl<SDValue> &InVals) const {
2067 // Assign locations to each value returned by this call.
2068 SmallVector<CCValAssign, 16> RVLocs;
2069 bool Is64Bit = Subtarget->is64Bit();
2070 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2072 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2074 // Copy all of the result registers out of their specified physreg.
2075 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2076 CCValAssign &VA = RVLocs[i];
2077 EVT CopyVT = VA.getValVT();
2079 // If this is x86-64, and we disabled SSE, we can't return FP values
2080 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2081 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2082 report_fatal_error("SSE register return with SSE disabled");
2085 // If we prefer to use the value in xmm registers, copy it out as f80 and
2086 // use a truncate to move it from fp stack reg to xmm reg.
2087 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2088 isScalarFPTypeInSSEReg(VA.getValVT()))
2091 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2092 CopyVT, InFlag).getValue(1);
2093 SDValue Val = Chain.getValue(0);
2095 if (CopyVT != VA.getValVT())
2096 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2097 // This truncation won't change the value.
2098 DAG.getIntPtrConstant(1));
2100 InFlag = Chain.getValue(2);
2101 InVals.push_back(Val);
2107 //===----------------------------------------------------------------------===//
2108 // C & StdCall & Fast Calling Convention implementation
2109 //===----------------------------------------------------------------------===//
2110 // StdCall calling convention seems to be standard for many Windows' API
2111 // routines and around. It differs from C calling convention just a little:
2112 // callee should clean up the stack, not caller. Symbols should be also
2113 // decorated in some fancy way :) It doesn't support any vector arguments.
2114 // For info on fast calling convention see Fast Calling Convention (tail call)
2115 // implementation LowerX86_32FastCCCallTo.
2117 /// CallIsStructReturn - Determines whether a call uses struct return
2119 enum StructReturnType {
2124 static StructReturnType
2125 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2127 return NotStructReturn;
2129 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2130 if (!Flags.isSRet())
2131 return NotStructReturn;
2132 if (Flags.isInReg())
2133 return RegStructReturn;
2134 return StackStructReturn;
2137 /// Determines whether a function uses struct return semantics.
2138 static StructReturnType
2139 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2141 return NotStructReturn;
2143 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2144 if (!Flags.isSRet())
2145 return NotStructReturn;
2146 if (Flags.isInReg())
2147 return RegStructReturn;
2148 return StackStructReturn;
2151 /// Make a copy of an aggregate at address specified by "Src" to address
2152 /// "Dst" with size and alignment information specified by the specific
2153 /// parameter attribute. The copy will be passed as a byval function parameter.
2155 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2156 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2158 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2160 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2161 /*isVolatile*/false, /*AlwaysInline=*/true,
2162 /*isTailCall*/false,
2163 MachinePointerInfo(), MachinePointerInfo());
2166 /// Return true if the calling convention is one that
2167 /// supports tail call optimization.
2168 static bool IsTailCallConvention(CallingConv::ID CC) {
2169 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2170 CC == CallingConv::HiPE);
2173 /// \brief Return true if the calling convention is a C calling convention.
2174 static bool IsCCallConvention(CallingConv::ID CC) {
2175 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2176 CC == CallingConv::X86_64_SysV);
2179 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2180 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2184 CallingConv::ID CalleeCC = CS.getCallingConv();
2185 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2191 /// Return true if the function is being made into
2192 /// a tailcall target by changing its ABI.
2193 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2194 bool GuaranteedTailCallOpt) {
2195 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2199 X86TargetLowering::LowerMemArgument(SDValue Chain,
2200 CallingConv::ID CallConv,
2201 const SmallVectorImpl<ISD::InputArg> &Ins,
2202 SDLoc dl, SelectionDAG &DAG,
2203 const CCValAssign &VA,
2204 MachineFrameInfo *MFI,
2206 // Create the nodes corresponding to a load from this parameter slot.
2207 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2208 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2209 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2210 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2213 // If value is passed by pointer we have address passed instead of the value
2215 if (VA.getLocInfo() == CCValAssign::Indirect)
2216 ValVT = VA.getLocVT();
2218 ValVT = VA.getValVT();
2220 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2221 // changed with more analysis.
2222 // In case of tail call optimization mark all arguments mutable. Since they
2223 // could be overwritten by lowering of arguments in case of a tail call.
2224 if (Flags.isByVal()) {
2225 unsigned Bytes = Flags.getByValSize();
2226 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2227 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2228 return DAG.getFrameIndex(FI, getPointerTy());
2230 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2231 VA.getLocMemOffset(), isImmutable);
2232 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2233 return DAG.getLoad(ValVT, dl, Chain, FIN,
2234 MachinePointerInfo::getFixedStack(FI),
2235 false, false, false, 0);
2239 // FIXME: Get this from tablegen.
2240 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2241 const X86Subtarget *Subtarget) {
2242 assert(Subtarget->is64Bit());
2244 if (Subtarget->isCallingConvWin64(CallConv)) {
2245 static const MCPhysReg GPR64ArgRegsWin64[] = {
2246 X86::RCX, X86::RDX, X86::R8, X86::R9
2248 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2251 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2252 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2254 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2257 // FIXME: Get this from tablegen.
2258 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2259 CallingConv::ID CallConv,
2260 const X86Subtarget *Subtarget) {
2261 assert(Subtarget->is64Bit());
2262 if (Subtarget->isCallingConvWin64(CallConv)) {
2263 // The XMM registers which might contain var arg parameters are shadowed
2264 // in their paired GPR. So we only need to save the GPR to their home
2266 // TODO: __vectorcall will change this.
2270 const Function *Fn = MF.getFunction();
2271 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2272 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2273 "SSE register cannot be used when SSE is disabled!");
2274 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2275 !Subtarget->hasSSE1())
2276 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2280 static const MCPhysReg XMMArgRegs64Bit[] = {
2281 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2282 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2284 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2288 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2289 CallingConv::ID CallConv,
2291 const SmallVectorImpl<ISD::InputArg> &Ins,
2294 SmallVectorImpl<SDValue> &InVals)
2296 MachineFunction &MF = DAG.getMachineFunction();
2297 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2298 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2300 const Function* Fn = MF.getFunction();
2301 if (Fn->hasExternalLinkage() &&
2302 Subtarget->isTargetCygMing() &&
2303 Fn->getName() == "main")
2304 FuncInfo->setForceFramePointer(true);
2306 MachineFrameInfo *MFI = MF.getFrameInfo();
2307 bool Is64Bit = Subtarget->is64Bit();
2308 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2310 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2311 "Var args not supported with calling convention fastcc, ghc or hipe");
2313 // Assign locations to all of the incoming arguments.
2314 SmallVector<CCValAssign, 16> ArgLocs;
2315 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2317 // Allocate shadow area for Win64
2319 CCInfo.AllocateStack(32, 8);
2321 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2323 unsigned LastVal = ~0U;
2325 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2326 CCValAssign &VA = ArgLocs[i];
2327 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2329 assert(VA.getValNo() != LastVal &&
2330 "Don't support value assigned to multiple locs yet");
2332 LastVal = VA.getValNo();
2334 if (VA.isRegLoc()) {
2335 EVT RegVT = VA.getLocVT();
2336 const TargetRegisterClass *RC;
2337 if (RegVT == MVT::i32)
2338 RC = &X86::GR32RegClass;
2339 else if (Is64Bit && RegVT == MVT::i64)
2340 RC = &X86::GR64RegClass;
2341 else if (RegVT == MVT::f32)
2342 RC = &X86::FR32RegClass;
2343 else if (RegVT == MVT::f64)
2344 RC = &X86::FR64RegClass;
2345 else if (RegVT.is512BitVector())
2346 RC = &X86::VR512RegClass;
2347 else if (RegVT.is256BitVector())
2348 RC = &X86::VR256RegClass;
2349 else if (RegVT.is128BitVector())
2350 RC = &X86::VR128RegClass;
2351 else if (RegVT == MVT::x86mmx)
2352 RC = &X86::VR64RegClass;
2353 else if (RegVT == MVT::i1)
2354 RC = &X86::VK1RegClass;
2355 else if (RegVT == MVT::v8i1)
2356 RC = &X86::VK8RegClass;
2357 else if (RegVT == MVT::v16i1)
2358 RC = &X86::VK16RegClass;
2359 else if (RegVT == MVT::v32i1)
2360 RC = &X86::VK32RegClass;
2361 else if (RegVT == MVT::v64i1)
2362 RC = &X86::VK64RegClass;
2364 llvm_unreachable("Unknown argument type!");
2366 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2367 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2369 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2370 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2372 if (VA.getLocInfo() == CCValAssign::SExt)
2373 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2374 DAG.getValueType(VA.getValVT()));
2375 else if (VA.getLocInfo() == CCValAssign::ZExt)
2376 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2377 DAG.getValueType(VA.getValVT()));
2378 else if (VA.getLocInfo() == CCValAssign::BCvt)
2379 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2381 if (VA.isExtInLoc()) {
2382 // Handle MMX values passed in XMM regs.
2383 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2384 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2386 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2389 assert(VA.isMemLoc());
2390 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2393 // If value is passed via pointer - do a load.
2394 if (VA.getLocInfo() == CCValAssign::Indirect)
2395 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2396 MachinePointerInfo(), false, false, false, 0);
2398 InVals.push_back(ArgValue);
2401 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2402 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2403 // The x86-64 ABIs require that for returning structs by value we copy
2404 // the sret argument into %rax/%eax (depending on ABI) for the return.
2405 // Win32 requires us to put the sret argument to %eax as well.
2406 // Save the argument into a virtual register so that we can access it
2407 // from the return points.
2408 if (Ins[i].Flags.isSRet()) {
2409 unsigned Reg = FuncInfo->getSRetReturnReg();
2411 MVT PtrTy = getPointerTy();
2412 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2413 FuncInfo->setSRetReturnReg(Reg);
2415 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2416 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2422 unsigned StackSize = CCInfo.getNextStackOffset();
2423 // Align stack specially for tail calls.
2424 if (FuncIsMadeTailCallSafe(CallConv,
2425 MF.getTarget().Options.GuaranteedTailCallOpt))
2426 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2428 // If the function takes variable number of arguments, make a frame index for
2429 // the start of the first vararg value... for expansion of llvm.va_start. We
2430 // can skip this if there are no va_start calls.
2431 if (MFI->hasVAStart() &&
2432 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2433 CallConv != CallingConv::X86_ThisCall))) {
2434 FuncInfo->setVarArgsFrameIndex(
2435 MFI->CreateFixedObject(1, StackSize, true));
2438 MachineModuleInfo &MMI = MF.getMMI();
2439 const Function *WinEHParent = nullptr;
2440 if (IsWin64 && MMI.hasWinEHFuncInfo(Fn))
2441 WinEHParent = MMI.getWinEHParent(Fn);
2442 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2443 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2445 // Figure out if XMM registers are in use.
2446 assert(!(MF.getTarget().Options.UseSoftFloat &&
2447 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2448 "SSE register cannot be used when SSE is disabled!");
2450 // 64-bit calling conventions support varargs and register parameters, so we
2451 // have to do extra work to spill them in the prologue.
2452 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2453 // Find the first unallocated argument registers.
2454 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2455 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2456 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2457 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2458 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2459 "SSE register cannot be used when SSE is disabled!");
2461 // Gather all the live in physical registers.
2462 SmallVector<SDValue, 6> LiveGPRs;
2463 SmallVector<SDValue, 8> LiveXMMRegs;
2465 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2466 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2468 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2470 if (!ArgXMMs.empty()) {
2471 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2472 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2473 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2474 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2475 LiveXMMRegs.push_back(
2476 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2481 // Get to the caller-allocated home save location. Add 8 to account
2482 // for the return address.
2483 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2484 FuncInfo->setRegSaveFrameIndex(
2485 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2486 // Fixup to set vararg frame on shadow area (4 x i64).
2488 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2490 // For X86-64, if there are vararg parameters that are passed via
2491 // registers, then we must store them to their spots on the stack so
2492 // they may be loaded by deferencing the result of va_next.
2493 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2494 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2495 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2496 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2499 // Store the integer parameter registers.
2500 SmallVector<SDValue, 8> MemOps;
2501 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2503 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2504 for (SDValue Val : LiveGPRs) {
2505 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2506 DAG.getIntPtrConstant(Offset));
2508 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2509 MachinePointerInfo::getFixedStack(
2510 FuncInfo->getRegSaveFrameIndex(), Offset),
2512 MemOps.push_back(Store);
2516 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2517 // Now store the XMM (fp + vector) parameter registers.
2518 SmallVector<SDValue, 12> SaveXMMOps;
2519 SaveXMMOps.push_back(Chain);
2520 SaveXMMOps.push_back(ALVal);
2521 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2522 FuncInfo->getRegSaveFrameIndex()));
2523 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2524 FuncInfo->getVarArgsFPOffset()));
2525 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2527 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2528 MVT::Other, SaveXMMOps));
2531 if (!MemOps.empty())
2532 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2533 } else if (IsWinEHOutlined) {
2534 // Get to the caller-allocated home save location. Add 8 to account
2535 // for the return address.
2536 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2537 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2538 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2540 MMI.getWinEHFuncInfo(Fn)
2541 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2542 FuncInfo->getRegSaveFrameIndex();
2544 // Store the second integer parameter (rdx) into rsp+16 relative to the
2545 // stack pointer at the entry of the function.
2547 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), getPointerTy());
2548 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2549 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2550 Chain = DAG.getStore(
2551 Val.getValue(1), dl, Val, RSFIN,
2552 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2553 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2556 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2557 // Find the largest legal vector type.
2558 MVT VecVT = MVT::Other;
2559 // FIXME: Only some x86_32 calling conventions support AVX512.
2560 if (Subtarget->hasAVX512() &&
2561 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2562 CallConv == CallingConv::Intel_OCL_BI)))
2563 VecVT = MVT::v16f32;
2564 else if (Subtarget->hasAVX())
2566 else if (Subtarget->hasSSE2())
2569 // We forward some GPRs and some vector types.
2570 SmallVector<MVT, 2> RegParmTypes;
2571 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2572 RegParmTypes.push_back(IntVT);
2573 if (VecVT != MVT::Other)
2574 RegParmTypes.push_back(VecVT);
2576 // Compute the set of forwarded registers. The rest are scratch.
2577 SmallVectorImpl<ForwardedRegister> &Forwards =
2578 FuncInfo->getForwardedMustTailRegParms();
2579 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2581 // Conservatively forward AL on x86_64, since it might be used for varargs.
2582 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2583 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2584 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2587 // Copy all forwards from physical to virtual registers.
2588 for (ForwardedRegister &F : Forwards) {
2589 // FIXME: Can we use a less constrained schedule?
2590 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2591 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2592 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2596 // Some CCs need callee pop.
2597 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2598 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2599 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2601 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2602 // If this is an sret function, the return should pop the hidden pointer.
2603 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2604 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2605 argsAreStructReturn(Ins) == StackStructReturn)
2606 FuncInfo->setBytesToPopOnReturn(4);
2610 // RegSaveFrameIndex is X86-64 only.
2611 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2612 if (CallConv == CallingConv::X86_FastCall ||
2613 CallConv == CallingConv::X86_ThisCall)
2614 // fastcc functions can't have varargs.
2615 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2618 FuncInfo->setArgumentStackSize(StackSize);
2620 if (IsWinEHParent) {
2621 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2622 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2623 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2624 SDValue Neg2 = DAG.getConstant(-2, MVT::i64);
2625 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2626 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2627 /*isVolatile=*/true,
2628 /*isNonTemporal=*/false, /*Alignment=*/0);
2635 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2636 SDValue StackPtr, SDValue Arg,
2637 SDLoc dl, SelectionDAG &DAG,
2638 const CCValAssign &VA,
2639 ISD::ArgFlagsTy Flags) const {
2640 unsigned LocMemOffset = VA.getLocMemOffset();
2641 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2642 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2643 if (Flags.isByVal())
2644 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2646 return DAG.getStore(Chain, dl, Arg, PtrOff,
2647 MachinePointerInfo::getStack(LocMemOffset),
2651 /// Emit a load of return address if tail call
2652 /// optimization is performed and it is required.
2654 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2655 SDValue &OutRetAddr, SDValue Chain,
2656 bool IsTailCall, bool Is64Bit,
2657 int FPDiff, SDLoc dl) const {
2658 // Adjust the Return address stack slot.
2659 EVT VT = getPointerTy();
2660 OutRetAddr = getReturnAddressFrameIndex(DAG);
2662 // Load the "old" Return address.
2663 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2664 false, false, false, 0);
2665 return SDValue(OutRetAddr.getNode(), 1);
2668 /// Emit a store of the return address if tail call
2669 /// optimization is performed and it is required (FPDiff!=0).
2670 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2671 SDValue Chain, SDValue RetAddrFrIdx,
2672 EVT PtrVT, unsigned SlotSize,
2673 int FPDiff, SDLoc dl) {
2674 // Store the return address to the appropriate stack slot.
2675 if (!FPDiff) return Chain;
2676 // Calculate the new stack slot for the return address.
2677 int NewReturnAddrFI =
2678 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2680 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2681 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2682 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2688 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2689 SmallVectorImpl<SDValue> &InVals) const {
2690 SelectionDAG &DAG = CLI.DAG;
2692 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2693 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2694 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2695 SDValue Chain = CLI.Chain;
2696 SDValue Callee = CLI.Callee;
2697 CallingConv::ID CallConv = CLI.CallConv;
2698 bool &isTailCall = CLI.IsTailCall;
2699 bool isVarArg = CLI.IsVarArg;
2701 MachineFunction &MF = DAG.getMachineFunction();
2702 bool Is64Bit = Subtarget->is64Bit();
2703 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2704 StructReturnType SR = callIsStructReturn(Outs);
2705 bool IsSibcall = false;
2706 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2708 if (MF.getTarget().Options.DisableTailCalls)
2711 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2713 // Force this to be a tail call. The verifier rules are enough to ensure
2714 // that we can lower this successfully without moving the return address
2717 } else if (isTailCall) {
2718 // Check if it's really possible to do a tail call.
2719 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2720 isVarArg, SR != NotStructReturn,
2721 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2722 Outs, OutVals, Ins, DAG);
2724 // Sibcalls are automatically detected tailcalls which do not require
2726 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2733 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2734 "Var args not supported with calling convention fastcc, ghc or hipe");
2736 // Analyze operands of the call, assigning locations to each operand.
2737 SmallVector<CCValAssign, 16> ArgLocs;
2738 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2740 // Allocate shadow area for Win64
2742 CCInfo.AllocateStack(32, 8);
2744 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2746 // Get a count of how many bytes are to be pushed on the stack.
2747 unsigned NumBytes = CCInfo.getNextStackOffset();
2749 // This is a sibcall. The memory operands are available in caller's
2750 // own caller's stack.
2752 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2753 IsTailCallConvention(CallConv))
2754 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2757 if (isTailCall && !IsSibcall && !IsMustTail) {
2758 // Lower arguments at fp - stackoffset + fpdiff.
2759 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2761 FPDiff = NumBytesCallerPushed - NumBytes;
2763 // Set the delta of movement of the returnaddr stackslot.
2764 // But only set if delta is greater than previous delta.
2765 if (FPDiff < X86Info->getTCReturnAddrDelta())
2766 X86Info->setTCReturnAddrDelta(FPDiff);
2769 unsigned NumBytesToPush = NumBytes;
2770 unsigned NumBytesToPop = NumBytes;
2772 // If we have an inalloca argument, all stack space has already been allocated
2773 // for us and be right at the top of the stack. We don't support multiple
2774 // arguments passed in memory when using inalloca.
2775 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2777 if (!ArgLocs.back().isMemLoc())
2778 report_fatal_error("cannot use inalloca attribute on a register "
2780 if (ArgLocs.back().getLocMemOffset() != 0)
2781 report_fatal_error("any parameter with the inalloca attribute must be "
2782 "the only memory argument");
2786 Chain = DAG.getCALLSEQ_START(
2787 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2789 SDValue RetAddrFrIdx;
2790 // Load return address for tail calls.
2791 if (isTailCall && FPDiff)
2792 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2793 Is64Bit, FPDiff, dl);
2795 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2796 SmallVector<SDValue, 8> MemOpChains;
2799 // Walk the register/memloc assignments, inserting copies/loads. In the case
2800 // of tail call optimization arguments are handle later.
2801 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2802 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2803 // Skip inalloca arguments, they have already been written.
2804 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2805 if (Flags.isInAlloca())
2808 CCValAssign &VA = ArgLocs[i];
2809 EVT RegVT = VA.getLocVT();
2810 SDValue Arg = OutVals[i];
2811 bool isByVal = Flags.isByVal();
2813 // Promote the value if needed.
2814 switch (VA.getLocInfo()) {
2815 default: llvm_unreachable("Unknown loc info!");
2816 case CCValAssign::Full: break;
2817 case CCValAssign::SExt:
2818 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2820 case CCValAssign::ZExt:
2821 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2823 case CCValAssign::AExt:
2824 if (RegVT.is128BitVector()) {
2825 // Special case: passing MMX values in XMM registers.
2826 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2827 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2828 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2830 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2832 case CCValAssign::BCvt:
2833 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2835 case CCValAssign::Indirect: {
2836 // Store the argument.
2837 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2838 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2839 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2840 MachinePointerInfo::getFixedStack(FI),
2847 if (VA.isRegLoc()) {
2848 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2849 if (isVarArg && IsWin64) {
2850 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2851 // shadow reg if callee is a varargs function.
2852 unsigned ShadowReg = 0;
2853 switch (VA.getLocReg()) {
2854 case X86::XMM0: ShadowReg = X86::RCX; break;
2855 case X86::XMM1: ShadowReg = X86::RDX; break;
2856 case X86::XMM2: ShadowReg = X86::R8; break;
2857 case X86::XMM3: ShadowReg = X86::R9; break;
2860 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2862 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2863 assert(VA.isMemLoc());
2864 if (!StackPtr.getNode())
2865 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2867 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2868 dl, DAG, VA, Flags));
2872 if (!MemOpChains.empty())
2873 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2875 if (Subtarget->isPICStyleGOT()) {
2876 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2879 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2880 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2882 // If we are tail calling and generating PIC/GOT style code load the
2883 // address of the callee into ECX. The value in ecx is used as target of
2884 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2885 // for tail calls on PIC/GOT architectures. Normally we would just put the
2886 // address of GOT into ebx and then call target@PLT. But for tail calls
2887 // ebx would be restored (since ebx is callee saved) before jumping to the
2890 // Note: The actual moving to ECX is done further down.
2891 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2892 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2893 !G->getGlobal()->hasProtectedVisibility())
2894 Callee = LowerGlobalAddress(Callee, DAG);
2895 else if (isa<ExternalSymbolSDNode>(Callee))
2896 Callee = LowerExternalSymbol(Callee, DAG);
2900 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2901 // From AMD64 ABI document:
2902 // For calls that may call functions that use varargs or stdargs
2903 // (prototype-less calls or calls to functions containing ellipsis (...) in
2904 // the declaration) %al is used as hidden argument to specify the number
2905 // of SSE registers used. The contents of %al do not need to match exactly
2906 // the number of registers, but must be an ubound on the number of SSE
2907 // registers used and is in the range 0 - 8 inclusive.
2909 // Count the number of XMM registers allocated.
2910 static const MCPhysReg XMMArgRegs[] = {
2911 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2912 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2914 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2915 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2916 && "SSE registers cannot be used when SSE is disabled");
2918 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2919 DAG.getConstant(NumXMMRegs, MVT::i8)));
2922 if (isVarArg && IsMustTail) {
2923 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
2924 for (const auto &F : Forwards) {
2925 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2926 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
2930 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2931 // don't need this because the eligibility check rejects calls that require
2932 // shuffling arguments passed in memory.
2933 if (!IsSibcall && isTailCall) {
2934 // Force all the incoming stack arguments to be loaded from the stack
2935 // before any new outgoing arguments are stored to the stack, because the
2936 // outgoing stack slots may alias the incoming argument stack slots, and
2937 // the alias isn't otherwise explicit. This is slightly more conservative
2938 // than necessary, because it means that each store effectively depends
2939 // on every argument instead of just those arguments it would clobber.
2940 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2942 SmallVector<SDValue, 8> MemOpChains2;
2945 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2946 CCValAssign &VA = ArgLocs[i];
2949 assert(VA.isMemLoc());
2950 SDValue Arg = OutVals[i];
2951 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2952 // Skip inalloca arguments. They don't require any work.
2953 if (Flags.isInAlloca())
2955 // Create frame index.
2956 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2957 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2958 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2959 FIN = DAG.getFrameIndex(FI, getPointerTy());
2961 if (Flags.isByVal()) {
2962 // Copy relative to framepointer.
2963 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2964 if (!StackPtr.getNode())
2965 StackPtr = DAG.getCopyFromReg(Chain, dl,
2966 RegInfo->getStackRegister(),
2968 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2970 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2974 // Store relative to framepointer.
2975 MemOpChains2.push_back(
2976 DAG.getStore(ArgChain, dl, Arg, FIN,
2977 MachinePointerInfo::getFixedStack(FI),
2982 if (!MemOpChains2.empty())
2983 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
2985 // Store the return address to the appropriate stack slot.
2986 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2987 getPointerTy(), RegInfo->getSlotSize(),
2991 // Build a sequence of copy-to-reg nodes chained together with token chain
2992 // and flag operands which copy the outgoing args into registers.
2994 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2995 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2996 RegsToPass[i].second, InFlag);
2997 InFlag = Chain.getValue(1);
3000 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3001 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3002 // In the 64-bit large code model, we have to make all calls
3003 // through a register, since the call instruction's 32-bit
3004 // pc-relative offset may not be large enough to hold the whole
3006 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3007 // If the callee is a GlobalAddress node (quite common, every direct call
3008 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3010 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3012 // We should use extra load for direct calls to dllimported functions in
3014 const GlobalValue *GV = G->getGlobal();
3015 if (!GV->hasDLLImportStorageClass()) {
3016 unsigned char OpFlags = 0;
3017 bool ExtraLoad = false;
3018 unsigned WrapperKind = ISD::DELETED_NODE;
3020 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3021 // external symbols most go through the PLT in PIC mode. If the symbol
3022 // has hidden or protected visibility, or if it is static or local, then
3023 // we don't need to use the PLT - we can directly call it.
3024 if (Subtarget->isTargetELF() &&
3025 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3026 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3027 OpFlags = X86II::MO_PLT;
3028 } else if (Subtarget->isPICStyleStubAny() &&
3029 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3030 (!Subtarget->getTargetTriple().isMacOSX() ||
3031 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3032 // PC-relative references to external symbols should go through $stub,
3033 // unless we're building with the leopard linker or later, which
3034 // automatically synthesizes these stubs.
3035 OpFlags = X86II::MO_DARWIN_STUB;
3036 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3037 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3038 // If the function is marked as non-lazy, generate an indirect call
3039 // which loads from the GOT directly. This avoids runtime overhead
3040 // at the cost of eager binding (and one extra byte of encoding).
3041 OpFlags = X86II::MO_GOTPCREL;
3042 WrapperKind = X86ISD::WrapperRIP;
3046 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3047 G->getOffset(), OpFlags);
3049 // Add a wrapper if needed.
3050 if (WrapperKind != ISD::DELETED_NODE)
3051 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3052 // Add extra indirection if needed.
3054 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3055 MachinePointerInfo::getGOT(),
3056 false, false, false, 0);
3058 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3059 unsigned char OpFlags = 0;
3061 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3062 // external symbols should go through the PLT.
3063 if (Subtarget->isTargetELF() &&
3064 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3065 OpFlags = X86II::MO_PLT;
3066 } else if (Subtarget->isPICStyleStubAny() &&
3067 (!Subtarget->getTargetTriple().isMacOSX() ||
3068 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3069 // PC-relative references to external symbols should go through $stub,
3070 // unless we're building with the leopard linker or later, which
3071 // automatically synthesizes these stubs.
3072 OpFlags = X86II::MO_DARWIN_STUB;
3075 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3077 } else if (Subtarget->isTarget64BitILP32() &&
3078 Callee->getValueType(0) == MVT::i32) {
3079 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3080 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3083 // Returns a chain & a flag for retval copy to use.
3084 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3085 SmallVector<SDValue, 8> Ops;
3087 if (!IsSibcall && isTailCall) {
3088 Chain = DAG.getCALLSEQ_END(Chain,
3089 DAG.getIntPtrConstant(NumBytesToPop, true),
3090 DAG.getIntPtrConstant(0, true), InFlag, dl);
3091 InFlag = Chain.getValue(1);
3094 Ops.push_back(Chain);
3095 Ops.push_back(Callee);
3098 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3100 // Add argument registers to the end of the list so that they are known live
3102 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3103 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3104 RegsToPass[i].second.getValueType()));
3106 // Add a register mask operand representing the call-preserved registers.
3107 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3108 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
3109 assert(Mask && "Missing call preserved mask for calling convention");
3110 Ops.push_back(DAG.getRegisterMask(Mask));
3112 if (InFlag.getNode())
3113 Ops.push_back(InFlag);
3117 //// If this is the first return lowered for this function, add the regs
3118 //// to the liveout set for the function.
3119 // This isn't right, although it's probably harmless on x86; liveouts
3120 // should be computed from returns not tail calls. Consider a void
3121 // function making a tail call to a function returning int.
3122 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3125 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3126 InFlag = Chain.getValue(1);
3128 // Create the CALLSEQ_END node.
3129 unsigned NumBytesForCalleeToPop;
3130 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3131 DAG.getTarget().Options.GuaranteedTailCallOpt))
3132 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3133 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3134 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3135 SR == StackStructReturn)
3136 // If this is a call to a struct-return function, the callee
3137 // pops the hidden struct pointer, so we have to push it back.
3138 // This is common for Darwin/X86, Linux & Mingw32 targets.
3139 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3140 NumBytesForCalleeToPop = 4;
3142 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3144 // Returns a flag for retval copy to use.
3146 Chain = DAG.getCALLSEQ_END(Chain,
3147 DAG.getIntPtrConstant(NumBytesToPop, true),
3148 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3151 InFlag = Chain.getValue(1);
3154 // Handle result values, copying them out of physregs into vregs that we
3156 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3157 Ins, dl, DAG, InVals);
3160 //===----------------------------------------------------------------------===//
3161 // Fast Calling Convention (tail call) implementation
3162 //===----------------------------------------------------------------------===//
3164 // Like std call, callee cleans arguments, convention except that ECX is
3165 // reserved for storing the tail called function address. Only 2 registers are
3166 // free for argument passing (inreg). Tail call optimization is performed
3168 // * tailcallopt is enabled
3169 // * caller/callee are fastcc
3170 // On X86_64 architecture with GOT-style position independent code only local
3171 // (within module) calls are supported at the moment.
3172 // To keep the stack aligned according to platform abi the function
3173 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3174 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3175 // If a tail called function callee has more arguments than the caller the
3176 // caller needs to make sure that there is room to move the RETADDR to. This is
3177 // achieved by reserving an area the size of the argument delta right after the
3178 // original RETADDR, but before the saved framepointer or the spilled registers
3179 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3191 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3192 /// for a 16 byte align requirement.
3194 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3195 SelectionDAG& DAG) const {
3196 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3197 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3198 unsigned StackAlignment = TFI.getStackAlignment();
3199 uint64_t AlignMask = StackAlignment - 1;
3200 int64_t Offset = StackSize;
3201 unsigned SlotSize = RegInfo->getSlotSize();
3202 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3203 // Number smaller than 12 so just add the difference.
3204 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3206 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3207 Offset = ((~AlignMask) & Offset) + StackAlignment +
3208 (StackAlignment-SlotSize);
3213 /// MatchingStackOffset - Return true if the given stack call argument is
3214 /// already available in the same position (relatively) of the caller's
3215 /// incoming argument stack.
3217 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3218 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3219 const X86InstrInfo *TII) {
3220 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3222 if (Arg.getOpcode() == ISD::CopyFromReg) {
3223 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3224 if (!TargetRegisterInfo::isVirtualRegister(VR))
3226 MachineInstr *Def = MRI->getVRegDef(VR);
3229 if (!Flags.isByVal()) {
3230 if (!TII->isLoadFromStackSlot(Def, FI))
3233 unsigned Opcode = Def->getOpcode();
3234 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3235 Opcode == X86::LEA64_32r) &&
3236 Def->getOperand(1).isFI()) {
3237 FI = Def->getOperand(1).getIndex();
3238 Bytes = Flags.getByValSize();
3242 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3243 if (Flags.isByVal())
3244 // ByVal argument is passed in as a pointer but it's now being
3245 // dereferenced. e.g.
3246 // define @foo(%struct.X* %A) {
3247 // tail call @bar(%struct.X* byval %A)
3250 SDValue Ptr = Ld->getBasePtr();
3251 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3254 FI = FINode->getIndex();
3255 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3256 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3257 FI = FINode->getIndex();
3258 Bytes = Flags.getByValSize();
3262 assert(FI != INT_MAX);
3263 if (!MFI->isFixedObjectIndex(FI))
3265 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3268 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3269 /// for tail call optimization. Targets which want to do tail call
3270 /// optimization should implement this function.
3272 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3273 CallingConv::ID CalleeCC,
3275 bool isCalleeStructRet,
3276 bool isCallerStructRet,
3278 const SmallVectorImpl<ISD::OutputArg> &Outs,
3279 const SmallVectorImpl<SDValue> &OutVals,
3280 const SmallVectorImpl<ISD::InputArg> &Ins,
3281 SelectionDAG &DAG) const {
3282 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3285 // If -tailcallopt is specified, make fastcc functions tail-callable.
3286 const MachineFunction &MF = DAG.getMachineFunction();
3287 const Function *CallerF = MF.getFunction();
3289 // If the function return type is x86_fp80 and the callee return type is not,
3290 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3291 // perform a tailcall optimization here.
3292 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3295 CallingConv::ID CallerCC = CallerF->getCallingConv();
3296 bool CCMatch = CallerCC == CalleeCC;
3297 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3298 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3300 // Win64 functions have extra shadow space for argument homing. Don't do the
3301 // sibcall if the caller and callee have mismatched expectations for this
3303 if (IsCalleeWin64 != IsCallerWin64)
3306 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3307 if (IsTailCallConvention(CalleeCC) && CCMatch)
3312 // Look for obvious safe cases to perform tail call optimization that do not
3313 // require ABI changes. This is what gcc calls sibcall.
3315 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3316 // emit a special epilogue.
3317 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3318 if (RegInfo->needsStackRealignment(MF))
3321 // Also avoid sibcall optimization if either caller or callee uses struct
3322 // return semantics.
3323 if (isCalleeStructRet || isCallerStructRet)
3326 // An stdcall/thiscall caller is expected to clean up its arguments; the
3327 // callee isn't going to do that.
3328 // FIXME: this is more restrictive than needed. We could produce a tailcall
3329 // when the stack adjustment matches. For example, with a thiscall that takes
3330 // only one argument.
3331 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3332 CallerCC == CallingConv::X86_ThisCall))
3335 // Do not sibcall optimize vararg calls unless all arguments are passed via
3337 if (isVarArg && !Outs.empty()) {
3339 // Optimizing for varargs on Win64 is unlikely to be safe without
3340 // additional testing.
3341 if (IsCalleeWin64 || IsCallerWin64)
3344 SmallVector<CCValAssign, 16> ArgLocs;
3345 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3348 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3349 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3350 if (!ArgLocs[i].isRegLoc())
3354 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3355 // stack. Therefore, if it's not used by the call it is not safe to optimize
3356 // this into a sibcall.
3357 bool Unused = false;
3358 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3365 SmallVector<CCValAssign, 16> RVLocs;
3366 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3368 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3369 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3370 CCValAssign &VA = RVLocs[i];
3371 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3376 // If the calling conventions do not match, then we'd better make sure the
3377 // results are returned in the same way as what the caller expects.
3379 SmallVector<CCValAssign, 16> RVLocs1;
3380 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3382 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3384 SmallVector<CCValAssign, 16> RVLocs2;
3385 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3387 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3389 if (RVLocs1.size() != RVLocs2.size())
3391 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3392 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3394 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3396 if (RVLocs1[i].isRegLoc()) {
3397 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3400 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3406 // If the callee takes no arguments then go on to check the results of the
3408 if (!Outs.empty()) {
3409 // Check if stack adjustment is needed. For now, do not do this if any
3410 // argument is passed on the stack.
3411 SmallVector<CCValAssign, 16> ArgLocs;
3412 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3415 // Allocate shadow area for Win64
3417 CCInfo.AllocateStack(32, 8);
3419 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3420 if (CCInfo.getNextStackOffset()) {
3421 MachineFunction &MF = DAG.getMachineFunction();
3422 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3425 // Check if the arguments are already laid out in the right way as
3426 // the caller's fixed stack objects.
3427 MachineFrameInfo *MFI = MF.getFrameInfo();
3428 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3429 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3430 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3431 CCValAssign &VA = ArgLocs[i];
3432 SDValue Arg = OutVals[i];
3433 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3434 if (VA.getLocInfo() == CCValAssign::Indirect)
3436 if (!VA.isRegLoc()) {
3437 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3444 // If the tailcall address may be in a register, then make sure it's
3445 // possible to register allocate for it. In 32-bit, the call address can
3446 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3447 // callee-saved registers are restored. These happen to be the same
3448 // registers used to pass 'inreg' arguments so watch out for those.
3449 if (!Subtarget->is64Bit() &&
3450 ((!isa<GlobalAddressSDNode>(Callee) &&
3451 !isa<ExternalSymbolSDNode>(Callee)) ||
3452 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3453 unsigned NumInRegs = 0;
3454 // In PIC we need an extra register to formulate the address computation
3456 unsigned MaxInRegs =
3457 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3459 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3460 CCValAssign &VA = ArgLocs[i];
3463 unsigned Reg = VA.getLocReg();
3466 case X86::EAX: case X86::EDX: case X86::ECX:
3467 if (++NumInRegs == MaxInRegs)
3479 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3480 const TargetLibraryInfo *libInfo) const {
3481 return X86::createFastISel(funcInfo, libInfo);
3484 //===----------------------------------------------------------------------===//
3485 // Other Lowering Hooks
3486 //===----------------------------------------------------------------------===//
3488 static bool MayFoldLoad(SDValue Op) {
3489 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3492 static bool MayFoldIntoStore(SDValue Op) {
3493 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3496 static bool isTargetShuffle(unsigned Opcode) {
3498 default: return false;
3499 case X86ISD::BLENDI:
3500 case X86ISD::PSHUFB:
3501 case X86ISD::PSHUFD:
3502 case X86ISD::PSHUFHW:
3503 case X86ISD::PSHUFLW:
3505 case X86ISD::PALIGNR:
3506 case X86ISD::MOVLHPS:
3507 case X86ISD::MOVLHPD:
3508 case X86ISD::MOVHLPS:
3509 case X86ISD::MOVLPS:
3510 case X86ISD::MOVLPD:
3511 case X86ISD::MOVSHDUP:
3512 case X86ISD::MOVSLDUP:
3513 case X86ISD::MOVDDUP:
3516 case X86ISD::UNPCKL:
3517 case X86ISD::UNPCKH:
3518 case X86ISD::VPERMILPI:
3519 case X86ISD::VPERM2X128:
3520 case X86ISD::VPERMI:
3525 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3526 SDValue V1, unsigned TargetMask,
3527 SelectionDAG &DAG) {
3529 default: llvm_unreachable("Unknown x86 shuffle node");
3530 case X86ISD::PSHUFD:
3531 case X86ISD::PSHUFHW:
3532 case X86ISD::PSHUFLW:
3533 case X86ISD::VPERMILPI:
3534 case X86ISD::VPERMI:
3535 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3539 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3540 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3542 default: llvm_unreachable("Unknown x86 shuffle node");
3543 case X86ISD::MOVLHPS:
3544 case X86ISD::MOVLHPD:
3545 case X86ISD::MOVHLPS:
3546 case X86ISD::MOVLPS:
3547 case X86ISD::MOVLPD:
3550 case X86ISD::UNPCKL:
3551 case X86ISD::UNPCKH:
3552 return DAG.getNode(Opc, dl, VT, V1, V2);
3556 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3557 MachineFunction &MF = DAG.getMachineFunction();
3558 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3559 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3560 int ReturnAddrIndex = FuncInfo->getRAIndex();
3562 if (ReturnAddrIndex == 0) {
3563 // Set up a frame object for the return address.
3564 unsigned SlotSize = RegInfo->getSlotSize();
3565 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3568 FuncInfo->setRAIndex(ReturnAddrIndex);
3571 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3574 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3575 bool hasSymbolicDisplacement) {
3576 // Offset should fit into 32 bit immediate field.
3577 if (!isInt<32>(Offset))
3580 // If we don't have a symbolic displacement - we don't have any extra
3582 if (!hasSymbolicDisplacement)
3585 // FIXME: Some tweaks might be needed for medium code model.
3586 if (M != CodeModel::Small && M != CodeModel::Kernel)
3589 // For small code model we assume that latest object is 16MB before end of 31
3590 // bits boundary. We may also accept pretty large negative constants knowing
3591 // that all objects are in the positive half of address space.
3592 if (M == CodeModel::Small && Offset < 16*1024*1024)
3595 // For kernel code model we know that all object resist in the negative half
3596 // of 32bits address space. We may not accept negative offsets, since they may
3597 // be just off and we may accept pretty large positive ones.
3598 if (M == CodeModel::Kernel && Offset >= 0)
3604 /// isCalleePop - Determines whether the callee is required to pop its
3605 /// own arguments. Callee pop is necessary to support tail calls.
3606 bool X86::isCalleePop(CallingConv::ID CallingConv,
3607 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3608 switch (CallingConv) {
3611 case CallingConv::X86_StdCall:
3612 case CallingConv::X86_FastCall:
3613 case CallingConv::X86_ThisCall:
3615 case CallingConv::Fast:
3616 case CallingConv::GHC:
3617 case CallingConv::HiPE:
3624 /// \brief Return true if the condition is an unsigned comparison operation.
3625 static bool isX86CCUnsigned(unsigned X86CC) {
3627 default: llvm_unreachable("Invalid integer condition!");
3628 case X86::COND_E: return true;
3629 case X86::COND_G: return false;
3630 case X86::COND_GE: return false;
3631 case X86::COND_L: return false;
3632 case X86::COND_LE: return false;
3633 case X86::COND_NE: return true;
3634 case X86::COND_B: return true;
3635 case X86::COND_A: return true;
3636 case X86::COND_BE: return true;
3637 case X86::COND_AE: return true;
3639 llvm_unreachable("covered switch fell through?!");
3642 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3643 /// specific condition code, returning the condition code and the LHS/RHS of the
3644 /// comparison to make.
3645 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3646 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3648 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3649 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3650 // X > -1 -> X == 0, jump !sign.
3651 RHS = DAG.getConstant(0, RHS.getValueType());
3652 return X86::COND_NS;
3654 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3655 // X < 0 -> X == 0, jump on sign.
3658 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3660 RHS = DAG.getConstant(0, RHS.getValueType());
3661 return X86::COND_LE;
3665 switch (SetCCOpcode) {
3666 default: llvm_unreachable("Invalid integer condition!");
3667 case ISD::SETEQ: return X86::COND_E;
3668 case ISD::SETGT: return X86::COND_G;
3669 case ISD::SETGE: return X86::COND_GE;
3670 case ISD::SETLT: return X86::COND_L;
3671 case ISD::SETLE: return X86::COND_LE;
3672 case ISD::SETNE: return X86::COND_NE;
3673 case ISD::SETULT: return X86::COND_B;
3674 case ISD::SETUGT: return X86::COND_A;
3675 case ISD::SETULE: return X86::COND_BE;
3676 case ISD::SETUGE: return X86::COND_AE;
3680 // First determine if it is required or is profitable to flip the operands.
3682 // If LHS is a foldable load, but RHS is not, flip the condition.
3683 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3684 !ISD::isNON_EXTLoad(RHS.getNode())) {
3685 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3686 std::swap(LHS, RHS);
3689 switch (SetCCOpcode) {
3695 std::swap(LHS, RHS);
3699 // On a floating point condition, the flags are set as follows:
3701 // 0 | 0 | 0 | X > Y
3702 // 0 | 0 | 1 | X < Y
3703 // 1 | 0 | 0 | X == Y
3704 // 1 | 1 | 1 | unordered
3705 switch (SetCCOpcode) {
3706 default: llvm_unreachable("Condcode should be pre-legalized away");
3708 case ISD::SETEQ: return X86::COND_E;
3709 case ISD::SETOLT: // flipped
3711 case ISD::SETGT: return X86::COND_A;
3712 case ISD::SETOLE: // flipped
3714 case ISD::SETGE: return X86::COND_AE;
3715 case ISD::SETUGT: // flipped
3717 case ISD::SETLT: return X86::COND_B;
3718 case ISD::SETUGE: // flipped
3720 case ISD::SETLE: return X86::COND_BE;
3722 case ISD::SETNE: return X86::COND_NE;
3723 case ISD::SETUO: return X86::COND_P;
3724 case ISD::SETO: return X86::COND_NP;
3726 case ISD::SETUNE: return X86::COND_INVALID;
3730 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3731 /// code. Current x86 isa includes the following FP cmov instructions:
3732 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3733 static bool hasFPCMov(unsigned X86CC) {
3749 /// isFPImmLegal - Returns true if the target can instruction select the
3750 /// specified FP immediate natively. If false, the legalizer will
3751 /// materialize the FP immediate as a load from a constant pool.
3752 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3753 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3754 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3760 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3761 ISD::LoadExtType ExtTy,
3763 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3764 // relocation target a movq or addq instruction: don't let the load shrink.
3765 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3766 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3767 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3768 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3772 /// \brief Returns true if it is beneficial to convert a load of a constant
3773 /// to just the constant itself.
3774 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3776 assert(Ty->isIntegerTy());
3778 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3779 if (BitSize == 0 || BitSize > 64)
3784 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3785 unsigned Index) const {
3786 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3789 return (Index == 0 || Index == ResVT.getVectorNumElements());
3792 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3793 // Speculate cttz only if we can directly use TZCNT.
3794 return Subtarget->hasBMI();
3797 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3798 // Speculate ctlz only if we can directly use LZCNT.
3799 return Subtarget->hasLZCNT();
3802 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3803 /// the specified range (L, H].
3804 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3805 return (Val < 0) || (Val >= Low && Val < Hi);
3808 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3809 /// specified value.
3810 static bool isUndefOrEqual(int Val, int CmpVal) {
3811 return (Val < 0 || Val == CmpVal);
3814 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3815 /// from position Pos and ending in Pos+Size, falls within the specified
3816 /// sequential range (Low, Low+Size]. or is undef.
3817 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3818 unsigned Pos, unsigned Size, int Low) {
3819 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3820 if (!isUndefOrEqual(Mask[i], Low))
3825 /// isVEXTRACTIndex - Return true if the specified
3826 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3827 /// suitable for instruction that extract 128 or 256 bit vectors
3828 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3829 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3830 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3833 // The index should be aligned on a vecWidth-bit boundary.
3835 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3837 MVT VT = N->getSimpleValueType(0);
3838 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3839 bool Result = (Index * ElSize) % vecWidth == 0;
3844 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3845 /// operand specifies a subvector insert that is suitable for input to
3846 /// insertion of 128 or 256-bit subvectors
3847 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3848 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3849 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3851 // The index should be aligned on a vecWidth-bit boundary.
3853 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3855 MVT VT = N->getSimpleValueType(0);
3856 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3857 bool Result = (Index * ElSize) % vecWidth == 0;
3862 bool X86::isVINSERT128Index(SDNode *N) {
3863 return isVINSERTIndex(N, 128);
3866 bool X86::isVINSERT256Index(SDNode *N) {
3867 return isVINSERTIndex(N, 256);
3870 bool X86::isVEXTRACT128Index(SDNode *N) {
3871 return isVEXTRACTIndex(N, 128);
3874 bool X86::isVEXTRACT256Index(SDNode *N) {
3875 return isVEXTRACTIndex(N, 256);
3878 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3879 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3880 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3881 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3884 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3886 MVT VecVT = N->getOperand(0).getSimpleValueType();
3887 MVT ElVT = VecVT.getVectorElementType();
3889 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3890 return Index / NumElemsPerChunk;
3893 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3894 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3895 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3896 llvm_unreachable("Illegal insert subvector for VINSERT");
3899 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3901 MVT VecVT = N->getSimpleValueType(0);
3902 MVT ElVT = VecVT.getVectorElementType();
3904 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3905 return Index / NumElemsPerChunk;
3908 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
3909 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3910 /// and VINSERTI128 instructions.
3911 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
3912 return getExtractVEXTRACTImmediate(N, 128);
3915 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
3916 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
3917 /// and VINSERTI64x4 instructions.
3918 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
3919 return getExtractVEXTRACTImmediate(N, 256);
3922 /// getInsertVINSERT128Immediate - Return the appropriate immediate
3923 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3924 /// and VINSERTI128 instructions.
3925 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
3926 return getInsertVINSERTImmediate(N, 128);
3929 /// getInsertVINSERT256Immediate - Return the appropriate immediate
3930 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
3931 /// and VINSERTI64x4 instructions.
3932 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
3933 return getInsertVINSERTImmediate(N, 256);
3936 /// isZero - Returns true if Elt is a constant integer zero
3937 static bool isZero(SDValue V) {
3938 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
3939 return C && C->isNullValue();
3942 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3944 bool X86::isZeroNode(SDValue Elt) {
3947 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
3948 return CFP->getValueAPF().isPosZero();
3952 /// getZeroVector - Returns a vector of specified type with all zero elements.
3954 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
3955 SelectionDAG &DAG, SDLoc dl) {
3956 assert(VT.isVector() && "Expected a vector type");
3958 // Always build SSE zero vectors as <4 x i32> bitcasted
3959 // to their dest type. This ensures they get CSE'd.
3961 if (VT.is128BitVector()) { // SSE
3962 if (Subtarget->hasSSE2()) { // SSE2
3963 SDValue Cst = DAG.getConstant(0, MVT::i32);
3964 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3966 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
3967 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3969 } else if (VT.is256BitVector()) { // AVX
3970 if (Subtarget->hasInt256()) { // AVX2
3971 SDValue Cst = DAG.getConstant(0, MVT::i32);
3972 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3973 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
3975 // 256-bit logic and arithmetic instructions in AVX are all
3976 // floating-point, no support for integer ops. Emit fp zeroed vectors.
3977 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
3978 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3979 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
3981 } else if (VT.is512BitVector()) { // AVX-512
3982 SDValue Cst = DAG.getConstant(0, MVT::i32);
3983 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
3984 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3985 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
3986 } else if (VT.getScalarType() == MVT::i1) {
3988 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
3989 && "Unexpected vector type");
3990 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
3991 && "Unexpected vector type");
3992 SDValue Cst = DAG.getConstant(0, MVT::i1);
3993 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
3994 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3996 llvm_unreachable("Unexpected vector type");
3998 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4001 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4002 SelectionDAG &DAG, SDLoc dl,
4003 unsigned vectorWidth) {
4004 assert((vectorWidth == 128 || vectorWidth == 256) &&
4005 "Unsupported vector width");
4006 EVT VT = Vec.getValueType();
4007 EVT ElVT = VT.getVectorElementType();
4008 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4009 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4010 VT.getVectorNumElements()/Factor);
4012 // Extract from UNDEF is UNDEF.
4013 if (Vec.getOpcode() == ISD::UNDEF)
4014 return DAG.getUNDEF(ResultVT);
4016 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4017 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4019 // This is the index of the first element of the vectorWidth-bit chunk
4021 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4024 // If the input is a buildvector just emit a smaller one.
4025 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4026 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4027 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4030 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
4031 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4034 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4035 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4036 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4037 /// instructions or a simple subregister reference. Idx is an index in the
4038 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4039 /// lowering EXTRACT_VECTOR_ELT operations easier.
4040 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4041 SelectionDAG &DAG, SDLoc dl) {
4042 assert((Vec.getValueType().is256BitVector() ||
4043 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4044 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4047 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4048 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4049 SelectionDAG &DAG, SDLoc dl) {
4050 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4051 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4054 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4055 unsigned IdxVal, SelectionDAG &DAG,
4056 SDLoc dl, unsigned vectorWidth) {
4057 assert((vectorWidth == 128 || vectorWidth == 256) &&
4058 "Unsupported vector width");
4059 // Inserting UNDEF is Result
4060 if (Vec.getOpcode() == ISD::UNDEF)
4062 EVT VT = Vec.getValueType();
4063 EVT ElVT = VT.getVectorElementType();
4064 EVT ResultVT = Result.getValueType();
4066 // Insert the relevant vectorWidth bits.
4067 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4069 // This is the index of the first element of the vectorWidth-bit chunk
4071 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4074 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
4075 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4078 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4079 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4080 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4081 /// simple superregister reference. Idx is an index in the 128 bits
4082 /// we want. It need not be aligned to a 128-bit boundary. That makes
4083 /// lowering INSERT_VECTOR_ELT operations easier.
4084 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4085 SelectionDAG &DAG, SDLoc dl) {
4086 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4088 // For insertion into the zero index (low half) of a 256-bit vector, it is
4089 // more efficient to generate a blend with immediate instead of an insert*128.
4090 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4091 // extend the subvector to the size of the result vector. Make sure that
4092 // we are not recursing on that node by checking for undef here.
4093 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4094 Result.getOpcode() != ISD::UNDEF) {
4095 EVT ResultVT = Result.getValueType();
4096 SDValue ZeroIndex = DAG.getIntPtrConstant(0);
4097 SDValue Undef = DAG.getUNDEF(ResultVT);
4098 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4101 // The blend instruction, and therefore its mask, depend on the data type.
4102 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4103 if (ScalarType.isFloatingPoint()) {
4104 // Choose either vblendps (float) or vblendpd (double).
4105 unsigned ScalarSize = ScalarType.getSizeInBits();
4106 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4107 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4108 SDValue Mask = DAG.getConstant(MaskVal, MVT::i8);
4109 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4112 const X86Subtarget &Subtarget =
4113 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4115 // AVX2 is needed for 256-bit integer blend support.
4116 // Integers must be cast to 32-bit because there is only vpblendd;
4117 // vpblendw can't be used for this because it has a handicapped mask.
4119 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4120 // is still more efficient than using the wrong domain vinsertf128 that
4121 // will be created by InsertSubVector().
4122 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4124 SDValue Mask = DAG.getConstant(0x0f, MVT::i8);
4125 Vec256 = DAG.getNode(ISD::BITCAST, dl, CastVT, Vec256);
4126 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4127 return DAG.getNode(ISD::BITCAST, dl, ResultVT, Vec256);
4130 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4133 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4134 SelectionDAG &DAG, SDLoc dl) {
4135 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4136 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4139 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4140 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4141 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4142 /// large BUILD_VECTORS.
4143 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4144 unsigned NumElems, SelectionDAG &DAG,
4146 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4147 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4150 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4151 unsigned NumElems, SelectionDAG &DAG,
4153 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4154 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4157 /// getOnesVector - Returns a vector of specified type with all bits set.
4158 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4159 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4160 /// Then bitcast to their original type, ensuring they get CSE'd.
4161 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4163 assert(VT.isVector() && "Expected a vector type");
4165 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
4167 if (VT.is256BitVector()) {
4168 if (HasInt256) { // AVX2
4169 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4170 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4172 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4173 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4175 } else if (VT.is128BitVector()) {
4176 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4178 llvm_unreachable("Unexpected vector type");
4180 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4183 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4184 /// operation of specified width.
4185 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4187 unsigned NumElems = VT.getVectorNumElements();
4188 SmallVector<int, 8> Mask;
4189 Mask.push_back(NumElems);
4190 for (unsigned i = 1; i != NumElems; ++i)
4192 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4195 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4196 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4198 unsigned NumElems = VT.getVectorNumElements();
4199 SmallVector<int, 8> Mask;
4200 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4202 Mask.push_back(i + NumElems);
4204 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4207 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4208 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4210 unsigned NumElems = VT.getVectorNumElements();
4211 SmallVector<int, 8> Mask;
4212 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4213 Mask.push_back(i + Half);
4214 Mask.push_back(i + NumElems + Half);
4216 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4219 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4220 /// vector of zero or undef vector. This produces a shuffle where the low
4221 /// element of V2 is swizzled into the zero/undef vector, landing at element
4222 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4223 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4225 const X86Subtarget *Subtarget,
4226 SelectionDAG &DAG) {
4227 MVT VT = V2.getSimpleValueType();
4229 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4230 unsigned NumElems = VT.getVectorNumElements();
4231 SmallVector<int, 16> MaskVec;
4232 for (unsigned i = 0; i != NumElems; ++i)
4233 // If this is the insertion idx, put the low elt of V2 here.
4234 MaskVec.push_back(i == Idx ? NumElems : i);
4235 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4238 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4239 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4240 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4241 /// shuffles which use a single input multiple times, and in those cases it will
4242 /// adjust the mask to only have indices within that single input.
4243 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4244 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4245 unsigned NumElems = VT.getVectorNumElements();
4249 bool IsFakeUnary = false;
4250 switch(N->getOpcode()) {
4251 case X86ISD::BLENDI:
4252 ImmN = N->getOperand(N->getNumOperands()-1);
4253 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4256 ImmN = N->getOperand(N->getNumOperands()-1);
4257 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4258 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4260 case X86ISD::UNPCKH:
4261 DecodeUNPCKHMask(VT, Mask);
4262 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4264 case X86ISD::UNPCKL:
4265 DecodeUNPCKLMask(VT, Mask);
4266 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4268 case X86ISD::MOVHLPS:
4269 DecodeMOVHLPSMask(NumElems, Mask);
4270 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4272 case X86ISD::MOVLHPS:
4273 DecodeMOVLHPSMask(NumElems, Mask);
4274 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4276 case X86ISD::PALIGNR:
4277 ImmN = N->getOperand(N->getNumOperands()-1);
4278 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4280 case X86ISD::PSHUFD:
4281 case X86ISD::VPERMILPI:
4282 ImmN = N->getOperand(N->getNumOperands()-1);
4283 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4286 case X86ISD::PSHUFHW:
4287 ImmN = N->getOperand(N->getNumOperands()-1);
4288 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4291 case X86ISD::PSHUFLW:
4292 ImmN = N->getOperand(N->getNumOperands()-1);
4293 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4296 case X86ISD::PSHUFB: {
4298 SDValue MaskNode = N->getOperand(1);
4299 while (MaskNode->getOpcode() == ISD::BITCAST)
4300 MaskNode = MaskNode->getOperand(0);
4302 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4303 // If we have a build-vector, then things are easy.
4304 EVT VT = MaskNode.getValueType();
4305 assert(VT.isVector() &&
4306 "Can't produce a non-vector with a build_vector!");
4307 if (!VT.isInteger())
4310 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4312 SmallVector<uint64_t, 32> RawMask;
4313 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4314 SDValue Op = MaskNode->getOperand(i);
4315 if (Op->getOpcode() == ISD::UNDEF) {
4316 RawMask.push_back((uint64_t)SM_SentinelUndef);
4319 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4322 APInt MaskElement = CN->getAPIntValue();
4324 // We now have to decode the element which could be any integer size and
4325 // extract each byte of it.
4326 for (int j = 0; j < NumBytesPerElement; ++j) {
4327 // Note that this is x86 and so always little endian: the low byte is
4328 // the first byte of the mask.
4329 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4330 MaskElement = MaskElement.lshr(8);
4333 DecodePSHUFBMask(RawMask, Mask);
4337 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4341 SDValue Ptr = MaskLoad->getBasePtr();
4342 if (Ptr->getOpcode() == X86ISD::Wrapper)
4343 Ptr = Ptr->getOperand(0);
4345 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4346 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4349 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4350 DecodePSHUFBMask(C, Mask);
4358 case X86ISD::VPERMI:
4359 ImmN = N->getOperand(N->getNumOperands()-1);
4360 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4365 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4367 case X86ISD::VPERM2X128:
4368 ImmN = N->getOperand(N->getNumOperands()-1);
4369 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4370 if (Mask.empty()) return false;
4372 case X86ISD::MOVSLDUP:
4373 DecodeMOVSLDUPMask(VT, Mask);
4376 case X86ISD::MOVSHDUP:
4377 DecodeMOVSHDUPMask(VT, Mask);
4380 case X86ISD::MOVDDUP:
4381 DecodeMOVDDUPMask(VT, Mask);
4384 case X86ISD::MOVLHPD:
4385 case X86ISD::MOVLPD:
4386 case X86ISD::MOVLPS:
4387 // Not yet implemented
4389 default: llvm_unreachable("unknown target shuffle node");
4392 // If we have a fake unary shuffle, the shuffle mask is spread across two
4393 // inputs that are actually the same node. Re-map the mask to always point
4394 // into the first input.
4397 if (M >= (int)Mask.size())
4403 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4404 /// element of the result of the vector shuffle.
4405 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4408 return SDValue(); // Limit search depth.
4410 SDValue V = SDValue(N, 0);
4411 EVT VT = V.getValueType();
4412 unsigned Opcode = V.getOpcode();
4414 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4415 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4416 int Elt = SV->getMaskElt(Index);
4419 return DAG.getUNDEF(VT.getVectorElementType());
4421 unsigned NumElems = VT.getVectorNumElements();
4422 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4423 : SV->getOperand(1);
4424 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4427 // Recurse into target specific vector shuffles to find scalars.
4428 if (isTargetShuffle(Opcode)) {
4429 MVT ShufVT = V.getSimpleValueType();
4430 unsigned NumElems = ShufVT.getVectorNumElements();
4431 SmallVector<int, 16> ShuffleMask;
4434 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4437 int Elt = ShuffleMask[Index];
4439 return DAG.getUNDEF(ShufVT.getVectorElementType());
4441 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4443 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4447 // Actual nodes that may contain scalar elements
4448 if (Opcode == ISD::BITCAST) {
4449 V = V.getOperand(0);
4450 EVT SrcVT = V.getValueType();
4451 unsigned NumElems = VT.getVectorNumElements();
4453 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4457 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4458 return (Index == 0) ? V.getOperand(0)
4459 : DAG.getUNDEF(VT.getVectorElementType());
4461 if (V.getOpcode() == ISD::BUILD_VECTOR)
4462 return V.getOperand(Index);
4467 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4469 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4470 unsigned NumNonZero, unsigned NumZero,
4472 const X86Subtarget* Subtarget,
4473 const TargetLowering &TLI) {
4481 // SSE4.1 - use PINSRB to insert each byte directly.
4482 if (Subtarget->hasSSE41()) {
4483 for (unsigned i = 0; i < 16; ++i) {
4484 bool isNonZero = (NonZeros & (1 << i)) != 0;
4488 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4490 V = DAG.getUNDEF(MVT::v16i8);
4493 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4494 MVT::v16i8, V, Op.getOperand(i),
4495 DAG.getIntPtrConstant(i));
4502 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4503 for (unsigned i = 0; i < 16; ++i) {
4504 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4505 if (ThisIsNonZero && First) {
4507 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4509 V = DAG.getUNDEF(MVT::v8i16);
4514 SDValue ThisElt, LastElt;
4515 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4516 if (LastIsNonZero) {
4517 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4518 MVT::i16, Op.getOperand(i-1));
4520 if (ThisIsNonZero) {
4521 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4522 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4523 ThisElt, DAG.getConstant(8, MVT::i8));
4525 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4529 if (ThisElt.getNode())
4530 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4531 DAG.getIntPtrConstant(i/2));
4535 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4538 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4540 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4541 unsigned NumNonZero, unsigned NumZero,
4543 const X86Subtarget* Subtarget,
4544 const TargetLowering &TLI) {
4551 for (unsigned i = 0; i < 8; ++i) {
4552 bool isNonZero = (NonZeros & (1 << i)) != 0;
4556 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4558 V = DAG.getUNDEF(MVT::v8i16);
4561 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4562 MVT::v8i16, V, Op.getOperand(i),
4563 DAG.getIntPtrConstant(i));
4570 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4571 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4572 const X86Subtarget *Subtarget,
4573 const TargetLowering &TLI) {
4574 // Find all zeroable elements.
4575 std::bitset<4> Zeroable;
4576 for (int i=0; i < 4; ++i) {
4577 SDValue Elt = Op->getOperand(i);
4578 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4580 assert(Zeroable.size() - Zeroable.count() > 1 &&
4581 "We expect at least two non-zero elements!");
4583 // We only know how to deal with build_vector nodes where elements are either
4584 // zeroable or extract_vector_elt with constant index.
4585 SDValue FirstNonZero;
4586 unsigned FirstNonZeroIdx;
4587 for (unsigned i=0; i < 4; ++i) {
4590 SDValue Elt = Op->getOperand(i);
4591 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4592 !isa<ConstantSDNode>(Elt.getOperand(1)))
4594 // Make sure that this node is extracting from a 128-bit vector.
4595 MVT VT = Elt.getOperand(0).getSimpleValueType();
4596 if (!VT.is128BitVector())
4598 if (!FirstNonZero.getNode()) {
4600 FirstNonZeroIdx = i;
4604 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4605 SDValue V1 = FirstNonZero.getOperand(0);
4606 MVT VT = V1.getSimpleValueType();
4608 // See if this build_vector can be lowered as a blend with zero.
4610 unsigned EltMaskIdx, EltIdx;
4612 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4613 if (Zeroable[EltIdx]) {
4614 // The zero vector will be on the right hand side.
4615 Mask[EltIdx] = EltIdx+4;
4619 Elt = Op->getOperand(EltIdx);
4620 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4621 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4622 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4624 Mask[EltIdx] = EltIdx;
4628 // Let the shuffle legalizer deal with blend operations.
4629 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4630 if (V1.getSimpleValueType() != VT)
4631 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4632 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4635 // See if we can lower this build_vector to a INSERTPS.
4636 if (!Subtarget->hasSSE41())
4639 SDValue V2 = Elt.getOperand(0);
4640 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4643 bool CanFold = true;
4644 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4648 SDValue Current = Op->getOperand(i);
4649 SDValue SrcVector = Current->getOperand(0);
4652 CanFold = SrcVector == V1 &&
4653 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4659 assert(V1.getNode() && "Expected at least two non-zero elements!");
4660 if (V1.getSimpleValueType() != MVT::v4f32)
4661 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4662 if (V2.getSimpleValueType() != MVT::v4f32)
4663 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4665 // Ok, we can emit an INSERTPS instruction.
4666 unsigned ZMask = Zeroable.to_ulong();
4668 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4669 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4670 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
4671 DAG.getIntPtrConstant(InsertPSMask));
4672 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
4675 /// Return a vector logical shift node.
4676 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4677 unsigned NumBits, SelectionDAG &DAG,
4678 const TargetLowering &TLI, SDLoc dl) {
4679 assert(VT.is128BitVector() && "Unknown type for VShift");
4680 MVT ShVT = MVT::v2i64;
4681 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4682 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4683 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4684 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4685 SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
4686 return DAG.getNode(ISD::BITCAST, dl, VT,
4687 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4691 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4693 // Check if the scalar load can be widened into a vector load. And if
4694 // the address is "base + cst" see if the cst can be "absorbed" into
4695 // the shuffle mask.
4696 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4697 SDValue Ptr = LD->getBasePtr();
4698 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4700 EVT PVT = LD->getValueType(0);
4701 if (PVT != MVT::i32 && PVT != MVT::f32)
4706 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4707 FI = FINode->getIndex();
4709 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4710 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4711 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4712 Offset = Ptr.getConstantOperandVal(1);
4713 Ptr = Ptr.getOperand(0);
4718 // FIXME: 256-bit vector instructions don't require a strict alignment,
4719 // improve this code to support it better.
4720 unsigned RequiredAlign = VT.getSizeInBits()/8;
4721 SDValue Chain = LD->getChain();
4722 // Make sure the stack object alignment is at least 16 or 32.
4723 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4724 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4725 if (MFI->isFixedObjectIndex(FI)) {
4726 // Can't change the alignment. FIXME: It's possible to compute
4727 // the exact stack offset and reference FI + adjust offset instead.
4728 // If someone *really* cares about this. That's the way to implement it.
4731 MFI->setObjectAlignment(FI, RequiredAlign);
4735 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4736 // Ptr + (Offset & ~15).
4739 if ((Offset % RequiredAlign) & 3)
4741 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4743 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
4744 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4746 int EltNo = (Offset - StartOffset) >> 2;
4747 unsigned NumElems = VT.getVectorNumElements();
4749 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4750 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4751 LD->getPointerInfo().getWithOffset(StartOffset),
4752 false, false, false, 0);
4754 SmallVector<int, 8> Mask(NumElems, EltNo);
4756 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4762 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4763 /// elements can be replaced by a single large load which has the same value as
4764 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4766 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4768 /// FIXME: we'd also like to handle the case where the last elements are zero
4769 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4770 /// There's even a handy isZeroNode for that purpose.
4771 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4772 SDLoc &DL, SelectionDAG &DAG,
4773 bool isAfterLegalize) {
4774 unsigned NumElems = Elts.size();
4776 LoadSDNode *LDBase = nullptr;
4777 unsigned LastLoadedElt = -1U;
4779 // For each element in the initializer, see if we've found a load or an undef.
4780 // If we don't find an initial load element, or later load elements are
4781 // non-consecutive, bail out.
4782 for (unsigned i = 0; i < NumElems; ++i) {
4783 SDValue Elt = Elts[i];
4784 // Look through a bitcast.
4785 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4786 Elt = Elt.getOperand(0);
4787 if (!Elt.getNode() ||
4788 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4791 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4793 LDBase = cast<LoadSDNode>(Elt.getNode());
4797 if (Elt.getOpcode() == ISD::UNDEF)
4800 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4801 EVT LdVT = Elt.getValueType();
4802 // Each loaded element must be the correct fractional portion of the
4803 // requested vector load.
4804 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4806 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4811 // If we have found an entire vector of loads and undefs, then return a large
4812 // load of the entire vector width starting at the base pointer. If we found
4813 // consecutive loads for the low half, generate a vzext_load node.
4814 if (LastLoadedElt == NumElems - 1) {
4815 assert(LDBase && "Did not find base load for merging consecutive loads");
4816 EVT EltVT = LDBase->getValueType(0);
4817 // Ensure that the input vector size for the merged loads matches the
4818 // cumulative size of the input elements.
4819 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4822 if (isAfterLegalize &&
4823 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4826 SDValue NewLd = SDValue();
4828 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4829 LDBase->getPointerInfo(), LDBase->isVolatile(),
4830 LDBase->isNonTemporal(), LDBase->isInvariant(),
4831 LDBase->getAlignment());
4833 if (LDBase->hasAnyUseOfValue(1)) {
4834 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4836 SDValue(NewLd.getNode(), 1));
4837 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4838 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4839 SDValue(NewLd.getNode(), 1));
4845 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4846 //of a v4i32 / v4f32. It's probably worth generalizing.
4847 EVT EltVT = VT.getVectorElementType();
4848 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4849 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4850 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4851 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4853 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4854 LDBase->getPointerInfo(),
4855 LDBase->getAlignment(),
4856 false/*isVolatile*/, true/*ReadMem*/,
4859 // Make sure the newly-created LOAD is in the same position as LDBase in
4860 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4861 // update uses of LDBase's output chain to use the TokenFactor.
4862 if (LDBase->hasAnyUseOfValue(1)) {
4863 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4864 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4865 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4866 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4867 SDValue(ResNode.getNode(), 1));
4870 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4875 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4876 /// to generate a splat value for the following cases:
4877 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4878 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4879 /// a scalar load, or a constant.
4880 /// The VBROADCAST node is returned when a pattern is found,
4881 /// or SDValue() otherwise.
4882 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4883 SelectionDAG &DAG) {
4884 // VBROADCAST requires AVX.
4885 // TODO: Splats could be generated for non-AVX CPUs using SSE
4886 // instructions, but there's less potential gain for only 128-bit vectors.
4887 if (!Subtarget->hasAVX())
4890 MVT VT = Op.getSimpleValueType();
4893 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4894 "Unsupported vector type for broadcast.");
4899 switch (Op.getOpcode()) {
4901 // Unknown pattern found.
4904 case ISD::BUILD_VECTOR: {
4905 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4906 BitVector UndefElements;
4907 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4909 // We need a splat of a single value to use broadcast, and it doesn't
4910 // make any sense if the value is only in one element of the vector.
4911 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4915 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4916 Ld.getOpcode() == ISD::ConstantFP);
4918 // Make sure that all of the users of a non-constant load are from the
4919 // BUILD_VECTOR node.
4920 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4925 case ISD::VECTOR_SHUFFLE: {
4926 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4928 // Shuffles must have a splat mask where the first element is
4930 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4933 SDValue Sc = Op.getOperand(0);
4934 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4935 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4937 if (!Subtarget->hasInt256())
4940 // Use the register form of the broadcast instruction available on AVX2.
4941 if (VT.getSizeInBits() >= 256)
4942 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4943 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
4946 Ld = Sc.getOperand(0);
4947 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4948 Ld.getOpcode() == ISD::ConstantFP);
4950 // The scalar_to_vector node and the suspected
4951 // load node must have exactly one user.
4952 // Constants may have multiple users.
4954 // AVX-512 has register version of the broadcast
4955 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
4956 Ld.getValueType().getSizeInBits() >= 32;
4957 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
4964 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
4965 bool IsGE256 = (VT.getSizeInBits() >= 256);
4967 // When optimizing for size, generate up to 5 extra bytes for a broadcast
4968 // instruction to save 8 or more bytes of constant pool data.
4969 // TODO: If multiple splats are generated to load the same constant,
4970 // it may be detrimental to overall size. There needs to be a way to detect
4971 // that condition to know if this is truly a size win.
4972 const Function *F = DAG.getMachineFunction().getFunction();
4973 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
4975 // Handle broadcasting a single constant scalar from the constant pool
4977 // On Sandybridge (no AVX2), it is still better to load a constant vector
4978 // from the constant pool and not to broadcast it from a scalar.
4979 // But override that restriction when optimizing for size.
4980 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
4981 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
4982 EVT CVT = Ld.getValueType();
4983 assert(!CVT.isVector() && "Must not broadcast a vector type");
4985 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
4986 // For size optimization, also splat v2f64 and v2i64, and for size opt
4987 // with AVX2, also splat i8 and i16.
4988 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
4989 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4990 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
4991 const Constant *C = nullptr;
4992 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
4993 C = CI->getConstantIntValue();
4994 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
4995 C = CF->getConstantFPValue();
4997 assert(C && "Invalid constant type");
4999 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5000 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5001 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5002 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5003 MachinePointerInfo::getConstantPool(),
5004 false, false, false, Alignment);
5006 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5010 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5012 // Handle AVX2 in-register broadcasts.
5013 if (!IsLoad && Subtarget->hasInt256() &&
5014 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5015 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5017 // The scalar source must be a normal load.
5021 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5022 (Subtarget->hasVLX() && ScalarSize == 64))
5023 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5025 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5026 // double since there is no vbroadcastsd xmm
5027 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5028 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5029 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5032 // Unsupported broadcast.
5036 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5037 /// underlying vector and index.
5039 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5041 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5043 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5044 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5047 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5049 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5051 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5052 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5055 // In this case the vector is the extract_subvector expression and the index
5056 // is 2, as specified by the shuffle.
5057 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5058 SDValue ShuffleVec = SVOp->getOperand(0);
5059 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5060 assert(ShuffleVecVT.getVectorElementType() ==
5061 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5063 int ShuffleIdx = SVOp->getMaskElt(Idx);
5064 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5065 ExtractedFromVec = ShuffleVec;
5071 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5072 MVT VT = Op.getSimpleValueType();
5074 // Skip if insert_vec_elt is not supported.
5075 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5076 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5080 unsigned NumElems = Op.getNumOperands();
5084 SmallVector<unsigned, 4> InsertIndices;
5085 SmallVector<int, 8> Mask(NumElems, -1);
5087 for (unsigned i = 0; i != NumElems; ++i) {
5088 unsigned Opc = Op.getOperand(i).getOpcode();
5090 if (Opc == ISD::UNDEF)
5093 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5094 // Quit if more than 1 elements need inserting.
5095 if (InsertIndices.size() > 1)
5098 InsertIndices.push_back(i);
5102 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5103 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5104 // Quit if non-constant index.
5105 if (!isa<ConstantSDNode>(ExtIdx))
5107 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5109 // Quit if extracted from vector of different type.
5110 if (ExtractedFromVec.getValueType() != VT)
5113 if (!VecIn1.getNode())
5114 VecIn1 = ExtractedFromVec;
5115 else if (VecIn1 != ExtractedFromVec) {
5116 if (!VecIn2.getNode())
5117 VecIn2 = ExtractedFromVec;
5118 else if (VecIn2 != ExtractedFromVec)
5119 // Quit if more than 2 vectors to shuffle
5123 if (ExtractedFromVec == VecIn1)
5125 else if (ExtractedFromVec == VecIn2)
5126 Mask[i] = Idx + NumElems;
5129 if (!VecIn1.getNode())
5132 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5133 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5134 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5135 unsigned Idx = InsertIndices[i];
5136 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5137 DAG.getIntPtrConstant(Idx));
5143 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5145 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5147 MVT VT = Op.getSimpleValueType();
5148 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5149 "Unexpected type in LowerBUILD_VECTORvXi1!");
5152 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5153 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5154 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5155 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5158 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5159 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5160 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5161 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5164 bool AllContants = true;
5165 uint64_t Immediate = 0;
5166 int NonConstIdx = -1;
5167 bool IsSplat = true;
5168 unsigned NumNonConsts = 0;
5169 unsigned NumConsts = 0;
5170 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5171 SDValue In = Op.getOperand(idx);
5172 if (In.getOpcode() == ISD::UNDEF)
5174 if (!isa<ConstantSDNode>(In)) {
5175 AllContants = false;
5180 if (cast<ConstantSDNode>(In)->getZExtValue())
5181 Immediate |= (1ULL << idx);
5183 if (In != Op.getOperand(0))
5188 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5189 DAG.getConstant(Immediate, MVT::i16));
5190 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5191 DAG.getIntPtrConstant(0));
5194 if (NumNonConsts == 1 && NonConstIdx != 0) {
5197 SDValue VecAsImm = DAG.getConstant(Immediate,
5198 MVT::getIntegerVT(VT.getSizeInBits()));
5199 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
5202 DstVec = DAG.getUNDEF(VT);
5203 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5204 Op.getOperand(NonConstIdx),
5205 DAG.getIntPtrConstant(NonConstIdx));
5207 if (!IsSplat && (NonConstIdx != 0))
5208 llvm_unreachable("Unsupported BUILD_VECTOR operation");
5209 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
5212 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5213 DAG.getConstant(-1, SelectVT),
5214 DAG.getConstant(0, SelectVT));
5216 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5217 DAG.getConstant((Immediate | 1), SelectVT),
5218 DAG.getConstant(Immediate, SelectVT));
5219 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
5222 /// \brief Return true if \p N implements a horizontal binop and return the
5223 /// operands for the horizontal binop into V0 and V1.
5225 /// This is a helper function of LowerToHorizontalOp().
5226 /// This function checks that the build_vector \p N in input implements a
5227 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5228 /// operation to match.
5229 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5230 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5231 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5234 /// This function only analyzes elements of \p N whose indices are
5235 /// in range [BaseIdx, LastIdx).
5236 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5238 unsigned BaseIdx, unsigned LastIdx,
5239 SDValue &V0, SDValue &V1) {
5240 EVT VT = N->getValueType(0);
5242 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5243 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5244 "Invalid Vector in input!");
5246 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5247 bool CanFold = true;
5248 unsigned ExpectedVExtractIdx = BaseIdx;
5249 unsigned NumElts = LastIdx - BaseIdx;
5250 V0 = DAG.getUNDEF(VT);
5251 V1 = DAG.getUNDEF(VT);
5253 // Check if N implements a horizontal binop.
5254 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5255 SDValue Op = N->getOperand(i + BaseIdx);
5258 if (Op->getOpcode() == ISD::UNDEF) {
5259 // Update the expected vector extract index.
5260 if (i * 2 == NumElts)
5261 ExpectedVExtractIdx = BaseIdx;
5262 ExpectedVExtractIdx += 2;
5266 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5271 SDValue Op0 = Op.getOperand(0);
5272 SDValue Op1 = Op.getOperand(1);
5274 // Try to match the following pattern:
5275 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5276 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5277 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5278 Op0.getOperand(0) == Op1.getOperand(0) &&
5279 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5280 isa<ConstantSDNode>(Op1.getOperand(1)));
5284 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5285 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5287 if (i * 2 < NumElts) {
5288 if (V0.getOpcode() == ISD::UNDEF) {
5289 V0 = Op0.getOperand(0);
5290 if (V0.getValueType() != VT)
5294 if (V1.getOpcode() == ISD::UNDEF) {
5295 V1 = Op0.getOperand(0);
5296 if (V1.getValueType() != VT)
5299 if (i * 2 == NumElts)
5300 ExpectedVExtractIdx = BaseIdx;
5303 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5304 if (I0 == ExpectedVExtractIdx)
5305 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5306 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5307 // Try to match the following dag sequence:
5308 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5309 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5313 ExpectedVExtractIdx += 2;
5319 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5320 /// a concat_vector.
5322 /// This is a helper function of LowerToHorizontalOp().
5323 /// This function expects two 256-bit vectors called V0 and V1.
5324 /// At first, each vector is split into two separate 128-bit vectors.
5325 /// Then, the resulting 128-bit vectors are used to implement two
5326 /// horizontal binary operations.
5328 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5330 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5331 /// the two new horizontal binop.
5332 /// When Mode is set, the first horizontal binop dag node would take as input
5333 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5334 /// horizontal binop dag node would take as input the lower 128-bit of V1
5335 /// and the upper 128-bit of V1.
5337 /// HADD V0_LO, V0_HI
5338 /// HADD V1_LO, V1_HI
5340 /// Otherwise, the first horizontal binop dag node takes as input the lower
5341 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5342 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5344 /// HADD V0_LO, V1_LO
5345 /// HADD V0_HI, V1_HI
5347 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5348 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5349 /// the upper 128-bits of the result.
5350 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5351 SDLoc DL, SelectionDAG &DAG,
5352 unsigned X86Opcode, bool Mode,
5353 bool isUndefLO, bool isUndefHI) {
5354 EVT VT = V0.getValueType();
5355 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5356 "Invalid nodes in input!");
5358 unsigned NumElts = VT.getVectorNumElements();
5359 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5360 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5361 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5362 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5363 EVT NewVT = V0_LO.getValueType();
5365 SDValue LO = DAG.getUNDEF(NewVT);
5366 SDValue HI = DAG.getUNDEF(NewVT);
5369 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5370 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5371 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5372 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5373 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5375 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5376 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5377 V1_LO->getOpcode() != ISD::UNDEF))
5378 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5380 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5381 V1_HI->getOpcode() != ISD::UNDEF))
5382 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5385 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5388 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5390 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5391 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5392 EVT VT = BV->getValueType(0);
5393 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5394 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5398 unsigned NumElts = VT.getVectorNumElements();
5399 SDValue InVec0 = DAG.getUNDEF(VT);
5400 SDValue InVec1 = DAG.getUNDEF(VT);
5402 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5403 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5405 // Odd-numbered elements in the input build vector are obtained from
5406 // adding two integer/float elements.
5407 // Even-numbered elements in the input build vector are obtained from
5408 // subtracting two integer/float elements.
5409 unsigned ExpectedOpcode = ISD::FSUB;
5410 unsigned NextExpectedOpcode = ISD::FADD;
5411 bool AddFound = false;
5412 bool SubFound = false;
5414 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5415 SDValue Op = BV->getOperand(i);
5417 // Skip 'undef' values.
5418 unsigned Opcode = Op.getOpcode();
5419 if (Opcode == ISD::UNDEF) {
5420 std::swap(ExpectedOpcode, NextExpectedOpcode);
5424 // Early exit if we found an unexpected opcode.
5425 if (Opcode != ExpectedOpcode)
5428 SDValue Op0 = Op.getOperand(0);
5429 SDValue Op1 = Op.getOperand(1);
5431 // Try to match the following pattern:
5432 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5433 // Early exit if we cannot match that sequence.
5434 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5435 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5436 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5437 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5438 Op0.getOperand(1) != Op1.getOperand(1))
5441 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5445 // We found a valid add/sub node. Update the information accordingly.
5451 // Update InVec0 and InVec1.
5452 if (InVec0.getOpcode() == ISD::UNDEF) {
5453 InVec0 = Op0.getOperand(0);
5454 if (InVec0.getValueType() != VT)
5457 if (InVec1.getOpcode() == ISD::UNDEF) {
5458 InVec1 = Op1.getOperand(0);
5459 if (InVec1.getValueType() != VT)
5463 // Make sure that operands in input to each add/sub node always
5464 // come from a same pair of vectors.
5465 if (InVec0 != Op0.getOperand(0)) {
5466 if (ExpectedOpcode == ISD::FSUB)
5469 // FADD is commutable. Try to commute the operands
5470 // and then test again.
5471 std::swap(Op0, Op1);
5472 if (InVec0 != Op0.getOperand(0))
5476 if (InVec1 != Op1.getOperand(0))
5479 // Update the pair of expected opcodes.
5480 std::swap(ExpectedOpcode, NextExpectedOpcode);
5483 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5484 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5485 InVec1.getOpcode() != ISD::UNDEF)
5486 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5491 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5492 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5493 const X86Subtarget *Subtarget,
5494 SelectionDAG &DAG) {
5495 EVT VT = BV->getValueType(0);
5496 unsigned NumElts = VT.getVectorNumElements();
5497 unsigned NumUndefsLO = 0;
5498 unsigned NumUndefsHI = 0;
5499 unsigned Half = NumElts/2;
5501 // Count the number of UNDEF operands in the build_vector in input.
5502 for (unsigned i = 0, e = Half; i != e; ++i)
5503 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5506 for (unsigned i = Half, e = NumElts; i != e; ++i)
5507 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5510 // Early exit if this is either a build_vector of all UNDEFs or all the
5511 // operands but one are UNDEF.
5512 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5516 SDValue InVec0, InVec1;
5517 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5518 // Try to match an SSE3 float HADD/HSUB.
5519 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5520 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5522 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5523 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5524 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5525 // Try to match an SSSE3 integer HADD/HSUB.
5526 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5527 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5529 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5530 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5533 if (!Subtarget->hasAVX())
5536 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5537 // Try to match an AVX horizontal add/sub of packed single/double
5538 // precision floating point values from 256-bit vectors.
5539 SDValue InVec2, InVec3;
5540 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5541 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5542 ((InVec0.getOpcode() == ISD::UNDEF ||
5543 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5544 ((InVec1.getOpcode() == ISD::UNDEF ||
5545 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5546 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5548 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5549 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5550 ((InVec0.getOpcode() == ISD::UNDEF ||
5551 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5552 ((InVec1.getOpcode() == ISD::UNDEF ||
5553 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5554 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5555 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5556 // Try to match an AVX2 horizontal add/sub of signed integers.
5557 SDValue InVec2, InVec3;
5559 bool CanFold = true;
5561 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5562 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5563 ((InVec0.getOpcode() == ISD::UNDEF ||
5564 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5565 ((InVec1.getOpcode() == ISD::UNDEF ||
5566 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5567 X86Opcode = X86ISD::HADD;
5568 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5569 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5570 ((InVec0.getOpcode() == ISD::UNDEF ||
5571 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5572 ((InVec1.getOpcode() == ISD::UNDEF ||
5573 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5574 X86Opcode = X86ISD::HSUB;
5579 // Fold this build_vector into a single horizontal add/sub.
5580 // Do this only if the target has AVX2.
5581 if (Subtarget->hasAVX2())
5582 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5584 // Do not try to expand this build_vector into a pair of horizontal
5585 // add/sub if we can emit a pair of scalar add/sub.
5586 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5589 // Convert this build_vector into a pair of horizontal binop followed by
5591 bool isUndefLO = NumUndefsLO == Half;
5592 bool isUndefHI = NumUndefsHI == Half;
5593 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5594 isUndefLO, isUndefHI);
5598 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5599 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5601 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5602 X86Opcode = X86ISD::HADD;
5603 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5604 X86Opcode = X86ISD::HSUB;
5605 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5606 X86Opcode = X86ISD::FHADD;
5607 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5608 X86Opcode = X86ISD::FHSUB;
5612 // Don't try to expand this build_vector into a pair of horizontal add/sub
5613 // if we can simply emit a pair of scalar add/sub.
5614 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5617 // Convert this build_vector into two horizontal add/sub followed by
5619 bool isUndefLO = NumUndefsLO == Half;
5620 bool isUndefHI = NumUndefsHI == Half;
5621 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5622 isUndefLO, isUndefHI);
5629 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5632 MVT VT = Op.getSimpleValueType();
5633 MVT ExtVT = VT.getVectorElementType();
5634 unsigned NumElems = Op.getNumOperands();
5636 // Generate vectors for predicate vectors.
5637 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5638 return LowerBUILD_VECTORvXi1(Op, DAG);
5640 // Vectors containing all zeros can be matched by pxor and xorps later
5641 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5642 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5643 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5644 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5647 return getZeroVector(VT, Subtarget, DAG, dl);
5650 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5651 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5652 // vpcmpeqd on 256-bit vectors.
5653 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5654 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5657 if (!VT.is512BitVector())
5658 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5661 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5662 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5664 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5665 return HorizontalOp;
5666 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5669 unsigned EVTBits = ExtVT.getSizeInBits();
5671 unsigned NumZero = 0;
5672 unsigned NumNonZero = 0;
5673 unsigned NonZeros = 0;
5674 bool IsAllConstants = true;
5675 SmallSet<SDValue, 8> Values;
5676 for (unsigned i = 0; i < NumElems; ++i) {
5677 SDValue Elt = Op.getOperand(i);
5678 if (Elt.getOpcode() == ISD::UNDEF)
5681 if (Elt.getOpcode() != ISD::Constant &&
5682 Elt.getOpcode() != ISD::ConstantFP)
5683 IsAllConstants = false;
5684 if (X86::isZeroNode(Elt))
5687 NonZeros |= (1 << i);
5692 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5693 if (NumNonZero == 0)
5694 return DAG.getUNDEF(VT);
5696 // Special case for single non-zero, non-undef, element.
5697 if (NumNonZero == 1) {
5698 unsigned Idx = countTrailingZeros(NonZeros);
5699 SDValue Item = Op.getOperand(Idx);
5701 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5702 // the value are obviously zero, truncate the value to i32 and do the
5703 // insertion that way. Only do this if the value is non-constant or if the
5704 // value is a constant being inserted into element 0. It is cheaper to do
5705 // a constant pool load than it is to do a movd + shuffle.
5706 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5707 (!IsAllConstants || Idx == 0)) {
5708 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5710 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5711 EVT VecVT = MVT::v4i32;
5713 // Truncate the value (which may itself be a constant) to i32, and
5714 // convert it to a vector with movd (S2V+shuffle to zero extend).
5715 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5716 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5718 ISD::BITCAST, dl, VT,
5719 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5723 // If we have a constant or non-constant insertion into the low element of
5724 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5725 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5726 // depending on what the source datatype is.
5729 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5731 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5732 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5733 if (VT.is512BitVector()) {
5734 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5735 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5736 Item, DAG.getIntPtrConstant(0));
5738 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5739 "Expected an SSE value type!");
5740 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5741 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5742 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5745 // We can't directly insert an i8 or i16 into a vector, so zero extend
5747 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5748 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5749 if (VT.is256BitVector()) {
5750 if (Subtarget->hasAVX()) {
5751 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5752 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5754 // Without AVX, we need to extend to a 128-bit vector and then
5755 // insert into the 256-bit vector.
5756 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5757 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5758 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5761 assert(VT.is128BitVector() && "Expected an SSE value type!");
5762 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5763 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5765 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5769 // Is it a vector logical left shift?
5770 if (NumElems == 2 && Idx == 1 &&
5771 X86::isZeroNode(Op.getOperand(0)) &&
5772 !X86::isZeroNode(Op.getOperand(1))) {
5773 unsigned NumBits = VT.getSizeInBits();
5774 return getVShift(true, VT,
5775 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5776 VT, Op.getOperand(1)),
5777 NumBits/2, DAG, *this, dl);
5780 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5783 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5784 // is a non-constant being inserted into an element other than the low one,
5785 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5786 // movd/movss) to move this into the low element, then shuffle it into
5788 if (EVTBits == 32) {
5789 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5790 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5794 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5795 if (Values.size() == 1) {
5796 if (EVTBits == 32) {
5797 // Instead of a shuffle like this:
5798 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5799 // Check if it's possible to issue this instead.
5800 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5801 unsigned Idx = countTrailingZeros(NonZeros);
5802 SDValue Item = Op.getOperand(Idx);
5803 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5804 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5809 // A vector full of immediates; various special cases are already
5810 // handled, so this is best done with a single constant-pool load.
5814 // For AVX-length vectors, see if we can use a vector load to get all of the
5815 // elements, otherwise build the individual 128-bit pieces and use
5816 // shuffles to put them in place.
5817 if (VT.is256BitVector() || VT.is512BitVector()) {
5818 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5820 // Check for a build vector of consecutive loads.
5821 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5824 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5826 // Build both the lower and upper subvector.
5827 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5828 makeArrayRef(&V[0], NumElems/2));
5829 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5830 makeArrayRef(&V[NumElems / 2], NumElems/2));
5832 // Recreate the wider vector with the lower and upper part.
5833 if (VT.is256BitVector())
5834 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5835 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5838 // Let legalizer expand 2-wide build_vectors.
5839 if (EVTBits == 64) {
5840 if (NumNonZero == 1) {
5841 // One half is zero or undef.
5842 unsigned Idx = countTrailingZeros(NonZeros);
5843 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5844 Op.getOperand(Idx));
5845 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5850 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5851 if (EVTBits == 8 && NumElems == 16)
5852 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5856 if (EVTBits == 16 && NumElems == 8)
5857 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5861 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5862 if (EVTBits == 32 && NumElems == 4)
5863 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
5866 // If element VT is == 32 bits, turn it into a number of shuffles.
5867 SmallVector<SDValue, 8> V(NumElems);
5868 if (NumElems == 4 && NumZero > 0) {
5869 for (unsigned i = 0; i < 4; ++i) {
5870 bool isZero = !(NonZeros & (1 << i));
5872 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5874 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5877 for (unsigned i = 0; i < 2; ++i) {
5878 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5881 V[i] = V[i*2]; // Must be a zero vector.
5884 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5887 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5890 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5895 bool Reverse1 = (NonZeros & 0x3) == 2;
5896 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5900 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5901 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5903 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5906 if (Values.size() > 1 && VT.is128BitVector()) {
5907 // Check for a build vector of consecutive loads.
5908 for (unsigned i = 0; i < NumElems; ++i)
5909 V[i] = Op.getOperand(i);
5911 // Check for elements which are consecutive loads.
5912 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5915 // Check for a build vector from mostly shuffle plus few inserting.
5916 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
5919 // For SSE 4.1, use insertps to put the high elements into the low element.
5920 if (Subtarget->hasSSE41()) {
5922 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5923 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5925 Result = DAG.getUNDEF(VT);
5927 for (unsigned i = 1; i < NumElems; ++i) {
5928 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5929 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5930 Op.getOperand(i), DAG.getIntPtrConstant(i));
5935 // Otherwise, expand into a number of unpckl*, start by extending each of
5936 // our (non-undef) elements to the full vector width with the element in the
5937 // bottom slot of the vector (which generates no code for SSE).
5938 for (unsigned i = 0; i < NumElems; ++i) {
5939 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5940 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5942 V[i] = DAG.getUNDEF(VT);
5945 // Next, we iteratively mix elements, e.g. for v4f32:
5946 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5947 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5948 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5949 unsigned EltStride = NumElems >> 1;
5950 while (EltStride != 0) {
5951 for (unsigned i = 0; i < EltStride; ++i) {
5952 // If V[i+EltStride] is undef and this is the first round of mixing,
5953 // then it is safe to just drop this shuffle: V[i] is already in the
5954 // right place, the one element (since it's the first round) being
5955 // inserted as undef can be dropped. This isn't safe for successive
5956 // rounds because they will permute elements within both vectors.
5957 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5958 EltStride == NumElems/2)
5961 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5970 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5971 // to create 256-bit vectors from two other 128-bit ones.
5972 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5974 MVT ResVT = Op.getSimpleValueType();
5976 assert((ResVT.is256BitVector() ||
5977 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
5979 SDValue V1 = Op.getOperand(0);
5980 SDValue V2 = Op.getOperand(1);
5981 unsigned NumElems = ResVT.getVectorNumElements();
5982 if (ResVT.is256BitVector())
5983 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5985 if (Op.getNumOperands() == 4) {
5986 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
5987 ResVT.getVectorNumElements()/2);
5988 SDValue V3 = Op.getOperand(2);
5989 SDValue V4 = Op.getOperand(3);
5990 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
5991 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
5993 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5996 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
5997 const X86Subtarget *Subtarget,
5998 SelectionDAG & DAG) {
6000 MVT ResVT = Op.getSimpleValueType();
6001 unsigned NumOfOperands = Op.getNumOperands();
6003 assert(isPowerOf2_32(NumOfOperands) &&
6004 "Unexpected number of operands in CONCAT_VECTORS");
6006 if (NumOfOperands > 2) {
6007 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6008 ResVT.getVectorNumElements()/2);
6009 SmallVector<SDValue, 2> Ops;
6010 for (unsigned i = 0; i < NumOfOperands/2; i++)
6011 Ops.push_back(Op.getOperand(i));
6012 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6014 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6015 Ops.push_back(Op.getOperand(i));
6016 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6017 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6020 SDValue V1 = Op.getOperand(0);
6021 SDValue V2 = Op.getOperand(1);
6022 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6023 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6025 if (IsZeroV1 && IsZeroV2)
6026 return getZeroVector(ResVT, Subtarget, DAG, dl);
6028 SDValue ZeroIdx = DAG.getIntPtrConstant(0);
6029 SDValue Undef = DAG.getUNDEF(ResVT);
6030 unsigned NumElems = ResVT.getVectorNumElements();
6031 SDValue ShiftBits = DAG.getConstant(NumElems/2, MVT::i8);
6033 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6034 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6038 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6039 // Zero the upper bits of V1
6040 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6041 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6044 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6047 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6048 const X86Subtarget *Subtarget,
6049 SelectionDAG &DAG) {
6050 MVT VT = Op.getSimpleValueType();
6051 if (VT.getVectorElementType() == MVT::i1)
6052 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6054 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6055 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6056 Op.getNumOperands() == 4)));
6058 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6059 // from two other 128-bit ones.
6061 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6062 return LowerAVXCONCAT_VECTORS(Op, DAG);
6066 //===----------------------------------------------------------------------===//
6067 // Vector shuffle lowering
6069 // This is an experimental code path for lowering vector shuffles on x86. It is
6070 // designed to handle arbitrary vector shuffles and blends, gracefully
6071 // degrading performance as necessary. It works hard to recognize idiomatic
6072 // shuffles and lower them to optimal instruction patterns without leaving
6073 // a framework that allows reasonably efficient handling of all vector shuffle
6075 //===----------------------------------------------------------------------===//
6077 /// \brief Tiny helper function to identify a no-op mask.
6079 /// This is a somewhat boring predicate function. It checks whether the mask
6080 /// array input, which is assumed to be a single-input shuffle mask of the kind
6081 /// used by the X86 shuffle instructions (not a fully general
6082 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6083 /// in-place shuffle are 'no-op's.
6084 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6085 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6086 if (Mask[i] != -1 && Mask[i] != i)
6091 /// \brief Helper function to classify a mask as a single-input mask.
6093 /// This isn't a generic single-input test because in the vector shuffle
6094 /// lowering we canonicalize single inputs to be the first input operand. This
6095 /// means we can more quickly test for a single input by only checking whether
6096 /// an input from the second operand exists. We also assume that the size of
6097 /// mask corresponds to the size of the input vectors which isn't true in the
6098 /// fully general case.
6099 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6101 if (M >= (int)Mask.size())
6106 /// \brief Test whether there are elements crossing 128-bit lanes in this
6109 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6110 /// and we routinely test for these.
6111 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6112 int LaneSize = 128 / VT.getScalarSizeInBits();
6113 int Size = Mask.size();
6114 for (int i = 0; i < Size; ++i)
6115 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6120 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6122 /// This checks a shuffle mask to see if it is performing the same
6123 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6124 /// that it is also not lane-crossing. It may however involve a blend from the
6125 /// same lane of a second vector.
6127 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6128 /// non-trivial to compute in the face of undef lanes. The representation is
6129 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6130 /// entries from both V1 and V2 inputs to the wider mask.
6132 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6133 SmallVectorImpl<int> &RepeatedMask) {
6134 int LaneSize = 128 / VT.getScalarSizeInBits();
6135 RepeatedMask.resize(LaneSize, -1);
6136 int Size = Mask.size();
6137 for (int i = 0; i < Size; ++i) {
6140 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6141 // This entry crosses lanes, so there is no way to model this shuffle.
6144 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6145 if (RepeatedMask[i % LaneSize] == -1)
6146 // This is the first non-undef entry in this slot of a 128-bit lane.
6147 RepeatedMask[i % LaneSize] =
6148 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6149 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6150 // Found a mismatch with the repeated mask.
6156 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6159 /// This is a fast way to test a shuffle mask against a fixed pattern:
6161 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6163 /// It returns true if the mask is exactly as wide as the argument list, and
6164 /// each element of the mask is either -1 (signifying undef) or the value given
6165 /// in the argument.
6166 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6167 ArrayRef<int> ExpectedMask) {
6168 if (Mask.size() != ExpectedMask.size())
6171 int Size = Mask.size();
6173 // If the values are build vectors, we can look through them to find
6174 // equivalent inputs that make the shuffles equivalent.
6175 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6176 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6178 for (int i = 0; i < Size; ++i)
6179 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6180 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6181 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6182 if (!MaskBV || !ExpectedBV ||
6183 MaskBV->getOperand(Mask[i] % Size) !=
6184 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6191 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6193 /// This helper function produces an 8-bit shuffle immediate corresponding to
6194 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6195 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6198 /// NB: We rely heavily on "undef" masks preserving the input lane.
6199 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
6200 SelectionDAG &DAG) {
6201 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6202 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6203 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6204 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6205 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6208 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6209 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6210 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6211 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6212 return DAG.getConstant(Imm, MVT::i8);
6215 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6217 /// This is used as a fallback approach when first class blend instructions are
6218 /// unavailable. Currently it is only suitable for integer vectors, but could
6219 /// be generalized for floating point vectors if desirable.
6220 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6221 SDValue V2, ArrayRef<int> Mask,
6222 SelectionDAG &DAG) {
6223 assert(VT.isInteger() && "Only supports integer vector types!");
6224 MVT EltVT = VT.getScalarType();
6225 int NumEltBits = EltVT.getSizeInBits();
6226 SDValue Zero = DAG.getConstant(0, EltVT);
6227 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
6228 SmallVector<SDValue, 16> MaskOps;
6229 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6230 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6231 return SDValue(); // Shuffled input!
6232 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6235 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6236 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6237 // We have to cast V2 around.
6238 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6239 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6240 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6241 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6242 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6243 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6246 /// \brief Try to emit a blend instruction for a shuffle.
6248 /// This doesn't do any checks for the availability of instructions for blending
6249 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6250 /// be matched in the backend with the type given. What it does check for is
6251 /// that the shuffle mask is in fact a blend.
6252 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6253 SDValue V2, ArrayRef<int> Mask,
6254 const X86Subtarget *Subtarget,
6255 SelectionDAG &DAG) {
6256 unsigned BlendMask = 0;
6257 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6258 if (Mask[i] >= Size) {
6259 if (Mask[i] != i + Size)
6260 return SDValue(); // Shuffled V2 input!
6261 BlendMask |= 1u << i;
6264 if (Mask[i] >= 0 && Mask[i] != i)
6265 return SDValue(); // Shuffled V1 input!
6267 switch (VT.SimpleTy) {
6272 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6273 DAG.getConstant(BlendMask, MVT::i8));
6277 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6281 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6282 // that instruction.
6283 if (Subtarget->hasAVX2()) {
6284 // Scale the blend by the number of 32-bit dwords per element.
6285 int Scale = VT.getScalarSizeInBits() / 32;
6287 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6288 if (Mask[i] >= Size)
6289 for (int j = 0; j < Scale; ++j)
6290 BlendMask |= 1u << (i * Scale + j);
6292 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6293 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6294 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6295 return DAG.getNode(ISD::BITCAST, DL, VT,
6296 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6297 DAG.getConstant(BlendMask, MVT::i8)));
6301 // For integer shuffles we need to expand the mask and cast the inputs to
6302 // v8i16s prior to blending.
6303 int Scale = 8 / VT.getVectorNumElements();
6305 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6306 if (Mask[i] >= Size)
6307 for (int j = 0; j < Scale; ++j)
6308 BlendMask |= 1u << (i * Scale + j);
6310 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6311 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6312 return DAG.getNode(ISD::BITCAST, DL, VT,
6313 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6314 DAG.getConstant(BlendMask, MVT::i8)));
6318 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6319 SmallVector<int, 8> RepeatedMask;
6320 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6321 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6322 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6324 for (int i = 0; i < 8; ++i)
6325 if (RepeatedMask[i] >= 16)
6326 BlendMask |= 1u << i;
6327 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6328 DAG.getConstant(BlendMask, MVT::i8));
6334 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6335 "256-bit byte-blends require AVX2 support!");
6337 // Scale the blend by the number of bytes per element.
6338 int Scale = VT.getScalarSizeInBits() / 8;
6340 // This form of blend is always done on bytes. Compute the byte vector
6342 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6344 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6345 // mix of LLVM's code generator and the x86 backend. We tell the code
6346 // generator that boolean values in the elements of an x86 vector register
6347 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6348 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6349 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6350 // of the element (the remaining are ignored) and 0 in that high bit would
6351 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6352 // the LLVM model for boolean values in vector elements gets the relevant
6353 // bit set, it is set backwards and over constrained relative to x86's
6355 SmallVector<SDValue, 32> VSELECTMask;
6356 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6357 for (int j = 0; j < Scale; ++j)
6358 VSELECTMask.push_back(
6359 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6360 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
6362 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6363 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6365 ISD::BITCAST, DL, VT,
6366 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6367 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6372 llvm_unreachable("Not a supported integer vector type!");
6376 /// \brief Try to lower as a blend of elements from two inputs followed by
6377 /// a single-input permutation.
6379 /// This matches the pattern where we can blend elements from two inputs and
6380 /// then reduce the shuffle to a single-input permutation.
6381 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6384 SelectionDAG &DAG) {
6385 // We build up the blend mask while checking whether a blend is a viable way
6386 // to reduce the shuffle.
6387 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6388 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6390 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6394 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6396 if (BlendMask[Mask[i] % Size] == -1)
6397 BlendMask[Mask[i] % Size] = Mask[i];
6398 else if (BlendMask[Mask[i] % Size] != Mask[i])
6399 return SDValue(); // Can't blend in the needed input!
6401 PermuteMask[i] = Mask[i] % Size;
6404 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6405 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6408 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6409 /// blends and permutes.
6411 /// This matches the extremely common pattern for handling combined
6412 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6413 /// operations. It will try to pick the best arrangement of shuffles and
6415 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6419 SelectionDAG &DAG) {
6420 // Shuffle the input elements into the desired positions in V1 and V2 and
6421 // blend them together.
6422 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6423 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6424 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6425 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6426 if (Mask[i] >= 0 && Mask[i] < Size) {
6427 V1Mask[i] = Mask[i];
6429 } else if (Mask[i] >= Size) {
6430 V2Mask[i] = Mask[i] - Size;
6431 BlendMask[i] = i + Size;
6434 // Try to lower with the simpler initial blend strategy unless one of the
6435 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6436 // shuffle may be able to fold with a load or other benefit. However, when
6437 // we'll have to do 2x as many shuffles in order to achieve this, blending
6438 // first is a better strategy.
6439 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6440 if (SDValue BlendPerm =
6441 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6444 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6445 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6446 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6449 /// \brief Try to lower a vector shuffle as a byte rotation.
6451 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6452 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6453 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6454 /// try to generically lower a vector shuffle through such an pattern. It
6455 /// does not check for the profitability of lowering either as PALIGNR or
6456 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6457 /// This matches shuffle vectors that look like:
6459 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6461 /// Essentially it concatenates V1 and V2, shifts right by some number of
6462 /// elements, and takes the low elements as the result. Note that while this is
6463 /// specified as a *right shift* because x86 is little-endian, it is a *left
6464 /// rotate* of the vector lanes.
6465 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6468 const X86Subtarget *Subtarget,
6469 SelectionDAG &DAG) {
6470 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6472 int NumElts = Mask.size();
6473 int NumLanes = VT.getSizeInBits() / 128;
6474 int NumLaneElts = NumElts / NumLanes;
6476 // We need to detect various ways of spelling a rotation:
6477 // [11, 12, 13, 14, 15, 0, 1, 2]
6478 // [-1, 12, 13, 14, -1, -1, 1, -1]
6479 // [-1, -1, -1, -1, -1, -1, 1, 2]
6480 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6481 // [-1, 4, 5, 6, -1, -1, 9, -1]
6482 // [-1, 4, 5, 6, -1, -1, -1, -1]
6485 for (int l = 0; l < NumElts; l += NumLaneElts) {
6486 for (int i = 0; i < NumLaneElts; ++i) {
6487 if (Mask[l + i] == -1)
6489 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6491 // Get the mod-Size index and lane correct it.
6492 int LaneIdx = (Mask[l + i] % NumElts) - l;
6493 // Make sure it was in this lane.
6494 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6497 // Determine where a rotated vector would have started.
6498 int StartIdx = i - LaneIdx;
6500 // The identity rotation isn't interesting, stop.
6503 // If we found the tail of a vector the rotation must be the missing
6504 // front. If we found the head of a vector, it must be how much of the
6506 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6509 Rotation = CandidateRotation;
6510 else if (Rotation != CandidateRotation)
6511 // The rotations don't match, so we can't match this mask.
6514 // Compute which value this mask is pointing at.
6515 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6517 // Compute which of the two target values this index should be assigned
6518 // to. This reflects whether the high elements are remaining or the low
6519 // elements are remaining.
6520 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6522 // Either set up this value if we've not encountered it before, or check
6523 // that it remains consistent.
6526 else if (TargetV != MaskV)
6527 // This may be a rotation, but it pulls from the inputs in some
6528 // unsupported interleaving.
6533 // Check that we successfully analyzed the mask, and normalize the results.
6534 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6535 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6541 // The actual rotate instruction rotates bytes, so we need to scale the
6542 // rotation based on how many bytes are in the vector lane.
6543 int Scale = 16 / NumLaneElts;
6545 // SSSE3 targets can use the palignr instruction.
6546 if (Subtarget->hasSSSE3()) {
6547 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6548 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6549 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6550 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6552 return DAG.getNode(ISD::BITCAST, DL, VT,
6553 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6554 DAG.getConstant(Rotation * Scale, MVT::i8)));
6557 assert(VT.getSizeInBits() == 128 &&
6558 "Rotate-based lowering only supports 128-bit lowering!");
6559 assert(Mask.size() <= 16 &&
6560 "Can shuffle at most 16 bytes in a 128-bit vector!");
6562 // Default SSE2 implementation
6563 int LoByteShift = 16 - Rotation * Scale;
6564 int HiByteShift = Rotation * Scale;
6566 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6567 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6568 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6570 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6571 DAG.getConstant(LoByteShift, MVT::i8));
6572 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6573 DAG.getConstant(HiByteShift, MVT::i8));
6574 return DAG.getNode(ISD::BITCAST, DL, VT,
6575 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6578 /// \brief Compute whether each element of a shuffle is zeroable.
6580 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6581 /// Either it is an undef element in the shuffle mask, the element of the input
6582 /// referenced is undef, or the element of the input referenced is known to be
6583 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6584 /// as many lanes with this technique as possible to simplify the remaining
6586 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6587 SDValue V1, SDValue V2) {
6588 SmallBitVector Zeroable(Mask.size(), false);
6590 while (V1.getOpcode() == ISD::BITCAST)
6591 V1 = V1->getOperand(0);
6592 while (V2.getOpcode() == ISD::BITCAST)
6593 V2 = V2->getOperand(0);
6595 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6596 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6598 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6600 // Handle the easy cases.
6601 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6606 // If this is an index into a build_vector node (which has the same number
6607 // of elements), dig out the input value and use it.
6608 SDValue V = M < Size ? V1 : V2;
6609 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6612 SDValue Input = V.getOperand(M % Size);
6613 // The UNDEF opcode check really should be dead code here, but not quite
6614 // worth asserting on (it isn't invalid, just unexpected).
6615 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6622 /// \brief Try to emit a bitmask instruction for a shuffle.
6624 /// This handles cases where we can model a blend exactly as a bitmask due to
6625 /// one of the inputs being zeroable.
6626 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6627 SDValue V2, ArrayRef<int> Mask,
6628 SelectionDAG &DAG) {
6629 MVT EltVT = VT.getScalarType();
6630 int NumEltBits = EltVT.getSizeInBits();
6631 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6632 SDValue Zero = DAG.getConstant(0, IntEltVT);
6633 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
6634 if (EltVT.isFloatingPoint()) {
6635 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6636 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6638 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6639 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6641 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6644 if (Mask[i] % Size != i)
6645 return SDValue(); // Not a blend.
6647 V = Mask[i] < Size ? V1 : V2;
6648 else if (V != (Mask[i] < Size ? V1 : V2))
6649 return SDValue(); // Can only let one input through the mask.
6651 VMaskOps[i] = AllOnes;
6654 return SDValue(); // No non-zeroable elements!
6656 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6657 V = DAG.getNode(VT.isFloatingPoint()
6658 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6663 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6665 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6666 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6667 /// matches elements from one of the input vectors shuffled to the left or
6668 /// right with zeroable elements 'shifted in'. It handles both the strictly
6669 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6672 /// PSHL : (little-endian) left bit shift.
6673 /// [ zz, 0, zz, 2 ]
6674 /// [ -1, 4, zz, -1 ]
6675 /// PSRL : (little-endian) right bit shift.
6677 /// [ -1, -1, 7, zz]
6678 /// PSLLDQ : (little-endian) left byte shift
6679 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6680 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6681 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6682 /// PSRLDQ : (little-endian) right byte shift
6683 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6684 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6685 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6686 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6687 SDValue V2, ArrayRef<int> Mask,
6688 SelectionDAG &DAG) {
6689 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6691 int Size = Mask.size();
6692 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6694 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6695 for (int i = 0; i < Size; i += Scale)
6696 for (int j = 0; j < Shift; ++j)
6697 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6703 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6704 for (int i = 0; i != Size; i += Scale) {
6705 unsigned Pos = Left ? i + Shift : i;
6706 unsigned Low = Left ? i : i + Shift;
6707 unsigned Len = Scale - Shift;
6708 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6709 Low + (V == V1 ? 0 : Size)))
6713 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6714 bool ByteShift = ShiftEltBits > 64;
6715 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6716 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6717 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6719 // Normalize the scale for byte shifts to still produce an i64 element
6721 Scale = ByteShift ? Scale / 2 : Scale;
6723 // We need to round trip through the appropriate type for the shift.
6724 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6725 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6726 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6727 "Illegal integer vector type");
6728 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6730 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
6731 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6734 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6735 // keep doubling the size of the integer elements up to that. We can
6736 // then shift the elements of the integer vector by whole multiples of
6737 // their width within the elements of the larger integer vector. Test each
6738 // multiple to see if we can find a match with the moved element indices
6739 // and that the shifted in elements are all zeroable.
6740 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6741 for (int Shift = 1; Shift != Scale; ++Shift)
6742 for (bool Left : {true, false})
6743 if (CheckZeros(Shift, Scale, Left))
6744 for (SDValue V : {V1, V2})
6745 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6752 /// \brief Lower a vector shuffle as a zero or any extension.
6754 /// Given a specific number of elements, element bit width, and extension
6755 /// stride, produce either a zero or any extension based on the available
6756 /// features of the subtarget.
6757 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6758 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6759 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6760 assert(Scale > 1 && "Need a scale to extend.");
6761 int NumElements = VT.getVectorNumElements();
6762 int EltBits = VT.getScalarSizeInBits();
6763 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6764 "Only 8, 16, and 32 bit elements can be extended.");
6765 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6767 // Found a valid zext mask! Try various lowering strategies based on the
6768 // input type and available ISA extensions.
6769 if (Subtarget->hasSSE41()) {
6770 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6771 NumElements / Scale);
6772 return DAG.getNode(ISD::BITCAST, DL, VT,
6773 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6776 // For any extends we can cheat for larger element sizes and use shuffle
6777 // instructions that can fold with a load and/or copy.
6778 if (AnyExt && EltBits == 32) {
6779 int PSHUFDMask[4] = {0, -1, 1, -1};
6781 ISD::BITCAST, DL, VT,
6782 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6783 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6784 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
6786 if (AnyExt && EltBits == 16 && Scale > 2) {
6787 int PSHUFDMask[4] = {0, -1, 0, -1};
6788 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6789 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6790 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
6791 int PSHUFHWMask[4] = {1, -1, -1, -1};
6793 ISD::BITCAST, DL, VT,
6794 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6795 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6796 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
6799 // If this would require more than 2 unpack instructions to expand, use
6800 // pshufb when available. We can only use more than 2 unpack instructions
6801 // when zero extending i8 elements which also makes it easier to use pshufb.
6802 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6803 assert(NumElements == 16 && "Unexpected byte vector width!");
6804 SDValue PSHUFBMask[16];
6805 for (int i = 0; i < 16; ++i)
6807 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
6808 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6809 return DAG.getNode(ISD::BITCAST, DL, VT,
6810 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6811 DAG.getNode(ISD::BUILD_VECTOR, DL,
6812 MVT::v16i8, PSHUFBMask)));
6815 // Otherwise emit a sequence of unpacks.
6817 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6818 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6819 : getZeroVector(InputVT, Subtarget, DAG, DL);
6820 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6821 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6825 } while (Scale > 1);
6826 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6829 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6831 /// This routine will try to do everything in its power to cleverly lower
6832 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6833 /// check for the profitability of this lowering, it tries to aggressively
6834 /// match this pattern. It will use all of the micro-architectural details it
6835 /// can to emit an efficient lowering. It handles both blends with all-zero
6836 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6837 /// masking out later).
6839 /// The reason we have dedicated lowering for zext-style shuffles is that they
6840 /// are both incredibly common and often quite performance sensitive.
6841 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6842 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6843 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6844 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6846 int Bits = VT.getSizeInBits();
6847 int NumElements = VT.getVectorNumElements();
6848 assert(VT.getScalarSizeInBits() <= 32 &&
6849 "Exceeds 32-bit integer zero extension limit");
6850 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6852 // Define a helper function to check a particular ext-scale and lower to it if
6854 auto Lower = [&](int Scale) -> SDValue {
6857 for (int i = 0; i < NumElements; ++i) {
6859 continue; // Valid anywhere but doesn't tell us anything.
6860 if (i % Scale != 0) {
6861 // Each of the extended elements need to be zeroable.
6865 // We no longer are in the anyext case.
6870 // Each of the base elements needs to be consecutive indices into the
6871 // same input vector.
6872 SDValue V = Mask[i] < NumElements ? V1 : V2;
6875 else if (InputV != V)
6876 return SDValue(); // Flip-flopping inputs.
6878 if (Mask[i] % NumElements != i / Scale)
6879 return SDValue(); // Non-consecutive strided elements.
6882 // If we fail to find an input, we have a zero-shuffle which should always
6883 // have already been handled.
6884 // FIXME: Maybe handle this here in case during blending we end up with one?
6888 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6889 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6892 // The widest scale possible for extending is to a 64-bit integer.
6893 assert(Bits % 64 == 0 &&
6894 "The number of bits in a vector must be divisible by 64 on x86!");
6895 int NumExtElements = Bits / 64;
6897 // Each iteration, try extending the elements half as much, but into twice as
6899 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6900 assert(NumElements % NumExtElements == 0 &&
6901 "The input vector size must be divisible by the extended size.");
6902 if (SDValue V = Lower(NumElements / NumExtElements))
6906 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6910 // Returns one of the source operands if the shuffle can be reduced to a
6911 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6912 auto CanZExtLowHalf = [&]() {
6913 for (int i = NumElements / 2; i != NumElements; ++i)
6916 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6918 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
6923 if (SDValue V = CanZExtLowHalf()) {
6924 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
6925 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
6926 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6929 // No viable ext lowering found.
6933 /// \brief Try to get a scalar value for a specific element of a vector.
6935 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
6936 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
6937 SelectionDAG &DAG) {
6938 MVT VT = V.getSimpleValueType();
6939 MVT EltVT = VT.getVectorElementType();
6940 while (V.getOpcode() == ISD::BITCAST)
6941 V = V.getOperand(0);
6942 // If the bitcasts shift the element size, we can't extract an equivalent
6944 MVT NewVT = V.getSimpleValueType();
6945 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
6948 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6949 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
6950 // Ensure the scalar operand is the same size as the destination.
6951 // FIXME: Add support for scalar truncation where possible.
6952 SDValue S = V.getOperand(Idx);
6953 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
6954 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
6960 /// \brief Helper to test for a load that can be folded with x86 shuffles.
6962 /// This is particularly important because the set of instructions varies
6963 /// significantly based on whether the operand is a load or not.
6964 static bool isShuffleFoldableLoad(SDValue V) {
6965 while (V.getOpcode() == ISD::BITCAST)
6966 V = V.getOperand(0);
6968 return ISD::isNON_EXTLoad(V.getNode());
6971 /// \brief Try to lower insertion of a single element into a zero vector.
6973 /// This is a common pattern that we have especially efficient patterns to lower
6974 /// across all subtarget feature sets.
6975 static SDValue lowerVectorShuffleAsElementInsertion(
6976 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6977 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6978 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6980 MVT EltVT = VT.getVectorElementType();
6982 int V2Index = std::find_if(Mask.begin(), Mask.end(),
6983 [&Mask](int M) { return M >= (int)Mask.size(); }) -
6985 bool IsV1Zeroable = true;
6986 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6987 if (i != V2Index && !Zeroable[i]) {
6988 IsV1Zeroable = false;
6992 // Check for a single input from a SCALAR_TO_VECTOR node.
6993 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
6994 // all the smarts here sunk into that routine. However, the current
6995 // lowering of BUILD_VECTOR makes that nearly impossible until the old
6996 // vector shuffle lowering is dead.
6997 if (SDValue V2S = getScalarValueForVectorElement(
6998 V2, Mask[V2Index] - Mask.size(), DAG)) {
6999 // We need to zext the scalar if it is smaller than an i32.
7000 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
7001 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7002 // Using zext to expand a narrow element won't work for non-zero
7007 // Zero-extend directly to i32.
7009 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7011 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7012 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7013 EltVT == MVT::i16) {
7014 // Either not inserting from the low element of the input or the input
7015 // element size is too small to use VZEXT_MOVL to clear the high bits.
7019 if (!IsV1Zeroable) {
7020 // If V1 can't be treated as a zero vector we have fewer options to lower
7021 // this. We can't support integer vectors or non-zero targets cheaply, and
7022 // the V1 elements can't be permuted in any way.
7023 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7024 if (!VT.isFloatingPoint() || V2Index != 0)
7026 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7027 V1Mask[V2Index] = -1;
7028 if (!isNoopShuffleMask(V1Mask))
7030 // This is essentially a special case blend operation, but if we have
7031 // general purpose blend operations, they are always faster. Bail and let
7032 // the rest of the lowering handle these as blends.
7033 if (Subtarget->hasSSE41())
7036 // Otherwise, use MOVSD or MOVSS.
7037 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7038 "Only two types of floating point element types to handle!");
7039 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7043 // This lowering only works for the low element with floating point vectors.
7044 if (VT.isFloatingPoint() && V2Index != 0)
7047 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7049 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7052 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7053 // the desired position. Otherwise it is more efficient to do a vector
7054 // shift left. We know that we can do a vector shift left because all
7055 // the inputs are zero.
7056 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7057 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7058 V2Shuffle[V2Index] = 0;
7059 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7061 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
7063 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7065 V2Index * EltVT.getSizeInBits()/8,
7066 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7067 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7073 /// \brief Try to lower broadcast of a single element.
7075 /// For convenience, this code also bundles all of the subtarget feature set
7076 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7077 /// a convenient way to factor it out.
7078 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7080 const X86Subtarget *Subtarget,
7081 SelectionDAG &DAG) {
7082 if (!Subtarget->hasAVX())
7084 if (VT.isInteger() && !Subtarget->hasAVX2())
7087 // Check that the mask is a broadcast.
7088 int BroadcastIdx = -1;
7090 if (M >= 0 && BroadcastIdx == -1)
7092 else if (M >= 0 && M != BroadcastIdx)
7095 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7096 "a sorted mask where the broadcast "
7099 // Go up the chain of (vector) values to find a scalar load that we can
7100 // combine with the broadcast.
7102 switch (V.getOpcode()) {
7103 case ISD::CONCAT_VECTORS: {
7104 int OperandSize = Mask.size() / V.getNumOperands();
7105 V = V.getOperand(BroadcastIdx / OperandSize);
7106 BroadcastIdx %= OperandSize;
7110 case ISD::INSERT_SUBVECTOR: {
7111 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7112 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7116 int BeginIdx = (int)ConstantIdx->getZExtValue();
7118 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7119 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7120 BroadcastIdx -= BeginIdx;
7131 // Check if this is a broadcast of a scalar. We special case lowering
7132 // for scalars so that we can more effectively fold with loads.
7133 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7134 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7135 V = V.getOperand(BroadcastIdx);
7137 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7138 // Only AVX2 has register broadcasts.
7139 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7141 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7142 // We can't broadcast from a vector register without AVX2, and we can only
7143 // broadcast from the zero-element of a vector register.
7147 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7150 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7151 // INSERTPS when the V1 elements are already in the correct locations
7152 // because otherwise we can just always use two SHUFPS instructions which
7153 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7154 // perform INSERTPS if a single V1 element is out of place and all V2
7155 // elements are zeroable.
7156 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7158 SelectionDAG &DAG) {
7159 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7160 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7161 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7162 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7164 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7167 int V1DstIndex = -1;
7168 int V2DstIndex = -1;
7169 bool V1UsedInPlace = false;
7171 for (int i = 0; i < 4; ++i) {
7172 // Synthesize a zero mask from the zeroable elements (includes undefs).
7178 // Flag if we use any V1 inputs in place.
7180 V1UsedInPlace = true;
7184 // We can only insert a single non-zeroable element.
7185 if (V1DstIndex != -1 || V2DstIndex != -1)
7189 // V1 input out of place for insertion.
7192 // V2 input for insertion.
7197 // Don't bother if we have no (non-zeroable) element for insertion.
7198 if (V1DstIndex == -1 && V2DstIndex == -1)
7201 // Determine element insertion src/dst indices. The src index is from the
7202 // start of the inserted vector, not the start of the concatenated vector.
7203 unsigned V2SrcIndex = 0;
7204 if (V1DstIndex != -1) {
7205 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7206 // and don't use the original V2 at all.
7207 V2SrcIndex = Mask[V1DstIndex];
7208 V2DstIndex = V1DstIndex;
7211 V2SrcIndex = Mask[V2DstIndex] - 4;
7214 // If no V1 inputs are used in place, then the result is created only from
7215 // the zero mask and the V2 insertion - so remove V1 dependency.
7217 V1 = DAG.getUNDEF(MVT::v4f32);
7219 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7220 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7222 // Insert the V2 element into the desired position.
7224 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7225 DAG.getConstant(InsertPSMask, MVT::i8));
7228 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7229 /// UNPCK instruction.
7231 /// This specifically targets cases where we end up with alternating between
7232 /// the two inputs, and so can permute them into something that feeds a single
7233 /// UNPCK instruction. Note that this routine only targets integer vectors
7234 /// because for floating point vectors we have a generalized SHUFPS lowering
7235 /// strategy that handles everything that doesn't *exactly* match an unpack,
7236 /// making this clever lowering unnecessary.
7237 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7238 SDValue V2, ArrayRef<int> Mask,
7239 SelectionDAG &DAG) {
7240 assert(!VT.isFloatingPoint() &&
7241 "This routine only supports integer vectors.");
7242 assert(!isSingleInputShuffleMask(Mask) &&
7243 "This routine should only be used when blending two inputs.");
7244 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7246 int Size = Mask.size();
7248 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7249 return M >= 0 && M % Size < Size / 2;
7251 int NumHiInputs = std::count_if(
7252 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7254 bool UnpackLo = NumLoInputs >= NumHiInputs;
7256 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7257 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7258 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7260 for (int i = 0; i < Size; ++i) {
7264 // Each element of the unpack contains Scale elements from this mask.
7265 int UnpackIdx = i / Scale;
7267 // We only handle the case where V1 feeds the first slots of the unpack.
7268 // We rely on canonicalization to ensure this is the case.
7269 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7272 // Setup the mask for this input. The indexing is tricky as we have to
7273 // handle the unpack stride.
7274 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7275 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7279 // If we will have to shuffle both inputs to use the unpack, check whether
7280 // we can just unpack first and shuffle the result. If so, skip this unpack.
7281 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7282 !isNoopShuffleMask(V2Mask))
7285 // Shuffle the inputs into place.
7286 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7287 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7289 // Cast the inputs to the type we will use to unpack them.
7290 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7291 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7293 // Unpack the inputs and cast the result back to the desired type.
7294 return DAG.getNode(ISD::BITCAST, DL, VT,
7295 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7296 DL, UnpackVT, V1, V2));
7299 // We try each unpack from the largest to the smallest to try and find one
7300 // that fits this mask.
7301 int OrigNumElements = VT.getVectorNumElements();
7302 int OrigScalarSize = VT.getScalarSizeInBits();
7303 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7304 int Scale = ScalarSize / OrigScalarSize;
7305 int NumElements = OrigNumElements / Scale;
7306 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7307 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7311 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7313 if (NumLoInputs == 0 || NumHiInputs == 0) {
7314 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7315 "We have to have *some* inputs!");
7316 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7318 // FIXME: We could consider the total complexity of the permute of each
7319 // possible unpacking. Or at the least we should consider how many
7320 // half-crossings are created.
7321 // FIXME: We could consider commuting the unpacks.
7323 SmallVector<int, 32> PermMask;
7324 PermMask.assign(Size, -1);
7325 for (int i = 0; i < Size; ++i) {
7329 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7332 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7334 return DAG.getVectorShuffle(
7335 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7337 DAG.getUNDEF(VT), PermMask);
7343 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7345 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7346 /// support for floating point shuffles but not integer shuffles. These
7347 /// instructions will incur a domain crossing penalty on some chips though so
7348 /// it is better to avoid lowering through this for integer vectors where
7350 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7351 const X86Subtarget *Subtarget,
7352 SelectionDAG &DAG) {
7354 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7355 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7356 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7357 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7358 ArrayRef<int> Mask = SVOp->getMask();
7359 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7361 if (isSingleInputShuffleMask(Mask)) {
7362 // Use low duplicate instructions for masks that match their pattern.
7363 if (Subtarget->hasSSE3())
7364 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7365 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7367 // Straight shuffle of a single input vector. Simulate this by using the
7368 // single input as both of the "inputs" to this instruction..
7369 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7371 if (Subtarget->hasAVX()) {
7372 // If we have AVX, we can use VPERMILPS which will allow folding a load
7373 // into the shuffle.
7374 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7375 DAG.getConstant(SHUFPDMask, MVT::i8));
7378 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
7379 DAG.getConstant(SHUFPDMask, MVT::i8));
7381 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7382 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7384 // If we have a single input, insert that into V1 if we can do so cheaply.
7385 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7386 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7387 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7389 // Try inverting the insertion since for v2 masks it is easy to do and we
7390 // can't reliably sort the mask one way or the other.
7391 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7392 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7393 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7394 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7398 // Try to use one of the special instruction patterns to handle two common
7399 // blend patterns if a zero-blend above didn't work.
7400 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7401 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7402 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7403 // We can either use a special instruction to load over the low double or
7404 // to move just the low double.
7406 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7408 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7410 if (Subtarget->hasSSE41())
7411 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7415 // Use dedicated unpack instructions for masks that match their pattern.
7416 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7417 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7418 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7419 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7421 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7422 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
7423 DAG.getConstant(SHUFPDMask, MVT::i8));
7426 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7428 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7429 /// the integer unit to minimize domain crossing penalties. However, for blends
7430 /// it falls back to the floating point shuffle operation with appropriate bit
7432 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7433 const X86Subtarget *Subtarget,
7434 SelectionDAG &DAG) {
7436 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7437 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7438 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7439 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7440 ArrayRef<int> Mask = SVOp->getMask();
7441 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7443 if (isSingleInputShuffleMask(Mask)) {
7444 // Check for being able to broadcast a single element.
7445 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7446 Mask, Subtarget, DAG))
7449 // Straight shuffle of a single input vector. For everything from SSE2
7450 // onward this has a single fast instruction with no scary immediates.
7451 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7452 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7453 int WidenedMask[4] = {
7454 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7455 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7457 ISD::BITCAST, DL, MVT::v2i64,
7458 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7459 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7461 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7462 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7463 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7464 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7466 // If we have a blend of two PACKUS operations an the blend aligns with the
7467 // low and half halves, we can just merge the PACKUS operations. This is
7468 // particularly important as it lets us merge shuffles that this routine itself
7470 auto GetPackNode = [](SDValue V) {
7471 while (V.getOpcode() == ISD::BITCAST)
7472 V = V.getOperand(0);
7474 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7476 if (SDValue V1Pack = GetPackNode(V1))
7477 if (SDValue V2Pack = GetPackNode(V2))
7478 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7479 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7480 Mask[0] == 0 ? V1Pack.getOperand(0)
7481 : V1Pack.getOperand(1),
7482 Mask[1] == 2 ? V2Pack.getOperand(0)
7483 : V2Pack.getOperand(1)));
7485 // Try to use shift instructions.
7487 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7490 // When loading a scalar and then shuffling it into a vector we can often do
7491 // the insertion cheaply.
7492 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7493 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7495 // Try inverting the insertion since for v2 masks it is easy to do and we
7496 // can't reliably sort the mask one way or the other.
7497 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7498 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7499 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7502 // We have different paths for blend lowering, but they all must use the
7503 // *exact* same predicate.
7504 bool IsBlendSupported = Subtarget->hasSSE41();
7505 if (IsBlendSupported)
7506 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7510 // Use dedicated unpack instructions for masks that match their pattern.
7511 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7512 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7513 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7514 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7516 // Try to use byte rotation instructions.
7517 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7518 if (Subtarget->hasSSSE3())
7519 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7520 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7523 // If we have direct support for blends, we should lower by decomposing into
7524 // a permute. That will be faster than the domain cross.
7525 if (IsBlendSupported)
7526 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7529 // We implement this with SHUFPD which is pretty lame because it will likely
7530 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7531 // However, all the alternatives are still more cycles and newer chips don't
7532 // have this problem. It would be really nice if x86 had better shuffles here.
7533 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7534 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7535 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7536 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7539 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7541 /// This is used to disable more specialized lowerings when the shufps lowering
7542 /// will happen to be efficient.
7543 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7544 // This routine only handles 128-bit shufps.
7545 assert(Mask.size() == 4 && "Unsupported mask size!");
7547 // To lower with a single SHUFPS we need to have the low half and high half
7548 // each requiring a single input.
7549 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7551 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7557 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7559 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7560 /// It makes no assumptions about whether this is the *best* lowering, it simply
7562 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7563 ArrayRef<int> Mask, SDValue V1,
7564 SDValue V2, SelectionDAG &DAG) {
7565 SDValue LowV = V1, HighV = V2;
7566 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7569 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7571 if (NumV2Elements == 1) {
7573 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7576 // Compute the index adjacent to V2Index and in the same half by toggling
7578 int V2AdjIndex = V2Index ^ 1;
7580 if (Mask[V2AdjIndex] == -1) {
7581 // Handles all the cases where we have a single V2 element and an undef.
7582 // This will only ever happen in the high lanes because we commute the
7583 // vector otherwise.
7585 std::swap(LowV, HighV);
7586 NewMask[V2Index] -= 4;
7588 // Handle the case where the V2 element ends up adjacent to a V1 element.
7589 // To make this work, blend them together as the first step.
7590 int V1Index = V2AdjIndex;
7591 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7592 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7593 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7595 // Now proceed to reconstruct the final blend as we have the necessary
7596 // high or low half formed.
7603 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7604 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7606 } else if (NumV2Elements == 2) {
7607 if (Mask[0] < 4 && Mask[1] < 4) {
7608 // Handle the easy case where we have V1 in the low lanes and V2 in the
7612 } else if (Mask[2] < 4 && Mask[3] < 4) {
7613 // We also handle the reversed case because this utility may get called
7614 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7615 // arrange things in the right direction.
7621 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7622 // trying to place elements directly, just blend them and set up the final
7623 // shuffle to place them.
7625 // The first two blend mask elements are for V1, the second two are for
7627 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7628 Mask[2] < 4 ? Mask[2] : Mask[3],
7629 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7630 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7631 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7632 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7634 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7637 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7638 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7639 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7640 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7643 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7644 getV4X86ShuffleImm8ForMask(NewMask, DAG));
7647 /// \brief Lower 4-lane 32-bit floating point shuffles.
7649 /// Uses instructions exclusively from the floating point unit to minimize
7650 /// domain crossing penalties, as these are sufficient to implement all v4f32
7652 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7653 const X86Subtarget *Subtarget,
7654 SelectionDAG &DAG) {
7656 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7657 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7658 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7660 ArrayRef<int> Mask = SVOp->getMask();
7661 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7664 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7666 if (NumV2Elements == 0) {
7667 // Check for being able to broadcast a single element.
7668 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7669 Mask, Subtarget, DAG))
7672 // Use even/odd duplicate instructions for masks that match their pattern.
7673 if (Subtarget->hasSSE3()) {
7674 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7675 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7676 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7677 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7680 if (Subtarget->hasAVX()) {
7681 // If we have AVX, we can use VPERMILPS which will allow folding a load
7682 // into the shuffle.
7683 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7684 getV4X86ShuffleImm8ForMask(Mask, DAG));
7687 // Otherwise, use a straight shuffle of a single input vector. We pass the
7688 // input vector to both operands to simulate this with a SHUFPS.
7689 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7690 getV4X86ShuffleImm8ForMask(Mask, DAG));
7693 // There are special ways we can lower some single-element blends. However, we
7694 // have custom ways we can lower more complex single-element blends below that
7695 // we defer to if both this and BLENDPS fail to match, so restrict this to
7696 // when the V2 input is targeting element 0 of the mask -- that is the fast
7698 if (NumV2Elements == 1 && Mask[0] >= 4)
7699 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7700 Mask, Subtarget, DAG))
7703 if (Subtarget->hasSSE41()) {
7704 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7708 // Use INSERTPS if we can complete the shuffle efficiently.
7709 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7712 if (!isSingleSHUFPSMask(Mask))
7713 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7714 DL, MVT::v4f32, V1, V2, Mask, DAG))
7718 // Use dedicated unpack instructions for masks that match their pattern.
7719 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7720 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7721 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7722 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7723 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7724 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7725 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7726 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7728 // Otherwise fall back to a SHUFPS lowering strategy.
7729 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7732 /// \brief Lower 4-lane i32 vector shuffles.
7734 /// We try to handle these with integer-domain shuffles where we can, but for
7735 /// blends we use the floating point domain blend instructions.
7736 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7737 const X86Subtarget *Subtarget,
7738 SelectionDAG &DAG) {
7740 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7741 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7742 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7743 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7744 ArrayRef<int> Mask = SVOp->getMask();
7745 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7747 // Whenever we can lower this as a zext, that instruction is strictly faster
7748 // than any alternative. It also allows us to fold memory operands into the
7749 // shuffle in many cases.
7750 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7751 Mask, Subtarget, DAG))
7755 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7757 if (NumV2Elements == 0) {
7758 // Check for being able to broadcast a single element.
7759 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7760 Mask, Subtarget, DAG))
7763 // Straight shuffle of a single input vector. For everything from SSE2
7764 // onward this has a single fast instruction with no scary immediates.
7765 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7766 // but we aren't actually going to use the UNPCK instruction because doing
7767 // so prevents folding a load into this instruction or making a copy.
7768 const int UnpackLoMask[] = {0, 0, 1, 1};
7769 const int UnpackHiMask[] = {2, 2, 3, 3};
7770 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7771 Mask = UnpackLoMask;
7772 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7773 Mask = UnpackHiMask;
7775 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7776 getV4X86ShuffleImm8ForMask(Mask, DAG));
7779 // Try to use shift instructions.
7781 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7784 // There are special ways we can lower some single-element blends.
7785 if (NumV2Elements == 1)
7786 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7787 Mask, Subtarget, DAG))
7790 // We have different paths for blend lowering, but they all must use the
7791 // *exact* same predicate.
7792 bool IsBlendSupported = Subtarget->hasSSE41();
7793 if (IsBlendSupported)
7794 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7798 if (SDValue Masked =
7799 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7802 // Use dedicated unpack instructions for masks that match their pattern.
7803 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7804 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7805 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7806 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7807 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7808 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7809 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7810 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7812 // Try to use byte rotation instructions.
7813 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7814 if (Subtarget->hasSSSE3())
7815 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7816 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7819 // If we have direct support for blends, we should lower by decomposing into
7820 // a permute. That will be faster than the domain cross.
7821 if (IsBlendSupported)
7822 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7825 // Try to lower by permuting the inputs into an unpack instruction.
7826 if (SDValue Unpack =
7827 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7830 // We implement this with SHUFPS because it can blend from two vectors.
7831 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7832 // up the inputs, bypassing domain shift penalties that we would encur if we
7833 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7835 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7836 DAG.getVectorShuffle(
7838 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7839 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7842 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7843 /// shuffle lowering, and the most complex part.
7845 /// The lowering strategy is to try to form pairs of input lanes which are
7846 /// targeted at the same half of the final vector, and then use a dword shuffle
7847 /// to place them onto the right half, and finally unpack the paired lanes into
7848 /// their final position.
7850 /// The exact breakdown of how to form these dword pairs and align them on the
7851 /// correct sides is really tricky. See the comments within the function for
7852 /// more of the details.
7854 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
7855 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
7856 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
7857 /// vector, form the analogous 128-bit 8-element Mask.
7858 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7859 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
7860 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7861 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
7862 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
7864 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
7865 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7866 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7868 SmallVector<int, 4> LoInputs;
7869 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7870 [](int M) { return M >= 0; });
7871 std::sort(LoInputs.begin(), LoInputs.end());
7872 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7873 SmallVector<int, 4> HiInputs;
7874 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7875 [](int M) { return M >= 0; });
7876 std::sort(HiInputs.begin(), HiInputs.end());
7877 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7879 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7880 int NumHToL = LoInputs.size() - NumLToL;
7882 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7883 int NumHToH = HiInputs.size() - NumLToH;
7884 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7885 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7886 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7887 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7889 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7890 // such inputs we can swap two of the dwords across the half mark and end up
7891 // with <=2 inputs to each half in each half. Once there, we can fall through
7892 // to the generic code below. For example:
7894 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7895 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7897 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7898 // and an existing 2-into-2 on the other half. In this case we may have to
7899 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7900 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7901 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7902 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7903 // half than the one we target for fixing) will be fixed when we re-enter this
7904 // path. We will also combine away any sequence of PSHUFD instructions that
7905 // result into a single instruction. Here is an example of the tricky case:
7907 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7908 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7910 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7912 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7913 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7915 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7916 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7918 // The result is fine to be handled by the generic logic.
7919 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
7920 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
7921 int AOffset, int BOffset) {
7922 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
7923 "Must call this with A having 3 or 1 inputs from the A half.");
7924 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
7925 "Must call this with B having 1 or 3 inputs from the B half.");
7926 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
7927 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
7929 // Compute the index of dword with only one word among the three inputs in
7930 // a half by taking the sum of the half with three inputs and subtracting
7931 // the sum of the actual three inputs. The difference is the remaining
7934 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
7935 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
7936 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
7937 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
7938 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
7939 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
7940 int TripleNonInputIdx =
7941 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
7942 TripleDWord = TripleNonInputIdx / 2;
7944 // We use xor with one to compute the adjacent DWord to whichever one the
7946 OneInputDWord = (OneInput / 2) ^ 1;
7948 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
7949 // and BToA inputs. If there is also such a problem with the BToB and AToB
7950 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
7951 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
7952 // is essential that we don't *create* a 3<-1 as then we might oscillate.
7953 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
7954 // Compute how many inputs will be flipped by swapping these DWords. We
7956 // to balance this to ensure we don't form a 3-1 shuffle in the other
7958 int NumFlippedAToBInputs =
7959 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
7960 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
7961 int NumFlippedBToBInputs =
7962 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
7963 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
7964 if ((NumFlippedAToBInputs == 1 &&
7965 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
7966 (NumFlippedBToBInputs == 1 &&
7967 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
7968 // We choose whether to fix the A half or B half based on whether that
7969 // half has zero flipped inputs. At zero, we may not be able to fix it
7970 // with that half. We also bias towards fixing the B half because that
7971 // will more commonly be the high half, and we have to bias one way.
7972 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
7973 ArrayRef<int> Inputs) {
7974 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
7975 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
7976 PinnedIdx ^ 1) != Inputs.end();
7977 // Determine whether the free index is in the flipped dword or the
7978 // unflipped dword based on where the pinned index is. We use this bit
7979 // in an xor to conditionally select the adjacent dword.
7980 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
7981 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7982 FixFreeIdx) != Inputs.end();
7983 if (IsFixIdxInput == IsFixFreeIdxInput)
7985 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7986 FixFreeIdx) != Inputs.end();
7987 assert(IsFixIdxInput != IsFixFreeIdxInput &&
7988 "We need to be changing the number of flipped inputs!");
7989 int PSHUFHalfMask[] = {0, 1, 2, 3};
7990 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
7991 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
7993 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
7996 if (M != -1 && M == FixIdx)
7998 else if (M != -1 && M == FixFreeIdx)
8001 if (NumFlippedBToBInputs != 0) {
8003 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8004 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8006 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8008 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8009 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8014 int PSHUFDMask[] = {0, 1, 2, 3};
8015 PSHUFDMask[ADWord] = BDWord;
8016 PSHUFDMask[BDWord] = ADWord;
8017 V = DAG.getNode(ISD::BITCAST, DL, VT,
8018 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8019 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8020 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8022 // Adjust the mask to match the new locations of A and B.
8024 if (M != -1 && M/2 == ADWord)
8025 M = 2 * BDWord + M % 2;
8026 else if (M != -1 && M/2 == BDWord)
8027 M = 2 * ADWord + M % 2;
8029 // Recurse back into this routine to re-compute state now that this isn't
8030 // a 3 and 1 problem.
8031 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8034 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8035 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8036 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8037 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8039 // At this point there are at most two inputs to the low and high halves from
8040 // each half. That means the inputs can always be grouped into dwords and
8041 // those dwords can then be moved to the correct half with a dword shuffle.
8042 // We use at most one low and one high word shuffle to collect these paired
8043 // inputs into dwords, and finally a dword shuffle to place them.
8044 int PSHUFLMask[4] = {-1, -1, -1, -1};
8045 int PSHUFHMask[4] = {-1, -1, -1, -1};
8046 int PSHUFDMask[4] = {-1, -1, -1, -1};
8048 // First fix the masks for all the inputs that are staying in their
8049 // original halves. This will then dictate the targets of the cross-half
8051 auto fixInPlaceInputs =
8052 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8053 MutableArrayRef<int> SourceHalfMask,
8054 MutableArrayRef<int> HalfMask, int HalfOffset) {
8055 if (InPlaceInputs.empty())
8057 if (InPlaceInputs.size() == 1) {
8058 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8059 InPlaceInputs[0] - HalfOffset;
8060 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8063 if (IncomingInputs.empty()) {
8064 // Just fix all of the in place inputs.
8065 for (int Input : InPlaceInputs) {
8066 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8067 PSHUFDMask[Input / 2] = Input / 2;
8072 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8073 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8074 InPlaceInputs[0] - HalfOffset;
8075 // Put the second input next to the first so that they are packed into
8076 // a dword. We find the adjacent index by toggling the low bit.
8077 int AdjIndex = InPlaceInputs[0] ^ 1;
8078 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8079 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8080 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8082 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8083 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8085 // Now gather the cross-half inputs and place them into a free dword of
8086 // their target half.
8087 // FIXME: This operation could almost certainly be simplified dramatically to
8088 // look more like the 3-1 fixing operation.
8089 auto moveInputsToRightHalf = [&PSHUFDMask](
8090 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8091 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8092 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8094 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8095 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8097 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8099 int LowWord = Word & ~1;
8100 int HighWord = Word | 1;
8101 return isWordClobbered(SourceHalfMask, LowWord) ||
8102 isWordClobbered(SourceHalfMask, HighWord);
8105 if (IncomingInputs.empty())
8108 if (ExistingInputs.empty()) {
8109 // Map any dwords with inputs from them into the right half.
8110 for (int Input : IncomingInputs) {
8111 // If the source half mask maps over the inputs, turn those into
8112 // swaps and use the swapped lane.
8113 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8114 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8115 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8116 Input - SourceOffset;
8117 // We have to swap the uses in our half mask in one sweep.
8118 for (int &M : HalfMask)
8119 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8121 else if (M == Input)
8122 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8124 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8125 Input - SourceOffset &&
8126 "Previous placement doesn't match!");
8128 // Note that this correctly re-maps both when we do a swap and when
8129 // we observe the other side of the swap above. We rely on that to
8130 // avoid swapping the members of the input list directly.
8131 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8134 // Map the input's dword into the correct half.
8135 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8136 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8138 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8140 "Previous placement doesn't match!");
8143 // And just directly shift any other-half mask elements to be same-half
8144 // as we will have mirrored the dword containing the element into the
8145 // same position within that half.
8146 for (int &M : HalfMask)
8147 if (M >= SourceOffset && M < SourceOffset + 4) {
8148 M = M - SourceOffset + DestOffset;
8149 assert(M >= 0 && "This should never wrap below zero!");
8154 // Ensure we have the input in a viable dword of its current half. This
8155 // is particularly tricky because the original position may be clobbered
8156 // by inputs being moved and *staying* in that half.
8157 if (IncomingInputs.size() == 1) {
8158 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8159 int InputFixed = std::find(std::begin(SourceHalfMask),
8160 std::end(SourceHalfMask), -1) -
8161 std::begin(SourceHalfMask) + SourceOffset;
8162 SourceHalfMask[InputFixed - SourceOffset] =
8163 IncomingInputs[0] - SourceOffset;
8164 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8166 IncomingInputs[0] = InputFixed;
8168 } else if (IncomingInputs.size() == 2) {
8169 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8170 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8171 // We have two non-adjacent or clobbered inputs we need to extract from
8172 // the source half. To do this, we need to map them into some adjacent
8173 // dword slot in the source mask.
8174 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8175 IncomingInputs[1] - SourceOffset};
8177 // If there is a free slot in the source half mask adjacent to one of
8178 // the inputs, place the other input in it. We use (Index XOR 1) to
8179 // compute an adjacent index.
8180 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8181 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8182 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8183 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8184 InputsFixed[1] = InputsFixed[0] ^ 1;
8185 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8186 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8187 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8188 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8189 InputsFixed[0] = InputsFixed[1] ^ 1;
8190 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8191 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8192 // The two inputs are in the same DWord but it is clobbered and the
8193 // adjacent DWord isn't used at all. Move both inputs to the free
8195 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8196 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8197 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8198 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8200 // The only way we hit this point is if there is no clobbering
8201 // (because there are no off-half inputs to this half) and there is no
8202 // free slot adjacent to one of the inputs. In this case, we have to
8203 // swap an input with a non-input.
8204 for (int i = 0; i < 4; ++i)
8205 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8206 "We can't handle any clobbers here!");
8207 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8208 "Cannot have adjacent inputs here!");
8210 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8211 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8213 // We also have to update the final source mask in this case because
8214 // it may need to undo the above swap.
8215 for (int &M : FinalSourceHalfMask)
8216 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8217 M = InputsFixed[1] + SourceOffset;
8218 else if (M == InputsFixed[1] + SourceOffset)
8219 M = (InputsFixed[0] ^ 1) + SourceOffset;
8221 InputsFixed[1] = InputsFixed[0] ^ 1;
8224 // Point everything at the fixed inputs.
8225 for (int &M : HalfMask)
8226 if (M == IncomingInputs[0])
8227 M = InputsFixed[0] + SourceOffset;
8228 else if (M == IncomingInputs[1])
8229 M = InputsFixed[1] + SourceOffset;
8231 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8232 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8235 llvm_unreachable("Unhandled input size!");
8238 // Now hoist the DWord down to the right half.
8239 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8240 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8241 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8242 for (int &M : HalfMask)
8243 for (int Input : IncomingInputs)
8245 M = FreeDWord * 2 + Input % 2;
8247 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8248 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8249 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8250 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8252 // Now enact all the shuffles we've computed to move the inputs into their
8254 if (!isNoopShuffleMask(PSHUFLMask))
8255 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8256 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
8257 if (!isNoopShuffleMask(PSHUFHMask))
8258 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8259 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
8260 if (!isNoopShuffleMask(PSHUFDMask))
8261 V = DAG.getNode(ISD::BITCAST, DL, VT,
8262 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8263 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8264 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8266 // At this point, each half should contain all its inputs, and we can then
8267 // just shuffle them into their final position.
8268 assert(std::count_if(LoMask.begin(), LoMask.end(),
8269 [](int M) { return M >= 4; }) == 0 &&
8270 "Failed to lift all the high half inputs to the low mask!");
8271 assert(std::count_if(HiMask.begin(), HiMask.end(),
8272 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8273 "Failed to lift all the low half inputs to the high mask!");
8275 // Do a half shuffle for the low mask.
8276 if (!isNoopShuffleMask(LoMask))
8277 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8278 getV4X86ShuffleImm8ForMask(LoMask, DAG));
8280 // Do a half shuffle with the high mask after shifting its values down.
8281 for (int &M : HiMask)
8284 if (!isNoopShuffleMask(HiMask))
8285 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8286 getV4X86ShuffleImm8ForMask(HiMask, DAG));
8291 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8292 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8293 SDValue V2, ArrayRef<int> Mask,
8294 SelectionDAG &DAG, bool &V1InUse,
8296 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8302 int Size = Mask.size();
8303 int Scale = 16 / Size;
8304 for (int i = 0; i < 16; ++i) {
8305 if (Mask[i / Scale] == -1) {
8306 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8308 const int ZeroMask = 0x80;
8309 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8311 int V2Idx = Mask[i / Scale] < Size
8313 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8314 if (Zeroable[i / Scale])
8315 V1Idx = V2Idx = ZeroMask;
8316 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
8317 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
8318 V1InUse |= (ZeroMask != V1Idx);
8319 V2InUse |= (ZeroMask != V2Idx);
8324 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8325 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8326 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8328 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8329 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8330 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8332 // If we need shuffled inputs from both, blend the two.
8334 if (V1InUse && V2InUse)
8335 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8337 V = V1InUse ? V1 : V2;
8339 // Cast the result back to the correct type.
8340 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8343 /// \brief Generic lowering of 8-lane i16 shuffles.
8345 /// This handles both single-input shuffles and combined shuffle/blends with
8346 /// two inputs. The single input shuffles are immediately delegated to
8347 /// a dedicated lowering routine.
8349 /// The blends are lowered in one of three fundamental ways. If there are few
8350 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8351 /// of the input is significantly cheaper when lowered as an interleaving of
8352 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8353 /// halves of the inputs separately (making them have relatively few inputs)
8354 /// and then concatenate them.
8355 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8356 const X86Subtarget *Subtarget,
8357 SelectionDAG &DAG) {
8359 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8360 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8361 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8362 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8363 ArrayRef<int> OrigMask = SVOp->getMask();
8364 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8365 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8366 MutableArrayRef<int> Mask(MaskStorage);
8368 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8370 // Whenever we can lower this as a zext, that instruction is strictly faster
8371 // than any alternative.
8372 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8373 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8376 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8378 auto isV2 = [](int M) { return M >= 8; };
8380 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8382 if (NumV2Inputs == 0) {
8383 // Check for being able to broadcast a single element.
8384 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8385 Mask, Subtarget, DAG))
8388 // Try to use shift instructions.
8390 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8393 // Use dedicated unpack instructions for masks that match their pattern.
8394 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8395 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8396 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8397 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8399 // Try to use byte rotation instructions.
8400 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8401 Mask, Subtarget, DAG))
8404 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8408 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8409 "All single-input shuffles should be canonicalized to be V1-input "
8412 // Try to use shift instructions.
8414 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8417 // There are special ways we can lower some single-element blends.
8418 if (NumV2Inputs == 1)
8419 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8420 Mask, Subtarget, DAG))
8423 // We have different paths for blend lowering, but they all must use the
8424 // *exact* same predicate.
8425 bool IsBlendSupported = Subtarget->hasSSE41();
8426 if (IsBlendSupported)
8427 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8431 if (SDValue Masked =
8432 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8435 // Use dedicated unpack instructions for masks that match their pattern.
8436 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8437 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8438 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8439 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8441 // Try to use byte rotation instructions.
8442 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8443 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8446 if (SDValue BitBlend =
8447 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8450 if (SDValue Unpack =
8451 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8454 // If we can't directly blend but can use PSHUFB, that will be better as it
8455 // can both shuffle and set up the inefficient blend.
8456 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8457 bool V1InUse, V2InUse;
8458 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8462 // We can always bit-blend if we have to so the fallback strategy is to
8463 // decompose into single-input permutes and blends.
8464 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8468 /// \brief Check whether a compaction lowering can be done by dropping even
8469 /// elements and compute how many times even elements must be dropped.
8471 /// This handles shuffles which take every Nth element where N is a power of
8472 /// two. Example shuffle masks:
8474 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8475 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8476 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8477 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8478 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8479 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8481 /// Any of these lanes can of course be undef.
8483 /// This routine only supports N <= 3.
8484 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8487 /// \returns N above, or the number of times even elements must be dropped if
8488 /// there is such a number. Otherwise returns zero.
8489 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8490 // Figure out whether we're looping over two inputs or just one.
8491 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8493 // The modulus for the shuffle vector entries is based on whether this is
8494 // a single input or not.
8495 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8496 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8497 "We should only be called with masks with a power-of-2 size!");
8499 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8501 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8502 // and 2^3 simultaneously. This is because we may have ambiguity with
8503 // partially undef inputs.
8504 bool ViableForN[3] = {true, true, true};
8506 for (int i = 0, e = Mask.size(); i < e; ++i) {
8507 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8512 bool IsAnyViable = false;
8513 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8514 if (ViableForN[j]) {
8517 // The shuffle mask must be equal to (i * 2^N) % M.
8518 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8521 ViableForN[j] = false;
8523 // Early exit if we exhaust the possible powers of two.
8528 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8532 // Return 0 as there is no viable power of two.
8536 /// \brief Generic lowering of v16i8 shuffles.
8538 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8539 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8540 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8541 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8543 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8544 const X86Subtarget *Subtarget,
8545 SelectionDAG &DAG) {
8547 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8548 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8549 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8550 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8551 ArrayRef<int> Mask = SVOp->getMask();
8552 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8554 // Try to use shift instructions.
8556 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8559 // Try to use byte rotation instructions.
8560 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8561 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8564 // Try to use a zext lowering.
8565 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8566 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8570 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8572 // For single-input shuffles, there are some nicer lowering tricks we can use.
8573 if (NumV2Elements == 0) {
8574 // Check for being able to broadcast a single element.
8575 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8576 Mask, Subtarget, DAG))
8579 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8580 // Notably, this handles splat and partial-splat shuffles more efficiently.
8581 // However, it only makes sense if the pre-duplication shuffle simplifies
8582 // things significantly. Currently, this means we need to be able to
8583 // express the pre-duplication shuffle as an i16 shuffle.
8585 // FIXME: We should check for other patterns which can be widened into an
8586 // i16 shuffle as well.
8587 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8588 for (int i = 0; i < 16; i += 2)
8589 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8594 auto tryToWidenViaDuplication = [&]() -> SDValue {
8595 if (!canWidenViaDuplication(Mask))
8597 SmallVector<int, 4> LoInputs;
8598 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8599 [](int M) { return M >= 0 && M < 8; });
8600 std::sort(LoInputs.begin(), LoInputs.end());
8601 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8603 SmallVector<int, 4> HiInputs;
8604 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8605 [](int M) { return M >= 8; });
8606 std::sort(HiInputs.begin(), HiInputs.end());
8607 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8610 bool TargetLo = LoInputs.size() >= HiInputs.size();
8611 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8612 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8614 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8615 SmallDenseMap<int, int, 8> LaneMap;
8616 for (int I : InPlaceInputs) {
8617 PreDupI16Shuffle[I/2] = I/2;
8620 int j = TargetLo ? 0 : 4, je = j + 4;
8621 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8622 // Check if j is already a shuffle of this input. This happens when
8623 // there are two adjacent bytes after we move the low one.
8624 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8625 // If we haven't yet mapped the input, search for a slot into which
8627 while (j < je && PreDupI16Shuffle[j] != -1)
8631 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8634 // Map this input with the i16 shuffle.
8635 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8638 // Update the lane map based on the mapping we ended up with.
8639 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8642 ISD::BITCAST, DL, MVT::v16i8,
8643 DAG.getVectorShuffle(MVT::v8i16, DL,
8644 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8645 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8647 // Unpack the bytes to form the i16s that will be shuffled into place.
8648 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8649 MVT::v16i8, V1, V1);
8651 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8652 for (int i = 0; i < 16; ++i)
8653 if (Mask[i] != -1) {
8654 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8655 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8656 if (PostDupI16Shuffle[i / 2] == -1)
8657 PostDupI16Shuffle[i / 2] = MappedMask;
8659 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8660 "Conflicting entrties in the original shuffle!");
8663 ISD::BITCAST, DL, MVT::v16i8,
8664 DAG.getVectorShuffle(MVT::v8i16, DL,
8665 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8666 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8668 if (SDValue V = tryToWidenViaDuplication())
8672 // Use dedicated unpack instructions for masks that match their pattern.
8673 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8674 0, 16, 1, 17, 2, 18, 3, 19,
8676 4, 20, 5, 21, 6, 22, 7, 23}))
8677 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8678 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8679 8, 24, 9, 25, 10, 26, 11, 27,
8681 12, 28, 13, 29, 14, 30, 15, 31}))
8682 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8684 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8685 // with PSHUFB. It is important to do this before we attempt to generate any
8686 // blends but after all of the single-input lowerings. If the single input
8687 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8688 // want to preserve that and we can DAG combine any longer sequences into
8689 // a PSHUFB in the end. But once we start blending from multiple inputs,
8690 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8691 // and there are *very* few patterns that would actually be faster than the
8692 // PSHUFB approach because of its ability to zero lanes.
8694 // FIXME: The only exceptions to the above are blends which are exact
8695 // interleavings with direct instructions supporting them. We currently don't
8696 // handle those well here.
8697 if (Subtarget->hasSSSE3()) {
8698 bool V1InUse = false;
8699 bool V2InUse = false;
8701 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8702 DAG, V1InUse, V2InUse);
8704 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8705 // do so. This avoids using them to handle blends-with-zero which is
8706 // important as a single pshufb is significantly faster for that.
8707 if (V1InUse && V2InUse) {
8708 if (Subtarget->hasSSE41())
8709 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8710 Mask, Subtarget, DAG))
8713 // We can use an unpack to do the blending rather than an or in some
8714 // cases. Even though the or may be (very minorly) more efficient, we
8715 // preference this lowering because there are common cases where part of
8716 // the complexity of the shuffles goes away when we do the final blend as
8718 // FIXME: It might be worth trying to detect if the unpack-feeding
8719 // shuffles will both be pshufb, in which case we shouldn't bother with
8721 if (SDValue Unpack =
8722 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8729 // There are special ways we can lower some single-element blends.
8730 if (NumV2Elements == 1)
8731 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8732 Mask, Subtarget, DAG))
8735 if (SDValue BitBlend =
8736 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8739 // Check whether a compaction lowering can be done. This handles shuffles
8740 // which take every Nth element for some even N. See the helper function for
8743 // We special case these as they can be particularly efficiently handled with
8744 // the PACKUSB instruction on x86 and they show up in common patterns of
8745 // rearranging bytes to truncate wide elements.
8746 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8747 // NumEvenDrops is the power of two stride of the elements. Another way of
8748 // thinking about it is that we need to drop the even elements this many
8749 // times to get the original input.
8750 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8752 // First we need to zero all the dropped bytes.
8753 assert(NumEvenDrops <= 3 &&
8754 "No support for dropping even elements more than 3 times.");
8755 // We use the mask type to pick which bytes are preserved based on how many
8756 // elements are dropped.
8757 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8758 SDValue ByteClearMask =
8759 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8760 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
8761 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8763 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8765 // Now pack things back together.
8766 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8767 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8768 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8769 for (int i = 1; i < NumEvenDrops; ++i) {
8770 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8771 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8777 // Handle multi-input cases by blending single-input shuffles.
8778 if (NumV2Elements > 0)
8779 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8782 // The fallback path for single-input shuffles widens this into two v8i16
8783 // vectors with unpacks, shuffles those, and then pulls them back together
8787 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8788 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8789 for (int i = 0; i < 16; ++i)
8791 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8793 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8795 SDValue VLoHalf, VHiHalf;
8796 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8797 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8799 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8800 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8801 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8802 [](int M) { return M >= 0 && M % 2 == 1; })) {
8803 // Use a mask to drop the high bytes.
8804 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8805 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8806 DAG.getConstant(0x00FF, MVT::v8i16));
8808 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8809 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8811 // Squash the masks to point directly into VLoHalf.
8812 for (int &M : LoBlendMask)
8815 for (int &M : HiBlendMask)
8819 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8820 // VHiHalf so that we can blend them as i16s.
8821 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8822 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8823 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8824 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8827 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8828 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8830 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8833 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8835 /// This routine breaks down the specific type of 128-bit shuffle and
8836 /// dispatches to the lowering routines accordingly.
8837 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8838 MVT VT, const X86Subtarget *Subtarget,
8839 SelectionDAG &DAG) {
8840 switch (VT.SimpleTy) {
8842 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8844 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8846 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8848 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8850 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8852 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8855 llvm_unreachable("Unimplemented!");
8859 /// \brief Helper function to test whether a shuffle mask could be
8860 /// simplified by widening the elements being shuffled.
8862 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8863 /// leaves it in an unspecified state.
8865 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8866 /// shuffle masks. The latter have the special property of a '-2' representing
8867 /// a zero-ed lane of a vector.
8868 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8869 SmallVectorImpl<int> &WidenedMask) {
8870 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8871 // If both elements are undef, its trivial.
8872 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8873 WidenedMask.push_back(SM_SentinelUndef);
8877 // Check for an undef mask and a mask value properly aligned to fit with
8878 // a pair of values. If we find such a case, use the non-undef mask's value.
8879 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8880 WidenedMask.push_back(Mask[i + 1] / 2);
8883 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8884 WidenedMask.push_back(Mask[i] / 2);
8888 // When zeroing, we need to spread the zeroing across both lanes to widen.
8889 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8890 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8891 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8892 WidenedMask.push_back(SM_SentinelZero);
8898 // Finally check if the two mask values are adjacent and aligned with
8900 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8901 WidenedMask.push_back(Mask[i] / 2);
8905 // Otherwise we can't safely widen the elements used in this shuffle.
8908 assert(WidenedMask.size() == Mask.size() / 2 &&
8909 "Incorrect size of mask after widening the elements!");
8914 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8916 /// This routine just extracts two subvectors, shuffles them independently, and
8917 /// then concatenates them back together. This should work effectively with all
8918 /// AVX vector shuffle types.
8919 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8920 SDValue V2, ArrayRef<int> Mask,
8921 SelectionDAG &DAG) {
8922 assert(VT.getSizeInBits() >= 256 &&
8923 "Only for 256-bit or wider vector shuffles!");
8924 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
8925 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
8927 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
8928 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
8930 int NumElements = VT.getVectorNumElements();
8931 int SplitNumElements = NumElements / 2;
8932 MVT ScalarVT = VT.getScalarType();
8933 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
8935 // Rather than splitting build-vectors, just build two narrower build
8936 // vectors. This helps shuffling with splats and zeros.
8937 auto SplitVector = [&](SDValue V) {
8938 while (V.getOpcode() == ISD::BITCAST)
8939 V = V->getOperand(0);
8941 MVT OrigVT = V.getSimpleValueType();
8942 int OrigNumElements = OrigVT.getVectorNumElements();
8943 int OrigSplitNumElements = OrigNumElements / 2;
8944 MVT OrigScalarVT = OrigVT.getScalarType();
8945 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
8949 auto *BV = dyn_cast<BuildVectorSDNode>(V);
8951 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8952 DAG.getIntPtrConstant(0));
8953 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8954 DAG.getIntPtrConstant(OrigSplitNumElements));
8957 SmallVector<SDValue, 16> LoOps, HiOps;
8958 for (int i = 0; i < OrigSplitNumElements; ++i) {
8959 LoOps.push_back(BV->getOperand(i));
8960 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
8962 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
8963 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
8965 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
8966 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
8969 SDValue LoV1, HiV1, LoV2, HiV2;
8970 std::tie(LoV1, HiV1) = SplitVector(V1);
8971 std::tie(LoV2, HiV2) = SplitVector(V2);
8973 // Now create two 4-way blends of these half-width vectors.
8974 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
8975 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
8976 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
8977 for (int i = 0; i < SplitNumElements; ++i) {
8978 int M = HalfMask[i];
8979 if (M >= NumElements) {
8980 if (M >= NumElements + SplitNumElements)
8984 V2BlendMask.push_back(M - NumElements);
8985 V1BlendMask.push_back(-1);
8986 BlendMask.push_back(SplitNumElements + i);
8987 } else if (M >= 0) {
8988 if (M >= SplitNumElements)
8992 V2BlendMask.push_back(-1);
8993 V1BlendMask.push_back(M);
8994 BlendMask.push_back(i);
8996 V2BlendMask.push_back(-1);
8997 V1BlendMask.push_back(-1);
8998 BlendMask.push_back(-1);
9002 // Because the lowering happens after all combining takes place, we need to
9003 // manually combine these blend masks as much as possible so that we create
9004 // a minimal number of high-level vector shuffle nodes.
9006 // First try just blending the halves of V1 or V2.
9007 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9008 return DAG.getUNDEF(SplitVT);
9009 if (!UseLoV2 && !UseHiV2)
9010 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9011 if (!UseLoV1 && !UseHiV1)
9012 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9014 SDValue V1Blend, V2Blend;
9015 if (UseLoV1 && UseHiV1) {
9017 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9019 // We only use half of V1 so map the usage down into the final blend mask.
9020 V1Blend = UseLoV1 ? LoV1 : HiV1;
9021 for (int i = 0; i < SplitNumElements; ++i)
9022 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9023 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9025 if (UseLoV2 && UseHiV2) {
9027 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9029 // We only use half of V2 so map the usage down into the final blend mask.
9030 V2Blend = UseLoV2 ? LoV2 : HiV2;
9031 for (int i = 0; i < SplitNumElements; ++i)
9032 if (BlendMask[i] >= SplitNumElements)
9033 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9035 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9037 SDValue Lo = HalfBlend(LoMask);
9038 SDValue Hi = HalfBlend(HiMask);
9039 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9042 /// \brief Either split a vector in halves or decompose the shuffles and the
9045 /// This is provided as a good fallback for many lowerings of non-single-input
9046 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9047 /// between splitting the shuffle into 128-bit components and stitching those
9048 /// back together vs. extracting the single-input shuffles and blending those
9050 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9051 SDValue V2, ArrayRef<int> Mask,
9052 SelectionDAG &DAG) {
9053 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9054 "lower single-input shuffles as it "
9055 "could then recurse on itself.");
9056 int Size = Mask.size();
9058 // If this can be modeled as a broadcast of two elements followed by a blend,
9059 // prefer that lowering. This is especially important because broadcasts can
9060 // often fold with memory operands.
9061 auto DoBothBroadcast = [&] {
9062 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9065 if (V2BroadcastIdx == -1)
9066 V2BroadcastIdx = M - Size;
9067 else if (M - Size != V2BroadcastIdx)
9069 } else if (M >= 0) {
9070 if (V1BroadcastIdx == -1)
9072 else if (M != V1BroadcastIdx)
9077 if (DoBothBroadcast())
9078 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9081 // If the inputs all stem from a single 128-bit lane of each input, then we
9082 // split them rather than blending because the split will decompose to
9083 // unusually few instructions.
9084 int LaneCount = VT.getSizeInBits() / 128;
9085 int LaneSize = Size / LaneCount;
9086 SmallBitVector LaneInputs[2];
9087 LaneInputs[0].resize(LaneCount, false);
9088 LaneInputs[1].resize(LaneCount, false);
9089 for (int i = 0; i < Size; ++i)
9091 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9092 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9093 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9095 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9096 // that the decomposed single-input shuffles don't end up here.
9097 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9100 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9101 /// a permutation and blend of those lanes.
9103 /// This essentially blends the out-of-lane inputs to each lane into the lane
9104 /// from a permuted copy of the vector. This lowering strategy results in four
9105 /// instructions in the worst case for a single-input cross lane shuffle which
9106 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9107 /// of. Special cases for each particular shuffle pattern should be handled
9108 /// prior to trying this lowering.
9109 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9110 SDValue V1, SDValue V2,
9112 SelectionDAG &DAG) {
9113 // FIXME: This should probably be generalized for 512-bit vectors as well.
9114 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9115 int LaneSize = Mask.size() / 2;
9117 // If there are only inputs from one 128-bit lane, splitting will in fact be
9118 // less expensive. The flags track whether the given lane contains an element
9119 // that crosses to another lane.
9120 bool LaneCrossing[2] = {false, false};
9121 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9122 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9123 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9124 if (!LaneCrossing[0] || !LaneCrossing[1])
9125 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9127 if (isSingleInputShuffleMask(Mask)) {
9128 SmallVector<int, 32> FlippedBlendMask;
9129 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9130 FlippedBlendMask.push_back(
9131 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9133 : Mask[i] % LaneSize +
9134 (i / LaneSize) * LaneSize + Size));
9136 // Flip the vector, and blend the results which should now be in-lane. The
9137 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9138 // 5 for the high source. The value 3 selects the high half of source 2 and
9139 // the value 2 selects the low half of source 2. We only use source 2 to
9140 // allow folding it into a memory operand.
9141 unsigned PERMMask = 3 | 2 << 4;
9142 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9143 V1, DAG.getConstant(PERMMask, MVT::i8));
9144 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9147 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9148 // will be handled by the above logic and a blend of the results, much like
9149 // other patterns in AVX.
9150 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9153 /// \brief Handle lowering 2-lane 128-bit shuffles.
9154 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9155 SDValue V2, ArrayRef<int> Mask,
9156 const X86Subtarget *Subtarget,
9157 SelectionDAG &DAG) {
9158 // TODO: If minimizing size and one of the inputs is a zero vector and the
9159 // the zero vector has only one use, we could use a VPERM2X128 to save the
9160 // instruction bytes needed to explicitly generate the zero vector.
9162 // Blends are faster and handle all the non-lane-crossing cases.
9163 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9167 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9168 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9170 // If either input operand is a zero vector, use VPERM2X128 because its mask
9171 // allows us to replace the zero input with an implicit zero.
9172 if (!IsV1Zero && !IsV2Zero) {
9173 // Check for patterns which can be matched with a single insert of a 128-bit
9175 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9176 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9177 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9178 VT.getVectorNumElements() / 2);
9179 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9180 DAG.getIntPtrConstant(0));
9181 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9182 OnlyUsesV1 ? V1 : V2, DAG.getIntPtrConstant(0));
9183 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9187 // Otherwise form a 128-bit permutation. After accounting for undefs,
9188 // convert the 64-bit shuffle mask selection values into 128-bit
9189 // selection bits by dividing the indexes by 2 and shifting into positions
9190 // defined by a vperm2*128 instruction's immediate control byte.
9192 // The immediate permute control byte looks like this:
9193 // [1:0] - select 128 bits from sources for low half of destination
9195 // [3] - zero low half of destination
9196 // [5:4] - select 128 bits from sources for high half of destination
9198 // [7] - zero high half of destination
9200 int MaskLO = Mask[0];
9201 if (MaskLO == SM_SentinelUndef)
9202 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9204 int MaskHI = Mask[2];
9205 if (MaskHI == SM_SentinelUndef)
9206 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9208 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9210 // If either input is a zero vector, replace it with an undef input.
9211 // Shuffle mask values < 4 are selecting elements of V1.
9212 // Shuffle mask values >= 4 are selecting elements of V2.
9213 // Adjust each half of the permute mask by clearing the half that was
9214 // selecting the zero vector and setting the zero mask bit.
9216 V1 = DAG.getUNDEF(VT);
9218 PermMask = (PermMask & 0xf0) | 0x08;
9220 PermMask = (PermMask & 0x0f) | 0x80;
9223 V2 = DAG.getUNDEF(VT);
9225 PermMask = (PermMask & 0xf0) | 0x08;
9227 PermMask = (PermMask & 0x0f) | 0x80;
9230 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9231 DAG.getConstant(PermMask, MVT::i8));
9234 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9235 /// shuffling each lane.
9237 /// This will only succeed when the result of fixing the 128-bit lanes results
9238 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9239 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9240 /// the lane crosses early and then use simpler shuffles within each lane.
9242 /// FIXME: It might be worthwhile at some point to support this without
9243 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9244 /// in x86 only floating point has interesting non-repeating shuffles, and even
9245 /// those are still *marginally* more expensive.
9246 static SDValue lowerVectorShuffleByMerging128BitLanes(
9247 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9248 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9249 assert(!isSingleInputShuffleMask(Mask) &&
9250 "This is only useful with multiple inputs.");
9252 int Size = Mask.size();
9253 int LaneSize = 128 / VT.getScalarSizeInBits();
9254 int NumLanes = Size / LaneSize;
9255 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9257 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9258 // check whether the in-128-bit lane shuffles share a repeating pattern.
9259 SmallVector<int, 4> Lanes;
9260 Lanes.resize(NumLanes, -1);
9261 SmallVector<int, 4> InLaneMask;
9262 InLaneMask.resize(LaneSize, -1);
9263 for (int i = 0; i < Size; ++i) {
9267 int j = i / LaneSize;
9270 // First entry we've seen for this lane.
9271 Lanes[j] = Mask[i] / LaneSize;
9272 } else if (Lanes[j] != Mask[i] / LaneSize) {
9273 // This doesn't match the lane selected previously!
9277 // Check that within each lane we have a consistent shuffle mask.
9278 int k = i % LaneSize;
9279 if (InLaneMask[k] < 0) {
9280 InLaneMask[k] = Mask[i] % LaneSize;
9281 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9282 // This doesn't fit a repeating in-lane mask.
9287 // First shuffle the lanes into place.
9288 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9289 VT.getSizeInBits() / 64);
9290 SmallVector<int, 8> LaneMask;
9291 LaneMask.resize(NumLanes * 2, -1);
9292 for (int i = 0; i < NumLanes; ++i)
9293 if (Lanes[i] >= 0) {
9294 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9295 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9298 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9299 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9300 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9302 // Cast it back to the type we actually want.
9303 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9305 // Now do a simple shuffle that isn't lane crossing.
9306 SmallVector<int, 8> NewMask;
9307 NewMask.resize(Size, -1);
9308 for (int i = 0; i < Size; ++i)
9310 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9311 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9312 "Must not introduce lane crosses at this point!");
9314 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9317 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9320 /// This returns true if the elements from a particular input are already in the
9321 /// slot required by the given mask and require no permutation.
9322 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9323 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9324 int Size = Mask.size();
9325 for (int i = 0; i < Size; ++i)
9326 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9332 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9334 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9335 /// isn't available.
9336 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9337 const X86Subtarget *Subtarget,
9338 SelectionDAG &DAG) {
9340 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9341 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9342 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9343 ArrayRef<int> Mask = SVOp->getMask();
9344 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9346 SmallVector<int, 4> WidenedMask;
9347 if (canWidenShuffleElements(Mask, WidenedMask))
9348 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9351 if (isSingleInputShuffleMask(Mask)) {
9352 // Check for being able to broadcast a single element.
9353 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9354 Mask, Subtarget, DAG))
9357 // Use low duplicate instructions for masks that match their pattern.
9358 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9359 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9361 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9362 // Non-half-crossing single input shuffles can be lowerid with an
9363 // interleaved permutation.
9364 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9365 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9366 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9367 DAG.getConstant(VPERMILPMask, MVT::i8));
9370 // With AVX2 we have direct support for this permutation.
9371 if (Subtarget->hasAVX2())
9372 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9373 getV4X86ShuffleImm8ForMask(Mask, DAG));
9375 // Otherwise, fall back.
9376 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9380 // X86 has dedicated unpack instructions that can handle specific blend
9381 // operations: UNPCKH and UNPCKL.
9382 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9383 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9384 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9385 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9386 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9387 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9388 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9389 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9391 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9395 // Check if the blend happens to exactly fit that of SHUFPD.
9396 if ((Mask[0] == -1 || Mask[0] < 2) &&
9397 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9398 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9399 (Mask[3] == -1 || Mask[3] >= 6)) {
9400 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9401 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9402 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9403 DAG.getConstant(SHUFPDMask, MVT::i8));
9405 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9406 (Mask[1] == -1 || Mask[1] < 2) &&
9407 (Mask[2] == -1 || Mask[2] >= 6) &&
9408 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9409 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9410 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9411 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9412 DAG.getConstant(SHUFPDMask, MVT::i8));
9415 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9416 // shuffle. However, if we have AVX2 and either inputs are already in place,
9417 // we will be able to shuffle even across lanes the other input in a single
9418 // instruction so skip this pattern.
9419 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9420 isShuffleMaskInputInPlace(1, Mask))))
9421 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9422 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9425 // If we have AVX2 then we always want to lower with a blend because an v4 we
9426 // can fully permute the elements.
9427 if (Subtarget->hasAVX2())
9428 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9431 // Otherwise fall back on generic lowering.
9432 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9435 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9437 /// This routine is only called when we have AVX2 and thus a reasonable
9438 /// instruction set for v4i64 shuffling..
9439 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9440 const X86Subtarget *Subtarget,
9441 SelectionDAG &DAG) {
9443 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9444 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9445 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9446 ArrayRef<int> Mask = SVOp->getMask();
9447 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9448 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9450 SmallVector<int, 4> WidenedMask;
9451 if (canWidenShuffleElements(Mask, WidenedMask))
9452 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9455 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9459 // Check for being able to broadcast a single element.
9460 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9461 Mask, Subtarget, DAG))
9464 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9465 // use lower latency instructions that will operate on both 128-bit lanes.
9466 SmallVector<int, 2> RepeatedMask;
9467 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9468 if (isSingleInputShuffleMask(Mask)) {
9469 int PSHUFDMask[] = {-1, -1, -1, -1};
9470 for (int i = 0; i < 2; ++i)
9471 if (RepeatedMask[i] >= 0) {
9472 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9473 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9476 ISD::BITCAST, DL, MVT::v4i64,
9477 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9478 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9479 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9483 // AVX2 provides a direct instruction for permuting a single input across
9485 if (isSingleInputShuffleMask(Mask))
9486 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9487 getV4X86ShuffleImm8ForMask(Mask, DAG));
9489 // Try to use shift instructions.
9491 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9494 // Use dedicated unpack instructions for masks that match their pattern.
9495 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9496 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9497 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9498 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9499 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9500 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9501 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9502 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9504 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9505 // shuffle. However, if we have AVX2 and either inputs are already in place,
9506 // we will be able to shuffle even across lanes the other input in a single
9507 // instruction so skip this pattern.
9508 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9509 isShuffleMaskInputInPlace(1, Mask))))
9510 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9511 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9514 // Otherwise fall back on generic blend lowering.
9515 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9519 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9521 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9522 /// isn't available.
9523 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9524 const X86Subtarget *Subtarget,
9525 SelectionDAG &DAG) {
9527 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9528 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9529 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9530 ArrayRef<int> Mask = SVOp->getMask();
9531 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9533 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9537 // Check for being able to broadcast a single element.
9538 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9539 Mask, Subtarget, DAG))
9542 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9543 // options to efficiently lower the shuffle.
9544 SmallVector<int, 4> RepeatedMask;
9545 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9546 assert(RepeatedMask.size() == 4 &&
9547 "Repeated masks must be half the mask width!");
9549 // Use even/odd duplicate instructions for masks that match their pattern.
9550 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9551 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9552 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9553 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9555 if (isSingleInputShuffleMask(Mask))
9556 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9557 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9559 // Use dedicated unpack instructions for masks that match their pattern.
9560 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9561 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9562 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9563 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9564 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9565 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9566 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9567 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9569 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9570 // have already handled any direct blends. We also need to squash the
9571 // repeated mask into a simulated v4f32 mask.
9572 for (int i = 0; i < 4; ++i)
9573 if (RepeatedMask[i] >= 8)
9574 RepeatedMask[i] -= 4;
9575 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9578 // If we have a single input shuffle with different shuffle patterns in the
9579 // two 128-bit lanes use the variable mask to VPERMILPS.
9580 if (isSingleInputShuffleMask(Mask)) {
9581 SDValue VPermMask[8];
9582 for (int i = 0; i < 8; ++i)
9583 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9584 : DAG.getConstant(Mask[i], MVT::i32);
9585 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9587 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9588 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9590 if (Subtarget->hasAVX2())
9591 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9592 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9593 DAG.getNode(ISD::BUILD_VECTOR, DL,
9594 MVT::v8i32, VPermMask)),
9597 // Otherwise, fall back.
9598 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9602 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9604 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9605 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9608 // If we have AVX2 then we always want to lower with a blend because at v8 we
9609 // can fully permute the elements.
9610 if (Subtarget->hasAVX2())
9611 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9614 // Otherwise fall back on generic lowering.
9615 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9618 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9620 /// This routine is only called when we have AVX2 and thus a reasonable
9621 /// instruction set for v8i32 shuffling..
9622 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9623 const X86Subtarget *Subtarget,
9624 SelectionDAG &DAG) {
9626 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9627 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9628 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9629 ArrayRef<int> Mask = SVOp->getMask();
9630 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9631 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9633 // Whenever we can lower this as a zext, that instruction is strictly faster
9634 // than any alternative. It also allows us to fold memory operands into the
9635 // shuffle in many cases.
9636 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9637 Mask, Subtarget, DAG))
9640 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9644 // Check for being able to broadcast a single element.
9645 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9646 Mask, Subtarget, DAG))
9649 // If the shuffle mask is repeated in each 128-bit lane we can use more
9650 // efficient instructions that mirror the shuffles across the two 128-bit
9652 SmallVector<int, 4> RepeatedMask;
9653 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9654 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9655 if (isSingleInputShuffleMask(Mask))
9656 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9657 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9659 // Use dedicated unpack instructions for masks that match their pattern.
9660 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9661 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9662 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9663 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9664 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9665 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9666 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9667 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9670 // Try to use shift instructions.
9672 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9675 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9676 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9679 // If the shuffle patterns aren't repeated but it is a single input, directly
9680 // generate a cross-lane VPERMD instruction.
9681 if (isSingleInputShuffleMask(Mask)) {
9682 SDValue VPermMask[8];
9683 for (int i = 0; i < 8; ++i)
9684 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9685 : DAG.getConstant(Mask[i], MVT::i32);
9687 X86ISD::VPERMV, DL, MVT::v8i32,
9688 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9691 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9693 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9694 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9697 // Otherwise fall back on generic blend lowering.
9698 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9702 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9704 /// This routine is only called when we have AVX2 and thus a reasonable
9705 /// instruction set for v16i16 shuffling..
9706 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9707 const X86Subtarget *Subtarget,
9708 SelectionDAG &DAG) {
9710 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9711 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9713 ArrayRef<int> Mask = SVOp->getMask();
9714 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9715 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9717 // Whenever we can lower this as a zext, that instruction is strictly faster
9718 // than any alternative. It also allows us to fold memory operands into the
9719 // shuffle in many cases.
9720 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9721 Mask, Subtarget, DAG))
9724 // Check for being able to broadcast a single element.
9725 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9726 Mask, Subtarget, DAG))
9729 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9733 // Use dedicated unpack instructions for masks that match their pattern.
9734 if (isShuffleEquivalent(V1, V2, Mask,
9735 {// First 128-bit lane:
9736 0, 16, 1, 17, 2, 18, 3, 19,
9737 // Second 128-bit lane:
9738 8, 24, 9, 25, 10, 26, 11, 27}))
9739 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9740 if (isShuffleEquivalent(V1, V2, Mask,
9741 {// First 128-bit lane:
9742 4, 20, 5, 21, 6, 22, 7, 23,
9743 // Second 128-bit lane:
9744 12, 28, 13, 29, 14, 30, 15, 31}))
9745 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9747 // Try to use shift instructions.
9749 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9752 // Try to use byte rotation instructions.
9753 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9754 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9757 if (isSingleInputShuffleMask(Mask)) {
9758 // There are no generalized cross-lane shuffle operations available on i16
9760 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9761 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9764 SmallVector<int, 8> RepeatedMask;
9765 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
9766 // As this is a single-input shuffle, the repeated mask should be
9767 // a strictly valid v8i16 mask that we can pass through to the v8i16
9768 // lowering to handle even the v16 case.
9769 return lowerV8I16GeneralSingleInputVectorShuffle(
9770 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
9773 SDValue PSHUFBMask[32];
9774 for (int i = 0; i < 16; ++i) {
9775 if (Mask[i] == -1) {
9776 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9780 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9781 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9782 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
9783 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
9786 ISD::BITCAST, DL, MVT::v16i16,
9788 X86ISD::PSHUFB, DL, MVT::v32i8,
9789 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9790 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9793 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9795 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9796 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9799 // Otherwise fall back on generic lowering.
9800 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9803 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9805 /// This routine is only called when we have AVX2 and thus a reasonable
9806 /// instruction set for v32i8 shuffling..
9807 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9808 const X86Subtarget *Subtarget,
9809 SelectionDAG &DAG) {
9811 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9812 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9813 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9814 ArrayRef<int> Mask = SVOp->getMask();
9815 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9816 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9818 // Whenever we can lower this as a zext, that instruction is strictly faster
9819 // than any alternative. It also allows us to fold memory operands into the
9820 // shuffle in many cases.
9821 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9822 Mask, Subtarget, DAG))
9825 // Check for being able to broadcast a single element.
9826 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9827 Mask, Subtarget, DAG))
9830 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9834 // Use dedicated unpack instructions for masks that match their pattern.
9835 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9837 if (isShuffleEquivalent(
9839 {// First 128-bit lane:
9840 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9841 // Second 128-bit lane:
9842 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9843 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9844 if (isShuffleEquivalent(
9846 {// First 128-bit lane:
9847 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9848 // Second 128-bit lane:
9849 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9850 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9852 // Try to use shift instructions.
9854 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9857 // Try to use byte rotation instructions.
9858 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9859 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9862 if (isSingleInputShuffleMask(Mask)) {
9863 // There are no generalized cross-lane shuffle operations available on i8
9865 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9866 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9869 SDValue PSHUFBMask[32];
9870 for (int i = 0; i < 32; ++i)
9873 ? DAG.getUNDEF(MVT::i8)
9874 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
9877 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9878 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9881 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9883 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9884 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9887 // Otherwise fall back on generic lowering.
9888 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9891 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9893 /// This routine either breaks down the specific type of a 256-bit x86 vector
9894 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9895 /// together based on the available instructions.
9896 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9897 MVT VT, const X86Subtarget *Subtarget,
9898 SelectionDAG &DAG) {
9900 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9901 ArrayRef<int> Mask = SVOp->getMask();
9903 // If we have a single input to the zero element, insert that into V1 if we
9904 // can do so cheaply.
9905 int NumElts = VT.getVectorNumElements();
9906 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
9907 return M >= NumElts;
9910 if (NumV2Elements == 1 && Mask[0] >= NumElts)
9911 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9912 DL, VT, V1, V2, Mask, Subtarget, DAG))
9915 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9916 // check for those subtargets here and avoid much of the subtarget querying in
9917 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9918 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9919 // floating point types there eventually, just immediately cast everything to
9920 // a float and operate entirely in that domain.
9921 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9922 int ElementBits = VT.getScalarSizeInBits();
9923 if (ElementBits < 32)
9924 // No floating point type available, decompose into 128-bit vectors.
9925 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9927 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9928 VT.getVectorNumElements());
9929 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9930 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9931 return DAG.getNode(ISD::BITCAST, DL, VT,
9932 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9935 switch (VT.SimpleTy) {
9937 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9939 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9941 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9943 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9945 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9947 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9950 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9954 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
9955 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9956 const X86Subtarget *Subtarget,
9957 SelectionDAG &DAG) {
9959 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9960 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9961 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9962 ArrayRef<int> Mask = SVOp->getMask();
9963 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9965 // X86 has dedicated unpack instructions that can handle specific blend
9966 // operations: UNPCKH and UNPCKL.
9967 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9968 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
9969 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9970 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
9972 // FIXME: Implement direct support for this type!
9973 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
9976 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
9977 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9978 const X86Subtarget *Subtarget,
9979 SelectionDAG &DAG) {
9981 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9982 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9983 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9984 ArrayRef<int> Mask = SVOp->getMask();
9985 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9987 // Use dedicated unpack instructions for masks that match their pattern.
9988 if (isShuffleEquivalent(V1, V2, Mask,
9989 {// First 128-bit lane.
9990 0, 16, 1, 17, 4, 20, 5, 21,
9991 // Second 128-bit lane.
9992 8, 24, 9, 25, 12, 28, 13, 29}))
9993 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
9994 if (isShuffleEquivalent(V1, V2, Mask,
9995 {// First 128-bit lane.
9996 2, 18, 3, 19, 6, 22, 7, 23,
9997 // Second 128-bit lane.
9998 10, 26, 11, 27, 14, 30, 15, 31}))
9999 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10001 // FIXME: Implement direct support for this type!
10002 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10005 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10006 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10007 const X86Subtarget *Subtarget,
10008 SelectionDAG &DAG) {
10010 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10011 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10012 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10013 ArrayRef<int> Mask = SVOp->getMask();
10014 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10016 // X86 has dedicated unpack instructions that can handle specific blend
10017 // operations: UNPCKH and UNPCKL.
10018 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10019 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10020 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10021 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10023 // FIXME: Implement direct support for this type!
10024 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10027 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10028 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10029 const X86Subtarget *Subtarget,
10030 SelectionDAG &DAG) {
10032 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10033 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10034 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10035 ArrayRef<int> Mask = SVOp->getMask();
10036 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10038 // Use dedicated unpack instructions for masks that match their pattern.
10039 if (isShuffleEquivalent(V1, V2, Mask,
10040 {// First 128-bit lane.
10041 0, 16, 1, 17, 4, 20, 5, 21,
10042 // Second 128-bit lane.
10043 8, 24, 9, 25, 12, 28, 13, 29}))
10044 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10045 if (isShuffleEquivalent(V1, V2, Mask,
10046 {// First 128-bit lane.
10047 2, 18, 3, 19, 6, 22, 7, 23,
10048 // Second 128-bit lane.
10049 10, 26, 11, 27, 14, 30, 15, 31}))
10050 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10052 // FIXME: Implement direct support for this type!
10053 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10056 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10057 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10058 const X86Subtarget *Subtarget,
10059 SelectionDAG &DAG) {
10061 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10062 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10063 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10064 ArrayRef<int> Mask = SVOp->getMask();
10065 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10066 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10068 // FIXME: Implement direct support for this type!
10069 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10072 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10073 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10074 const X86Subtarget *Subtarget,
10075 SelectionDAG &DAG) {
10077 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10078 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10079 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10080 ArrayRef<int> Mask = SVOp->getMask();
10081 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10082 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10084 // FIXME: Implement direct support for this type!
10085 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10088 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10090 /// This routine either breaks down the specific type of a 512-bit x86 vector
10091 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10092 /// together based on the available instructions.
10093 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10094 MVT VT, const X86Subtarget *Subtarget,
10095 SelectionDAG &DAG) {
10097 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10098 ArrayRef<int> Mask = SVOp->getMask();
10099 assert(Subtarget->hasAVX512() &&
10100 "Cannot lower 512-bit vectors w/ basic ISA!");
10102 // Check for being able to broadcast a single element.
10103 if (SDValue Broadcast =
10104 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10107 // Dispatch to each element type for lowering. If we don't have supprot for
10108 // specific element type shuffles at 512 bits, immediately split them and
10109 // lower them. Each lowering routine of a given type is allowed to assume that
10110 // the requisite ISA extensions for that element type are available.
10111 switch (VT.SimpleTy) {
10113 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10115 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10117 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10119 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10121 if (Subtarget->hasBWI())
10122 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10125 if (Subtarget->hasBWI())
10126 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10130 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10133 // Otherwise fall back on splitting.
10134 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10137 /// \brief Top-level lowering for x86 vector shuffles.
10139 /// This handles decomposition, canonicalization, and lowering of all x86
10140 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10141 /// above in helper routines. The canonicalization attempts to widen shuffles
10142 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10143 /// s.t. only one of the two inputs needs to be tested, etc.
10144 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10145 SelectionDAG &DAG) {
10146 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10147 ArrayRef<int> Mask = SVOp->getMask();
10148 SDValue V1 = Op.getOperand(0);
10149 SDValue V2 = Op.getOperand(1);
10150 MVT VT = Op.getSimpleValueType();
10151 int NumElements = VT.getVectorNumElements();
10154 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10156 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10157 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10158 if (V1IsUndef && V2IsUndef)
10159 return DAG.getUNDEF(VT);
10161 // When we create a shuffle node we put the UNDEF node to second operand,
10162 // but in some cases the first operand may be transformed to UNDEF.
10163 // In this case we should just commute the node.
10165 return DAG.getCommutedVectorShuffle(*SVOp);
10167 // Check for non-undef masks pointing at an undef vector and make the masks
10168 // undef as well. This makes it easier to match the shuffle based solely on
10172 if (M >= NumElements) {
10173 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10174 for (int &M : NewMask)
10175 if (M >= NumElements)
10177 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10180 // We actually see shuffles that are entirely re-arrangements of a set of
10181 // zero inputs. This mostly happens while decomposing complex shuffles into
10182 // simple ones. Directly lower these as a buildvector of zeros.
10183 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10184 if (Zeroable.all())
10185 return getZeroVector(VT, Subtarget, DAG, dl);
10187 // Try to collapse shuffles into using a vector type with fewer elements but
10188 // wider element types. We cap this to not form integers or floating point
10189 // elements wider than 64 bits, but it might be interesting to form i128
10190 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10191 SmallVector<int, 16> WidenedMask;
10192 if (VT.getScalarSizeInBits() < 64 &&
10193 canWidenShuffleElements(Mask, WidenedMask)) {
10194 MVT NewEltVT = VT.isFloatingPoint()
10195 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10196 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10197 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10198 // Make sure that the new vector type is legal. For example, v2f64 isn't
10200 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10201 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
10202 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
10203 return DAG.getNode(ISD::BITCAST, dl, VT,
10204 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10208 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10209 for (int M : SVOp->getMask())
10211 ++NumUndefElements;
10212 else if (M < NumElements)
10217 // Commute the shuffle as needed such that more elements come from V1 than
10218 // V2. This allows us to match the shuffle pattern strictly on how many
10219 // elements come from V1 without handling the symmetric cases.
10220 if (NumV2Elements > NumV1Elements)
10221 return DAG.getCommutedVectorShuffle(*SVOp);
10223 // When the number of V1 and V2 elements are the same, try to minimize the
10224 // number of uses of V2 in the low half of the vector. When that is tied,
10225 // ensure that the sum of indices for V1 is equal to or lower than the sum
10226 // indices for V2. When those are equal, try to ensure that the number of odd
10227 // indices for V1 is lower than the number of odd indices for V2.
10228 if (NumV1Elements == NumV2Elements) {
10229 int LowV1Elements = 0, LowV2Elements = 0;
10230 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10231 if (M >= NumElements)
10235 if (LowV2Elements > LowV1Elements) {
10236 return DAG.getCommutedVectorShuffle(*SVOp);
10237 } else if (LowV2Elements == LowV1Elements) {
10238 int SumV1Indices = 0, SumV2Indices = 0;
10239 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10240 if (SVOp->getMask()[i] >= NumElements)
10242 else if (SVOp->getMask()[i] >= 0)
10244 if (SumV2Indices < SumV1Indices) {
10245 return DAG.getCommutedVectorShuffle(*SVOp);
10246 } else if (SumV2Indices == SumV1Indices) {
10247 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10248 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10249 if (SVOp->getMask()[i] >= NumElements)
10250 NumV2OddIndices += i % 2;
10251 else if (SVOp->getMask()[i] >= 0)
10252 NumV1OddIndices += i % 2;
10253 if (NumV2OddIndices < NumV1OddIndices)
10254 return DAG.getCommutedVectorShuffle(*SVOp);
10259 // For each vector width, delegate to a specialized lowering routine.
10260 if (VT.getSizeInBits() == 128)
10261 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10263 if (VT.getSizeInBits() == 256)
10264 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10266 // Force AVX-512 vectors to be scalarized for now.
10267 // FIXME: Implement AVX-512 support!
10268 if (VT.getSizeInBits() == 512)
10269 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10271 llvm_unreachable("Unimplemented!");
10274 // This function assumes its argument is a BUILD_VECTOR of constants or
10275 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10277 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10278 unsigned &MaskValue) {
10280 unsigned NumElems = BuildVector->getNumOperands();
10281 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10282 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10283 unsigned NumElemsInLane = NumElems / NumLanes;
10285 // Blend for v16i16 should be symetric for the both lanes.
10286 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10287 SDValue EltCond = BuildVector->getOperand(i);
10288 SDValue SndLaneEltCond =
10289 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10291 int Lane1Cond = -1, Lane2Cond = -1;
10292 if (isa<ConstantSDNode>(EltCond))
10293 Lane1Cond = !isZero(EltCond);
10294 if (isa<ConstantSDNode>(SndLaneEltCond))
10295 Lane2Cond = !isZero(SndLaneEltCond);
10297 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10298 // Lane1Cond != 0, means we want the first argument.
10299 // Lane1Cond == 0, means we want the second argument.
10300 // The encoding of this argument is 0 for the first argument, 1
10301 // for the second. Therefore, invert the condition.
10302 MaskValue |= !Lane1Cond << i;
10303 else if (Lane1Cond < 0)
10304 MaskValue |= !Lane2Cond << i;
10311 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10312 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10313 const X86Subtarget *Subtarget,
10314 SelectionDAG &DAG) {
10315 SDValue Cond = Op.getOperand(0);
10316 SDValue LHS = Op.getOperand(1);
10317 SDValue RHS = Op.getOperand(2);
10319 MVT VT = Op.getSimpleValueType();
10321 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10323 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10325 // Only non-legal VSELECTs reach this lowering, convert those into generic
10326 // shuffles and re-use the shuffle lowering path for blends.
10327 SmallVector<int, 32> Mask;
10328 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10329 SDValue CondElt = CondBV->getOperand(i);
10331 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10333 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10336 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10337 // A vselect where all conditions and data are constants can be optimized into
10338 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10339 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10340 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10341 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10344 // Try to lower this to a blend-style vector shuffle. This can handle all
10345 // constant condition cases.
10346 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10349 // Variable blends are only legal from SSE4.1 onward.
10350 if (!Subtarget->hasSSE41())
10353 // Only some types will be legal on some subtargets. If we can emit a legal
10354 // VSELECT-matching blend, return Op, and but if we need to expand, return
10356 switch (Op.getSimpleValueType().SimpleTy) {
10358 // Most of the vector types have blends past SSE4.1.
10362 // The byte blends for AVX vectors were introduced only in AVX2.
10363 if (Subtarget->hasAVX2())
10370 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10371 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10374 // FIXME: We should custom lower this by fixing the condition and using i8
10380 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10381 MVT VT = Op.getSimpleValueType();
10384 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10387 if (VT.getSizeInBits() == 8) {
10388 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10389 Op.getOperand(0), Op.getOperand(1));
10390 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10391 DAG.getValueType(VT));
10392 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10395 if (VT.getSizeInBits() == 16) {
10396 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10397 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10399 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10400 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10401 DAG.getNode(ISD::BITCAST, dl,
10404 Op.getOperand(1)));
10405 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10406 Op.getOperand(0), Op.getOperand(1));
10407 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10408 DAG.getValueType(VT));
10409 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10412 if (VT == MVT::f32) {
10413 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10414 // the result back to FR32 register. It's only worth matching if the
10415 // result has a single use which is a store or a bitcast to i32. And in
10416 // the case of a store, it's not worth it if the index is a constant 0,
10417 // because a MOVSSmr can be used instead, which is smaller and faster.
10418 if (!Op.hasOneUse())
10420 SDNode *User = *Op.getNode()->use_begin();
10421 if ((User->getOpcode() != ISD::STORE ||
10422 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10423 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10424 (User->getOpcode() != ISD::BITCAST ||
10425 User->getValueType(0) != MVT::i32))
10427 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10428 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10431 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10434 if (VT == MVT::i32 || VT == MVT::i64) {
10435 // ExtractPS/pextrq works with constant index.
10436 if (isa<ConstantSDNode>(Op.getOperand(1)))
10442 /// Extract one bit from mask vector, like v16i1 or v8i1.
10443 /// AVX-512 feature.
10445 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10446 SDValue Vec = Op.getOperand(0);
10448 MVT VecVT = Vec.getSimpleValueType();
10449 SDValue Idx = Op.getOperand(1);
10450 MVT EltVT = Op.getSimpleValueType();
10452 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10453 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10454 "Unexpected vector type in ExtractBitFromMaskVector");
10456 // variable index can't be handled in mask registers,
10457 // extend vector to VR512
10458 if (!isa<ConstantSDNode>(Idx)) {
10459 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10460 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10461 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10462 ExtVT.getVectorElementType(), Ext, Idx);
10463 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10466 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10467 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10468 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10469 rc = getRegClassFor(MVT::v16i1);
10470 unsigned MaxSift = rc->getSize()*8 - 1;
10471 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10472 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10473 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10474 DAG.getConstant(MaxSift, MVT::i8));
10475 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10476 DAG.getIntPtrConstant(0));
10480 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10481 SelectionDAG &DAG) const {
10483 SDValue Vec = Op.getOperand(0);
10484 MVT VecVT = Vec.getSimpleValueType();
10485 SDValue Idx = Op.getOperand(1);
10487 if (Op.getSimpleValueType() == MVT::i1)
10488 return ExtractBitFromMaskVector(Op, DAG);
10490 if (!isa<ConstantSDNode>(Idx)) {
10491 if (VecVT.is512BitVector() ||
10492 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10493 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10496 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10497 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10498 MaskEltVT.getSizeInBits());
10500 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10501 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10502 getZeroVector(MaskVT, Subtarget, DAG, dl),
10503 Idx, DAG.getConstant(0, getPointerTy()));
10504 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10505 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10506 Perm, DAG.getConstant(0, getPointerTy()));
10511 // If this is a 256-bit vector result, first extract the 128-bit vector and
10512 // then extract the element from the 128-bit vector.
10513 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10515 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10516 // Get the 128-bit vector.
10517 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10518 MVT EltVT = VecVT.getVectorElementType();
10520 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10522 //if (IdxVal >= NumElems/2)
10523 // IdxVal -= NumElems/2;
10524 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10525 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10526 DAG.getConstant(IdxVal, MVT::i32));
10529 assert(VecVT.is128BitVector() && "Unexpected vector length");
10531 if (Subtarget->hasSSE41()) {
10532 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10537 MVT VT = Op.getSimpleValueType();
10538 // TODO: handle v16i8.
10539 if (VT.getSizeInBits() == 16) {
10540 SDValue Vec = Op.getOperand(0);
10541 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10543 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10544 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10545 DAG.getNode(ISD::BITCAST, dl,
10547 Op.getOperand(1)));
10548 // Transform it so it match pextrw which produces a 32-bit result.
10549 MVT EltVT = MVT::i32;
10550 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10551 Op.getOperand(0), Op.getOperand(1));
10552 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10553 DAG.getValueType(VT));
10554 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10557 if (VT.getSizeInBits() == 32) {
10558 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10562 // SHUFPS the element to the lowest double word, then movss.
10563 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10564 MVT VVT = Op.getOperand(0).getSimpleValueType();
10565 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10566 DAG.getUNDEF(VVT), Mask);
10567 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10568 DAG.getIntPtrConstant(0));
10571 if (VT.getSizeInBits() == 64) {
10572 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10573 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10574 // to match extract_elt for f64.
10575 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10579 // UNPCKHPD the element to the lowest double word, then movsd.
10580 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10581 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10582 int Mask[2] = { 1, -1 };
10583 MVT VVT = Op.getOperand(0).getSimpleValueType();
10584 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10585 DAG.getUNDEF(VVT), Mask);
10586 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10587 DAG.getIntPtrConstant(0));
10593 /// Insert one bit to mask vector, like v16i1 or v8i1.
10594 /// AVX-512 feature.
10596 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10598 SDValue Vec = Op.getOperand(0);
10599 SDValue Elt = Op.getOperand(1);
10600 SDValue Idx = Op.getOperand(2);
10601 MVT VecVT = Vec.getSimpleValueType();
10603 if (!isa<ConstantSDNode>(Idx)) {
10604 // Non constant index. Extend source and destination,
10605 // insert element and then truncate the result.
10606 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10607 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10608 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10609 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10610 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10611 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10614 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10615 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10616 if (Vec.getOpcode() == ISD::UNDEF)
10617 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10618 DAG.getConstant(IdxVal, MVT::i8));
10619 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10620 unsigned MaxSift = rc->getSize()*8 - 1;
10621 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10622 DAG.getConstant(MaxSift, MVT::i8));
10623 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10624 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10625 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10628 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10629 SelectionDAG &DAG) const {
10630 MVT VT = Op.getSimpleValueType();
10631 MVT EltVT = VT.getVectorElementType();
10633 if (EltVT == MVT::i1)
10634 return InsertBitToMaskVector(Op, DAG);
10637 SDValue N0 = Op.getOperand(0);
10638 SDValue N1 = Op.getOperand(1);
10639 SDValue N2 = Op.getOperand(2);
10640 if (!isa<ConstantSDNode>(N2))
10642 auto *N2C = cast<ConstantSDNode>(N2);
10643 unsigned IdxVal = N2C->getZExtValue();
10645 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10646 // into that, and then insert the subvector back into the result.
10647 if (VT.is256BitVector() || VT.is512BitVector()) {
10648 // With a 256-bit vector, we can insert into the zero element efficiently
10649 // using a blend if we have AVX or AVX2 and the right data type.
10650 if (VT.is256BitVector() && IdxVal == 0) {
10651 // TODO: It is worthwhile to cast integer to floating point and back
10652 // and incur a domain crossing penalty if that's what we'll end up
10653 // doing anyway after extracting to a 128-bit vector.
10654 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
10655 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
10656 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
10657 N2 = DAG.getIntPtrConstant(1);
10658 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
10662 // Get the desired 128-bit vector chunk.
10663 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10665 // Insert the element into the desired chunk.
10666 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10667 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10669 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10670 DAG.getConstant(IdxIn128, MVT::i32));
10672 // Insert the changed part back into the bigger vector
10673 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10675 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10677 if (Subtarget->hasSSE41()) {
10678 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10680 if (VT == MVT::v8i16) {
10681 Opc = X86ISD::PINSRW;
10683 assert(VT == MVT::v16i8);
10684 Opc = X86ISD::PINSRB;
10687 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10689 if (N1.getValueType() != MVT::i32)
10690 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10691 if (N2.getValueType() != MVT::i32)
10692 N2 = DAG.getIntPtrConstant(IdxVal);
10693 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10696 if (EltVT == MVT::f32) {
10697 // Bits [7:6] of the constant are the source select. This will always be
10698 // zero here. The DAG Combiner may combine an extract_elt index into
10699 // these bits. For example (insert (extract, 3), 2) could be matched by
10700 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
10701 // Bits [5:4] of the constant are the destination select. This is the
10702 // value of the incoming immediate.
10703 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10704 // combine either bitwise AND or insert of float 0.0 to set these bits.
10706 const Function *F = DAG.getMachineFunction().getFunction();
10707 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
10708 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
10709 // If this is an insertion of 32-bits into the low 32-bits of
10710 // a vector, we prefer to generate a blend with immediate rather
10711 // than an insertps. Blends are simpler operations in hardware and so
10712 // will always have equal or better performance than insertps.
10713 // But if optimizing for size and there's a load folding opportunity,
10714 // generate insertps because blendps does not have a 32-bit memory
10716 N2 = DAG.getIntPtrConstant(1);
10717 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10718 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
10720 N2 = DAG.getIntPtrConstant(IdxVal << 4);
10721 // Create this as a scalar to vector..
10722 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10723 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10726 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10727 // PINSR* works with constant index.
10732 if (EltVT == MVT::i8)
10735 if (EltVT.getSizeInBits() == 16) {
10736 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10737 // as its second argument.
10738 if (N1.getValueType() != MVT::i32)
10739 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10740 if (N2.getValueType() != MVT::i32)
10741 N2 = DAG.getIntPtrConstant(IdxVal);
10742 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10747 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10749 MVT OpVT = Op.getSimpleValueType();
10751 // If this is a 256-bit vector result, first insert into a 128-bit
10752 // vector and then insert into the 256-bit vector.
10753 if (!OpVT.is128BitVector()) {
10754 // Insert into a 128-bit vector.
10755 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10756 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10757 OpVT.getVectorNumElements() / SizeFactor);
10759 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10761 // Insert the 128-bit vector.
10762 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10765 if (OpVT == MVT::v1i64 &&
10766 Op.getOperand(0).getValueType() == MVT::i64)
10767 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10769 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10770 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10771 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10772 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10775 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10776 // a simple subregister reference or explicit instructions to grab
10777 // upper bits of a vector.
10778 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10779 SelectionDAG &DAG) {
10781 SDValue In = Op.getOperand(0);
10782 SDValue Idx = Op.getOperand(1);
10783 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10784 MVT ResVT = Op.getSimpleValueType();
10785 MVT InVT = In.getSimpleValueType();
10787 if (Subtarget->hasFp256()) {
10788 if (ResVT.is128BitVector() &&
10789 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10790 isa<ConstantSDNode>(Idx)) {
10791 return Extract128BitVector(In, IdxVal, DAG, dl);
10793 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10794 isa<ConstantSDNode>(Idx)) {
10795 return Extract256BitVector(In, IdxVal, DAG, dl);
10801 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10802 // simple superregister reference or explicit instructions to insert
10803 // the upper bits of a vector.
10804 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10805 SelectionDAG &DAG) {
10806 if (!Subtarget->hasAVX())
10810 SDValue Vec = Op.getOperand(0);
10811 SDValue SubVec = Op.getOperand(1);
10812 SDValue Idx = Op.getOperand(2);
10814 if (!isa<ConstantSDNode>(Idx))
10817 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10818 MVT OpVT = Op.getSimpleValueType();
10819 MVT SubVecVT = SubVec.getSimpleValueType();
10821 // Fold two 16-byte subvector loads into one 32-byte load:
10822 // (insert_subvector (insert_subvector undef, (load addr), 0),
10823 // (load addr + 16), Elts/2)
10825 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10826 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10827 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10828 !Subtarget->isUnalignedMem32Slow()) {
10829 SDValue SubVec2 = Vec.getOperand(1);
10830 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10831 if (Idx2->getZExtValue() == 0) {
10832 SDValue Ops[] = { SubVec2, SubVec };
10833 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10840 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10841 SubVecVT.is128BitVector())
10842 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10844 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10845 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10847 if (OpVT.getVectorElementType() == MVT::i1) {
10848 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
10850 SDValue ZeroIdx = DAG.getIntPtrConstant(0);
10851 SDValue Undef = DAG.getUNDEF(OpVT);
10852 unsigned NumElems = OpVT.getVectorNumElements();
10853 SDValue ShiftBits = DAG.getConstant(NumElems/2, MVT::i8);
10855 if (IdxVal == OpVT.getVectorNumElements() / 2) {
10856 // Zero upper bits of the Vec
10857 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10858 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10860 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10862 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10863 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10866 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10868 // Zero upper bits of the Vec2
10869 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10870 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
10871 // Zero lower bits of the Vec
10872 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10873 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10874 // Merge them together
10875 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10881 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10882 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10883 // one of the above mentioned nodes. It has to be wrapped because otherwise
10884 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10885 // be used to form addressing mode. These wrapped nodes will be selected
10888 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10889 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10891 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10892 // global base reg.
10893 unsigned char OpFlag = 0;
10894 unsigned WrapperKind = X86ISD::Wrapper;
10895 CodeModel::Model M = DAG.getTarget().getCodeModel();
10897 if (Subtarget->isPICStyleRIPRel() &&
10898 (M == CodeModel::Small || M == CodeModel::Kernel))
10899 WrapperKind = X86ISD::WrapperRIP;
10900 else if (Subtarget->isPICStyleGOT())
10901 OpFlag = X86II::MO_GOTOFF;
10902 else if (Subtarget->isPICStyleStubPIC())
10903 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10905 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10906 CP->getAlignment(),
10907 CP->getOffset(), OpFlag);
10909 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10910 // With PIC, the address is actually $g + Offset.
10912 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10913 DAG.getNode(X86ISD::GlobalBaseReg,
10914 SDLoc(), getPointerTy()),
10921 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10922 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10924 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10925 // global base reg.
10926 unsigned char OpFlag = 0;
10927 unsigned WrapperKind = X86ISD::Wrapper;
10928 CodeModel::Model M = DAG.getTarget().getCodeModel();
10930 if (Subtarget->isPICStyleRIPRel() &&
10931 (M == CodeModel::Small || M == CodeModel::Kernel))
10932 WrapperKind = X86ISD::WrapperRIP;
10933 else if (Subtarget->isPICStyleGOT())
10934 OpFlag = X86II::MO_GOTOFF;
10935 else if (Subtarget->isPICStyleStubPIC())
10936 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10938 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10941 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10943 // With PIC, the address is actually $g + Offset.
10945 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10946 DAG.getNode(X86ISD::GlobalBaseReg,
10947 SDLoc(), getPointerTy()),
10954 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10955 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10957 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10958 // global base reg.
10959 unsigned char OpFlag = 0;
10960 unsigned WrapperKind = X86ISD::Wrapper;
10961 CodeModel::Model M = DAG.getTarget().getCodeModel();
10963 if (Subtarget->isPICStyleRIPRel() &&
10964 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10965 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10966 OpFlag = X86II::MO_GOTPCREL;
10967 WrapperKind = X86ISD::WrapperRIP;
10968 } else if (Subtarget->isPICStyleGOT()) {
10969 OpFlag = X86II::MO_GOT;
10970 } else if (Subtarget->isPICStyleStubPIC()) {
10971 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10972 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10973 OpFlag = X86II::MO_DARWIN_NONLAZY;
10976 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
10979 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10981 // With PIC, the address is actually $g + Offset.
10982 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
10983 !Subtarget->is64Bit()) {
10984 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10985 DAG.getNode(X86ISD::GlobalBaseReg,
10986 SDLoc(), getPointerTy()),
10990 // For symbols that require a load from a stub to get the address, emit the
10992 if (isGlobalStubReference(OpFlag))
10993 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
10994 MachinePointerInfo::getGOT(), false, false, false, 0);
11000 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11001 // Create the TargetBlockAddressAddress node.
11002 unsigned char OpFlags =
11003 Subtarget->ClassifyBlockAddressReference();
11004 CodeModel::Model M = DAG.getTarget().getCodeModel();
11005 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11006 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11008 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
11011 if (Subtarget->isPICStyleRIPRel() &&
11012 (M == CodeModel::Small || M == CodeModel::Kernel))
11013 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11015 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11017 // With PIC, the address is actually $g + Offset.
11018 if (isGlobalRelativeToPICBase(OpFlags)) {
11019 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11020 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11028 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11029 int64_t Offset, SelectionDAG &DAG) const {
11030 // Create the TargetGlobalAddress node, folding in the constant
11031 // offset if it is legal.
11032 unsigned char OpFlags =
11033 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11034 CodeModel::Model M = DAG.getTarget().getCodeModel();
11036 if (OpFlags == X86II::MO_NO_FLAG &&
11037 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11038 // A direct static reference to a global.
11039 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
11042 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
11045 if (Subtarget->isPICStyleRIPRel() &&
11046 (M == CodeModel::Small || M == CodeModel::Kernel))
11047 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11049 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11051 // With PIC, the address is actually $g + Offset.
11052 if (isGlobalRelativeToPICBase(OpFlags)) {
11053 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11054 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11058 // For globals that require a load from a stub to get the address, emit the
11060 if (isGlobalStubReference(OpFlags))
11061 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
11062 MachinePointerInfo::getGOT(), false, false, false, 0);
11064 // If there was a non-zero offset that we didn't fold, create an explicit
11065 // addition for it.
11067 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
11068 DAG.getConstant(Offset, getPointerTy()));
11074 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11075 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11076 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11077 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11081 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11082 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11083 unsigned char OperandFlags, bool LocalDynamic = false) {
11084 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11085 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11087 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11088 GA->getValueType(0),
11092 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11096 SDValue Ops[] = { Chain, TGA, *InFlag };
11097 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11099 SDValue Ops[] = { Chain, TGA };
11100 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11103 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11104 MFI->setAdjustsStack(true);
11105 MFI->setHasCalls(true);
11107 SDValue Flag = Chain.getValue(1);
11108 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11111 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11113 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11116 SDLoc dl(GA); // ? function entry point might be better
11117 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11118 DAG.getNode(X86ISD::GlobalBaseReg,
11119 SDLoc(), PtrVT), InFlag);
11120 InFlag = Chain.getValue(1);
11122 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11125 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11127 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11129 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11130 X86::RAX, X86II::MO_TLSGD);
11133 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11139 // Get the start address of the TLS block for this module.
11140 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11141 .getInfo<X86MachineFunctionInfo>();
11142 MFI->incNumLocalDynamicTLSAccesses();
11146 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11147 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11150 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11151 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11152 InFlag = Chain.getValue(1);
11153 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11154 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11157 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11161 unsigned char OperandFlags = X86II::MO_DTPOFF;
11162 unsigned WrapperKind = X86ISD::Wrapper;
11163 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11164 GA->getValueType(0),
11165 GA->getOffset(), OperandFlags);
11166 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11168 // Add x@dtpoff with the base.
11169 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11172 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11173 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11174 const EVT PtrVT, TLSModel::Model model,
11175 bool is64Bit, bool isPIC) {
11178 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11179 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11180 is64Bit ? 257 : 256));
11182 SDValue ThreadPointer =
11183 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
11184 MachinePointerInfo(Ptr), false, false, false, 0);
11186 unsigned char OperandFlags = 0;
11187 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11189 unsigned WrapperKind = X86ISD::Wrapper;
11190 if (model == TLSModel::LocalExec) {
11191 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11192 } else if (model == TLSModel::InitialExec) {
11194 OperandFlags = X86II::MO_GOTTPOFF;
11195 WrapperKind = X86ISD::WrapperRIP;
11197 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11200 llvm_unreachable("Unexpected model");
11203 // emit "addl x@ntpoff,%eax" (local exec)
11204 // or "addl x@indntpoff,%eax" (initial exec)
11205 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11207 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11208 GA->getOffset(), OperandFlags);
11209 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11211 if (model == TLSModel::InitialExec) {
11212 if (isPIC && !is64Bit) {
11213 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11214 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11218 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11219 MachinePointerInfo::getGOT(), false, false, false, 0);
11222 // The address of the thread local variable is the add of the thread
11223 // pointer with the offset of the variable.
11224 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11228 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11230 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11231 const GlobalValue *GV = GA->getGlobal();
11233 if (Subtarget->isTargetELF()) {
11234 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11237 case TLSModel::GeneralDynamic:
11238 if (Subtarget->is64Bit())
11239 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
11240 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
11241 case TLSModel::LocalDynamic:
11242 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
11243 Subtarget->is64Bit());
11244 case TLSModel::InitialExec:
11245 case TLSModel::LocalExec:
11246 return LowerToTLSExecModel(
11247 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
11248 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
11250 llvm_unreachable("Unknown TLS model.");
11253 if (Subtarget->isTargetDarwin()) {
11254 // Darwin only has one model of TLS. Lower to that.
11255 unsigned char OpFlag = 0;
11256 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11257 X86ISD::WrapperRIP : X86ISD::Wrapper;
11259 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11260 // global base reg.
11261 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11262 !Subtarget->is64Bit();
11264 OpFlag = X86II::MO_TLVP_PIC_BASE;
11266 OpFlag = X86II::MO_TLVP;
11268 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11269 GA->getValueType(0),
11270 GA->getOffset(), OpFlag);
11271 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11273 // With PIC32, the address is actually $g + Offset.
11275 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11276 DAG.getNode(X86ISD::GlobalBaseReg,
11277 SDLoc(), getPointerTy()),
11280 // Lowering the machine isd will make sure everything is in the right
11282 SDValue Chain = DAG.getEntryNode();
11283 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11284 SDValue Args[] = { Chain, Offset };
11285 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11287 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11288 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11289 MFI->setAdjustsStack(true);
11291 // And our return value (tls address) is in the standard call return value
11293 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11294 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11295 Chain.getValue(1));
11298 if (Subtarget->isTargetKnownWindowsMSVC() ||
11299 Subtarget->isTargetWindowsGNU()) {
11300 // Just use the implicit TLS architecture
11301 // Need to generate someting similar to:
11302 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11304 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11305 // mov rcx, qword [rdx+rcx*8]
11306 // mov eax, .tls$:tlsvar
11307 // [rax+rcx] contains the address
11308 // Windows 64bit: gs:0x58
11309 // Windows 32bit: fs:__tls_array
11312 SDValue Chain = DAG.getEntryNode();
11314 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11315 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11316 // use its literal value of 0x2C.
11317 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11318 ? Type::getInt8PtrTy(*DAG.getContext(),
11320 : Type::getInt32PtrTy(*DAG.getContext(),
11324 Subtarget->is64Bit()
11325 ? DAG.getIntPtrConstant(0x58)
11326 : (Subtarget->isTargetWindowsGNU()
11327 ? DAG.getIntPtrConstant(0x2C)
11328 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11330 SDValue ThreadPointer =
11331 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11332 MachinePointerInfo(Ptr), false, false, false, 0);
11334 // Load the _tls_index variable
11335 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11336 if (Subtarget->is64Bit())
11337 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
11338 IDX, MachinePointerInfo(), MVT::i32,
11339 false, false, false, 0);
11341 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11342 false, false, false, 0);
11344 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
11346 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11348 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11349 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11350 false, false, false, 0);
11352 // Get the offset of start of .tls section
11353 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11354 GA->getValueType(0),
11355 GA->getOffset(), X86II::MO_SECREL);
11356 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11358 // The address of the thread local variable is the add of the thread
11359 // pointer with the offset of the variable.
11360 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11363 llvm_unreachable("TLS not implemented for this target.");
11366 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11367 /// and take a 2 x i32 value to shift plus a shift amount.
11368 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11369 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11370 MVT VT = Op.getSimpleValueType();
11371 unsigned VTBits = VT.getSizeInBits();
11373 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11374 SDValue ShOpLo = Op.getOperand(0);
11375 SDValue ShOpHi = Op.getOperand(1);
11376 SDValue ShAmt = Op.getOperand(2);
11377 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11378 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11380 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11381 DAG.getConstant(VTBits - 1, MVT::i8));
11382 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11383 DAG.getConstant(VTBits - 1, MVT::i8))
11384 : DAG.getConstant(0, VT);
11386 SDValue Tmp2, Tmp3;
11387 if (Op.getOpcode() == ISD::SHL_PARTS) {
11388 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11389 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11391 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11392 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11395 // If the shift amount is larger or equal than the width of a part we can't
11396 // rely on the results of shld/shrd. Insert a test and select the appropriate
11397 // values for large shift amounts.
11398 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11399 DAG.getConstant(VTBits, MVT::i8));
11400 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11401 AndNode, DAG.getConstant(0, MVT::i8));
11404 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
11405 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11406 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11408 if (Op.getOpcode() == ISD::SHL_PARTS) {
11409 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11410 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11412 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11413 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11416 SDValue Ops[2] = { Lo, Hi };
11417 return DAG.getMergeValues(Ops, dl);
11420 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11421 SelectionDAG &DAG) const {
11422 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11425 if (SrcVT.isVector()) {
11426 if (SrcVT.getVectorElementType() == MVT::i1) {
11427 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11428 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11429 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11430 Op.getOperand(0)));
11435 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11436 "Unknown SINT_TO_FP to lower!");
11438 // These are really Legal; return the operand so the caller accepts it as
11440 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11442 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11443 Subtarget->is64Bit()) {
11447 unsigned Size = SrcVT.getSizeInBits()/8;
11448 MachineFunction &MF = DAG.getMachineFunction();
11449 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11450 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11451 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11453 MachinePointerInfo::getFixedStack(SSFI),
11455 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11458 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11460 SelectionDAG &DAG) const {
11464 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11466 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11468 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11470 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11472 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11473 MachineMemOperand *MMO;
11475 int SSFI = FI->getIndex();
11477 DAG.getMachineFunction()
11478 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11479 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11481 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11482 StackSlot = StackSlot.getOperand(1);
11484 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11485 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11487 Tys, Ops, SrcVT, MMO);
11490 Chain = Result.getValue(1);
11491 SDValue InFlag = Result.getValue(2);
11493 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11494 // shouldn't be necessary except that RFP cannot be live across
11495 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11496 MachineFunction &MF = DAG.getMachineFunction();
11497 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11498 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11499 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11500 Tys = DAG.getVTList(MVT::Other);
11502 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11504 MachineMemOperand *MMO =
11505 DAG.getMachineFunction()
11506 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11507 MachineMemOperand::MOStore, SSFISize, SSFISize);
11509 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11510 Ops, Op.getValueType(), MMO);
11511 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11512 MachinePointerInfo::getFixedStack(SSFI),
11513 false, false, false, 0);
11519 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11520 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11521 SelectionDAG &DAG) const {
11522 // This algorithm is not obvious. Here it is what we're trying to output:
11525 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11526 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11528 haddpd %xmm0, %xmm0
11530 pshufd $0x4e, %xmm0, %xmm1
11536 LLVMContext *Context = DAG.getContext();
11538 // Build some magic constants.
11539 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11540 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11541 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11543 SmallVector<Constant*,2> CV1;
11545 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11546 APInt(64, 0x4330000000000000ULL))));
11548 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11549 APInt(64, 0x4530000000000000ULL))));
11550 Constant *C1 = ConstantVector::get(CV1);
11551 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11553 // Load the 64-bit value into an XMM register.
11554 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11556 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11557 MachinePointerInfo::getConstantPool(),
11558 false, false, false, 16);
11559 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11560 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11563 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11564 MachinePointerInfo::getConstantPool(),
11565 false, false, false, 16);
11566 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11567 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11570 if (Subtarget->hasSSE3()) {
11571 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11572 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11574 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11575 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11577 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11578 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11582 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11583 DAG.getIntPtrConstant(0));
11586 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11587 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11588 SelectionDAG &DAG) const {
11590 // FP constant to bias correct the final result.
11591 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
11594 // Load the 32-bit value into an XMM register.
11595 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11598 // Zero out the upper parts of the register.
11599 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11601 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11602 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11603 DAG.getIntPtrConstant(0));
11605 // Or the load with the bias.
11606 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11607 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11608 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11609 MVT::v2f64, Load)),
11610 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11611 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11612 MVT::v2f64, Bias)));
11613 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11614 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11615 DAG.getIntPtrConstant(0));
11617 // Subtract the bias.
11618 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11620 // Handle final rounding.
11621 EVT DestVT = Op.getValueType();
11623 if (DestVT.bitsLT(MVT::f64))
11624 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11625 DAG.getIntPtrConstant(0));
11626 if (DestVT.bitsGT(MVT::f64))
11627 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11629 // Handle final rounding.
11633 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11634 const X86Subtarget &Subtarget) {
11635 // The algorithm is the following:
11636 // #ifdef __SSE4_1__
11637 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11638 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11639 // (uint4) 0x53000000, 0xaa);
11641 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11642 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11644 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11645 // return (float4) lo + fhi;
11648 SDValue V = Op->getOperand(0);
11649 EVT VecIntVT = V.getValueType();
11650 bool Is128 = VecIntVT == MVT::v4i32;
11651 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11652 // If we convert to something else than the supported type, e.g., to v4f64,
11654 if (VecFloatVT != Op->getValueType(0))
11657 unsigned NumElts = VecIntVT.getVectorNumElements();
11658 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11659 "Unsupported custom type");
11660 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11662 // In the #idef/#else code, we have in common:
11663 // - The vector of constants:
11669 // Create the splat vector for 0x4b000000.
11670 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
11671 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11672 CstLow, CstLow, CstLow, CstLow};
11673 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11674 makeArrayRef(&CstLowArray[0], NumElts));
11675 // Create the splat vector for 0x53000000.
11676 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
11677 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11678 CstHigh, CstHigh, CstHigh, CstHigh};
11679 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11680 makeArrayRef(&CstHighArray[0], NumElts));
11682 // Create the right shift.
11683 SDValue CstShift = DAG.getConstant(16, MVT::i32);
11684 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11685 CstShift, CstShift, CstShift, CstShift};
11686 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11687 makeArrayRef(&CstShiftArray[0], NumElts));
11688 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11691 if (Subtarget.hasSSE41()) {
11692 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11693 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11694 SDValue VecCstLowBitcast =
11695 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11696 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11697 // Low will be bitcasted right away, so do not bother bitcasting back to its
11699 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11700 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
11701 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11702 // (uint4) 0x53000000, 0xaa);
11703 SDValue VecCstHighBitcast =
11704 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11705 SDValue VecShiftBitcast =
11706 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11707 // High will be bitcasted right away, so do not bother bitcasting back to
11708 // its original type.
11709 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11710 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
11712 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
11713 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11714 CstMask, CstMask, CstMask);
11715 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11716 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11717 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11719 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11720 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11723 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11724 SDValue CstFAdd = DAG.getConstantFP(
11725 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
11726 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11727 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11728 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11729 makeArrayRef(&CstFAddArray[0], NumElts));
11731 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11732 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11734 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11735 // return (float4) lo + fhi;
11736 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11737 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11740 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11741 SelectionDAG &DAG) const {
11742 SDValue N0 = Op.getOperand(0);
11743 MVT SVT = N0.getSimpleValueType();
11746 switch (SVT.SimpleTy) {
11748 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11753 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11754 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11755 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11759 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11761 llvm_unreachable(nullptr);
11764 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11765 SelectionDAG &DAG) const {
11766 SDValue N0 = Op.getOperand(0);
11769 if (Op.getValueType().isVector())
11770 return lowerUINT_TO_FP_vec(Op, DAG);
11772 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11773 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11774 // the optimization here.
11775 if (DAG.SignBitIsZero(N0))
11776 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11778 MVT SrcVT = N0.getSimpleValueType();
11779 MVT DstVT = Op.getSimpleValueType();
11780 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11781 return LowerUINT_TO_FP_i64(Op, DAG);
11782 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11783 return LowerUINT_TO_FP_i32(Op, DAG);
11784 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11787 // Make a 64-bit buffer, and use it to build an FILD.
11788 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11789 if (SrcVT == MVT::i32) {
11790 SDValue WordOff = DAG.getConstant(4, getPointerTy());
11791 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11792 getPointerTy(), StackSlot, WordOff);
11793 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11794 StackSlot, MachinePointerInfo(),
11796 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
11797 OffsetSlot, MachinePointerInfo(),
11799 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11803 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11804 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11805 StackSlot, MachinePointerInfo(),
11807 // For i64 source, we need to add the appropriate power of 2 if the input
11808 // was negative. This is the same as the optimization in
11809 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11810 // we must be careful to do the computation in x87 extended precision, not
11811 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11812 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11813 MachineMemOperand *MMO =
11814 DAG.getMachineFunction()
11815 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11816 MachineMemOperand::MOLoad, 8, 8);
11818 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11819 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11820 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11823 APInt FF(32, 0x5F800000ULL);
11825 // Check whether the sign bit is set.
11826 SDValue SignSet = DAG.getSetCC(dl,
11827 getSetCCResultType(*DAG.getContext(), MVT::i64),
11828 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
11831 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11832 SDValue FudgePtr = DAG.getConstantPool(
11833 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11836 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11837 SDValue Zero = DAG.getIntPtrConstant(0);
11838 SDValue Four = DAG.getIntPtrConstant(4);
11839 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11841 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11843 // Load the value out, extending it from f32 to f80.
11844 // FIXME: Avoid the extend by constructing the right constant pool?
11845 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11846 FudgePtr, MachinePointerInfo::getConstantPool(),
11847 MVT::f32, false, false, false, 4);
11848 // Extend everything to 80 bits to force it to be done on x87.
11849 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11850 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
11853 std::pair<SDValue,SDValue>
11854 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11855 bool IsSigned, bool IsReplace) const {
11858 EVT DstTy = Op.getValueType();
11860 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11861 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11865 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11866 DstTy.getSimpleVT() >= MVT::i16 &&
11867 "Unknown FP_TO_INT to lower!");
11869 // These are really Legal.
11870 if (DstTy == MVT::i32 &&
11871 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11872 return std::make_pair(SDValue(), SDValue());
11873 if (Subtarget->is64Bit() &&
11874 DstTy == MVT::i64 &&
11875 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11876 return std::make_pair(SDValue(), SDValue());
11878 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11879 // stack slot, or into the FTOL runtime function.
11880 MachineFunction &MF = DAG.getMachineFunction();
11881 unsigned MemSize = DstTy.getSizeInBits()/8;
11882 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11883 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11886 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11887 Opc = X86ISD::WIN_FTOL;
11889 switch (DstTy.getSimpleVT().SimpleTy) {
11890 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11891 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11892 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11893 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11896 SDValue Chain = DAG.getEntryNode();
11897 SDValue Value = Op.getOperand(0);
11898 EVT TheVT = Op.getOperand(0).getValueType();
11899 // FIXME This causes a redundant load/store if the SSE-class value is already
11900 // in memory, such as if it is on the callstack.
11901 if (isScalarFPTypeInSSEReg(TheVT)) {
11902 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11903 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11904 MachinePointerInfo::getFixedStack(SSFI),
11906 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11908 Chain, StackSlot, DAG.getValueType(TheVT)
11911 MachineMemOperand *MMO =
11912 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11913 MachineMemOperand::MOLoad, MemSize, MemSize);
11914 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11915 Chain = Value.getValue(1);
11916 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11917 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11920 MachineMemOperand *MMO =
11921 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11922 MachineMemOperand::MOStore, MemSize, MemSize);
11924 if (Opc != X86ISD::WIN_FTOL) {
11925 // Build the FP_TO_INT*_IN_MEM
11926 SDValue Ops[] = { Chain, Value, StackSlot };
11927 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11929 return std::make_pair(FIST, StackSlot);
11931 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11932 DAG.getVTList(MVT::Other, MVT::Glue),
11934 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11935 MVT::i32, ftol.getValue(1));
11936 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11937 MVT::i32, eax.getValue(2));
11938 SDValue Ops[] = { eax, edx };
11939 SDValue pair = IsReplace
11940 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11941 : DAG.getMergeValues(Ops, DL);
11942 return std::make_pair(pair, SDValue());
11946 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11947 const X86Subtarget *Subtarget) {
11948 MVT VT = Op->getSimpleValueType(0);
11949 SDValue In = Op->getOperand(0);
11950 MVT InVT = In.getSimpleValueType();
11953 // Optimize vectors in AVX mode:
11956 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11957 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11958 // Concat upper and lower parts.
11961 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11962 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11963 // Concat upper and lower parts.
11966 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11967 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11968 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
11971 if (Subtarget->hasInt256())
11972 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
11974 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
11975 SDValue Undef = DAG.getUNDEF(InVT);
11976 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
11977 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11978 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11980 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
11981 VT.getVectorNumElements()/2);
11983 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
11984 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
11986 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
11989 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
11990 SelectionDAG &DAG) {
11991 MVT VT = Op->getSimpleValueType(0);
11992 SDValue In = Op->getOperand(0);
11993 MVT InVT = In.getSimpleValueType();
11995 unsigned int NumElts = VT.getVectorNumElements();
11996 if (NumElts != 8 && NumElts != 16)
11999 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12000 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12002 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
12003 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12004 // Now we have only mask extension
12005 assert(InVT.getVectorElementType() == MVT::i1);
12006 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
12007 const Constant *C = cast<ConstantSDNode>(Cst)->getConstantIntValue();
12008 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
12009 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
12010 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
12011 MachinePointerInfo::getConstantPool(),
12012 false, false, false, Alignment);
12014 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
12015 if (VT.is512BitVector())
12017 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
12020 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12021 SelectionDAG &DAG) {
12022 if (Subtarget->hasFp256()) {
12023 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12031 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12032 SelectionDAG &DAG) {
12034 MVT VT = Op.getSimpleValueType();
12035 SDValue In = Op.getOperand(0);
12036 MVT SVT = In.getSimpleValueType();
12038 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12039 return LowerZERO_EXTEND_AVX512(Op, DAG);
12041 if (Subtarget->hasFp256()) {
12042 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12047 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12048 VT.getVectorNumElements() != SVT.getVectorNumElements());
12052 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12054 MVT VT = Op.getSimpleValueType();
12055 SDValue In = Op.getOperand(0);
12056 MVT InVT = In.getSimpleValueType();
12058 if (VT == MVT::i1) {
12059 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12060 "Invalid scalar TRUNCATE operation");
12061 if (InVT.getSizeInBits() >= 32)
12063 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12064 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12066 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12067 "Invalid TRUNCATE operation");
12069 // move vector to mask - truncate solution for SKX
12070 if (VT.getVectorElementType() == MVT::i1) {
12071 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12072 Subtarget->hasBWI())
12073 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12074 if ((InVT.is256BitVector() || InVT.is128BitVector())
12075 && InVT.getScalarSizeInBits() <= 16 &&
12076 Subtarget->hasBWI() && Subtarget->hasVLX())
12077 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12078 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12079 Subtarget->hasDQI())
12080 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12081 if ((InVT.is256BitVector() || InVT.is128BitVector())
12082 && InVT.getScalarSizeInBits() >= 32 &&
12083 Subtarget->hasDQI() && Subtarget->hasVLX())
12084 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12086 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
12087 if (VT.getVectorElementType().getSizeInBits() >=8)
12088 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12090 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12091 unsigned NumElts = InVT.getVectorNumElements();
12092 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12093 if (InVT.getSizeInBits() < 512) {
12094 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12095 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12099 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
12100 const Constant *C = cast<ConstantSDNode>(Cst)->getConstantIntValue();
12101 SDValue CP = DAG.getConstantPool(C, getPointerTy());
12102 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
12103 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
12104 MachinePointerInfo::getConstantPool(),
12105 false, false, false, Alignment);
12106 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
12107 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12108 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12111 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12112 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12113 if (Subtarget->hasInt256()) {
12114 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12115 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
12116 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12118 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12119 DAG.getIntPtrConstant(0));
12122 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12123 DAG.getIntPtrConstant(0));
12124 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12125 DAG.getIntPtrConstant(2));
12126 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12127 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12128 static const int ShufMask[] = {0, 2, 4, 6};
12129 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12132 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12133 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12134 if (Subtarget->hasInt256()) {
12135 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
12137 SmallVector<SDValue,32> pshufbMask;
12138 for (unsigned i = 0; i < 2; ++i) {
12139 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
12140 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
12141 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
12142 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
12143 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
12144 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
12145 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
12146 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
12147 for (unsigned j = 0; j < 8; ++j)
12148 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
12150 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12151 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12152 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
12154 static const int ShufMask[] = {0, 2, -1, -1};
12155 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12157 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12158 DAG.getIntPtrConstant(0));
12159 return DAG.getNode(ISD::BITCAST, DL, VT, In);
12162 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12163 DAG.getIntPtrConstant(0));
12165 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12166 DAG.getIntPtrConstant(4));
12168 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
12169 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
12171 // The PSHUFB mask:
12172 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12173 -1, -1, -1, -1, -1, -1, -1, -1};
12175 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12176 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12177 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12179 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12180 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12182 // The MOVLHPS Mask:
12183 static const int ShufMask2[] = {0, 1, 4, 5};
12184 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12185 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
12188 // Handle truncation of V256 to V128 using shuffles.
12189 if (!VT.is128BitVector() || !InVT.is256BitVector())
12192 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12194 unsigned NumElems = VT.getVectorNumElements();
12195 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12197 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12198 // Prepare truncation shuffle mask
12199 for (unsigned i = 0; i != NumElems; ++i)
12200 MaskVec[i] = i * 2;
12201 SDValue V = DAG.getVectorShuffle(NVT, DL,
12202 DAG.getNode(ISD::BITCAST, DL, NVT, In),
12203 DAG.getUNDEF(NVT), &MaskVec[0]);
12204 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12205 DAG.getIntPtrConstant(0));
12208 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12209 SelectionDAG &DAG) const {
12210 assert(!Op.getSimpleValueType().isVector());
12212 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12213 /*IsSigned=*/ true, /*IsReplace=*/ false);
12214 SDValue FIST = Vals.first, StackSlot = Vals.second;
12215 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12216 if (!FIST.getNode()) return Op;
12218 if (StackSlot.getNode())
12219 // Load the result.
12220 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12221 FIST, StackSlot, MachinePointerInfo(),
12222 false, false, false, 0);
12224 // The node is the result.
12228 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12229 SelectionDAG &DAG) const {
12230 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12231 /*IsSigned=*/ false, /*IsReplace=*/ false);
12232 SDValue FIST = Vals.first, StackSlot = Vals.second;
12233 assert(FIST.getNode() && "Unexpected failure");
12235 if (StackSlot.getNode())
12236 // Load the result.
12237 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12238 FIST, StackSlot, MachinePointerInfo(),
12239 false, false, false, 0);
12241 // The node is the result.
12245 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12247 MVT VT = Op.getSimpleValueType();
12248 SDValue In = Op.getOperand(0);
12249 MVT SVT = In.getSimpleValueType();
12251 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12253 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12254 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12255 In, DAG.getUNDEF(SVT)));
12258 /// The only differences between FABS and FNEG are the mask and the logic op.
12259 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12260 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12261 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12262 "Wrong opcode for lowering FABS or FNEG.");
12264 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12266 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12267 // into an FNABS. We'll lower the FABS after that if it is still in use.
12269 for (SDNode *User : Op->uses())
12270 if (User->getOpcode() == ISD::FNEG)
12273 SDValue Op0 = Op.getOperand(0);
12274 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12277 MVT VT = Op.getSimpleValueType();
12278 // Assume scalar op for initialization; update for vector if needed.
12279 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
12280 // generate a 16-byte vector constant and logic op even for the scalar case.
12281 // Using a 16-byte mask allows folding the load of the mask with
12282 // the logic op, so it can save (~4 bytes) on code size.
12284 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
12285 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12286 // decide if we should generate a 16-byte constant mask when we only need 4 or
12287 // 8 bytes for the scalar case.
12288 if (VT.isVector()) {
12289 EltVT = VT.getVectorElementType();
12290 NumElts = VT.getVectorNumElements();
12293 unsigned EltBits = EltVT.getSizeInBits();
12294 LLVMContext *Context = DAG.getContext();
12295 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12297 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12298 Constant *C = ConstantInt::get(*Context, MaskElt);
12299 C = ConstantVector::getSplat(NumElts, C);
12300 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12301 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
12302 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12303 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12304 MachinePointerInfo::getConstantPool(),
12305 false, false, false, Alignment);
12307 if (VT.isVector()) {
12308 // For a vector, cast operands to a vector type, perform the logic op,
12309 // and cast the result back to the original value type.
12310 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12311 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12312 SDValue Operand = IsFNABS ?
12313 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12314 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12315 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12316 return DAG.getNode(ISD::BITCAST, dl, VT,
12317 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12320 // If not vector, then scalar.
12321 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12322 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12323 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12326 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12327 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12328 LLVMContext *Context = DAG.getContext();
12329 SDValue Op0 = Op.getOperand(0);
12330 SDValue Op1 = Op.getOperand(1);
12332 MVT VT = Op.getSimpleValueType();
12333 MVT SrcVT = Op1.getSimpleValueType();
12335 // If second operand is smaller, extend it first.
12336 if (SrcVT.bitsLT(VT)) {
12337 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12340 // And if it is bigger, shrink it first.
12341 if (SrcVT.bitsGT(VT)) {
12342 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
12346 // At this point the operands and the result should have the same
12347 // type, and that won't be f80 since that is not custom lowered.
12349 const fltSemantics &Sem =
12350 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12351 const unsigned SizeInBits = VT.getSizeInBits();
12353 SmallVector<Constant *, 4> CV(
12354 VT == MVT::f64 ? 2 : 4,
12355 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12357 // First, clear all bits but the sign bit from the second operand (sign).
12358 CV[0] = ConstantFP::get(*Context,
12359 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12360 Constant *C = ConstantVector::get(CV);
12361 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12362 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12363 MachinePointerInfo::getConstantPool(),
12364 false, false, false, 16);
12365 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12367 // Next, clear the sign bit from the first operand (magnitude).
12368 // If it's a constant, we can clear it here.
12369 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12370 APFloat APF = Op0CN->getValueAPF();
12371 // If the magnitude is a positive zero, the sign bit alone is enough.
12372 if (APF.isPosZero())
12375 CV[0] = ConstantFP::get(*Context, APF);
12377 CV[0] = ConstantFP::get(
12379 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12381 C = ConstantVector::get(CV);
12382 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12383 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12384 MachinePointerInfo::getConstantPool(),
12385 false, false, false, 16);
12386 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12387 if (!isa<ConstantFPSDNode>(Op0))
12388 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12390 // OR the magnitude value with the sign bit.
12391 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12394 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12395 SDValue N0 = Op.getOperand(0);
12397 MVT VT = Op.getSimpleValueType();
12399 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12400 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12401 DAG.getConstant(1, VT));
12402 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
12405 // Check whether an OR'd tree is PTEST-able.
12406 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12407 SelectionDAG &DAG) {
12408 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12410 if (!Subtarget->hasSSE41())
12413 if (!Op->hasOneUse())
12416 SDNode *N = Op.getNode();
12419 SmallVector<SDValue, 8> Opnds;
12420 DenseMap<SDValue, unsigned> VecInMap;
12421 SmallVector<SDValue, 8> VecIns;
12422 EVT VT = MVT::Other;
12424 // Recognize a special case where a vector is casted into wide integer to
12426 Opnds.push_back(N->getOperand(0));
12427 Opnds.push_back(N->getOperand(1));
12429 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12430 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12431 // BFS traverse all OR'd operands.
12432 if (I->getOpcode() == ISD::OR) {
12433 Opnds.push_back(I->getOperand(0));
12434 Opnds.push_back(I->getOperand(1));
12435 // Re-evaluate the number of nodes to be traversed.
12436 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12440 // Quit if a non-EXTRACT_VECTOR_ELT
12441 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12444 // Quit if without a constant index.
12445 SDValue Idx = I->getOperand(1);
12446 if (!isa<ConstantSDNode>(Idx))
12449 SDValue ExtractedFromVec = I->getOperand(0);
12450 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12451 if (M == VecInMap.end()) {
12452 VT = ExtractedFromVec.getValueType();
12453 // Quit if not 128/256-bit vector.
12454 if (!VT.is128BitVector() && !VT.is256BitVector())
12456 // Quit if not the same type.
12457 if (VecInMap.begin() != VecInMap.end() &&
12458 VT != VecInMap.begin()->first.getValueType())
12460 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12461 VecIns.push_back(ExtractedFromVec);
12463 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12466 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12467 "Not extracted from 128-/256-bit vector.");
12469 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12471 for (DenseMap<SDValue, unsigned>::const_iterator
12472 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12473 // Quit if not all elements are used.
12474 if (I->second != FullMask)
12478 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12480 // Cast all vectors into TestVT for PTEST.
12481 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12482 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12484 // If more than one full vectors are evaluated, OR them first before PTEST.
12485 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12486 // Each iteration will OR 2 nodes and append the result until there is only
12487 // 1 node left, i.e. the final OR'd value of all vectors.
12488 SDValue LHS = VecIns[Slot];
12489 SDValue RHS = VecIns[Slot + 1];
12490 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12493 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12494 VecIns.back(), VecIns.back());
12497 /// \brief return true if \c Op has a use that doesn't just read flags.
12498 static bool hasNonFlagsUse(SDValue Op) {
12499 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12501 SDNode *User = *UI;
12502 unsigned UOpNo = UI.getOperandNo();
12503 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12504 // Look pass truncate.
12505 UOpNo = User->use_begin().getOperandNo();
12506 User = *User->use_begin();
12509 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12510 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12516 /// Emit nodes that will be selected as "test Op0,Op0", or something
12518 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12519 SelectionDAG &DAG) const {
12520 if (Op.getValueType() == MVT::i1) {
12521 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12522 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12523 DAG.getConstant(0, MVT::i8));
12525 // CF and OF aren't always set the way we want. Determine which
12526 // of these we need.
12527 bool NeedCF = false;
12528 bool NeedOF = false;
12531 case X86::COND_A: case X86::COND_AE:
12532 case X86::COND_B: case X86::COND_BE:
12535 case X86::COND_G: case X86::COND_GE:
12536 case X86::COND_L: case X86::COND_LE:
12537 case X86::COND_O: case X86::COND_NO: {
12538 // Check if we really need to set the
12539 // Overflow flag. If NoSignedWrap is present
12540 // that is not actually needed.
12541 switch (Op->getOpcode()) {
12546 const BinaryWithFlagsSDNode *BinNode =
12547 cast<BinaryWithFlagsSDNode>(Op.getNode());
12548 if (BinNode->hasNoSignedWrap())
12558 // See if we can use the EFLAGS value from the operand instead of
12559 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12560 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12561 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12562 // Emit a CMP with 0, which is the TEST pattern.
12563 //if (Op.getValueType() == MVT::i1)
12564 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12565 // DAG.getConstant(0, MVT::i1));
12566 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12567 DAG.getConstant(0, Op.getValueType()));
12569 unsigned Opcode = 0;
12570 unsigned NumOperands = 0;
12572 // Truncate operations may prevent the merge of the SETCC instruction
12573 // and the arithmetic instruction before it. Attempt to truncate the operands
12574 // of the arithmetic instruction and use a reduced bit-width instruction.
12575 bool NeedTruncation = false;
12576 SDValue ArithOp = Op;
12577 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12578 SDValue Arith = Op->getOperand(0);
12579 // Both the trunc and the arithmetic op need to have one user each.
12580 if (Arith->hasOneUse())
12581 switch (Arith.getOpcode()) {
12588 NeedTruncation = true;
12594 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12595 // which may be the result of a CAST. We use the variable 'Op', which is the
12596 // non-casted variable when we check for possible users.
12597 switch (ArithOp.getOpcode()) {
12599 // Due to an isel shortcoming, be conservative if this add is likely to be
12600 // selected as part of a load-modify-store instruction. When the root node
12601 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12602 // uses of other nodes in the match, such as the ADD in this case. This
12603 // leads to the ADD being left around and reselected, with the result being
12604 // two adds in the output. Alas, even if none our users are stores, that
12605 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12606 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12607 // climbing the DAG back to the root, and it doesn't seem to be worth the
12609 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12610 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12611 if (UI->getOpcode() != ISD::CopyToReg &&
12612 UI->getOpcode() != ISD::SETCC &&
12613 UI->getOpcode() != ISD::STORE)
12616 if (ConstantSDNode *C =
12617 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12618 // An add of one will be selected as an INC.
12619 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12620 Opcode = X86ISD::INC;
12625 // An add of negative one (subtract of one) will be selected as a DEC.
12626 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12627 Opcode = X86ISD::DEC;
12633 // Otherwise use a regular EFLAGS-setting add.
12634 Opcode = X86ISD::ADD;
12639 // If we have a constant logical shift that's only used in a comparison
12640 // against zero turn it into an equivalent AND. This allows turning it into
12641 // a TEST instruction later.
12642 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12643 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12644 EVT VT = Op.getValueType();
12645 unsigned BitWidth = VT.getSizeInBits();
12646 unsigned ShAmt = Op->getConstantOperandVal(1);
12647 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12649 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12650 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12651 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12652 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12654 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12655 DAG.getConstant(Mask, VT));
12656 DAG.ReplaceAllUsesWith(Op, New);
12662 // If the primary and result isn't used, don't bother using X86ISD::AND,
12663 // because a TEST instruction will be better.
12664 if (!hasNonFlagsUse(Op))
12670 // Due to the ISEL shortcoming noted above, be conservative if this op is
12671 // likely to be selected as part of a load-modify-store instruction.
12672 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12673 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12674 if (UI->getOpcode() == ISD::STORE)
12677 // Otherwise use a regular EFLAGS-setting instruction.
12678 switch (ArithOp.getOpcode()) {
12679 default: llvm_unreachable("unexpected operator!");
12680 case ISD::SUB: Opcode = X86ISD::SUB; break;
12681 case ISD::XOR: Opcode = X86ISD::XOR; break;
12682 case ISD::AND: Opcode = X86ISD::AND; break;
12684 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12685 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12686 if (EFLAGS.getNode())
12689 Opcode = X86ISD::OR;
12703 return SDValue(Op.getNode(), 1);
12709 // If we found that truncation is beneficial, perform the truncation and
12711 if (NeedTruncation) {
12712 EVT VT = Op.getValueType();
12713 SDValue WideVal = Op->getOperand(0);
12714 EVT WideVT = WideVal.getValueType();
12715 unsigned ConvertedOp = 0;
12716 // Use a target machine opcode to prevent further DAGCombine
12717 // optimizations that may separate the arithmetic operations
12718 // from the setcc node.
12719 switch (WideVal.getOpcode()) {
12721 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12722 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12723 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12724 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12725 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12729 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12730 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12731 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12732 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12733 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12739 // Emit a CMP with 0, which is the TEST pattern.
12740 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12741 DAG.getConstant(0, Op.getValueType()));
12743 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12744 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12746 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12747 DAG.ReplaceAllUsesWith(Op, New);
12748 return SDValue(New.getNode(), 1);
12751 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12753 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12754 SDLoc dl, SelectionDAG &DAG) const {
12755 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12756 if (C->getAPIntValue() == 0)
12757 return EmitTest(Op0, X86CC, dl, DAG);
12759 if (Op0.getValueType() == MVT::i1)
12760 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12763 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12764 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12765 // Do the comparison at i32 if it's smaller, besides the Atom case.
12766 // This avoids subregister aliasing issues. Keep the smaller reference
12767 // if we're optimizing for size, however, as that'll allow better folding
12768 // of memory operations.
12769 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12770 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12771 Attribute::MinSize) &&
12772 !Subtarget->isAtom()) {
12773 unsigned ExtendOp =
12774 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12775 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12776 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12778 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12779 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12780 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12782 return SDValue(Sub.getNode(), 1);
12784 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12787 /// Convert a comparison if required by the subtarget.
12788 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12789 SelectionDAG &DAG) const {
12790 // If the subtarget does not support the FUCOMI instruction, floating-point
12791 // comparisons have to be converted.
12792 if (Subtarget->hasCMov() ||
12793 Cmp.getOpcode() != X86ISD::CMP ||
12794 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12795 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12798 // The instruction selector will select an FUCOM instruction instead of
12799 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12800 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12801 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12803 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12804 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12805 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12806 DAG.getConstant(8, MVT::i8));
12807 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12808 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12811 /// The minimum architected relative accuracy is 2^-12. We need one
12812 /// Newton-Raphson step to have a good float result (24 bits of precision).
12813 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12814 DAGCombinerInfo &DCI,
12815 unsigned &RefinementSteps,
12816 bool &UseOneConstNR) const {
12817 // FIXME: We should use instruction latency models to calculate the cost of
12818 // each potential sequence, but this is very hard to do reliably because
12819 // at least Intel's Core* chips have variable timing based on the number of
12820 // significant digits in the divisor and/or sqrt operand.
12821 if (!Subtarget->useSqrtEst())
12824 EVT VT = Op.getValueType();
12826 // SSE1 has rsqrtss and rsqrtps.
12827 // TODO: Add support for AVX512 (v16f32).
12828 // It is likely not profitable to do this for f64 because a double-precision
12829 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12830 // instructions: convert to single, rsqrtss, convert back to double, refine
12831 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12832 // along with FMA, this could be a throughput win.
12833 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12834 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12835 RefinementSteps = 1;
12836 UseOneConstNR = false;
12837 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12842 /// The minimum architected relative accuracy is 2^-12. We need one
12843 /// Newton-Raphson step to have a good float result (24 bits of precision).
12844 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12845 DAGCombinerInfo &DCI,
12846 unsigned &RefinementSteps) const {
12847 // FIXME: We should use instruction latency models to calculate the cost of
12848 // each potential sequence, but this is very hard to do reliably because
12849 // at least Intel's Core* chips have variable timing based on the number of
12850 // significant digits in the divisor.
12851 if (!Subtarget->useReciprocalEst())
12854 EVT VT = Op.getValueType();
12856 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12857 // TODO: Add support for AVX512 (v16f32).
12858 // It is likely not profitable to do this for f64 because a double-precision
12859 // reciprocal estimate with refinement on x86 prior to FMA requires
12860 // 15 instructions: convert to single, rcpss, convert back to double, refine
12861 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12862 // along with FMA, this could be a throughput win.
12863 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12864 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12865 RefinementSteps = ReciprocalEstimateRefinementSteps;
12866 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12871 /// If we have at least two divisions that use the same divisor, convert to
12872 /// multplication by a reciprocal. This may need to be adjusted for a given
12873 /// CPU if a division's cost is not at least twice the cost of a multiplication.
12874 /// This is because we still need one division to calculate the reciprocal and
12875 /// then we need two multiplies by that reciprocal as replacements for the
12876 /// original divisions.
12877 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
12878 return NumUsers > 1;
12881 static bool isAllOnes(SDValue V) {
12882 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12883 return C && C->isAllOnesValue();
12886 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12887 /// if it's possible.
12888 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12889 SDLoc dl, SelectionDAG &DAG) const {
12890 SDValue Op0 = And.getOperand(0);
12891 SDValue Op1 = And.getOperand(1);
12892 if (Op0.getOpcode() == ISD::TRUNCATE)
12893 Op0 = Op0.getOperand(0);
12894 if (Op1.getOpcode() == ISD::TRUNCATE)
12895 Op1 = Op1.getOperand(0);
12898 if (Op1.getOpcode() == ISD::SHL)
12899 std::swap(Op0, Op1);
12900 if (Op0.getOpcode() == ISD::SHL) {
12901 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12902 if (And00C->getZExtValue() == 1) {
12903 // If we looked past a truncate, check that it's only truncating away
12905 unsigned BitWidth = Op0.getValueSizeInBits();
12906 unsigned AndBitWidth = And.getValueSizeInBits();
12907 if (BitWidth > AndBitWidth) {
12909 DAG.computeKnownBits(Op0, Zeros, Ones);
12910 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12914 RHS = Op0.getOperand(1);
12916 } else if (Op1.getOpcode() == ISD::Constant) {
12917 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
12918 uint64_t AndRHSVal = AndRHS->getZExtValue();
12919 SDValue AndLHS = Op0;
12921 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
12922 LHS = AndLHS.getOperand(0);
12923 RHS = AndLHS.getOperand(1);
12926 // Use BT if the immediate can't be encoded in a TEST instruction.
12927 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
12929 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
12933 if (LHS.getNode()) {
12934 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12935 // instruction. Since the shift amount is in-range-or-undefined, we know
12936 // that doing a bittest on the i32 value is ok. We extend to i32 because
12937 // the encoding for the i16 version is larger than the i32 version.
12938 // Also promote i16 to i32 for performance / code size reason.
12939 if (LHS.getValueType() == MVT::i8 ||
12940 LHS.getValueType() == MVT::i16)
12941 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12943 // If the operand types disagree, extend the shift amount to match. Since
12944 // BT ignores high bits (like shifts) we can use anyextend.
12945 if (LHS.getValueType() != RHS.getValueType())
12946 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12948 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12949 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12950 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12951 DAG.getConstant(Cond, MVT::i8), BT);
12957 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12959 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12964 // SSE Condition code mapping:
12973 switch (SetCCOpcode) {
12974 default: llvm_unreachable("Unexpected SETCC condition");
12976 case ISD::SETEQ: SSECC = 0; break;
12978 case ISD::SETGT: Swap = true; // Fallthrough
12980 case ISD::SETOLT: SSECC = 1; break;
12982 case ISD::SETGE: Swap = true; // Fallthrough
12984 case ISD::SETOLE: SSECC = 2; break;
12985 case ISD::SETUO: SSECC = 3; break;
12987 case ISD::SETNE: SSECC = 4; break;
12988 case ISD::SETULE: Swap = true; // Fallthrough
12989 case ISD::SETUGE: SSECC = 5; break;
12990 case ISD::SETULT: Swap = true; // Fallthrough
12991 case ISD::SETUGT: SSECC = 6; break;
12992 case ISD::SETO: SSECC = 7; break;
12994 case ISD::SETONE: SSECC = 8; break;
12997 std::swap(Op0, Op1);
13002 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13003 // ones, and then concatenate the result back.
13004 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13005 MVT VT = Op.getSimpleValueType();
13007 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13008 "Unsupported value type for operation");
13010 unsigned NumElems = VT.getVectorNumElements();
13012 SDValue CC = Op.getOperand(2);
13014 // Extract the LHS vectors
13015 SDValue LHS = Op.getOperand(0);
13016 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13017 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13019 // Extract the RHS vectors
13020 SDValue RHS = Op.getOperand(1);
13021 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13022 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13024 // Issue the operation on the smaller types and concatenate the result back
13025 MVT EltVT = VT.getVectorElementType();
13026 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13027 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13028 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13029 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13032 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13033 SDValue Op0 = Op.getOperand(0);
13034 SDValue Op1 = Op.getOperand(1);
13035 SDValue CC = Op.getOperand(2);
13036 MVT VT = Op.getSimpleValueType();
13039 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13040 "Unexpected type for boolean compare operation");
13041 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13042 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13043 DAG.getConstant(-1, VT));
13044 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13045 DAG.getConstant(-1, VT));
13046 switch (SetCCOpcode) {
13047 default: llvm_unreachable("Unexpected SETCC condition");
13049 // (x != y) -> ~(x ^ y)
13050 return DAG.getNode(ISD::XOR, dl, VT,
13051 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13052 DAG.getConstant(-1, VT));
13054 // (x == y) -> (x ^ y)
13055 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13058 // (x > y) -> (x & ~y)
13059 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13062 // (x < y) -> (~x & y)
13063 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13066 // (x <= y) -> (~x | y)
13067 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13070 // (x >=y) -> (x | ~y)
13071 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13075 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13076 const X86Subtarget *Subtarget) {
13077 SDValue Op0 = Op.getOperand(0);
13078 SDValue Op1 = Op.getOperand(1);
13079 SDValue CC = Op.getOperand(2);
13080 MVT VT = Op.getSimpleValueType();
13083 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13084 Op.getValueType().getScalarType() == MVT::i1 &&
13085 "Cannot set masked compare for this operation");
13087 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13089 bool Unsigned = false;
13092 switch (SetCCOpcode) {
13093 default: llvm_unreachable("Unexpected SETCC condition");
13094 case ISD::SETNE: SSECC = 4; break;
13095 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13096 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13097 case ISD::SETLT: Swap = true; //fall-through
13098 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13099 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13100 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13101 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13102 case ISD::SETULE: Unsigned = true; //fall-through
13103 case ISD::SETLE: SSECC = 2; break;
13107 std::swap(Op0, Op1);
13109 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13110 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13111 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13112 DAG.getConstant(SSECC, MVT::i8));
13115 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13116 /// operand \p Op1. If non-trivial (for example because it's not constant)
13117 /// return an empty value.
13118 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13120 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13124 MVT VT = Op1.getSimpleValueType();
13125 MVT EVT = VT.getVectorElementType();
13126 unsigned n = VT.getVectorNumElements();
13127 SmallVector<SDValue, 8> ULTOp1;
13129 for (unsigned i = 0; i < n; ++i) {
13130 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13131 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13134 // Avoid underflow.
13135 APInt Val = Elt->getAPIntValue();
13139 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
13142 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13145 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13146 SelectionDAG &DAG) {
13147 SDValue Op0 = Op.getOperand(0);
13148 SDValue Op1 = Op.getOperand(1);
13149 SDValue CC = Op.getOperand(2);
13150 MVT VT = Op.getSimpleValueType();
13151 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13152 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13157 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13158 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13161 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13162 unsigned Opc = X86ISD::CMPP;
13163 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13164 assert(VT.getVectorNumElements() <= 16);
13165 Opc = X86ISD::CMPM;
13167 // In the two special cases we can't handle, emit two comparisons.
13170 unsigned CombineOpc;
13171 if (SetCCOpcode == ISD::SETUEQ) {
13172 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13174 assert(SetCCOpcode == ISD::SETONE);
13175 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13178 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13179 DAG.getConstant(CC0, MVT::i8));
13180 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13181 DAG.getConstant(CC1, MVT::i8));
13182 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13184 // Handle all other FP comparisons here.
13185 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13186 DAG.getConstant(SSECC, MVT::i8));
13189 // Break 256-bit integer vector compare into smaller ones.
13190 if (VT.is256BitVector() && !Subtarget->hasInt256())
13191 return Lower256IntVSETCC(Op, DAG);
13193 EVT OpVT = Op1.getValueType();
13194 if (OpVT.getVectorElementType() == MVT::i1)
13195 return LowerBoolVSETCC_AVX512(Op, DAG);
13197 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13198 if (Subtarget->hasAVX512()) {
13199 if (Op1.getValueType().is512BitVector() ||
13200 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13201 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13202 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13204 // In AVX-512 architecture setcc returns mask with i1 elements,
13205 // But there is no compare instruction for i8 and i16 elements in KNL.
13206 // We are not talking about 512-bit operands in this case, these
13207 // types are illegal.
13209 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13210 OpVT.getVectorElementType().getSizeInBits() >= 8))
13211 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13212 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13215 // We are handling one of the integer comparisons here. Since SSE only has
13216 // GT and EQ comparisons for integer, swapping operands and multiple
13217 // operations may be required for some comparisons.
13219 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13220 bool Subus = false;
13222 switch (SetCCOpcode) {
13223 default: llvm_unreachable("Unexpected SETCC condition");
13224 case ISD::SETNE: Invert = true;
13225 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13226 case ISD::SETLT: Swap = true;
13227 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13228 case ISD::SETGE: Swap = true;
13229 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13230 Invert = true; break;
13231 case ISD::SETULT: Swap = true;
13232 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13233 FlipSigns = true; break;
13234 case ISD::SETUGE: Swap = true;
13235 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13236 FlipSigns = true; Invert = true; break;
13239 // Special case: Use min/max operations for SETULE/SETUGE
13240 MVT VET = VT.getVectorElementType();
13242 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13243 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13246 switch (SetCCOpcode) {
13248 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
13249 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
13252 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13255 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13256 if (!MinMax && hasSubus) {
13257 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13259 // t = psubus Op0, Op1
13260 // pcmpeq t, <0..0>
13261 switch (SetCCOpcode) {
13263 case ISD::SETULT: {
13264 // If the comparison is against a constant we can turn this into a
13265 // setule. With psubus, setule does not require a swap. This is
13266 // beneficial because the constant in the register is no longer
13267 // destructed as the destination so it can be hoisted out of a loop.
13268 // Only do this pre-AVX since vpcmp* is no longer destructive.
13269 if (Subtarget->hasAVX())
13271 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13272 if (ULEOp1.getNode()) {
13274 Subus = true; Invert = false; Swap = false;
13278 // Psubus is better than flip-sign because it requires no inversion.
13279 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13280 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13284 Opc = X86ISD::SUBUS;
13290 std::swap(Op0, Op1);
13292 // Check that the operation in question is available (most are plain SSE2,
13293 // but PCMPGTQ and PCMPEQQ have different requirements).
13294 if (VT == MVT::v2i64) {
13295 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13296 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13298 // First cast everything to the right type.
13299 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13300 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13302 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13303 // bits of the inputs before performing those operations. The lower
13304 // compare is always unsigned.
13307 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
13309 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
13310 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
13311 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13312 Sign, Zero, Sign, Zero);
13314 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13315 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13317 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13318 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13319 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13321 // Create masks for only the low parts/high parts of the 64 bit integers.
13322 static const int MaskHi[] = { 1, 1, 3, 3 };
13323 static const int MaskLo[] = { 0, 0, 2, 2 };
13324 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13325 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13326 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13328 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13329 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13332 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13334 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13337 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13338 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13339 // pcmpeqd + pshufd + pand.
13340 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13342 // First cast everything to the right type.
13343 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13344 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13347 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13349 // Make sure the lower and upper halves are both all-ones.
13350 static const int Mask[] = { 1, 0, 3, 2 };
13351 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13352 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13355 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13357 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13361 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13362 // bits of the inputs before performing those operations.
13364 EVT EltVT = VT.getVectorElementType();
13365 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
13366 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13367 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13370 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13372 // If the logical-not of the result is required, perform that now.
13374 Result = DAG.getNOT(dl, Result, VT);
13377 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13380 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13381 getZeroVector(VT, Subtarget, DAG, dl));
13386 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13388 MVT VT = Op.getSimpleValueType();
13390 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13392 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13393 && "SetCC type must be 8-bit or 1-bit integer");
13394 SDValue Op0 = Op.getOperand(0);
13395 SDValue Op1 = Op.getOperand(1);
13397 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13399 // Optimize to BT if possible.
13400 // Lower (X & (1 << N)) == 0 to BT(X, N).
13401 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13402 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13403 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13404 Op1.getOpcode() == ISD::Constant &&
13405 cast<ConstantSDNode>(Op1)->isNullValue() &&
13406 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13407 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13408 if (NewSetCC.getNode()) {
13410 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13415 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13417 if (Op1.getOpcode() == ISD::Constant &&
13418 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13419 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13420 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13422 // If the input is a setcc, then reuse the input setcc or use a new one with
13423 // the inverted condition.
13424 if (Op0.getOpcode() == X86ISD::SETCC) {
13425 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13426 bool Invert = (CC == ISD::SETNE) ^
13427 cast<ConstantSDNode>(Op1)->isNullValue();
13431 CCode = X86::GetOppositeBranchCondition(CCode);
13432 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13433 DAG.getConstant(CCode, MVT::i8),
13434 Op0.getOperand(1));
13436 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13440 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13441 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13442 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13444 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13445 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
13448 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13449 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
13450 if (X86CC == X86::COND_INVALID)
13453 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13454 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13455 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13456 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
13458 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13462 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13463 static bool isX86LogicalCmp(SDValue Op) {
13464 unsigned Opc = Op.getNode()->getOpcode();
13465 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13466 Opc == X86ISD::SAHF)
13468 if (Op.getResNo() == 1 &&
13469 (Opc == X86ISD::ADD ||
13470 Opc == X86ISD::SUB ||
13471 Opc == X86ISD::ADC ||
13472 Opc == X86ISD::SBB ||
13473 Opc == X86ISD::SMUL ||
13474 Opc == X86ISD::UMUL ||
13475 Opc == X86ISD::INC ||
13476 Opc == X86ISD::DEC ||
13477 Opc == X86ISD::OR ||
13478 Opc == X86ISD::XOR ||
13479 Opc == X86ISD::AND))
13482 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13488 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13489 if (V.getOpcode() != ISD::TRUNCATE)
13492 SDValue VOp0 = V.getOperand(0);
13493 unsigned InBits = VOp0.getValueSizeInBits();
13494 unsigned Bits = V.getValueSizeInBits();
13495 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13498 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13499 bool addTest = true;
13500 SDValue Cond = Op.getOperand(0);
13501 SDValue Op1 = Op.getOperand(1);
13502 SDValue Op2 = Op.getOperand(2);
13504 EVT VT = Op1.getValueType();
13507 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13508 // are available or VBLENDV if AVX is available.
13509 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13510 if (Cond.getOpcode() == ISD::SETCC &&
13511 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13512 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13513 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13514 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13515 int SSECC = translateX86FSETCC(
13516 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13519 if (Subtarget->hasAVX512()) {
13520 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13521 DAG.getConstant(SSECC, MVT::i8));
13522 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13525 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13526 DAG.getConstant(SSECC, MVT::i8));
13528 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13529 // of 3 logic instructions for size savings and potentially speed.
13530 // Unfortunately, there is no scalar form of VBLENDV.
13532 // If either operand is a constant, don't try this. We can expect to
13533 // optimize away at least one of the logic instructions later in that
13534 // case, so that sequence would be faster than a variable blend.
13536 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13537 // uses XMM0 as the selection register. That may need just as many
13538 // instructions as the AND/ANDN/OR sequence due to register moves, so
13541 if (Subtarget->hasAVX() &&
13542 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13544 // Convert to vectors, do a VSELECT, and convert back to scalar.
13545 // All of the conversions should be optimized away.
13547 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13548 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13549 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13550 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13552 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13553 VCmp = DAG.getNode(ISD::BITCAST, DL, VCmpVT, VCmp);
13555 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13557 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13558 VSel, DAG.getIntPtrConstant(0));
13560 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13561 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13562 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13566 if (Cond.getOpcode() == ISD::SETCC) {
13567 SDValue NewCond = LowerSETCC(Cond, DAG);
13568 if (NewCond.getNode())
13572 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13573 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13574 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13575 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13576 if (Cond.getOpcode() == X86ISD::SETCC &&
13577 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13578 isZero(Cond.getOperand(1).getOperand(1))) {
13579 SDValue Cmp = Cond.getOperand(1);
13581 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13583 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13584 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13585 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13587 SDValue CmpOp0 = Cmp.getOperand(0);
13588 // Apply further optimizations for special cases
13589 // (select (x != 0), -1, 0) -> neg & sbb
13590 // (select (x == 0), 0, -1) -> neg & sbb
13591 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13592 if (YC->isNullValue() &&
13593 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13594 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13595 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13596 DAG.getConstant(0, CmpOp0.getValueType()),
13598 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13599 DAG.getConstant(X86::COND_B, MVT::i8),
13600 SDValue(Neg.getNode(), 1));
13604 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13605 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
13606 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13608 SDValue Res = // Res = 0 or -1.
13609 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13610 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
13612 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13613 Res = DAG.getNOT(DL, Res, Res.getValueType());
13615 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13616 if (!N2C || !N2C->isNullValue())
13617 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13622 // Look past (and (setcc_carry (cmp ...)), 1).
13623 if (Cond.getOpcode() == ISD::AND &&
13624 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13625 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13626 if (C && C->getAPIntValue() == 1)
13627 Cond = Cond.getOperand(0);
13630 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13631 // setting operand in place of the X86ISD::SETCC.
13632 unsigned CondOpcode = Cond.getOpcode();
13633 if (CondOpcode == X86ISD::SETCC ||
13634 CondOpcode == X86ISD::SETCC_CARRY) {
13635 CC = Cond.getOperand(0);
13637 SDValue Cmp = Cond.getOperand(1);
13638 unsigned Opc = Cmp.getOpcode();
13639 MVT VT = Op.getSimpleValueType();
13641 bool IllegalFPCMov = false;
13642 if (VT.isFloatingPoint() && !VT.isVector() &&
13643 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13644 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13646 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13647 Opc == X86ISD::BT) { // FIXME
13651 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13652 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13653 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13654 Cond.getOperand(0).getValueType() != MVT::i8)) {
13655 SDValue LHS = Cond.getOperand(0);
13656 SDValue RHS = Cond.getOperand(1);
13657 unsigned X86Opcode;
13660 switch (CondOpcode) {
13661 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13662 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13663 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13664 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13665 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13666 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13667 default: llvm_unreachable("unexpected overflowing operator");
13669 if (CondOpcode == ISD::UMULO)
13670 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13673 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13675 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13677 if (CondOpcode == ISD::UMULO)
13678 Cond = X86Op.getValue(2);
13680 Cond = X86Op.getValue(1);
13682 CC = DAG.getConstant(X86Cond, MVT::i8);
13687 // Look pass the truncate if the high bits are known zero.
13688 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13689 Cond = Cond.getOperand(0);
13691 // We know the result of AND is compared against zero. Try to match
13693 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13694 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13695 if (NewSetCC.getNode()) {
13696 CC = NewSetCC.getOperand(0);
13697 Cond = NewSetCC.getOperand(1);
13704 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13705 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13708 // a < b ? -1 : 0 -> RES = ~setcc_carry
13709 // a < b ? 0 : -1 -> RES = setcc_carry
13710 // a >= b ? -1 : 0 -> RES = setcc_carry
13711 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13712 if (Cond.getOpcode() == X86ISD::SUB) {
13713 Cond = ConvertCmpIfNecessary(Cond, DAG);
13714 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13716 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13717 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13718 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13719 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
13720 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13721 return DAG.getNOT(DL, Res, Res.getValueType());
13726 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13727 // widen the cmov and push the truncate through. This avoids introducing a new
13728 // branch during isel and doesn't add any extensions.
13729 if (Op.getValueType() == MVT::i8 &&
13730 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13731 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13732 if (T1.getValueType() == T2.getValueType() &&
13733 // Blacklist CopyFromReg to avoid partial register stalls.
13734 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13735 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13736 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13737 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13741 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13742 // condition is true.
13743 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13744 SDValue Ops[] = { Op2, Op1, CC, Cond };
13745 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13748 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13749 SelectionDAG &DAG) {
13750 MVT VT = Op->getSimpleValueType(0);
13751 SDValue In = Op->getOperand(0);
13752 MVT InVT = In.getSimpleValueType();
13753 MVT VTElt = VT.getVectorElementType();
13754 MVT InVTElt = InVT.getVectorElementType();
13758 if ((InVTElt == MVT::i1) &&
13759 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13760 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13762 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13763 VTElt.getSizeInBits() <= 16)) ||
13765 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13766 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13768 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13769 VTElt.getSizeInBits() >= 32))))
13770 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13772 unsigned int NumElts = VT.getVectorNumElements();
13774 if (NumElts != 8 && NumElts != 16)
13777 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13778 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13779 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13780 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13783 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13784 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13786 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
13787 Constant *C = ConstantInt::get(*DAG.getContext(),
13788 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
13790 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
13791 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13792 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
13793 MachinePointerInfo::getConstantPool(),
13794 false, false, false, Alignment);
13795 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
13796 if (VT.is512BitVector())
13798 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
13801 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13802 SelectionDAG &DAG) {
13803 MVT VT = Op->getSimpleValueType(0);
13804 SDValue In = Op->getOperand(0);
13805 MVT InVT = In.getSimpleValueType();
13808 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13809 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13811 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13812 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13813 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13816 if (Subtarget->hasInt256())
13817 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13819 // Optimize vectors in AVX mode
13820 // Sign extend v8i16 to v8i32 and
13823 // Divide input vector into two parts
13824 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13825 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13826 // concat the vectors to original VT
13828 unsigned NumElems = InVT.getVectorNumElements();
13829 SDValue Undef = DAG.getUNDEF(InVT);
13831 SmallVector<int,8> ShufMask1(NumElems, -1);
13832 for (unsigned i = 0; i != NumElems/2; ++i)
13835 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13837 SmallVector<int,8> ShufMask2(NumElems, -1);
13838 for (unsigned i = 0; i != NumElems/2; ++i)
13839 ShufMask2[i] = i + NumElems/2;
13841 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13843 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13844 VT.getVectorNumElements()/2);
13846 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13847 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13849 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13852 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13853 // may emit an illegal shuffle but the expansion is still better than scalar
13854 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13855 // we'll emit a shuffle and a arithmetic shift.
13856 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13857 // TODO: It is possible to support ZExt by zeroing the undef values during
13858 // the shuffle phase or after the shuffle.
13859 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13860 SelectionDAG &DAG) {
13861 MVT RegVT = Op.getSimpleValueType();
13862 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13863 assert(RegVT.isInteger() &&
13864 "We only custom lower integer vector sext loads.");
13866 // Nothing useful we can do without SSE2 shuffles.
13867 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13869 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13871 EVT MemVT = Ld->getMemoryVT();
13872 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13873 unsigned RegSz = RegVT.getSizeInBits();
13875 ISD::LoadExtType Ext = Ld->getExtensionType();
13877 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13878 && "Only anyext and sext are currently implemented.");
13879 assert(MemVT != RegVT && "Cannot extend to the same type");
13880 assert(MemVT.isVector() && "Must load a vector from memory");
13882 unsigned NumElems = RegVT.getVectorNumElements();
13883 unsigned MemSz = MemVT.getSizeInBits();
13884 assert(RegSz > MemSz && "Register size must be greater than the mem size");
13886 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
13887 // The only way in which we have a legal 256-bit vector result but not the
13888 // integer 256-bit operations needed to directly lower a sextload is if we
13889 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
13890 // a 128-bit vector and a normal sign_extend to 256-bits that should get
13891 // correctly legalized. We do this late to allow the canonical form of
13892 // sextload to persist throughout the rest of the DAG combiner -- it wants
13893 // to fold together any extensions it can, and so will fuse a sign_extend
13894 // of an sextload into a sextload targeting a wider value.
13896 if (MemSz == 128) {
13897 // Just switch this to a normal load.
13898 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
13899 "it must be a legal 128-bit vector "
13901 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
13902 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
13903 Ld->isInvariant(), Ld->getAlignment());
13905 assert(MemSz < 128 &&
13906 "Can't extend a type wider than 128 bits to a 256 bit vector!");
13907 // Do an sext load to a 128-bit vector type. We want to use the same
13908 // number of elements, but elements half as wide. This will end up being
13909 // recursively lowered by this routine, but will succeed as we definitely
13910 // have all the necessary features if we're using AVX1.
13912 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
13913 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
13915 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
13916 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
13917 Ld->isNonTemporal(), Ld->isInvariant(),
13918 Ld->getAlignment());
13921 // Replace chain users with the new chain.
13922 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
13923 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
13925 // Finally, do a normal sign-extend to the desired register.
13926 return DAG.getSExtOrTrunc(Load, dl, RegVT);
13929 // All sizes must be a power of two.
13930 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
13931 "Non-power-of-two elements are not custom lowered!");
13933 // Attempt to load the original value using scalar loads.
13934 // Find the largest scalar type that divides the total loaded size.
13935 MVT SclrLoadTy = MVT::i8;
13936 for (MVT Tp : MVT::integer_valuetypes()) {
13937 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
13942 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
13943 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
13945 SclrLoadTy = MVT::f64;
13947 // Calculate the number of scalar loads that we need to perform
13948 // in order to load our vector from memory.
13949 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
13951 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
13952 "Can only lower sext loads with a single scalar load!");
13954 unsigned loadRegZize = RegSz;
13955 if (Ext == ISD::SEXTLOAD && RegSz == 256)
13958 // Represent our vector as a sequence of elements which are the
13959 // largest scalar that we can load.
13960 EVT LoadUnitVecVT = EVT::getVectorVT(
13961 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
13963 // Represent the data using the same element type that is stored in
13964 // memory. In practice, we ''widen'' MemVT.
13966 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
13967 loadRegZize / MemVT.getScalarType().getSizeInBits());
13969 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
13970 "Invalid vector type");
13972 // We can't shuffle using an illegal type.
13973 assert(TLI.isTypeLegal(WideVecVT) &&
13974 "We only lower types that form legal widened vector types");
13976 SmallVector<SDValue, 8> Chains;
13977 SDValue Ptr = Ld->getBasePtr();
13978 SDValue Increment =
13979 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
13980 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
13982 for (unsigned i = 0; i < NumLoads; ++i) {
13983 // Perform a single load.
13984 SDValue ScalarLoad =
13985 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
13986 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
13987 Ld->getAlignment());
13988 Chains.push_back(ScalarLoad.getValue(1));
13989 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
13990 // another round of DAGCombining.
13992 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
13994 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
13995 ScalarLoad, DAG.getIntPtrConstant(i));
13997 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14000 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14002 // Bitcast the loaded value to a vector of the original element type, in
14003 // the size of the target vector type.
14004 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
14005 unsigned SizeRatio = RegSz / MemSz;
14007 if (Ext == ISD::SEXTLOAD) {
14008 // If we have SSE4.1, we can directly emit a VSEXT node.
14009 if (Subtarget->hasSSE41()) {
14010 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14011 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14015 // Otherwise we'll shuffle the small elements in the high bits of the
14016 // larger type and perform an arithmetic shift. If the shift is not legal
14017 // it's better to scalarize.
14018 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14019 "We can't implement a sext load without an arithmetic right shift!");
14021 // Redistribute the loaded elements into the different locations.
14022 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14023 for (unsigned i = 0; i != NumElems; ++i)
14024 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14026 SDValue Shuff = DAG.getVectorShuffle(
14027 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14029 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14031 // Build the arithmetic shift.
14032 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14033 MemVT.getVectorElementType().getSizeInBits();
14035 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
14037 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14041 // Redistribute the loaded elements into the different locations.
14042 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14043 for (unsigned i = 0; i != NumElems; ++i)
14044 ShuffleVec[i * SizeRatio] = i;
14046 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14047 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14049 // Bitcast to the requested type.
14050 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14051 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14055 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14056 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14057 // from the AND / OR.
14058 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14059 Opc = Op.getOpcode();
14060 if (Opc != ISD::OR && Opc != ISD::AND)
14062 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14063 Op.getOperand(0).hasOneUse() &&
14064 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14065 Op.getOperand(1).hasOneUse());
14068 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14069 // 1 and that the SETCC node has a single use.
14070 static bool isXor1OfSetCC(SDValue Op) {
14071 if (Op.getOpcode() != ISD::XOR)
14073 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14074 if (N1C && N1C->getAPIntValue() == 1) {
14075 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14076 Op.getOperand(0).hasOneUse();
14081 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14082 bool addTest = true;
14083 SDValue Chain = Op.getOperand(0);
14084 SDValue Cond = Op.getOperand(1);
14085 SDValue Dest = Op.getOperand(2);
14088 bool Inverted = false;
14090 if (Cond.getOpcode() == ISD::SETCC) {
14091 // Check for setcc([su]{add,sub,mul}o == 0).
14092 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14093 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14094 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14095 Cond.getOperand(0).getResNo() == 1 &&
14096 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14097 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14098 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14099 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14100 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14101 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14103 Cond = Cond.getOperand(0);
14105 SDValue NewCond = LowerSETCC(Cond, DAG);
14106 if (NewCond.getNode())
14111 // FIXME: LowerXALUO doesn't handle these!!
14112 else if (Cond.getOpcode() == X86ISD::ADD ||
14113 Cond.getOpcode() == X86ISD::SUB ||
14114 Cond.getOpcode() == X86ISD::SMUL ||
14115 Cond.getOpcode() == X86ISD::UMUL)
14116 Cond = LowerXALUO(Cond, DAG);
14119 // Look pass (and (setcc_carry (cmp ...)), 1).
14120 if (Cond.getOpcode() == ISD::AND &&
14121 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14122 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14123 if (C && C->getAPIntValue() == 1)
14124 Cond = Cond.getOperand(0);
14127 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14128 // setting operand in place of the X86ISD::SETCC.
14129 unsigned CondOpcode = Cond.getOpcode();
14130 if (CondOpcode == X86ISD::SETCC ||
14131 CondOpcode == X86ISD::SETCC_CARRY) {
14132 CC = Cond.getOperand(0);
14134 SDValue Cmp = Cond.getOperand(1);
14135 unsigned Opc = Cmp.getOpcode();
14136 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14137 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14141 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14145 // These can only come from an arithmetic instruction with overflow,
14146 // e.g. SADDO, UADDO.
14147 Cond = Cond.getNode()->getOperand(1);
14153 CondOpcode = Cond.getOpcode();
14154 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14155 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14156 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14157 Cond.getOperand(0).getValueType() != MVT::i8)) {
14158 SDValue LHS = Cond.getOperand(0);
14159 SDValue RHS = Cond.getOperand(1);
14160 unsigned X86Opcode;
14163 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14164 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14166 switch (CondOpcode) {
14167 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14169 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14171 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14174 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14175 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14177 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14179 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14182 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14183 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14184 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14185 default: llvm_unreachable("unexpected overflowing operator");
14188 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14189 if (CondOpcode == ISD::UMULO)
14190 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14193 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14195 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14197 if (CondOpcode == ISD::UMULO)
14198 Cond = X86Op.getValue(2);
14200 Cond = X86Op.getValue(1);
14202 CC = DAG.getConstant(X86Cond, MVT::i8);
14206 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14207 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14208 if (CondOpc == ISD::OR) {
14209 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14210 // two branches instead of an explicit OR instruction with a
14212 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14213 isX86LogicalCmp(Cmp)) {
14214 CC = Cond.getOperand(0).getOperand(0);
14215 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14216 Chain, Dest, CC, Cmp);
14217 CC = Cond.getOperand(1).getOperand(0);
14221 } else { // ISD::AND
14222 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14223 // two branches instead of an explicit AND instruction with a
14224 // separate test. However, we only do this if this block doesn't
14225 // have a fall-through edge, because this requires an explicit
14226 // jmp when the condition is false.
14227 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14228 isX86LogicalCmp(Cmp) &&
14229 Op.getNode()->hasOneUse()) {
14230 X86::CondCode CCode =
14231 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14232 CCode = X86::GetOppositeBranchCondition(CCode);
14233 CC = DAG.getConstant(CCode, MVT::i8);
14234 SDNode *User = *Op.getNode()->use_begin();
14235 // Look for an unconditional branch following this conditional branch.
14236 // We need this because we need to reverse the successors in order
14237 // to implement FCMP_OEQ.
14238 if (User->getOpcode() == ISD::BR) {
14239 SDValue FalseBB = User->getOperand(1);
14241 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14242 assert(NewBR == User);
14246 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14247 Chain, Dest, CC, Cmp);
14248 X86::CondCode CCode =
14249 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14250 CCode = X86::GetOppositeBranchCondition(CCode);
14251 CC = DAG.getConstant(CCode, MVT::i8);
14257 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14258 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14259 // It should be transformed during dag combiner except when the condition
14260 // is set by a arithmetics with overflow node.
14261 X86::CondCode CCode =
14262 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14263 CCode = X86::GetOppositeBranchCondition(CCode);
14264 CC = DAG.getConstant(CCode, MVT::i8);
14265 Cond = Cond.getOperand(0).getOperand(1);
14267 } else if (Cond.getOpcode() == ISD::SETCC &&
14268 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14269 // For FCMP_OEQ, we can emit
14270 // two branches instead of an explicit AND instruction with a
14271 // separate test. However, we only do this if this block doesn't
14272 // have a fall-through edge, because this requires an explicit
14273 // jmp when the condition is false.
14274 if (Op.getNode()->hasOneUse()) {
14275 SDNode *User = *Op.getNode()->use_begin();
14276 // Look for an unconditional branch following this conditional branch.
14277 // We need this because we need to reverse the successors in order
14278 // to implement FCMP_OEQ.
14279 if (User->getOpcode() == ISD::BR) {
14280 SDValue FalseBB = User->getOperand(1);
14282 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14283 assert(NewBR == User);
14287 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14288 Cond.getOperand(0), Cond.getOperand(1));
14289 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14290 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
14291 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14292 Chain, Dest, CC, Cmp);
14293 CC = DAG.getConstant(X86::COND_P, MVT::i8);
14298 } else if (Cond.getOpcode() == ISD::SETCC &&
14299 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14300 // For FCMP_UNE, we can emit
14301 // two branches instead of an explicit AND instruction with a
14302 // separate test. However, we only do this if this block doesn't
14303 // have a fall-through edge, because this requires an explicit
14304 // jmp when the condition is false.
14305 if (Op.getNode()->hasOneUse()) {
14306 SDNode *User = *Op.getNode()->use_begin();
14307 // Look for an unconditional branch following this conditional branch.
14308 // We need this because we need to reverse the successors in order
14309 // to implement FCMP_UNE.
14310 if (User->getOpcode() == ISD::BR) {
14311 SDValue FalseBB = User->getOperand(1);
14313 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14314 assert(NewBR == User);
14317 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14318 Cond.getOperand(0), Cond.getOperand(1));
14319 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14320 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
14321 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14322 Chain, Dest, CC, Cmp);
14323 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
14333 // Look pass the truncate if the high bits are known zero.
14334 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14335 Cond = Cond.getOperand(0);
14337 // We know the result of AND is compared against zero. Try to match
14339 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14340 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14341 if (NewSetCC.getNode()) {
14342 CC = NewSetCC.getOperand(0);
14343 Cond = NewSetCC.getOperand(1);
14350 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14351 CC = DAG.getConstant(X86Cond, MVT::i8);
14352 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14354 Cond = ConvertCmpIfNecessary(Cond, DAG);
14355 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14356 Chain, Dest, CC, Cond);
14359 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14360 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14361 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14362 // that the guard pages used by the OS virtual memory manager are allocated in
14363 // correct sequence.
14365 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14366 SelectionDAG &DAG) const {
14367 MachineFunction &MF = DAG.getMachineFunction();
14368 bool SplitStack = MF.shouldSplitStack();
14369 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14374 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14375 SDNode* Node = Op.getNode();
14377 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14378 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14379 " not tell us which reg is the stack pointer!");
14380 EVT VT = Node->getValueType(0);
14381 SDValue Tmp1 = SDValue(Node, 0);
14382 SDValue Tmp2 = SDValue(Node, 1);
14383 SDValue Tmp3 = Node->getOperand(2);
14384 SDValue Chain = Tmp1.getOperand(0);
14386 // Chain the dynamic stack allocation so that it doesn't modify the stack
14387 // pointer when other instructions are using the stack.
14388 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
14391 SDValue Size = Tmp2.getOperand(1);
14392 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14393 Chain = SP.getValue(1);
14394 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14395 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14396 unsigned StackAlign = TFI.getStackAlignment();
14397 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14398 if (Align > StackAlign)
14399 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14400 DAG.getConstant(-(uint64_t)Align, VT));
14401 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14403 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
14404 DAG.getIntPtrConstant(0, true), SDValue(),
14407 SDValue Ops[2] = { Tmp1, Tmp2 };
14408 return DAG.getMergeValues(Ops, dl);
14412 SDValue Chain = Op.getOperand(0);
14413 SDValue Size = Op.getOperand(1);
14414 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14415 EVT VT = Op.getNode()->getValueType(0);
14417 bool Is64Bit = Subtarget->is64Bit();
14418 EVT SPTy = getPointerTy();
14421 MachineRegisterInfo &MRI = MF.getRegInfo();
14424 // The 64 bit implementation of segmented stacks needs to clobber both r10
14425 // r11. This makes it impossible to use it along with nested parameters.
14426 const Function *F = MF.getFunction();
14428 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14430 if (I->hasNestAttr())
14431 report_fatal_error("Cannot use segmented stacks with functions that "
14432 "have nested arguments.");
14435 const TargetRegisterClass *AddrRegClass =
14436 getRegClassFor(getPointerTy());
14437 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14438 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14439 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14440 DAG.getRegister(Vreg, SPTy));
14441 SDValue Ops1[2] = { Value, Chain };
14442 return DAG.getMergeValues(Ops1, dl);
14445 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14447 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14448 Flag = Chain.getValue(1);
14449 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14451 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14453 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14454 unsigned SPReg = RegInfo->getStackRegister();
14455 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14456 Chain = SP.getValue(1);
14459 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14460 DAG.getConstant(-(uint64_t)Align, VT));
14461 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14464 SDValue Ops1[2] = { SP, Chain };
14465 return DAG.getMergeValues(Ops1, dl);
14469 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14470 MachineFunction &MF = DAG.getMachineFunction();
14471 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14473 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14476 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14477 // vastart just stores the address of the VarArgsFrameIndex slot into the
14478 // memory location argument.
14479 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14481 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14482 MachinePointerInfo(SV), false, false, 0);
14486 // gp_offset (0 - 6 * 8)
14487 // fp_offset (48 - 48 + 8 * 16)
14488 // overflow_arg_area (point to parameters coming in memory).
14490 SmallVector<SDValue, 8> MemOps;
14491 SDValue FIN = Op.getOperand(1);
14493 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14494 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14496 FIN, MachinePointerInfo(SV), false, false, 0);
14497 MemOps.push_back(Store);
14500 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14501 FIN, DAG.getIntPtrConstant(4));
14502 Store = DAG.getStore(Op.getOperand(0), DL,
14503 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
14505 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14506 MemOps.push_back(Store);
14508 // Store ptr to overflow_arg_area
14509 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14510 FIN, DAG.getIntPtrConstant(4));
14511 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14513 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14514 MachinePointerInfo(SV, 8),
14516 MemOps.push_back(Store);
14518 // Store ptr to reg_save_area.
14519 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14520 FIN, DAG.getIntPtrConstant(8));
14521 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14523 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14524 MachinePointerInfo(SV, 16), false, false, 0);
14525 MemOps.push_back(Store);
14526 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14529 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14530 assert(Subtarget->is64Bit() &&
14531 "LowerVAARG only handles 64-bit va_arg!");
14532 assert((Subtarget->isTargetLinux() ||
14533 Subtarget->isTargetDarwin()) &&
14534 "Unhandled target in LowerVAARG");
14535 assert(Op.getNode()->getNumOperands() == 4);
14536 SDValue Chain = Op.getOperand(0);
14537 SDValue SrcPtr = Op.getOperand(1);
14538 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14539 unsigned Align = Op.getConstantOperandVal(3);
14542 EVT ArgVT = Op.getNode()->getValueType(0);
14543 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14544 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14547 // Decide which area this value should be read from.
14548 // TODO: Implement the AMD64 ABI in its entirety. This simple
14549 // selection mechanism works only for the basic types.
14550 if (ArgVT == MVT::f80) {
14551 llvm_unreachable("va_arg for f80 not yet implemented");
14552 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14553 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14554 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14555 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14557 llvm_unreachable("Unhandled argument type in LowerVAARG");
14560 if (ArgMode == 2) {
14561 // Sanity Check: Make sure using fp_offset makes sense.
14562 assert(!DAG.getTarget().Options.UseSoftFloat &&
14563 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14564 Attribute::NoImplicitFloat)) &&
14565 Subtarget->hasSSE1());
14568 // Insert VAARG_64 node into the DAG
14569 // VAARG_64 returns two values: Variable Argument Address, Chain
14570 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, MVT::i32),
14571 DAG.getConstant(ArgMode, MVT::i8),
14572 DAG.getConstant(Align, MVT::i32)};
14573 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14574 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14575 VTs, InstOps, MVT::i64,
14576 MachinePointerInfo(SV),
14578 /*Volatile=*/false,
14580 /*WriteMem=*/true);
14581 Chain = VAARG.getValue(1);
14583 // Load the next argument and return it
14584 return DAG.getLoad(ArgVT, dl,
14587 MachinePointerInfo(),
14588 false, false, false, 0);
14591 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14592 SelectionDAG &DAG) {
14593 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14594 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14595 SDValue Chain = Op.getOperand(0);
14596 SDValue DstPtr = Op.getOperand(1);
14597 SDValue SrcPtr = Op.getOperand(2);
14598 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14599 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14602 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14603 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
14605 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14608 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14609 // amount is a constant. Takes immediate version of shift as input.
14610 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14611 SDValue SrcOp, uint64_t ShiftAmt,
14612 SelectionDAG &DAG) {
14613 MVT ElementType = VT.getVectorElementType();
14615 // Fold this packed shift into its first operand if ShiftAmt is 0.
14619 // Check for ShiftAmt >= element width
14620 if (ShiftAmt >= ElementType.getSizeInBits()) {
14621 if (Opc == X86ISD::VSRAI)
14622 ShiftAmt = ElementType.getSizeInBits() - 1;
14624 return DAG.getConstant(0, VT);
14627 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14628 && "Unknown target vector shift-by-constant node");
14630 // Fold this packed vector shift into a build vector if SrcOp is a
14631 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14632 if (VT == SrcOp.getSimpleValueType() &&
14633 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14634 SmallVector<SDValue, 8> Elts;
14635 unsigned NumElts = SrcOp->getNumOperands();
14636 ConstantSDNode *ND;
14639 default: llvm_unreachable(nullptr);
14640 case X86ISD::VSHLI:
14641 for (unsigned i=0; i!=NumElts; ++i) {
14642 SDValue CurrentOp = SrcOp->getOperand(i);
14643 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14644 Elts.push_back(CurrentOp);
14647 ND = cast<ConstantSDNode>(CurrentOp);
14648 const APInt &C = ND->getAPIntValue();
14649 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
14652 case X86ISD::VSRLI:
14653 for (unsigned i=0; i!=NumElts; ++i) {
14654 SDValue CurrentOp = SrcOp->getOperand(i);
14655 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14656 Elts.push_back(CurrentOp);
14659 ND = cast<ConstantSDNode>(CurrentOp);
14660 const APInt &C = ND->getAPIntValue();
14661 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
14664 case X86ISD::VSRAI:
14665 for (unsigned i=0; i!=NumElts; ++i) {
14666 SDValue CurrentOp = SrcOp->getOperand(i);
14667 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14668 Elts.push_back(CurrentOp);
14671 ND = cast<ConstantSDNode>(CurrentOp);
14672 const APInt &C = ND->getAPIntValue();
14673 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
14678 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14681 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
14684 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14685 // may or may not be a constant. Takes immediate version of shift as input.
14686 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14687 SDValue SrcOp, SDValue ShAmt,
14688 SelectionDAG &DAG) {
14689 MVT SVT = ShAmt.getSimpleValueType();
14690 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14692 // Catch shift-by-constant.
14693 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14694 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14695 CShAmt->getZExtValue(), DAG);
14697 // Change opcode to non-immediate version
14699 default: llvm_unreachable("Unknown target vector shift node");
14700 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14701 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14702 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14705 const X86Subtarget &Subtarget =
14706 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14707 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14708 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14709 // Let the shuffle legalizer expand this shift amount node.
14710 SDValue Op0 = ShAmt.getOperand(0);
14711 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14712 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14714 // Need to build a vector containing shift amount.
14715 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14716 SmallVector<SDValue, 4> ShOps;
14717 ShOps.push_back(ShAmt);
14718 if (SVT == MVT::i32) {
14719 ShOps.push_back(DAG.getConstant(0, SVT));
14720 ShOps.push_back(DAG.getUNDEF(SVT));
14722 ShOps.push_back(DAG.getUNDEF(SVT));
14724 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14725 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14728 // The return type has to be a 128-bit type with the same element
14729 // type as the input type.
14730 MVT EltVT = VT.getVectorElementType();
14731 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14733 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14734 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14737 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14738 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14739 /// necessary casting for \p Mask when lowering masking intrinsics.
14740 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14741 SDValue PreservedSrc,
14742 const X86Subtarget *Subtarget,
14743 SelectionDAG &DAG) {
14744 EVT VT = Op.getValueType();
14745 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14746 MVT::i1, VT.getVectorNumElements());
14747 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14748 Mask.getValueType().getSizeInBits());
14751 assert(MaskVT.isSimple() && "invalid mask type");
14753 if (isAllOnes(Mask))
14756 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14757 // are extracted by EXTRACT_SUBVECTOR.
14758 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14759 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14760 DAG.getIntPtrConstant(0));
14762 switch (Op.getOpcode()) {
14764 case X86ISD::PCMPEQM:
14765 case X86ISD::PCMPGTM:
14767 case X86ISD::CMPMU:
14768 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14770 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14771 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14772 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14775 /// \brief Creates an SDNode for a predicated scalar operation.
14776 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14777 /// The mask is comming as MVT::i8 and it should be truncated
14778 /// to MVT::i1 while lowering masking intrinsics.
14779 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14780 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14781 /// a scalar instruction.
14782 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14783 SDValue PreservedSrc,
14784 const X86Subtarget *Subtarget,
14785 SelectionDAG &DAG) {
14786 if (isAllOnes(Mask))
14789 EVT VT = Op.getValueType();
14791 // The mask should be of type MVT::i1
14792 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14794 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14795 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14796 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14799 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14800 SelectionDAG &DAG) {
14802 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14803 EVT VT = Op.getValueType();
14804 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14806 switch(IntrData->Type) {
14807 case INTR_TYPE_1OP:
14808 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14809 case INTR_TYPE_2OP:
14810 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14812 case INTR_TYPE_3OP:
14813 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14814 Op.getOperand(2), Op.getOperand(3));
14815 case INTR_TYPE_1OP_MASK_RM: {
14816 SDValue Src = Op.getOperand(1);
14817 SDValue Src0 = Op.getOperand(2);
14818 SDValue Mask = Op.getOperand(3);
14819 SDValue RoundingMode = Op.getOperand(4);
14820 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14822 Mask, Src0, Subtarget, DAG);
14824 case INTR_TYPE_SCALAR_MASK_RM: {
14825 SDValue Src1 = Op.getOperand(1);
14826 SDValue Src2 = Op.getOperand(2);
14827 SDValue Src0 = Op.getOperand(3);
14828 SDValue Mask = Op.getOperand(4);
14829 // There are 2 kinds of intrinsics in this group:
14830 // (1) With supress-all-exceptions (sae) - 6 operands
14831 // (2) With rounding mode and sae - 7 operands.
14832 if (Op.getNumOperands() == 6) {
14833 SDValue Sae = Op.getOperand(5);
14834 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14836 Mask, Src0, Subtarget, DAG);
14838 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
14839 SDValue RoundingMode = Op.getOperand(5);
14840 SDValue Sae = Op.getOperand(6);
14841 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14842 RoundingMode, Sae),
14843 Mask, Src0, Subtarget, DAG);
14845 case INTR_TYPE_2OP_MASK: {
14846 SDValue Src1 = Op.getOperand(1);
14847 SDValue Src2 = Op.getOperand(2);
14848 SDValue PassThru = Op.getOperand(3);
14849 SDValue Mask = Op.getOperand(4);
14850 // We specify 2 possible opcodes for intrinsics with rounding modes.
14851 // First, we check if the intrinsic may have non-default rounding mode,
14852 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14853 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14854 if (IntrWithRoundingModeOpcode != 0) {
14855 SDValue Rnd = Op.getOperand(5);
14856 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14857 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14858 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14859 dl, Op.getValueType(),
14861 Mask, PassThru, Subtarget, DAG);
14864 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14866 Mask, PassThru, Subtarget, DAG);
14868 case FMA_OP_MASK: {
14869 SDValue Src1 = Op.getOperand(1);
14870 SDValue Src2 = Op.getOperand(2);
14871 SDValue Src3 = Op.getOperand(3);
14872 SDValue Mask = Op.getOperand(4);
14873 // We specify 2 possible opcodes for intrinsics with rounding modes.
14874 // First, we check if the intrinsic may have non-default rounding mode,
14875 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14876 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14877 if (IntrWithRoundingModeOpcode != 0) {
14878 SDValue Rnd = Op.getOperand(5);
14879 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14880 X86::STATIC_ROUNDING::CUR_DIRECTION)
14881 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14882 dl, Op.getValueType(),
14883 Src1, Src2, Src3, Rnd),
14884 Mask, Src1, Subtarget, DAG);
14886 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
14887 dl, Op.getValueType(),
14889 Mask, Src1, Subtarget, DAG);
14892 case CMP_MASK_CC: {
14893 // Comparison intrinsics with masks.
14894 // Example of transformation:
14895 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
14896 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
14898 // (v8i1 (insert_subvector undef,
14899 // (v2i1 (and (PCMPEQM %a, %b),
14900 // (extract_subvector
14901 // (v8i1 (bitcast %mask)), 0))), 0))))
14902 EVT VT = Op.getOperand(1).getValueType();
14903 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14904 VT.getVectorNumElements());
14905 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
14906 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14907 Mask.getValueType().getSizeInBits());
14909 if (IntrData->Type == CMP_MASK_CC) {
14910 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14911 Op.getOperand(2), Op.getOperand(3));
14913 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
14914 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14917 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
14918 DAG.getTargetConstant(0, MaskVT),
14920 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
14921 DAG.getUNDEF(BitcastVT), CmpMask,
14922 DAG.getIntPtrConstant(0));
14923 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
14925 case COMI: { // Comparison intrinsics
14926 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
14927 SDValue LHS = Op.getOperand(1);
14928 SDValue RHS = Op.getOperand(2);
14929 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
14930 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
14931 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
14932 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14933 DAG.getConstant(X86CC, MVT::i8), Cond);
14934 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14937 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
14938 Op.getOperand(1), Op.getOperand(2), DAG);
14940 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
14941 Op.getSimpleValueType(),
14943 Op.getOperand(2), DAG),
14944 Op.getOperand(4), Op.getOperand(3), Subtarget,
14946 case COMPRESS_EXPAND_IN_REG: {
14947 SDValue Mask = Op.getOperand(3);
14948 SDValue DataToCompress = Op.getOperand(1);
14949 SDValue PassThru = Op.getOperand(2);
14950 if (isAllOnes(Mask)) // return data as is
14951 return Op.getOperand(1);
14952 EVT VT = Op.getValueType();
14953 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14954 VT.getVectorNumElements());
14955 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14956 Mask.getValueType().getSizeInBits());
14958 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14959 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14960 DAG.getIntPtrConstant(0));
14962 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
14966 SDValue Mask = Op.getOperand(3);
14967 EVT VT = Op.getValueType();
14968 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14969 VT.getVectorNumElements());
14970 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14971 Mask.getValueType().getSizeInBits());
14973 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14974 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14975 DAG.getIntPtrConstant(0));
14976 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
14985 default: return SDValue(); // Don't custom lower most intrinsics.
14987 case Intrinsic::x86_avx2_permd:
14988 case Intrinsic::x86_avx2_permps:
14989 // Operands intentionally swapped. Mask is last operand to intrinsic,
14990 // but second operand for node/instruction.
14991 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
14992 Op.getOperand(2), Op.getOperand(1));
14994 case Intrinsic::x86_avx512_mask_valign_q_512:
14995 case Intrinsic::x86_avx512_mask_valign_d_512:
14996 // Vector source operands are swapped.
14997 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
14998 Op.getValueType(), Op.getOperand(2),
15001 Op.getOperand(5), Op.getOperand(4),
15004 // ptest and testp intrinsics. The intrinsic these come from are designed to
15005 // return an integer value, not just an instruction so lower it to the ptest
15006 // or testp pattern and a setcc for the result.
15007 case Intrinsic::x86_sse41_ptestz:
15008 case Intrinsic::x86_sse41_ptestc:
15009 case Intrinsic::x86_sse41_ptestnzc:
15010 case Intrinsic::x86_avx_ptestz_256:
15011 case Intrinsic::x86_avx_ptestc_256:
15012 case Intrinsic::x86_avx_ptestnzc_256:
15013 case Intrinsic::x86_avx_vtestz_ps:
15014 case Intrinsic::x86_avx_vtestc_ps:
15015 case Intrinsic::x86_avx_vtestnzc_ps:
15016 case Intrinsic::x86_avx_vtestz_pd:
15017 case Intrinsic::x86_avx_vtestc_pd:
15018 case Intrinsic::x86_avx_vtestnzc_pd:
15019 case Intrinsic::x86_avx_vtestz_ps_256:
15020 case Intrinsic::x86_avx_vtestc_ps_256:
15021 case Intrinsic::x86_avx_vtestnzc_ps_256:
15022 case Intrinsic::x86_avx_vtestz_pd_256:
15023 case Intrinsic::x86_avx_vtestc_pd_256:
15024 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15025 bool IsTestPacked = false;
15028 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15029 case Intrinsic::x86_avx_vtestz_ps:
15030 case Intrinsic::x86_avx_vtestz_pd:
15031 case Intrinsic::x86_avx_vtestz_ps_256:
15032 case Intrinsic::x86_avx_vtestz_pd_256:
15033 IsTestPacked = true; // Fallthrough
15034 case Intrinsic::x86_sse41_ptestz:
15035 case Intrinsic::x86_avx_ptestz_256:
15037 X86CC = X86::COND_E;
15039 case Intrinsic::x86_avx_vtestc_ps:
15040 case Intrinsic::x86_avx_vtestc_pd:
15041 case Intrinsic::x86_avx_vtestc_ps_256:
15042 case Intrinsic::x86_avx_vtestc_pd_256:
15043 IsTestPacked = true; // Fallthrough
15044 case Intrinsic::x86_sse41_ptestc:
15045 case Intrinsic::x86_avx_ptestc_256:
15047 X86CC = X86::COND_B;
15049 case Intrinsic::x86_avx_vtestnzc_ps:
15050 case Intrinsic::x86_avx_vtestnzc_pd:
15051 case Intrinsic::x86_avx_vtestnzc_ps_256:
15052 case Intrinsic::x86_avx_vtestnzc_pd_256:
15053 IsTestPacked = true; // Fallthrough
15054 case Intrinsic::x86_sse41_ptestnzc:
15055 case Intrinsic::x86_avx_ptestnzc_256:
15057 X86CC = X86::COND_A;
15061 SDValue LHS = Op.getOperand(1);
15062 SDValue RHS = Op.getOperand(2);
15063 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15064 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15065 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
15066 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15067 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15069 case Intrinsic::x86_avx512_kortestz_w:
15070 case Intrinsic::x86_avx512_kortestc_w: {
15071 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15072 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
15073 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
15074 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
15075 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15076 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15077 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15080 case Intrinsic::x86_sse42_pcmpistria128:
15081 case Intrinsic::x86_sse42_pcmpestria128:
15082 case Intrinsic::x86_sse42_pcmpistric128:
15083 case Intrinsic::x86_sse42_pcmpestric128:
15084 case Intrinsic::x86_sse42_pcmpistrio128:
15085 case Intrinsic::x86_sse42_pcmpestrio128:
15086 case Intrinsic::x86_sse42_pcmpistris128:
15087 case Intrinsic::x86_sse42_pcmpestris128:
15088 case Intrinsic::x86_sse42_pcmpistriz128:
15089 case Intrinsic::x86_sse42_pcmpestriz128: {
15093 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15094 case Intrinsic::x86_sse42_pcmpistria128:
15095 Opcode = X86ISD::PCMPISTRI;
15096 X86CC = X86::COND_A;
15098 case Intrinsic::x86_sse42_pcmpestria128:
15099 Opcode = X86ISD::PCMPESTRI;
15100 X86CC = X86::COND_A;
15102 case Intrinsic::x86_sse42_pcmpistric128:
15103 Opcode = X86ISD::PCMPISTRI;
15104 X86CC = X86::COND_B;
15106 case Intrinsic::x86_sse42_pcmpestric128:
15107 Opcode = X86ISD::PCMPESTRI;
15108 X86CC = X86::COND_B;
15110 case Intrinsic::x86_sse42_pcmpistrio128:
15111 Opcode = X86ISD::PCMPISTRI;
15112 X86CC = X86::COND_O;
15114 case Intrinsic::x86_sse42_pcmpestrio128:
15115 Opcode = X86ISD::PCMPESTRI;
15116 X86CC = X86::COND_O;
15118 case Intrinsic::x86_sse42_pcmpistris128:
15119 Opcode = X86ISD::PCMPISTRI;
15120 X86CC = X86::COND_S;
15122 case Intrinsic::x86_sse42_pcmpestris128:
15123 Opcode = X86ISD::PCMPESTRI;
15124 X86CC = X86::COND_S;
15126 case Intrinsic::x86_sse42_pcmpistriz128:
15127 Opcode = X86ISD::PCMPISTRI;
15128 X86CC = X86::COND_E;
15130 case Intrinsic::x86_sse42_pcmpestriz128:
15131 Opcode = X86ISD::PCMPESTRI;
15132 X86CC = X86::COND_E;
15135 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15136 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15137 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15138 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15139 DAG.getConstant(X86CC, MVT::i8),
15140 SDValue(PCMP.getNode(), 1));
15141 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15144 case Intrinsic::x86_sse42_pcmpistri128:
15145 case Intrinsic::x86_sse42_pcmpestri128: {
15147 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15148 Opcode = X86ISD::PCMPISTRI;
15150 Opcode = X86ISD::PCMPESTRI;
15152 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15153 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15154 return DAG.getNode(Opcode, dl, VTs, NewOps);
15159 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15160 SDValue Src, SDValue Mask, SDValue Base,
15161 SDValue Index, SDValue ScaleOp, SDValue Chain,
15162 const X86Subtarget * Subtarget) {
15164 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15165 assert(C && "Invalid scale type");
15166 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
15167 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15168 Index.getSimpleValueType().getVectorNumElements());
15170 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15172 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
15174 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15175 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15176 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
15177 SDValue Segment = DAG.getRegister(0, MVT::i32);
15178 if (Src.getOpcode() == ISD::UNDEF)
15179 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15180 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15181 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15182 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15183 return DAG.getMergeValues(RetOps, dl);
15186 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15187 SDValue Src, SDValue Mask, SDValue Base,
15188 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15190 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15191 assert(C && "Invalid scale type");
15192 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
15193 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
15194 SDValue Segment = DAG.getRegister(0, MVT::i32);
15195 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15196 Index.getSimpleValueType().getVectorNumElements());
15198 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15200 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
15202 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15203 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15204 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15205 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15206 return SDValue(Res, 1);
15209 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15210 SDValue Mask, SDValue Base, SDValue Index,
15211 SDValue ScaleOp, SDValue Chain) {
15213 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15214 assert(C && "Invalid scale type");
15215 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
15216 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
15217 SDValue Segment = DAG.getRegister(0, MVT::i32);
15219 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15221 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15223 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
15225 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15226 //SDVTList VTs = DAG.getVTList(MVT::Other);
15227 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15228 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15229 return SDValue(Res, 0);
15232 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15233 // read performance monitor counters (x86_rdpmc).
15234 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15235 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15236 SmallVectorImpl<SDValue> &Results) {
15237 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15238 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15241 // The ECX register is used to select the index of the performance counter
15243 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
15245 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
15247 // Reads the content of a 64-bit performance counter and returns it in the
15248 // registers EDX:EAX.
15249 if (Subtarget->is64Bit()) {
15250 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15251 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15254 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15255 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15258 Chain = HI.getValue(1);
15260 if (Subtarget->is64Bit()) {
15261 // The EAX register is loaded with the low-order 32 bits. The EDX register
15262 // is loaded with the supported high-order bits of the counter.
15263 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15264 DAG.getConstant(32, MVT::i8));
15265 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15266 Results.push_back(Chain);
15270 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15271 SDValue Ops[] = { LO, HI };
15272 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15273 Results.push_back(Pair);
15274 Results.push_back(Chain);
15277 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
15278 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
15279 // also used to custom lower READCYCLECOUNTER nodes.
15280 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
15281 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15282 SmallVectorImpl<SDValue> &Results) {
15283 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15284 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
15287 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
15288 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
15289 // and the EAX register is loaded with the low-order 32 bits.
15290 if (Subtarget->is64Bit()) {
15291 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15292 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15295 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15296 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15299 SDValue Chain = HI.getValue(1);
15301 if (Opcode == X86ISD::RDTSCP_DAG) {
15302 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15304 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
15305 // the ECX register. Add 'ecx' explicitly to the chain.
15306 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
15308 // Explicitly store the content of ECX at the location passed in input
15309 // to the 'rdtscp' intrinsic.
15310 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
15311 MachinePointerInfo(), false, false, 0);
15314 if (Subtarget->is64Bit()) {
15315 // The EDX register is loaded with the high-order 32 bits of the MSR, and
15316 // the EAX register is loaded with the low-order 32 bits.
15317 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15318 DAG.getConstant(32, MVT::i8));
15319 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15320 Results.push_back(Chain);
15324 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15325 SDValue Ops[] = { LO, HI };
15326 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15327 Results.push_back(Pair);
15328 Results.push_back(Chain);
15331 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
15332 SelectionDAG &DAG) {
15333 SmallVector<SDValue, 2> Results;
15335 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
15337 return DAG.getMergeValues(Results, DL);
15341 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15342 SelectionDAG &DAG) {
15343 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
15345 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
15350 switch(IntrData->Type) {
15352 llvm_unreachable("Unknown Intrinsic Type");
15356 // Emit the node with the right value type.
15357 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
15358 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15360 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
15361 // Otherwise return the value from Rand, which is always 0, casted to i32.
15362 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
15363 DAG.getConstant(1, Op->getValueType(1)),
15364 DAG.getConstant(X86::COND_B, MVT::i32),
15365 SDValue(Result.getNode(), 1) };
15366 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
15367 DAG.getVTList(Op->getValueType(1), MVT::Glue),
15370 // Return { result, isValid, chain }.
15371 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
15372 SDValue(Result.getNode(), 2));
15375 //gather(v1, mask, index, base, scale);
15376 SDValue Chain = Op.getOperand(0);
15377 SDValue Src = Op.getOperand(2);
15378 SDValue Base = Op.getOperand(3);
15379 SDValue Index = Op.getOperand(4);
15380 SDValue Mask = Op.getOperand(5);
15381 SDValue Scale = Op.getOperand(6);
15382 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
15386 //scatter(base, mask, index, v1, scale);
15387 SDValue Chain = Op.getOperand(0);
15388 SDValue Base = Op.getOperand(2);
15389 SDValue Mask = Op.getOperand(3);
15390 SDValue Index = Op.getOperand(4);
15391 SDValue Src = Op.getOperand(5);
15392 SDValue Scale = Op.getOperand(6);
15393 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
15396 SDValue Hint = Op.getOperand(6);
15397 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
15398 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
15399 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
15400 SDValue Chain = Op.getOperand(0);
15401 SDValue Mask = Op.getOperand(2);
15402 SDValue Index = Op.getOperand(3);
15403 SDValue Base = Op.getOperand(4);
15404 SDValue Scale = Op.getOperand(5);
15405 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15407 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15409 SmallVector<SDValue, 2> Results;
15410 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
15411 return DAG.getMergeValues(Results, dl);
15413 // Read Performance Monitoring Counters.
15415 SmallVector<SDValue, 2> Results;
15416 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15417 return DAG.getMergeValues(Results, dl);
15419 // XTEST intrinsics.
15421 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15422 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15423 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15424 DAG.getConstant(X86::COND_NE, MVT::i8),
15426 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15427 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15428 Ret, SDValue(InTrans.getNode(), 1));
15432 SmallVector<SDValue, 2> Results;
15433 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15434 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15435 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15436 DAG.getConstant(-1, MVT::i8));
15437 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15438 Op.getOperand(4), GenCF.getValue(1));
15439 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15440 Op.getOperand(5), MachinePointerInfo(),
15442 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15443 DAG.getConstant(X86::COND_B, MVT::i8),
15445 Results.push_back(SetCC);
15446 Results.push_back(Store);
15447 return DAG.getMergeValues(Results, dl);
15449 case COMPRESS_TO_MEM: {
15451 SDValue Mask = Op.getOperand(4);
15452 SDValue DataToCompress = Op.getOperand(3);
15453 SDValue Addr = Op.getOperand(2);
15454 SDValue Chain = Op.getOperand(0);
15456 if (isAllOnes(Mask)) // return just a store
15457 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15458 MachinePointerInfo(), false, false, 0);
15460 EVT VT = DataToCompress.getValueType();
15461 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15462 VT.getVectorNumElements());
15463 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15464 Mask.getValueType().getSizeInBits());
15465 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15466 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15467 DAG.getIntPtrConstant(0));
15469 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15470 DataToCompress, DAG.getUNDEF(VT));
15471 return DAG.getStore(Chain, dl, Compressed, Addr,
15472 MachinePointerInfo(), false, false, 0);
15474 case EXPAND_FROM_MEM: {
15476 SDValue Mask = Op.getOperand(4);
15477 SDValue PathThru = Op.getOperand(3);
15478 SDValue Addr = Op.getOperand(2);
15479 SDValue Chain = Op.getOperand(0);
15480 EVT VT = Op.getValueType();
15482 if (isAllOnes(Mask)) // return just a load
15483 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15485 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15486 VT.getVectorNumElements());
15487 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15488 Mask.getValueType().getSizeInBits());
15489 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15490 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15491 DAG.getIntPtrConstant(0));
15493 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15494 false, false, false, 0);
15496 SDValue Results[] = {
15497 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15499 return DAG.getMergeValues(Results, dl);
15504 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15505 SelectionDAG &DAG) const {
15506 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15507 MFI->setReturnAddressIsTaken(true);
15509 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15512 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15514 EVT PtrVT = getPointerTy();
15517 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15518 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15519 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
15520 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15521 DAG.getNode(ISD::ADD, dl, PtrVT,
15522 FrameAddr, Offset),
15523 MachinePointerInfo(), false, false, false, 0);
15526 // Just load the return address.
15527 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15528 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15529 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15532 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15533 MachineFunction &MF = DAG.getMachineFunction();
15534 MachineFrameInfo *MFI = MF.getFrameInfo();
15535 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15536 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15537 EVT VT = Op.getValueType();
15539 MFI->setFrameAddressIsTaken(true);
15541 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15542 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15543 // is not possible to crawl up the stack without looking at the unwind codes
15545 int FrameAddrIndex = FuncInfo->getFAIndex();
15546 if (!FrameAddrIndex) {
15547 // Set up a frame object for the return address.
15548 unsigned SlotSize = RegInfo->getSlotSize();
15549 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15550 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
15551 FuncInfo->setFAIndex(FrameAddrIndex);
15553 return DAG.getFrameIndex(FrameAddrIndex, VT);
15556 unsigned FrameReg =
15557 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15558 SDLoc dl(Op); // FIXME probably not meaningful
15559 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15560 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15561 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15562 "Invalid Frame Register!");
15563 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15565 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15566 MachinePointerInfo(),
15567 false, false, false, 0);
15571 // FIXME? Maybe this could be a TableGen attribute on some registers and
15572 // this table could be generated automatically from RegInfo.
15573 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15575 unsigned Reg = StringSwitch<unsigned>(RegName)
15576 .Case("esp", X86::ESP)
15577 .Case("rsp", X86::RSP)
15581 report_fatal_error("Invalid register name global variable");
15584 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15585 SelectionDAG &DAG) const {
15586 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15587 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
15590 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15591 SDValue Chain = Op.getOperand(0);
15592 SDValue Offset = Op.getOperand(1);
15593 SDValue Handler = Op.getOperand(2);
15596 EVT PtrVT = getPointerTy();
15597 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15598 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15599 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15600 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15601 "Invalid Frame Register!");
15602 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15603 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15605 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15606 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
15607 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15608 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15610 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15612 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15613 DAG.getRegister(StoreAddrReg, PtrVT));
15616 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15617 SelectionDAG &DAG) const {
15619 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15620 DAG.getVTList(MVT::i32, MVT::Other),
15621 Op.getOperand(0), Op.getOperand(1));
15624 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15625 SelectionDAG &DAG) const {
15627 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15628 Op.getOperand(0), Op.getOperand(1));
15631 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15632 return Op.getOperand(0);
15635 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15636 SelectionDAG &DAG) const {
15637 SDValue Root = Op.getOperand(0);
15638 SDValue Trmp = Op.getOperand(1); // trampoline
15639 SDValue FPtr = Op.getOperand(2); // nested function
15640 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15643 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15644 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15646 if (Subtarget->is64Bit()) {
15647 SDValue OutChains[6];
15649 // Large code-model.
15650 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15651 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15653 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15654 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15656 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15658 // Load the pointer to the nested function into R11.
15659 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15660 SDValue Addr = Trmp;
15661 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15662 Addr, MachinePointerInfo(TrmpAddr),
15665 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15666 DAG.getConstant(2, MVT::i64));
15667 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15668 MachinePointerInfo(TrmpAddr, 2),
15671 // Load the 'nest' parameter value into R10.
15672 // R10 is specified in X86CallingConv.td
15673 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15674 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15675 DAG.getConstant(10, MVT::i64));
15676 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15677 Addr, MachinePointerInfo(TrmpAddr, 10),
15680 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15681 DAG.getConstant(12, MVT::i64));
15682 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15683 MachinePointerInfo(TrmpAddr, 12),
15686 // Jump to the nested function.
15687 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15688 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15689 DAG.getConstant(20, MVT::i64));
15690 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15691 Addr, MachinePointerInfo(TrmpAddr, 20),
15694 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15695 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15696 DAG.getConstant(22, MVT::i64));
15697 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
15698 MachinePointerInfo(TrmpAddr, 22),
15701 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15703 const Function *Func =
15704 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15705 CallingConv::ID CC = Func->getCallingConv();
15710 llvm_unreachable("Unsupported calling convention");
15711 case CallingConv::C:
15712 case CallingConv::X86_StdCall: {
15713 // Pass 'nest' parameter in ECX.
15714 // Must be kept in sync with X86CallingConv.td
15715 NestReg = X86::ECX;
15717 // Check that ECX wasn't needed by an 'inreg' parameter.
15718 FunctionType *FTy = Func->getFunctionType();
15719 const AttributeSet &Attrs = Func->getAttributes();
15721 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15722 unsigned InRegCount = 0;
15725 for (FunctionType::param_iterator I = FTy->param_begin(),
15726 E = FTy->param_end(); I != E; ++I, ++Idx)
15727 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15728 // FIXME: should only count parameters that are lowered to integers.
15729 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15731 if (InRegCount > 2) {
15732 report_fatal_error("Nest register in use - reduce number of inreg"
15738 case CallingConv::X86_FastCall:
15739 case CallingConv::X86_ThisCall:
15740 case CallingConv::Fast:
15741 // Pass 'nest' parameter in EAX.
15742 // Must be kept in sync with X86CallingConv.td
15743 NestReg = X86::EAX;
15747 SDValue OutChains[4];
15748 SDValue Addr, Disp;
15750 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15751 DAG.getConstant(10, MVT::i32));
15752 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15754 // This is storing the opcode for MOV32ri.
15755 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15756 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15757 OutChains[0] = DAG.getStore(Root, dl,
15758 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
15759 Trmp, MachinePointerInfo(TrmpAddr),
15762 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15763 DAG.getConstant(1, MVT::i32));
15764 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15765 MachinePointerInfo(TrmpAddr, 1),
15768 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15769 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15770 DAG.getConstant(5, MVT::i32));
15771 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
15772 MachinePointerInfo(TrmpAddr, 5),
15775 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15776 DAG.getConstant(6, MVT::i32));
15777 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15778 MachinePointerInfo(TrmpAddr, 6),
15781 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15785 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15786 SelectionDAG &DAG) const {
15788 The rounding mode is in bits 11:10 of FPSR, and has the following
15790 00 Round to nearest
15795 FLT_ROUNDS, on the other hand, expects the following:
15802 To perform the conversion, we do:
15803 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15806 MachineFunction &MF = DAG.getMachineFunction();
15807 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15808 unsigned StackAlignment = TFI.getStackAlignment();
15809 MVT VT = Op.getSimpleValueType();
15812 // Save FP Control Word to stack slot
15813 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15814 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15816 MachineMemOperand *MMO =
15817 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15818 MachineMemOperand::MOStore, 2, 2);
15820 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15821 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15822 DAG.getVTList(MVT::Other),
15823 Ops, MVT::i16, MMO);
15825 // Load FP Control Word from stack slot
15826 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15827 MachinePointerInfo(), false, false, false, 0);
15829 // Transform as necessary
15831 DAG.getNode(ISD::SRL, DL, MVT::i16,
15832 DAG.getNode(ISD::AND, DL, MVT::i16,
15833 CWD, DAG.getConstant(0x800, MVT::i16)),
15834 DAG.getConstant(11, MVT::i8));
15836 DAG.getNode(ISD::SRL, DL, MVT::i16,
15837 DAG.getNode(ISD::AND, DL, MVT::i16,
15838 CWD, DAG.getConstant(0x400, MVT::i16)),
15839 DAG.getConstant(9, MVT::i8));
15842 DAG.getNode(ISD::AND, DL, MVT::i16,
15843 DAG.getNode(ISD::ADD, DL, MVT::i16,
15844 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
15845 DAG.getConstant(1, MVT::i16)),
15846 DAG.getConstant(3, MVT::i16));
15848 return DAG.getNode((VT.getSizeInBits() < 16 ?
15849 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
15852 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
15853 MVT VT = Op.getSimpleValueType();
15855 unsigned NumBits = VT.getSizeInBits();
15858 Op = Op.getOperand(0);
15859 if (VT == MVT::i8) {
15860 // Zero extend to i32 since there is not an i8 bsr.
15862 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15865 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
15866 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15867 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15869 // If src is zero (i.e. bsr sets ZF), returns NumBits.
15872 DAG.getConstant(NumBits+NumBits-1, OpVT),
15873 DAG.getConstant(X86::COND_E, MVT::i8),
15876 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
15878 // Finally xor with NumBits-1.
15879 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15882 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15886 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
15887 MVT VT = Op.getSimpleValueType();
15889 unsigned NumBits = VT.getSizeInBits();
15892 Op = Op.getOperand(0);
15893 if (VT == MVT::i8) {
15894 // Zero extend to i32 since there is not an i8 bsr.
15896 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15899 // Issue a bsr (scan bits in reverse).
15900 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15901 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15903 // And xor with NumBits-1.
15904 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15907 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15911 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
15912 MVT VT = Op.getSimpleValueType();
15913 unsigned NumBits = VT.getSizeInBits();
15915 Op = Op.getOperand(0);
15917 // Issue a bsf (scan bits forward) which also sets EFLAGS.
15918 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
15919 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
15921 // If src is zero (i.e. bsf sets ZF), returns NumBits.
15924 DAG.getConstant(NumBits, VT),
15925 DAG.getConstant(X86::COND_E, MVT::i8),
15928 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
15931 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
15932 // ones, and then concatenate the result back.
15933 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
15934 MVT VT = Op.getSimpleValueType();
15936 assert(VT.is256BitVector() && VT.isInteger() &&
15937 "Unsupported value type for operation");
15939 unsigned NumElems = VT.getVectorNumElements();
15942 // Extract the LHS vectors
15943 SDValue LHS = Op.getOperand(0);
15944 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15945 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15947 // Extract the RHS vectors
15948 SDValue RHS = Op.getOperand(1);
15949 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15950 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15952 MVT EltVT = VT.getVectorElementType();
15953 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15955 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15956 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15957 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15960 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15961 assert(Op.getSimpleValueType().is256BitVector() &&
15962 Op.getSimpleValueType().isInteger() &&
15963 "Only handle AVX 256-bit vector integer operation");
15964 return Lower256IntArith(Op, DAG);
15967 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15968 assert(Op.getSimpleValueType().is256BitVector() &&
15969 Op.getSimpleValueType().isInteger() &&
15970 "Only handle AVX 256-bit vector integer operation");
15971 return Lower256IntArith(Op, DAG);
15974 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
15975 SelectionDAG &DAG) {
15977 MVT VT = Op.getSimpleValueType();
15979 // Decompose 256-bit ops into smaller 128-bit ops.
15980 if (VT.is256BitVector() && !Subtarget->hasInt256())
15981 return Lower256IntArith(Op, DAG);
15983 SDValue A = Op.getOperand(0);
15984 SDValue B = Op.getOperand(1);
15986 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
15987 // pairs, multiply and truncate.
15988 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
15989 if (Subtarget->hasInt256()) {
15990 if (VT == MVT::v32i8) {
15991 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
15992 SDValue Lo = DAG.getIntPtrConstant(0);
15993 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2);
15994 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
15995 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
15996 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
15997 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
15998 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15999 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16000 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16003 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16004 return DAG.getNode(
16005 ISD::TRUNCATE, dl, VT,
16006 DAG.getNode(ISD::MUL, dl, ExVT,
16007 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16008 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16011 assert(VT == MVT::v16i8 &&
16012 "Pre-AVX2 support only supports v16i8 multiplication");
16013 MVT ExVT = MVT::v8i16;
16015 // Extract the lo parts and sign extend to i16
16017 if (Subtarget->hasSSE41()) {
16018 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16019 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16021 const int ShufMask[] = {0, -1, 1, -1, 2, -1, 3, -1,
16022 4, -1, 5, -1, 6, -1, 7, -1};
16023 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16024 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16025 ALo = DAG.getNode(ISD::BITCAST, dl, ExVT, ALo);
16026 BLo = DAG.getNode(ISD::BITCAST, dl, ExVT, BLo);
16027 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, ExVT));
16028 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, ExVT));
16031 // Extract the hi parts and sign extend to i16
16033 if (Subtarget->hasSSE41()) {
16034 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16035 -1, -1, -1, -1, -1, -1, -1, -1};
16036 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16037 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16038 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16039 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16041 const int ShufMask[] = {8, -1, 9, -1, 10, -1, 11, -1,
16042 12, -1, 13, -1, 14, -1, 15, -1};
16043 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16044 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16045 AHi = DAG.getNode(ISD::BITCAST, dl, ExVT, AHi);
16046 BHi = DAG.getNode(ISD::BITCAST, dl, ExVT, BHi);
16047 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, ExVT));
16048 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, ExVT));
16051 // Multiply, mask the lower 8bits of the lo/hi results and pack
16052 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16053 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16054 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, ExVT));
16055 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, ExVT));
16056 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16059 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16060 if (VT == MVT::v4i32) {
16061 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16062 "Should not custom lower when pmuldq is available!");
16064 // Extract the odd parts.
16065 static const int UnpackMask[] = { 1, -1, 3, -1 };
16066 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16067 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16069 // Multiply the even parts.
16070 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16071 // Now multiply odd parts.
16072 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16074 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
16075 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
16077 // Merge the two vectors back together with a shuffle. This expands into 2
16079 static const int ShufMask[] = { 0, 4, 2, 6 };
16080 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16083 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16084 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16086 // Ahi = psrlqi(a, 32);
16087 // Bhi = psrlqi(b, 32);
16089 // AloBlo = pmuludq(a, b);
16090 // AloBhi = pmuludq(a, Bhi);
16091 // AhiBlo = pmuludq(Ahi, b);
16093 // AloBhi = psllqi(AloBhi, 32);
16094 // AhiBlo = psllqi(AhiBlo, 32);
16095 // return AloBlo + AloBhi + AhiBlo;
16097 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16098 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16100 // Bit cast to 32-bit vectors for MULUDQ
16101 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16102 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16103 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
16104 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
16105 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
16106 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
16108 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16109 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16110 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16112 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16113 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16115 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16116 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16119 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16120 assert(Subtarget->isTargetWin64() && "Unexpected target");
16121 EVT VT = Op.getValueType();
16122 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16123 "Unexpected return type for lowering");
16127 switch (Op->getOpcode()) {
16128 default: llvm_unreachable("Unexpected request for libcall!");
16129 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16130 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16131 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16132 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16133 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16134 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16138 SDValue InChain = DAG.getEntryNode();
16140 TargetLowering::ArgListTy Args;
16141 TargetLowering::ArgListEntry Entry;
16142 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16143 EVT ArgVT = Op->getOperand(i).getValueType();
16144 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16145 "Unexpected argument type for lowering");
16146 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16147 Entry.Node = StackPtr;
16148 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16150 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16151 Entry.Ty = PointerType::get(ArgTy,0);
16152 Entry.isSExt = false;
16153 Entry.isZExt = false;
16154 Args.push_back(Entry);
16157 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16160 TargetLowering::CallLoweringInfo CLI(DAG);
16161 CLI.setDebugLoc(dl).setChain(InChain)
16162 .setCallee(getLibcallCallingConv(LC),
16163 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16164 Callee, std::move(Args), 0)
16165 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16167 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16168 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
16171 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16172 SelectionDAG &DAG) {
16173 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16174 EVT VT = Op0.getValueType();
16177 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16178 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16180 // PMULxD operations multiply each even value (starting at 0) of LHS with
16181 // the related value of RHS and produce a widen result.
16182 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16183 // => <2 x i64> <ae|cg>
16185 // In other word, to have all the results, we need to perform two PMULxD:
16186 // 1. one with the even values.
16187 // 2. one with the odd values.
16188 // To achieve #2, with need to place the odd values at an even position.
16190 // Place the odd value at an even position (basically, shift all values 1
16191 // step to the left):
16192 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16193 // <a|b|c|d> => <b|undef|d|undef>
16194 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16195 // <e|f|g|h> => <f|undef|h|undef>
16196 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16198 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16200 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16201 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16203 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16204 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16205 // => <2 x i64> <ae|cg>
16206 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
16207 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
16208 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
16209 // => <2 x i64> <bf|dh>
16210 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
16211 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
16213 // Shuffle it back into the right order.
16214 SDValue Highs, Lows;
16215 if (VT == MVT::v8i32) {
16216 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
16217 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16218 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
16219 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16221 const int HighMask[] = {1, 5, 3, 7};
16222 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16223 const int LowMask[] = {0, 4, 2, 6};
16224 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16227 // If we have a signed multiply but no PMULDQ fix up the high parts of a
16228 // unsigned multiply.
16229 if (IsSigned && !Subtarget->hasSSE41()) {
16231 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
16232 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
16233 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
16234 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
16235 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
16237 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
16238 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
16241 // The first result of MUL_LOHI is actually the low value, followed by the
16243 SDValue Ops[] = {Lows, Highs};
16244 return DAG.getMergeValues(Ops, dl);
16247 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
16248 const X86Subtarget *Subtarget) {
16249 MVT VT = Op.getSimpleValueType();
16251 SDValue R = Op.getOperand(0);
16252 SDValue Amt = Op.getOperand(1);
16254 // Optimize shl/srl/sra with constant shift amount.
16255 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
16256 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
16257 uint64_t ShiftAmt = ShiftConst->getZExtValue();
16259 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
16260 (Subtarget->hasInt256() &&
16261 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
16262 (Subtarget->hasAVX512() &&
16263 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
16264 if (Op.getOpcode() == ISD::SHL)
16265 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
16267 if (Op.getOpcode() == ISD::SRL)
16268 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
16270 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
16271 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
16275 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
16276 unsigned NumElts = VT.getVectorNumElements();
16277 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
16279 if (Op.getOpcode() == ISD::SHL) {
16280 // Make a large shift.
16281 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
16283 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
16284 // Zero out the rightmost bits.
16285 SmallVector<SDValue, 32> V(
16286 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), MVT::i8));
16287 return DAG.getNode(ISD::AND, dl, VT, SHL,
16288 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16290 if (Op.getOpcode() == ISD::SRL) {
16291 // Make a large shift.
16292 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
16294 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
16295 // Zero out the leftmost bits.
16296 SmallVector<SDValue, 32> V(
16297 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, MVT::i8));
16298 return DAG.getNode(ISD::AND, dl, VT, SRL,
16299 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16301 if (Op.getOpcode() == ISD::SRA) {
16302 if (ShiftAmt == 7) {
16303 // R s>> 7 === R s< 0
16304 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
16305 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
16308 // R s>> a === ((R u>> a) ^ m) - m
16309 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
16310 SmallVector<SDValue, 32> V(NumElts,
16311 DAG.getConstant(128 >> ShiftAmt, MVT::i8));
16312 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
16313 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
16314 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
16317 llvm_unreachable("Unknown shift opcode.");
16322 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16323 if (!Subtarget->is64Bit() &&
16324 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
16325 Amt.getOpcode() == ISD::BITCAST &&
16326 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16327 Amt = Amt.getOperand(0);
16328 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16329 VT.getVectorNumElements();
16330 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
16331 uint64_t ShiftAmt = 0;
16332 for (unsigned i = 0; i != Ratio; ++i) {
16333 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
16337 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
16339 // Check remaining shift amounts.
16340 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16341 uint64_t ShAmt = 0;
16342 for (unsigned j = 0; j != Ratio; ++j) {
16343 ConstantSDNode *C =
16344 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
16348 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
16350 if (ShAmt != ShiftAmt)
16353 switch (Op.getOpcode()) {
16355 llvm_unreachable("Unknown shift opcode!");
16357 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
16360 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
16363 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
16371 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
16372 const X86Subtarget* Subtarget) {
16373 MVT VT = Op.getSimpleValueType();
16375 SDValue R = Op.getOperand(0);
16376 SDValue Amt = Op.getOperand(1);
16378 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
16379 VT == MVT::v4i32 || VT == MVT::v8i16 ||
16380 (Subtarget->hasInt256() &&
16381 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
16382 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
16383 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
16385 EVT EltVT = VT.getVectorElementType();
16387 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
16388 // Check if this build_vector node is doing a splat.
16389 // If so, then set BaseShAmt equal to the splat value.
16390 BaseShAmt = BV->getSplatValue();
16391 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
16392 BaseShAmt = SDValue();
16394 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
16395 Amt = Amt.getOperand(0);
16397 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
16398 if (SVN && SVN->isSplat()) {
16399 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
16400 SDValue InVec = Amt.getOperand(0);
16401 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
16402 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
16403 "Unexpected shuffle index found!");
16404 BaseShAmt = InVec.getOperand(SplatIdx);
16405 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
16406 if (ConstantSDNode *C =
16407 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
16408 if (C->getZExtValue() == SplatIdx)
16409 BaseShAmt = InVec.getOperand(1);
16414 // Avoid introducing an extract element from a shuffle.
16415 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
16416 DAG.getIntPtrConstant(SplatIdx));
16420 if (BaseShAmt.getNode()) {
16421 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
16422 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
16423 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
16424 else if (EltVT.bitsLT(MVT::i32))
16425 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
16427 switch (Op.getOpcode()) {
16429 llvm_unreachable("Unknown shift opcode!");
16431 switch (VT.SimpleTy) {
16432 default: return SDValue();
16441 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
16444 switch (VT.SimpleTy) {
16445 default: return SDValue();
16452 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
16455 switch (VT.SimpleTy) {
16456 default: return SDValue();
16465 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
16471 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16472 if (!Subtarget->is64Bit() &&
16473 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
16474 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
16475 Amt.getOpcode() == ISD::BITCAST &&
16476 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16477 Amt = Amt.getOperand(0);
16478 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16479 VT.getVectorNumElements();
16480 std::vector<SDValue> Vals(Ratio);
16481 for (unsigned i = 0; i != Ratio; ++i)
16482 Vals[i] = Amt.getOperand(i);
16483 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16484 for (unsigned j = 0; j != Ratio; ++j)
16485 if (Vals[j] != Amt.getOperand(i + j))
16488 switch (Op.getOpcode()) {
16490 llvm_unreachable("Unknown shift opcode!");
16492 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
16494 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
16496 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
16503 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16504 SelectionDAG &DAG) {
16505 MVT VT = Op.getSimpleValueType();
16507 SDValue R = Op.getOperand(0);
16508 SDValue Amt = Op.getOperand(1);
16510 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16511 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16513 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
16516 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
16519 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
16522 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
16523 if (Subtarget->hasInt256()) {
16524 if (Op.getOpcode() == ISD::SRL &&
16525 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16526 VT == MVT::v4i64 || VT == MVT::v8i32))
16528 if (Op.getOpcode() == ISD::SHL &&
16529 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16530 VT == MVT::v4i64 || VT == MVT::v8i32))
16532 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
16536 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
16537 // shifts per-lane and then shuffle the partial results back together.
16538 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
16539 // Splat the shift amounts so the scalar shifts above will catch it.
16540 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
16541 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
16542 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
16543 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
16544 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
16547 // If possible, lower this packed shift into a vector multiply instead of
16548 // expanding it into a sequence of scalar shifts.
16549 // Do this only if the vector shift count is a constant build_vector.
16550 if (Op.getOpcode() == ISD::SHL &&
16551 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16552 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16553 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16554 SmallVector<SDValue, 8> Elts;
16555 EVT SVT = VT.getScalarType();
16556 unsigned SVTBits = SVT.getSizeInBits();
16557 const APInt &One = APInt(SVTBits, 1);
16558 unsigned NumElems = VT.getVectorNumElements();
16560 for (unsigned i=0; i !=NumElems; ++i) {
16561 SDValue Op = Amt->getOperand(i);
16562 if (Op->getOpcode() == ISD::UNDEF) {
16563 Elts.push_back(Op);
16567 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16568 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16569 uint64_t ShAmt = C.getZExtValue();
16570 if (ShAmt >= SVTBits) {
16571 Elts.push_back(DAG.getUNDEF(SVT));
16574 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
16576 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16577 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16580 // Lower SHL with variable shift amount.
16581 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16582 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
16584 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
16585 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16586 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16587 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16590 // If possible, lower this shift as a sequence of two shifts by
16591 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16593 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16595 // Could be rewritten as:
16596 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16598 // The advantage is that the two shifts from the example would be
16599 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16600 // the vector shift into four scalar shifts plus four pairs of vector
16602 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16603 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16604 unsigned TargetOpcode = X86ISD::MOVSS;
16605 bool CanBeSimplified;
16606 // The splat value for the first packed shift (the 'X' from the example).
16607 SDValue Amt1 = Amt->getOperand(0);
16608 // The splat value for the second packed shift (the 'Y' from the example).
16609 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16610 Amt->getOperand(2);
16612 // See if it is possible to replace this node with a sequence of
16613 // two shifts followed by a MOVSS/MOVSD
16614 if (VT == MVT::v4i32) {
16615 // Check if it is legal to use a MOVSS.
16616 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16617 Amt2 == Amt->getOperand(3);
16618 if (!CanBeSimplified) {
16619 // Otherwise, check if we can still simplify this node using a MOVSD.
16620 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16621 Amt->getOperand(2) == Amt->getOperand(3);
16622 TargetOpcode = X86ISD::MOVSD;
16623 Amt2 = Amt->getOperand(2);
16626 // Do similar checks for the case where the machine value type
16628 CanBeSimplified = Amt1 == Amt->getOperand(1);
16629 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16630 CanBeSimplified = Amt2 == Amt->getOperand(i);
16632 if (!CanBeSimplified) {
16633 TargetOpcode = X86ISD::MOVSD;
16634 CanBeSimplified = true;
16635 Amt2 = Amt->getOperand(4);
16636 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16637 CanBeSimplified = Amt1 == Amt->getOperand(i);
16638 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16639 CanBeSimplified = Amt2 == Amt->getOperand(j);
16643 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16644 isa<ConstantSDNode>(Amt2)) {
16645 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16646 EVT CastVT = MVT::v4i32;
16648 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
16649 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16651 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
16652 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16653 if (TargetOpcode == X86ISD::MOVSD)
16654 CastVT = MVT::v2i64;
16655 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16656 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16657 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16659 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16663 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16664 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
16667 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
16668 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
16670 // Turn 'a' into a mask suitable for VSELECT
16671 SDValue VSelM = DAG.getConstant(0x80, VT);
16672 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16673 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16675 SDValue CM1 = DAG.getConstant(0x0f, VT);
16676 SDValue CM2 = DAG.getConstant(0x3f, VT);
16678 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
16679 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
16680 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
16681 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16682 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16685 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16686 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16687 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16689 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
16690 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
16691 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
16692 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16693 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16696 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16697 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16698 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16700 // return VSELECT(r, r+r, a);
16701 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16702 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16706 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16707 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16708 // solution better.
16709 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16710 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16712 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16713 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16714 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16715 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16716 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16719 // Decompose 256-bit shifts into smaller 128-bit shifts.
16720 if (VT.is256BitVector()) {
16721 unsigned NumElems = VT.getVectorNumElements();
16722 MVT EltVT = VT.getVectorElementType();
16723 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16725 // Extract the two vectors
16726 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16727 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16729 // Recreate the shift amount vectors
16730 SDValue Amt1, Amt2;
16731 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16732 // Constant shift amount
16733 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
16734 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
16735 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
16737 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16738 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16740 // Variable shift amount
16741 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16742 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16745 // Issue new vector shifts for the smaller types
16746 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16747 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16749 // Concatenate the result back
16750 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16756 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16757 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16758 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16759 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16760 // has only one use.
16761 SDNode *N = Op.getNode();
16762 SDValue LHS = N->getOperand(0);
16763 SDValue RHS = N->getOperand(1);
16764 unsigned BaseOp = 0;
16767 switch (Op.getOpcode()) {
16768 default: llvm_unreachable("Unknown ovf instruction!");
16770 // A subtract of one will be selected as a INC. Note that INC doesn't
16771 // set CF, so we can't do this for UADDO.
16772 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16774 BaseOp = X86ISD::INC;
16775 Cond = X86::COND_O;
16778 BaseOp = X86ISD::ADD;
16779 Cond = X86::COND_O;
16782 BaseOp = X86ISD::ADD;
16783 Cond = X86::COND_B;
16786 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16787 // set CF, so we can't do this for USUBO.
16788 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16790 BaseOp = X86ISD::DEC;
16791 Cond = X86::COND_O;
16794 BaseOp = X86ISD::SUB;
16795 Cond = X86::COND_O;
16798 BaseOp = X86ISD::SUB;
16799 Cond = X86::COND_B;
16802 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16803 Cond = X86::COND_O;
16805 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16806 if (N->getValueType(0) == MVT::i8) {
16807 BaseOp = X86ISD::UMUL8;
16808 Cond = X86::COND_O;
16811 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16813 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16816 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16817 DAG.getConstant(X86::COND_O, MVT::i32),
16818 SDValue(Sum.getNode(), 2));
16820 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16824 // Also sets EFLAGS.
16825 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16826 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16829 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16830 DAG.getConstant(Cond, MVT::i32),
16831 SDValue(Sum.getNode(), 1));
16833 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16836 /// Returns true if the operand type is exactly twice the native width, and
16837 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16838 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16839 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16840 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16841 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16844 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16845 else if (OpWidth == 128)
16846 return Subtarget->hasCmpxchg16b();
16851 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16852 return needsCmpXchgNb(SI->getValueOperand()->getType());
16855 // Note: this turns large loads into lock cmpxchg8b/16b.
16856 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16857 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16858 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16859 return needsCmpXchgNb(PTy->getElementType());
16862 TargetLoweringBase::AtomicRMWExpansionKind
16863 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16864 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16865 const Type *MemType = AI->getType();
16867 // If the operand is too big, we must see if cmpxchg8/16b is available
16868 // and default to library calls otherwise.
16869 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
16870 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
16871 : AtomicRMWExpansionKind::None;
16874 AtomicRMWInst::BinOp Op = AI->getOperation();
16877 llvm_unreachable("Unknown atomic operation");
16878 case AtomicRMWInst::Xchg:
16879 case AtomicRMWInst::Add:
16880 case AtomicRMWInst::Sub:
16881 // It's better to use xadd, xsub or xchg for these in all cases.
16882 return AtomicRMWExpansionKind::None;
16883 case AtomicRMWInst::Or:
16884 case AtomicRMWInst::And:
16885 case AtomicRMWInst::Xor:
16886 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16887 // prefix to a normal instruction for these operations.
16888 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
16889 : AtomicRMWExpansionKind::None;
16890 case AtomicRMWInst::Nand:
16891 case AtomicRMWInst::Max:
16892 case AtomicRMWInst::Min:
16893 case AtomicRMWInst::UMax:
16894 case AtomicRMWInst::UMin:
16895 // These always require a non-trivial set of data operations on x86. We must
16896 // use a cmpxchg loop.
16897 return AtomicRMWExpansionKind::CmpXChg;
16901 static bool hasMFENCE(const X86Subtarget& Subtarget) {
16902 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
16903 // no-sse2). There isn't any reason to disable it if the target processor
16905 return Subtarget.hasSSE2() || Subtarget.is64Bit();
16909 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
16910 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16911 const Type *MemType = AI->getType();
16912 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
16913 // there is no benefit in turning such RMWs into loads, and it is actually
16914 // harmful as it introduces a mfence.
16915 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16918 auto Builder = IRBuilder<>(AI);
16919 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16920 auto SynchScope = AI->getSynchScope();
16921 // We must restrict the ordering to avoid generating loads with Release or
16922 // ReleaseAcquire orderings.
16923 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
16924 auto Ptr = AI->getPointerOperand();
16926 // Before the load we need a fence. Here is an example lifted from
16927 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
16930 // x.store(1, relaxed);
16931 // r1 = y.fetch_add(0, release);
16933 // y.fetch_add(42, acquire);
16934 // r2 = x.load(relaxed);
16935 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
16936 // lowered to just a load without a fence. A mfence flushes the store buffer,
16937 // making the optimization clearly correct.
16938 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
16939 // otherwise, we might be able to be more agressive on relaxed idempotent
16940 // rmw. In practice, they do not look useful, so we don't try to be
16941 // especially clever.
16942 if (SynchScope == SingleThread) {
16943 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
16944 // the IR level, so we must wrap it in an intrinsic.
16946 } else if (hasMFENCE(*Subtarget)) {
16947 Function *MFence = llvm::Intrinsic::getDeclaration(M,
16948 Intrinsic::x86_sse2_mfence);
16949 Builder.CreateCall(MFence);
16951 // FIXME: it might make sense to use a locked operation here but on a
16952 // different cache-line to prevent cache-line bouncing. In practice it
16953 // is probably a small win, and x86 processors without mfence are rare
16954 // enough that we do not bother.
16958 // Finally we can emit the atomic load.
16959 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
16960 AI->getType()->getPrimitiveSizeInBits());
16961 Loaded->setAtomic(Order, SynchScope);
16962 AI->replaceAllUsesWith(Loaded);
16963 AI->eraseFromParent();
16967 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
16968 SelectionDAG &DAG) {
16970 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
16971 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
16972 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
16973 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
16975 // The only fence that needs an instruction is a sequentially-consistent
16976 // cross-thread fence.
16977 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
16978 if (hasMFENCE(*Subtarget))
16979 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
16981 SDValue Chain = Op.getOperand(0);
16982 SDValue Zero = DAG.getConstant(0, MVT::i32);
16984 DAG.getRegister(X86::ESP, MVT::i32), // Base
16985 DAG.getTargetConstant(1, MVT::i8), // Scale
16986 DAG.getRegister(0, MVT::i32), // Index
16987 DAG.getTargetConstant(0, MVT::i32), // Disp
16988 DAG.getRegister(0, MVT::i32), // Segment.
16992 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
16993 return SDValue(Res, 0);
16996 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
16997 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17000 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17001 SelectionDAG &DAG) {
17002 MVT T = Op.getSimpleValueType();
17006 switch(T.SimpleTy) {
17007 default: llvm_unreachable("Invalid value type!");
17008 case MVT::i8: Reg = X86::AL; size = 1; break;
17009 case MVT::i16: Reg = X86::AX; size = 2; break;
17010 case MVT::i32: Reg = X86::EAX; size = 4; break;
17012 assert(Subtarget->is64Bit() && "Node not type legal!");
17013 Reg = X86::RAX; size = 8;
17016 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17017 Op.getOperand(2), SDValue());
17018 SDValue Ops[] = { cpIn.getValue(0),
17021 DAG.getTargetConstant(size, MVT::i8),
17022 cpIn.getValue(1) };
17023 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17024 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17025 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17029 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17030 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17031 MVT::i32, cpOut.getValue(2));
17032 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17033 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17035 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17036 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17037 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17041 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17042 SelectionDAG &DAG) {
17043 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17044 MVT DstVT = Op.getSimpleValueType();
17046 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17047 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17048 if (DstVT != MVT::f64)
17049 // This conversion needs to be expanded.
17052 SDValue InVec = Op->getOperand(0);
17054 unsigned NumElts = SrcVT.getVectorNumElements();
17055 EVT SVT = SrcVT.getVectorElementType();
17057 // Widen the vector in input in the case of MVT::v2i32.
17058 // Example: from MVT::v2i32 to MVT::v4i32.
17059 SmallVector<SDValue, 16> Elts;
17060 for (unsigned i = 0, e = NumElts; i != e; ++i)
17061 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17062 DAG.getIntPtrConstant(i)));
17064 // Explicitly mark the extra elements as Undef.
17065 Elts.append(NumElts, DAG.getUNDEF(SVT));
17067 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17068 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17069 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
17070 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17071 DAG.getIntPtrConstant(0));
17074 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17075 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17076 assert((DstVT == MVT::i64 ||
17077 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17078 "Unexpected custom BITCAST");
17079 // i64 <=> MMX conversions are Legal.
17080 if (SrcVT==MVT::i64 && DstVT.isVector())
17082 if (DstVT==MVT::i64 && SrcVT.isVector())
17084 // MMX <=> MMX conversions are Legal.
17085 if (SrcVT.isVector() && DstVT.isVector())
17087 // All other conversions need to be expanded.
17091 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17092 SelectionDAG &DAG) {
17093 SDNode *Node = Op.getNode();
17096 Op = Op.getOperand(0);
17097 EVT VT = Op.getValueType();
17098 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17099 "CTPOP lowering only implemented for 128/256-bit wide vector types");
17101 unsigned NumElts = VT.getVectorNumElements();
17102 EVT EltVT = VT.getVectorElementType();
17103 unsigned Len = EltVT.getSizeInBits();
17105 // This is the vectorized version of the "best" algorithm from
17106 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
17107 // with a minor tweak to use a series of adds + shifts instead of vector
17108 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
17110 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
17111 // v8i32 => Always profitable
17113 // FIXME: There a couple of possible improvements:
17115 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
17116 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
17118 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
17119 "CTPOP not implemented for this vector element type.");
17121 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
17122 // extra legalization.
17123 bool NeedsBitcast = EltVT == MVT::i32;
17124 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
17126 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
17127 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
17128 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
17130 // v = v - ((v >> 1) & 0x55555555...)
17131 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
17132 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
17133 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
17135 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17137 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
17138 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
17140 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
17142 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
17143 if (VT != And.getValueType())
17144 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17145 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
17147 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
17148 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
17149 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
17150 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
17151 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
17153 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
17154 if (NeedsBitcast) {
17155 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17156 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
17157 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
17160 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
17161 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
17162 if (VT != AndRHS.getValueType()) {
17163 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
17164 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
17166 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
17168 // v = (v + (v >> 4)) & 0x0F0F0F0F...
17169 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
17170 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
17171 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
17172 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17174 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
17175 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
17176 if (NeedsBitcast) {
17177 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17178 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
17180 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
17181 if (VT != And.getValueType())
17182 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17184 // The algorithm mentioned above uses:
17185 // v = (v * 0x01010101...) >> (Len - 8)
17187 // Change it to use vector adds + vector shifts which yield faster results on
17188 // Haswell than using vector integer multiplication.
17190 // For i32 elements:
17191 // v = v + (v >> 8)
17192 // v = v + (v >> 16)
17194 // For i64 elements:
17195 // v = v + (v >> 8)
17196 // v = v + (v >> 16)
17197 // v = v + (v >> 32)
17200 SmallVector<SDValue, 8> Csts;
17201 for (unsigned i = 8; i <= Len/2; i *= 2) {
17202 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
17203 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
17204 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
17205 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17209 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
17210 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
17211 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
17212 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
17213 if (NeedsBitcast) {
17214 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17215 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
17217 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
17218 if (VT != And.getValueType())
17219 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17224 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
17225 SDNode *Node = Op.getNode();
17227 EVT T = Node->getValueType(0);
17228 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
17229 DAG.getConstant(0, T), Node->getOperand(2));
17230 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
17231 cast<AtomicSDNode>(Node)->getMemoryVT(),
17232 Node->getOperand(0),
17233 Node->getOperand(1), negOp,
17234 cast<AtomicSDNode>(Node)->getMemOperand(),
17235 cast<AtomicSDNode>(Node)->getOrdering(),
17236 cast<AtomicSDNode>(Node)->getSynchScope());
17239 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
17240 SDNode *Node = Op.getNode();
17242 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
17244 // Convert seq_cst store -> xchg
17245 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
17246 // FIXME: On 32-bit, store -> fist or movq would be more efficient
17247 // (The only way to get a 16-byte store is cmpxchg16b)
17248 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
17249 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
17250 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17251 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
17252 cast<AtomicSDNode>(Node)->getMemoryVT(),
17253 Node->getOperand(0),
17254 Node->getOperand(1), Node->getOperand(2),
17255 cast<AtomicSDNode>(Node)->getMemOperand(),
17256 cast<AtomicSDNode>(Node)->getOrdering(),
17257 cast<AtomicSDNode>(Node)->getSynchScope());
17258 return Swap.getValue(1);
17260 // Other atomic stores have a simple pattern.
17264 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
17265 EVT VT = Op.getNode()->getSimpleValueType(0);
17267 // Let legalize expand this if it isn't a legal type yet.
17268 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17271 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17274 bool ExtraOp = false;
17275 switch (Op.getOpcode()) {
17276 default: llvm_unreachable("Invalid code");
17277 case ISD::ADDC: Opc = X86ISD::ADD; break;
17278 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
17279 case ISD::SUBC: Opc = X86ISD::SUB; break;
17280 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
17284 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17286 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17287 Op.getOperand(1), Op.getOperand(2));
17290 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
17291 SelectionDAG &DAG) {
17292 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
17294 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
17295 // which returns the values as { float, float } (in XMM0) or
17296 // { double, double } (which is returned in XMM0, XMM1).
17298 SDValue Arg = Op.getOperand(0);
17299 EVT ArgVT = Arg.getValueType();
17300 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17302 TargetLowering::ArgListTy Args;
17303 TargetLowering::ArgListEntry Entry;
17307 Entry.isSExt = false;
17308 Entry.isZExt = false;
17309 Args.push_back(Entry);
17311 bool isF64 = ArgVT == MVT::f64;
17312 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
17313 // the small struct {f32, f32} is returned in (eax, edx). For f64,
17314 // the results are returned via SRet in memory.
17315 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
17316 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17317 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
17319 Type *RetTy = isF64
17320 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
17321 : (Type*)VectorType::get(ArgTy, 4);
17323 TargetLowering::CallLoweringInfo CLI(DAG);
17324 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
17325 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
17327 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
17330 // Returned in xmm0 and xmm1.
17331 return CallResult.first;
17333 // Returned in bits 0:31 and 32:64 xmm0.
17334 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17335 CallResult.first, DAG.getIntPtrConstant(0));
17336 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17337 CallResult.first, DAG.getIntPtrConstant(1));
17338 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
17339 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
17342 /// LowerOperation - Provide custom lowering hooks for some operations.
17344 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
17345 switch (Op.getOpcode()) {
17346 default: llvm_unreachable("Should not custom lower this!");
17347 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
17348 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
17349 return LowerCMP_SWAP(Op, Subtarget, DAG);
17350 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
17351 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
17352 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
17353 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
17354 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
17355 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
17356 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
17357 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
17358 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
17359 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
17360 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
17361 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
17362 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
17363 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
17364 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
17365 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
17366 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
17367 case ISD::SHL_PARTS:
17368 case ISD::SRA_PARTS:
17369 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
17370 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
17371 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
17372 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
17373 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
17374 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
17375 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
17376 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
17377 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
17378 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
17379 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
17381 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
17382 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
17383 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
17384 case ISD::SETCC: return LowerSETCC(Op, DAG);
17385 case ISD::SELECT: return LowerSELECT(Op, DAG);
17386 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
17387 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
17388 case ISD::VASTART: return LowerVASTART(Op, DAG);
17389 case ISD::VAARG: return LowerVAARG(Op, DAG);
17390 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
17391 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
17392 case ISD::INTRINSIC_VOID:
17393 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
17394 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
17395 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
17396 case ISD::FRAME_TO_ARGS_OFFSET:
17397 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
17398 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
17399 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
17400 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
17401 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
17402 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
17403 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
17404 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
17405 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
17406 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
17407 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
17408 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
17409 case ISD::UMUL_LOHI:
17410 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
17413 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
17419 case ISD::UMULO: return LowerXALUO(Op, DAG);
17420 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
17421 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
17425 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
17426 case ISD::ADD: return LowerADD(Op, DAG);
17427 case ISD::SUB: return LowerSUB(Op, DAG);
17428 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
17432 /// ReplaceNodeResults - Replace a node with an illegal result type
17433 /// with a new node built out of custom code.
17434 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
17435 SmallVectorImpl<SDValue>&Results,
17436 SelectionDAG &DAG) const {
17438 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17439 switch (N->getOpcode()) {
17441 llvm_unreachable("Do not know how to custom type legalize this operation!");
17442 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17443 case X86ISD::FMINC:
17445 case X86ISD::FMAXC:
17446 case X86ISD::FMAX: {
17447 EVT VT = N->getValueType(0);
17448 if (VT != MVT::v2f32)
17449 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17450 SDValue UNDEF = DAG.getUNDEF(VT);
17451 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17452 N->getOperand(0), UNDEF);
17453 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17454 N->getOperand(1), UNDEF);
17455 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17458 case ISD::SIGN_EXTEND_INREG:
17463 // We don't want to expand or promote these.
17470 case ISD::UDIVREM: {
17471 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17472 Results.push_back(V);
17475 case ISD::FP_TO_SINT:
17476 case ISD::FP_TO_UINT: {
17477 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17479 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17482 std::pair<SDValue,SDValue> Vals =
17483 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17484 SDValue FIST = Vals.first, StackSlot = Vals.second;
17485 if (FIST.getNode()) {
17486 EVT VT = N->getValueType(0);
17487 // Return a load from the stack slot.
17488 if (StackSlot.getNode())
17489 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17490 MachinePointerInfo(),
17491 false, false, false, 0));
17493 Results.push_back(FIST);
17497 case ISD::UINT_TO_FP: {
17498 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17499 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17500 N->getValueType(0) != MVT::v2f32)
17502 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17504 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
17506 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17507 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17508 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17509 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17510 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17511 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17514 case ISD::FP_ROUND: {
17515 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17517 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17518 Results.push_back(V);
17521 case ISD::INTRINSIC_W_CHAIN: {
17522 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17524 default : llvm_unreachable("Do not know how to custom type "
17525 "legalize this intrinsic operation!");
17526 case Intrinsic::x86_rdtsc:
17527 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17529 case Intrinsic::x86_rdtscp:
17530 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17532 case Intrinsic::x86_rdpmc:
17533 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17536 case ISD::READCYCLECOUNTER: {
17537 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17540 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17541 EVT T = N->getValueType(0);
17542 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17543 bool Regs64bit = T == MVT::i128;
17544 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17545 SDValue cpInL, cpInH;
17546 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17547 DAG.getConstant(0, HalfT));
17548 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17549 DAG.getConstant(1, HalfT));
17550 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17551 Regs64bit ? X86::RAX : X86::EAX,
17553 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17554 Regs64bit ? X86::RDX : X86::EDX,
17555 cpInH, cpInL.getValue(1));
17556 SDValue swapInL, swapInH;
17557 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17558 DAG.getConstant(0, HalfT));
17559 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17560 DAG.getConstant(1, HalfT));
17561 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17562 Regs64bit ? X86::RBX : X86::EBX,
17563 swapInL, cpInH.getValue(1));
17564 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17565 Regs64bit ? X86::RCX : X86::ECX,
17566 swapInH, swapInL.getValue(1));
17567 SDValue Ops[] = { swapInH.getValue(0),
17569 swapInH.getValue(1) };
17570 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17571 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17572 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17573 X86ISD::LCMPXCHG8_DAG;
17574 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17575 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17576 Regs64bit ? X86::RAX : X86::EAX,
17577 HalfT, Result.getValue(1));
17578 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17579 Regs64bit ? X86::RDX : X86::EDX,
17580 HalfT, cpOutL.getValue(2));
17581 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17583 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17584 MVT::i32, cpOutH.getValue(2));
17586 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17587 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17588 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17590 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17591 Results.push_back(Success);
17592 Results.push_back(EFLAGS.getValue(1));
17595 case ISD::ATOMIC_SWAP:
17596 case ISD::ATOMIC_LOAD_ADD:
17597 case ISD::ATOMIC_LOAD_SUB:
17598 case ISD::ATOMIC_LOAD_AND:
17599 case ISD::ATOMIC_LOAD_OR:
17600 case ISD::ATOMIC_LOAD_XOR:
17601 case ISD::ATOMIC_LOAD_NAND:
17602 case ISD::ATOMIC_LOAD_MIN:
17603 case ISD::ATOMIC_LOAD_MAX:
17604 case ISD::ATOMIC_LOAD_UMIN:
17605 case ISD::ATOMIC_LOAD_UMAX:
17606 case ISD::ATOMIC_LOAD: {
17607 // Delegate to generic TypeLegalization. Situations we can really handle
17608 // should have already been dealt with by AtomicExpandPass.cpp.
17611 case ISD::BITCAST: {
17612 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17613 EVT DstVT = N->getValueType(0);
17614 EVT SrcVT = N->getOperand(0)->getValueType(0);
17616 if (SrcVT != MVT::f64 ||
17617 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17620 unsigned NumElts = DstVT.getVectorNumElements();
17621 EVT SVT = DstVT.getVectorElementType();
17622 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17623 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17624 MVT::v2f64, N->getOperand(0));
17625 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17627 if (ExperimentalVectorWideningLegalization) {
17628 // If we are legalizing vectors by widening, we already have the desired
17629 // legal vector type, just return it.
17630 Results.push_back(ToVecInt);
17634 SmallVector<SDValue, 8> Elts;
17635 for (unsigned i = 0, e = NumElts; i != e; ++i)
17636 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17637 ToVecInt, DAG.getIntPtrConstant(i)));
17639 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17644 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17646 default: return nullptr;
17647 case X86ISD::BSF: return "X86ISD::BSF";
17648 case X86ISD::BSR: return "X86ISD::BSR";
17649 case X86ISD::SHLD: return "X86ISD::SHLD";
17650 case X86ISD::SHRD: return "X86ISD::SHRD";
17651 case X86ISD::FAND: return "X86ISD::FAND";
17652 case X86ISD::FANDN: return "X86ISD::FANDN";
17653 case X86ISD::FOR: return "X86ISD::FOR";
17654 case X86ISD::FXOR: return "X86ISD::FXOR";
17655 case X86ISD::FSRL: return "X86ISD::FSRL";
17656 case X86ISD::FILD: return "X86ISD::FILD";
17657 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17658 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17659 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17660 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17661 case X86ISD::FLD: return "X86ISD::FLD";
17662 case X86ISD::FST: return "X86ISD::FST";
17663 case X86ISD::CALL: return "X86ISD::CALL";
17664 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17665 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17666 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17667 case X86ISD::BT: return "X86ISD::BT";
17668 case X86ISD::CMP: return "X86ISD::CMP";
17669 case X86ISD::COMI: return "X86ISD::COMI";
17670 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17671 case X86ISD::CMPM: return "X86ISD::CMPM";
17672 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17673 case X86ISD::SETCC: return "X86ISD::SETCC";
17674 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17675 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17676 case X86ISD::CMOV: return "X86ISD::CMOV";
17677 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17678 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17679 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17680 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17681 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17682 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17683 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17684 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17685 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17686 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17687 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17688 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17689 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17690 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17691 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17692 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17693 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17694 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17695 case X86ISD::HADD: return "X86ISD::HADD";
17696 case X86ISD::HSUB: return "X86ISD::HSUB";
17697 case X86ISD::FHADD: return "X86ISD::FHADD";
17698 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17699 case X86ISD::UMAX: return "X86ISD::UMAX";
17700 case X86ISD::UMIN: return "X86ISD::UMIN";
17701 case X86ISD::SMAX: return "X86ISD::SMAX";
17702 case X86ISD::SMIN: return "X86ISD::SMIN";
17703 case X86ISD::FMAX: return "X86ISD::FMAX";
17704 case X86ISD::FMIN: return "X86ISD::FMIN";
17705 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17706 case X86ISD::FMINC: return "X86ISD::FMINC";
17707 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17708 case X86ISD::FRCP: return "X86ISD::FRCP";
17709 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17710 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17711 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17712 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17713 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17714 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17715 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17716 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17717 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17718 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17719 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17720 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17721 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17722 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17723 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17724 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17725 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17726 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17727 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17728 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17729 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17730 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17731 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17732 case X86ISD::VSHL: return "X86ISD::VSHL";
17733 case X86ISD::VSRL: return "X86ISD::VSRL";
17734 case X86ISD::VSRA: return "X86ISD::VSRA";
17735 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17736 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17737 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17738 case X86ISD::CMPP: return "X86ISD::CMPP";
17739 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17740 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17741 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17742 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17743 case X86ISD::ADD: return "X86ISD::ADD";
17744 case X86ISD::SUB: return "X86ISD::SUB";
17745 case X86ISD::ADC: return "X86ISD::ADC";
17746 case X86ISD::SBB: return "X86ISD::SBB";
17747 case X86ISD::SMUL: return "X86ISD::SMUL";
17748 case X86ISD::UMUL: return "X86ISD::UMUL";
17749 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17750 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17751 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17752 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17753 case X86ISD::INC: return "X86ISD::INC";
17754 case X86ISD::DEC: return "X86ISD::DEC";
17755 case X86ISD::OR: return "X86ISD::OR";
17756 case X86ISD::XOR: return "X86ISD::XOR";
17757 case X86ISD::AND: return "X86ISD::AND";
17758 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17759 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17760 case X86ISD::PTEST: return "X86ISD::PTEST";
17761 case X86ISD::TESTP: return "X86ISD::TESTP";
17762 case X86ISD::TESTM: return "X86ISD::TESTM";
17763 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17764 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17765 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17766 case X86ISD::PACKUS: return "X86ISD::PACKUS";
17767 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
17768 case X86ISD::VALIGN: return "X86ISD::VALIGN";
17769 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
17770 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
17771 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
17772 case X86ISD::SHUFP: return "X86ISD::SHUFP";
17773 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
17774 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
17775 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
17776 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
17777 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
17778 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
17779 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
17780 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
17781 case X86ISD::MOVSD: return "X86ISD::MOVSD";
17782 case X86ISD::MOVSS: return "X86ISD::MOVSS";
17783 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
17784 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
17785 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
17786 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
17787 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
17788 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
17789 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
17790 case X86ISD::VPERMV: return "X86ISD::VPERMV";
17791 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
17792 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
17793 case X86ISD::VPERMI: return "X86ISD::VPERMI";
17794 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
17795 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
17796 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
17797 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
17798 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
17799 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
17800 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
17801 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
17802 case X86ISD::SAHF: return "X86ISD::SAHF";
17803 case X86ISD::RDRAND: return "X86ISD::RDRAND";
17804 case X86ISD::RDSEED: return "X86ISD::RDSEED";
17805 case X86ISD::FMADD: return "X86ISD::FMADD";
17806 case X86ISD::FMSUB: return "X86ISD::FMSUB";
17807 case X86ISD::FNMADD: return "X86ISD::FNMADD";
17808 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
17809 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
17810 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
17811 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
17812 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
17813 case X86ISD::XTEST: return "X86ISD::XTEST";
17814 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
17815 case X86ISD::EXPAND: return "X86ISD::EXPAND";
17816 case X86ISD::SELECT: return "X86ISD::SELECT";
17817 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
17818 case X86ISD::RCP28: return "X86ISD::RCP28";
17819 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
17820 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
17821 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
17822 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
17823 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
17827 // isLegalAddressingMode - Return true if the addressing mode represented
17828 // by AM is legal for this target, for a load/store of the specified type.
17829 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
17831 // X86 supports extremely general addressing modes.
17832 CodeModel::Model M = getTargetMachine().getCodeModel();
17833 Reloc::Model R = getTargetMachine().getRelocationModel();
17835 // X86 allows a sign-extended 32-bit immediate field as a displacement.
17836 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
17841 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
17843 // If a reference to this global requires an extra load, we can't fold it.
17844 if (isGlobalStubReference(GVFlags))
17847 // If BaseGV requires a register for the PIC base, we cannot also have a
17848 // BaseReg specified.
17849 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
17852 // If lower 4G is not available, then we must use rip-relative addressing.
17853 if ((M != CodeModel::Small || R != Reloc::Static) &&
17854 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
17858 switch (AM.Scale) {
17864 // These scales always work.
17869 // These scales are formed with basereg+scalereg. Only accept if there is
17874 default: // Other stuff never works.
17881 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
17882 unsigned Bits = Ty->getScalarSizeInBits();
17884 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
17885 // particularly cheaper than those without.
17889 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
17890 // variable shifts just as cheap as scalar ones.
17891 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
17894 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
17895 // fully general vector.
17899 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
17900 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17902 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
17903 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
17904 return NumBits1 > NumBits2;
17907 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
17908 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17911 if (!isTypeLegal(EVT::getEVT(Ty1)))
17914 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
17916 // Assuming the caller doesn't have a zeroext or signext return parameter,
17917 // truncation all the way down to i1 is valid.
17921 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
17922 return isInt<32>(Imm);
17925 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
17926 // Can also use sub to handle negated immediates.
17927 return isInt<32>(Imm);
17930 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
17931 if (!VT1.isInteger() || !VT2.isInteger())
17933 unsigned NumBits1 = VT1.getSizeInBits();
17934 unsigned NumBits2 = VT2.getSizeInBits();
17935 return NumBits1 > NumBits2;
17938 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
17939 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17940 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
17943 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
17944 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17945 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
17948 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
17949 EVT VT1 = Val.getValueType();
17950 if (isZExtFree(VT1, VT2))
17953 if (Val.getOpcode() != ISD::LOAD)
17956 if (!VT1.isSimple() || !VT1.isInteger() ||
17957 !VT2.isSimple() || !VT2.isInteger())
17960 switch (VT1.getSimpleVT().SimpleTy) {
17965 // X86 has 8, 16, and 32-bit zero-extending loads.
17972 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
17975 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
17976 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
17979 VT = VT.getScalarType();
17981 if (!VT.isSimple())
17984 switch (VT.getSimpleVT().SimpleTy) {
17995 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
17996 // i16 instructions are longer (0x66 prefix) and potentially slower.
17997 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18000 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18001 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18002 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18003 /// are assumed to be legal.
18005 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18007 if (!VT.isSimple())
18010 // Very little shuffling can be done for 64-bit vectors right now.
18011 if (VT.getSizeInBits() == 64)
18014 // We only care that the types being shuffled are legal. The lowering can
18015 // handle any possible shuffle mask that results.
18016 return isTypeLegal(VT.getSimpleVT());
18020 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18022 // Just delegate to the generic legality, clear masks aren't special.
18023 return isShuffleMaskLegal(Mask, VT);
18026 //===----------------------------------------------------------------------===//
18027 // X86 Scheduler Hooks
18028 //===----------------------------------------------------------------------===//
18030 /// Utility function to emit xbegin specifying the start of an RTM region.
18031 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18032 const TargetInstrInfo *TII) {
18033 DebugLoc DL = MI->getDebugLoc();
18035 const BasicBlock *BB = MBB->getBasicBlock();
18036 MachineFunction::iterator I = MBB;
18039 // For the v = xbegin(), we generate
18050 MachineBasicBlock *thisMBB = MBB;
18051 MachineFunction *MF = MBB->getParent();
18052 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18053 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18054 MF->insert(I, mainMBB);
18055 MF->insert(I, sinkMBB);
18057 // Transfer the remainder of BB and its successor edges to sinkMBB.
18058 sinkMBB->splice(sinkMBB->begin(), MBB,
18059 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18060 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18064 // # fallthrough to mainMBB
18065 // # abortion to sinkMBB
18066 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18067 thisMBB->addSuccessor(mainMBB);
18068 thisMBB->addSuccessor(sinkMBB);
18072 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18073 mainMBB->addSuccessor(sinkMBB);
18076 // EAX is live into the sinkMBB
18077 sinkMBB->addLiveIn(X86::EAX);
18078 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18079 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18082 MI->eraseFromParent();
18086 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18087 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18088 // in the .td file.
18089 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18090 const TargetInstrInfo *TII) {
18092 switch (MI->getOpcode()) {
18093 default: llvm_unreachable("illegal opcode!");
18094 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18095 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18096 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18097 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18098 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18099 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18100 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18101 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18104 DebugLoc dl = MI->getDebugLoc();
18105 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18107 unsigned NumArgs = MI->getNumOperands();
18108 for (unsigned i = 1; i < NumArgs; ++i) {
18109 MachineOperand &Op = MI->getOperand(i);
18110 if (!(Op.isReg() && Op.isImplicit()))
18111 MIB.addOperand(Op);
18113 if (MI->hasOneMemOperand())
18114 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18116 BuildMI(*BB, MI, dl,
18117 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18118 .addReg(X86::XMM0);
18120 MI->eraseFromParent();
18124 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18125 // defs in an instruction pattern
18126 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18127 const TargetInstrInfo *TII) {
18129 switch (MI->getOpcode()) {
18130 default: llvm_unreachable("illegal opcode!");
18131 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18132 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18133 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18134 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18135 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18136 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18137 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18138 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18141 DebugLoc dl = MI->getDebugLoc();
18142 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18144 unsigned NumArgs = MI->getNumOperands(); // remove the results
18145 for (unsigned i = 1; i < NumArgs; ++i) {
18146 MachineOperand &Op = MI->getOperand(i);
18147 if (!(Op.isReg() && Op.isImplicit()))
18148 MIB.addOperand(Op);
18150 if (MI->hasOneMemOperand())
18151 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18153 BuildMI(*BB, MI, dl,
18154 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18157 MI->eraseFromParent();
18161 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18162 const X86Subtarget *Subtarget) {
18163 DebugLoc dl = MI->getDebugLoc();
18164 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18165 // Address into RAX/EAX, other two args into ECX, EDX.
18166 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18167 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18168 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18169 for (int i = 0; i < X86::AddrNumOperands; ++i)
18170 MIB.addOperand(MI->getOperand(i));
18172 unsigned ValOps = X86::AddrNumOperands;
18173 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18174 .addReg(MI->getOperand(ValOps).getReg());
18175 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18176 .addReg(MI->getOperand(ValOps+1).getReg());
18178 // The instruction doesn't actually take any operands though.
18179 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18181 MI->eraseFromParent(); // The pseudo is gone now.
18185 MachineBasicBlock *
18186 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
18187 MachineBasicBlock *MBB) const {
18188 // Emit va_arg instruction on X86-64.
18190 // Operands to this pseudo-instruction:
18191 // 0 ) Output : destination address (reg)
18192 // 1-5) Input : va_list address (addr, i64mem)
18193 // 6 ) ArgSize : Size (in bytes) of vararg type
18194 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18195 // 8 ) Align : Alignment of type
18196 // 9 ) EFLAGS (implicit-def)
18198 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18199 static_assert(X86::AddrNumOperands == 5,
18200 "VAARG_64 assumes 5 address operands");
18202 unsigned DestReg = MI->getOperand(0).getReg();
18203 MachineOperand &Base = MI->getOperand(1);
18204 MachineOperand &Scale = MI->getOperand(2);
18205 MachineOperand &Index = MI->getOperand(3);
18206 MachineOperand &Disp = MI->getOperand(4);
18207 MachineOperand &Segment = MI->getOperand(5);
18208 unsigned ArgSize = MI->getOperand(6).getImm();
18209 unsigned ArgMode = MI->getOperand(7).getImm();
18210 unsigned Align = MI->getOperand(8).getImm();
18212 // Memory Reference
18213 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
18214 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18215 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18217 // Machine Information
18218 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18219 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
18220 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
18221 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
18222 DebugLoc DL = MI->getDebugLoc();
18224 // struct va_list {
18227 // i64 overflow_area (address)
18228 // i64 reg_save_area (address)
18230 // sizeof(va_list) = 24
18231 // alignment(va_list) = 8
18233 unsigned TotalNumIntRegs = 6;
18234 unsigned TotalNumXMMRegs = 8;
18235 bool UseGPOffset = (ArgMode == 1);
18236 bool UseFPOffset = (ArgMode == 2);
18237 unsigned MaxOffset = TotalNumIntRegs * 8 +
18238 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
18240 /* Align ArgSize to a multiple of 8 */
18241 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
18242 bool NeedsAlign = (Align > 8);
18244 MachineBasicBlock *thisMBB = MBB;
18245 MachineBasicBlock *overflowMBB;
18246 MachineBasicBlock *offsetMBB;
18247 MachineBasicBlock *endMBB;
18249 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
18250 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
18251 unsigned OffsetReg = 0;
18253 if (!UseGPOffset && !UseFPOffset) {
18254 // If we only pull from the overflow region, we don't create a branch.
18255 // We don't need to alter control flow.
18256 OffsetDestReg = 0; // unused
18257 OverflowDestReg = DestReg;
18259 offsetMBB = nullptr;
18260 overflowMBB = thisMBB;
18263 // First emit code to check if gp_offset (or fp_offset) is below the bound.
18264 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
18265 // If not, pull from overflow_area. (branch to overflowMBB)
18270 // offsetMBB overflowMBB
18275 // Registers for the PHI in endMBB
18276 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
18277 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
18279 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18280 MachineFunction *MF = MBB->getParent();
18281 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18282 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18283 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18285 MachineFunction::iterator MBBIter = MBB;
18288 // Insert the new basic blocks
18289 MF->insert(MBBIter, offsetMBB);
18290 MF->insert(MBBIter, overflowMBB);
18291 MF->insert(MBBIter, endMBB);
18293 // Transfer the remainder of MBB and its successor edges to endMBB.
18294 endMBB->splice(endMBB->begin(), thisMBB,
18295 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
18296 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
18298 // Make offsetMBB and overflowMBB successors of thisMBB
18299 thisMBB->addSuccessor(offsetMBB);
18300 thisMBB->addSuccessor(overflowMBB);
18302 // endMBB is a successor of both offsetMBB and overflowMBB
18303 offsetMBB->addSuccessor(endMBB);
18304 overflowMBB->addSuccessor(endMBB);
18306 // Load the offset value into a register
18307 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18308 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
18312 .addDisp(Disp, UseFPOffset ? 4 : 0)
18313 .addOperand(Segment)
18314 .setMemRefs(MMOBegin, MMOEnd);
18316 // Check if there is enough room left to pull this argument.
18317 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
18319 .addImm(MaxOffset + 8 - ArgSizeA8);
18321 // Branch to "overflowMBB" if offset >= max
18322 // Fall through to "offsetMBB" otherwise
18323 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
18324 .addMBB(overflowMBB);
18327 // In offsetMBB, emit code to use the reg_save_area.
18329 assert(OffsetReg != 0);
18331 // Read the reg_save_area address.
18332 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
18333 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
18338 .addOperand(Segment)
18339 .setMemRefs(MMOBegin, MMOEnd);
18341 // Zero-extend the offset
18342 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
18343 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
18346 .addImm(X86::sub_32bit);
18348 // Add the offset to the reg_save_area to get the final address.
18349 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
18350 .addReg(OffsetReg64)
18351 .addReg(RegSaveReg);
18353 // Compute the offset for the next argument
18354 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18355 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
18357 .addImm(UseFPOffset ? 16 : 8);
18359 // Store it back into the va_list.
18360 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
18364 .addDisp(Disp, UseFPOffset ? 4 : 0)
18365 .addOperand(Segment)
18366 .addReg(NextOffsetReg)
18367 .setMemRefs(MMOBegin, MMOEnd);
18370 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
18375 // Emit code to use overflow area
18378 // Load the overflow_area address into a register.
18379 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
18380 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
18385 .addOperand(Segment)
18386 .setMemRefs(MMOBegin, MMOEnd);
18388 // If we need to align it, do so. Otherwise, just copy the address
18389 // to OverflowDestReg.
18391 // Align the overflow address
18392 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
18393 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
18395 // aligned_addr = (addr + (align-1)) & ~(align-1)
18396 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
18397 .addReg(OverflowAddrReg)
18400 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
18402 .addImm(~(uint64_t)(Align-1));
18404 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
18405 .addReg(OverflowAddrReg);
18408 // Compute the next overflow address after this argument.
18409 // (the overflow address should be kept 8-byte aligned)
18410 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
18411 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
18412 .addReg(OverflowDestReg)
18413 .addImm(ArgSizeA8);
18415 // Store the new overflow address.
18416 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
18421 .addOperand(Segment)
18422 .addReg(NextAddrReg)
18423 .setMemRefs(MMOBegin, MMOEnd);
18425 // If we branched, emit the PHI to the front of endMBB.
18427 BuildMI(*endMBB, endMBB->begin(), DL,
18428 TII->get(X86::PHI), DestReg)
18429 .addReg(OffsetDestReg).addMBB(offsetMBB)
18430 .addReg(OverflowDestReg).addMBB(overflowMBB);
18433 // Erase the pseudo instruction
18434 MI->eraseFromParent();
18439 MachineBasicBlock *
18440 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18442 MachineBasicBlock *MBB) const {
18443 // Emit code to save XMM registers to the stack. The ABI says that the
18444 // number of registers to save is given in %al, so it's theoretically
18445 // possible to do an indirect jump trick to avoid saving all of them,
18446 // however this code takes a simpler approach and just executes all
18447 // of the stores if %al is non-zero. It's less code, and it's probably
18448 // easier on the hardware branch predictor, and stores aren't all that
18449 // expensive anyway.
18451 // Create the new basic blocks. One block contains all the XMM stores,
18452 // and one block is the final destination regardless of whether any
18453 // stores were performed.
18454 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18455 MachineFunction *F = MBB->getParent();
18456 MachineFunction::iterator MBBIter = MBB;
18458 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
18459 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
18460 F->insert(MBBIter, XMMSaveMBB);
18461 F->insert(MBBIter, EndMBB);
18463 // Transfer the remainder of MBB and its successor edges to EndMBB.
18464 EndMBB->splice(EndMBB->begin(), MBB,
18465 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18466 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
18468 // The original block will now fall through to the XMM save block.
18469 MBB->addSuccessor(XMMSaveMBB);
18470 // The XMMSaveMBB will fall through to the end block.
18471 XMMSaveMBB->addSuccessor(EndMBB);
18473 // Now add the instructions.
18474 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18475 DebugLoc DL = MI->getDebugLoc();
18477 unsigned CountReg = MI->getOperand(0).getReg();
18478 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
18479 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
18481 if (!Subtarget->isTargetWin64()) {
18482 // If %al is 0, branch around the XMM save block.
18483 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
18484 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18485 MBB->addSuccessor(EndMBB);
18488 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18489 // that was just emitted, but clearly shouldn't be "saved".
18490 assert((MI->getNumOperands() <= 3 ||
18491 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18492 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18493 && "Expected last argument to be EFLAGS");
18494 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18495 // In the XMM save block, save all the XMM argument registers.
18496 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18497 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18498 MachineMemOperand *MMO =
18499 F->getMachineMemOperand(
18500 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18501 MachineMemOperand::MOStore,
18502 /*Size=*/16, /*Align=*/16);
18503 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18504 .addFrameIndex(RegSaveFrameIndex)
18505 .addImm(/*Scale=*/1)
18506 .addReg(/*IndexReg=*/0)
18507 .addImm(/*Disp=*/Offset)
18508 .addReg(/*Segment=*/0)
18509 .addReg(MI->getOperand(i).getReg())
18510 .addMemOperand(MMO);
18513 MI->eraseFromParent(); // The pseudo instruction is gone now.
18518 // The EFLAGS operand of SelectItr might be missing a kill marker
18519 // because there were multiple uses of EFLAGS, and ISel didn't know
18520 // which to mark. Figure out whether SelectItr should have had a
18521 // kill marker, and set it if it should. Returns the correct kill
18523 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18524 MachineBasicBlock* BB,
18525 const TargetRegisterInfo* TRI) {
18526 // Scan forward through BB for a use/def of EFLAGS.
18527 MachineBasicBlock::iterator miI(std::next(SelectItr));
18528 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18529 const MachineInstr& mi = *miI;
18530 if (mi.readsRegister(X86::EFLAGS))
18532 if (mi.definesRegister(X86::EFLAGS))
18533 break; // Should have kill-flag - update below.
18536 // If we hit the end of the block, check whether EFLAGS is live into a
18538 if (miI == BB->end()) {
18539 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18540 sEnd = BB->succ_end();
18541 sItr != sEnd; ++sItr) {
18542 MachineBasicBlock* succ = *sItr;
18543 if (succ->isLiveIn(X86::EFLAGS))
18548 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18549 // out. SelectMI should have a kill flag on EFLAGS.
18550 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18554 MachineBasicBlock *
18555 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18556 MachineBasicBlock *BB) const {
18557 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18558 DebugLoc DL = MI->getDebugLoc();
18560 // To "insert" a SELECT_CC instruction, we actually have to insert the
18561 // diamond control-flow pattern. The incoming instruction knows the
18562 // destination vreg to set, the condition code register to branch on, the
18563 // true/false values to select between, and a branch opcode to use.
18564 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18565 MachineFunction::iterator It = BB;
18571 // cmpTY ccX, r1, r2
18573 // fallthrough --> copy0MBB
18574 MachineBasicBlock *thisMBB = BB;
18575 MachineFunction *F = BB->getParent();
18577 // We also lower double CMOVs:
18578 // (CMOV (CMOV F, T, cc1), T, cc2)
18579 // to two successives branches. For that, we look for another CMOV as the
18580 // following instruction.
18582 // Without this, we would add a PHI between the two jumps, which ends up
18583 // creating a few copies all around. For instance, for
18585 // (sitofp (zext (fcmp une)))
18587 // we would generate:
18589 // ucomiss %xmm1, %xmm0
18590 // movss <1.0f>, %xmm0
18591 // movaps %xmm0, %xmm1
18593 // xorps %xmm1, %xmm1
18596 // movaps %xmm1, %xmm0
18600 // because this custom-inserter would have generated:
18612 // A: X = ...; Y = ...
18614 // C: Z = PHI [X, A], [Y, B]
18616 // E: PHI [X, C], [Z, D]
18618 // If we lower both CMOVs in a single step, we can instead generate:
18630 // A: X = ...; Y = ...
18632 // E: PHI [X, A], [X, C], [Y, D]
18634 // Which, in our sitofp/fcmp example, gives us something like:
18636 // ucomiss %xmm1, %xmm0
18637 // movss <1.0f>, %xmm0
18640 // xorps %xmm0, %xmm0
18644 MachineInstr *NextCMOV = nullptr;
18645 MachineBasicBlock::iterator NextMIIt =
18646 std::next(MachineBasicBlock::iterator(MI));
18647 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
18648 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
18649 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
18650 NextCMOV = &*NextMIIt;
18652 MachineBasicBlock *jcc1MBB = nullptr;
18654 // If we have a double CMOV, we lower it to two successive branches to
18655 // the same block. EFLAGS is used by both, so mark it as live in the second.
18657 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
18658 F->insert(It, jcc1MBB);
18659 jcc1MBB->addLiveIn(X86::EFLAGS);
18662 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18663 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18664 F->insert(It, copy0MBB);
18665 F->insert(It, sinkMBB);
18667 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18668 // live into the sink and copy blocks.
18669 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18671 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
18672 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
18673 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
18674 copy0MBB->addLiveIn(X86::EFLAGS);
18675 sinkMBB->addLiveIn(X86::EFLAGS);
18678 // Transfer the remainder of BB and its successor edges to sinkMBB.
18679 sinkMBB->splice(sinkMBB->begin(), BB,
18680 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18681 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18683 // Add the true and fallthrough blocks as its successors.
18685 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
18686 BB->addSuccessor(jcc1MBB);
18688 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
18689 // jump to the sinkMBB.
18690 jcc1MBB->addSuccessor(copy0MBB);
18691 jcc1MBB->addSuccessor(sinkMBB);
18693 BB->addSuccessor(copy0MBB);
18696 // The true block target of the first (or only) branch is always sinkMBB.
18697 BB->addSuccessor(sinkMBB);
18699 // Create the conditional branch instruction.
18701 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18702 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18705 unsigned Opc2 = X86::GetCondBranchFromCond(
18706 (X86::CondCode)NextCMOV->getOperand(3).getImm());
18707 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
18711 // %FalseValue = ...
18712 // # fallthrough to sinkMBB
18713 copy0MBB->addSuccessor(sinkMBB);
18716 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18718 MachineInstrBuilder MIB =
18719 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
18720 MI->getOperand(0).getReg())
18721 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18722 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18724 // If we have a double CMOV, the second Jcc provides the same incoming
18725 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
18727 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
18728 // Copy the PHI result to the register defined by the second CMOV.
18729 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
18730 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
18731 .addReg(MI->getOperand(0).getReg());
18732 NextCMOV->eraseFromParent();
18735 MI->eraseFromParent(); // The pseudo instruction is gone now.
18739 MachineBasicBlock *
18740 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18741 MachineBasicBlock *BB) const {
18742 MachineFunction *MF = BB->getParent();
18743 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18744 DebugLoc DL = MI->getDebugLoc();
18745 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18747 assert(MF->shouldSplitStack());
18749 const bool Is64Bit = Subtarget->is64Bit();
18750 const bool IsLP64 = Subtarget->isTarget64BitLP64();
18752 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
18753 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
18756 // ... [Till the alloca]
18757 // If stacklet is not large enough, jump to mallocMBB
18760 // Allocate by subtracting from RSP
18761 // Jump to continueMBB
18764 // Allocate by call to runtime
18768 // [rest of original BB]
18771 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18772 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18773 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18775 MachineRegisterInfo &MRI = MF->getRegInfo();
18776 const TargetRegisterClass *AddrRegClass =
18777 getRegClassFor(getPointerTy());
18779 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18780 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18781 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
18782 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
18783 sizeVReg = MI->getOperand(1).getReg(),
18784 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
18786 MachineFunction::iterator MBBIter = BB;
18789 MF->insert(MBBIter, bumpMBB);
18790 MF->insert(MBBIter, mallocMBB);
18791 MF->insert(MBBIter, continueMBB);
18793 continueMBB->splice(continueMBB->begin(), BB,
18794 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18795 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
18797 // Add code to the main basic block to check if the stack limit has been hit,
18798 // and if so, jump to mallocMBB otherwise to bumpMBB.
18799 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
18800 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
18801 .addReg(tmpSPVReg).addReg(sizeVReg);
18802 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
18803 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
18804 .addReg(SPLimitVReg);
18805 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
18807 // bumpMBB simply decreases the stack pointer, since we know the current
18808 // stacklet has enough space.
18809 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
18810 .addReg(SPLimitVReg);
18811 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
18812 .addReg(SPLimitVReg);
18813 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18815 // Calls into a routine in libgcc to allocate more space from the heap.
18816 const uint32_t *RegMask =
18817 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
18819 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
18821 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18822 .addExternalSymbol("__morestack_allocate_stack_space")
18823 .addRegMask(RegMask)
18824 .addReg(X86::RDI, RegState::Implicit)
18825 .addReg(X86::RAX, RegState::ImplicitDefine);
18826 } else if (Is64Bit) {
18827 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
18829 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18830 .addExternalSymbol("__morestack_allocate_stack_space")
18831 .addRegMask(RegMask)
18832 .addReg(X86::EDI, RegState::Implicit)
18833 .addReg(X86::EAX, RegState::ImplicitDefine);
18835 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
18837 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
18838 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
18839 .addExternalSymbol("__morestack_allocate_stack_space")
18840 .addRegMask(RegMask)
18841 .addReg(X86::EAX, RegState::ImplicitDefine);
18845 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
18848 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
18849 .addReg(IsLP64 ? X86::RAX : X86::EAX);
18850 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18852 // Set up the CFG correctly.
18853 BB->addSuccessor(bumpMBB);
18854 BB->addSuccessor(mallocMBB);
18855 mallocMBB->addSuccessor(continueMBB);
18856 bumpMBB->addSuccessor(continueMBB);
18858 // Take care of the PHI nodes.
18859 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
18860 MI->getOperand(0).getReg())
18861 .addReg(mallocPtrVReg).addMBB(mallocMBB)
18862 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
18864 // Delete the original pseudo instruction.
18865 MI->eraseFromParent();
18868 return continueMBB;
18871 MachineBasicBlock *
18872 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
18873 MachineBasicBlock *BB) const {
18874 DebugLoc DL = MI->getDebugLoc();
18876 assert(!Subtarget->isTargetMachO());
18878 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
18880 MI->eraseFromParent(); // The pseudo instruction is gone now.
18884 MachineBasicBlock *
18885 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
18886 MachineBasicBlock *BB) const {
18887 // This is pretty easy. We're taking the value that we received from
18888 // our load from the relocation, sticking it in either RDI (x86-64)
18889 // or EAX and doing an indirect call. The return value will then
18890 // be in the normal return register.
18891 MachineFunction *F = BB->getParent();
18892 const X86InstrInfo *TII = Subtarget->getInstrInfo();
18893 DebugLoc DL = MI->getDebugLoc();
18895 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
18896 assert(MI->getOperand(3).isGlobal() && "This should be a global");
18898 // Get a register mask for the lowered call.
18899 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
18900 // proper register mask.
18901 const uint32_t *RegMask =
18902 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
18903 if (Subtarget->is64Bit()) {
18904 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18905 TII->get(X86::MOV64rm), X86::RDI)
18907 .addImm(0).addReg(0)
18908 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18909 MI->getOperand(3).getTargetFlags())
18911 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
18912 addDirectMem(MIB, X86::RDI);
18913 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
18914 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
18915 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18916 TII->get(X86::MOV32rm), X86::EAX)
18918 .addImm(0).addReg(0)
18919 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18920 MI->getOperand(3).getTargetFlags())
18922 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18923 addDirectMem(MIB, X86::EAX);
18924 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18926 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18927 TII->get(X86::MOV32rm), X86::EAX)
18928 .addReg(TII->getGlobalBaseReg(F))
18929 .addImm(0).addReg(0)
18930 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18931 MI->getOperand(3).getTargetFlags())
18933 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18934 addDirectMem(MIB, X86::EAX);
18935 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18938 MI->eraseFromParent(); // The pseudo instruction is gone now.
18942 MachineBasicBlock *
18943 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
18944 MachineBasicBlock *MBB) const {
18945 DebugLoc DL = MI->getDebugLoc();
18946 MachineFunction *MF = MBB->getParent();
18947 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18948 MachineRegisterInfo &MRI = MF->getRegInfo();
18950 const BasicBlock *BB = MBB->getBasicBlock();
18951 MachineFunction::iterator I = MBB;
18954 // Memory Reference
18955 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18956 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18959 unsigned MemOpndSlot = 0;
18961 unsigned CurOp = 0;
18963 DstReg = MI->getOperand(CurOp++).getReg();
18964 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
18965 assert(RC->hasType(MVT::i32) && "Invalid destination!");
18966 unsigned mainDstReg = MRI.createVirtualRegister(RC);
18967 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
18969 MemOpndSlot = CurOp;
18971 MVT PVT = getPointerTy();
18972 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18973 "Invalid Pointer Size!");
18975 // For v = setjmp(buf), we generate
18978 // buf[LabelOffset] = restoreMBB
18979 // SjLjSetup restoreMBB
18985 // v = phi(main, restore)
18988 // if base pointer being used, load it from frame
18991 MachineBasicBlock *thisMBB = MBB;
18992 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18993 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18994 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
18995 MF->insert(I, mainMBB);
18996 MF->insert(I, sinkMBB);
18997 MF->push_back(restoreMBB);
18999 MachineInstrBuilder MIB;
19001 // Transfer the remainder of BB and its successor edges to sinkMBB.
19002 sinkMBB->splice(sinkMBB->begin(), MBB,
19003 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19004 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19007 unsigned PtrStoreOpc = 0;
19008 unsigned LabelReg = 0;
19009 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19010 Reloc::Model RM = MF->getTarget().getRelocationModel();
19011 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19012 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19014 // Prepare IP either in reg or imm.
19015 if (!UseImmLabel) {
19016 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19017 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19018 LabelReg = MRI.createVirtualRegister(PtrRC);
19019 if (Subtarget->is64Bit()) {
19020 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19024 .addMBB(restoreMBB)
19027 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19028 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19029 .addReg(XII->getGlobalBaseReg(MF))
19032 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19036 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19038 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19039 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19040 if (i == X86::AddrDisp)
19041 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19043 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19046 MIB.addReg(LabelReg);
19048 MIB.addMBB(restoreMBB);
19049 MIB.setMemRefs(MMOBegin, MMOEnd);
19051 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19052 .addMBB(restoreMBB);
19054 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19055 MIB.addRegMask(RegInfo->getNoPreservedMask());
19056 thisMBB->addSuccessor(mainMBB);
19057 thisMBB->addSuccessor(restoreMBB);
19061 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19062 mainMBB->addSuccessor(sinkMBB);
19065 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19066 TII->get(X86::PHI), DstReg)
19067 .addReg(mainDstReg).addMBB(mainMBB)
19068 .addReg(restoreDstReg).addMBB(restoreMBB);
19071 if (RegInfo->hasBasePointer(*MF)) {
19072 const bool Uses64BitFramePtr =
19073 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
19074 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
19075 X86FI->setRestoreBasePointer(MF);
19076 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
19077 unsigned BasePtr = RegInfo->getBaseRegister();
19078 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
19079 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
19080 FramePtr, true, X86FI->getRestoreBasePointerOffset())
19081 .setMIFlag(MachineInstr::FrameSetup);
19083 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19084 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
19085 restoreMBB->addSuccessor(sinkMBB);
19087 MI->eraseFromParent();
19091 MachineBasicBlock *
19092 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19093 MachineBasicBlock *MBB) const {
19094 DebugLoc DL = MI->getDebugLoc();
19095 MachineFunction *MF = MBB->getParent();
19096 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19097 MachineRegisterInfo &MRI = MF->getRegInfo();
19099 // Memory Reference
19100 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19101 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19103 MVT PVT = getPointerTy();
19104 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19105 "Invalid Pointer Size!");
19107 const TargetRegisterClass *RC =
19108 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19109 unsigned Tmp = MRI.createVirtualRegister(RC);
19110 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19111 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19112 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19113 unsigned SP = RegInfo->getStackRegister();
19115 MachineInstrBuilder MIB;
19117 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19118 const int64_t SPOffset = 2 * PVT.getStoreSize();
19120 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19121 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19124 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19125 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19126 MIB.addOperand(MI->getOperand(i));
19127 MIB.setMemRefs(MMOBegin, MMOEnd);
19129 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19130 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19131 if (i == X86::AddrDisp)
19132 MIB.addDisp(MI->getOperand(i), LabelOffset);
19134 MIB.addOperand(MI->getOperand(i));
19136 MIB.setMemRefs(MMOBegin, MMOEnd);
19138 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19139 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19140 if (i == X86::AddrDisp)
19141 MIB.addDisp(MI->getOperand(i), SPOffset);
19143 MIB.addOperand(MI->getOperand(i));
19145 MIB.setMemRefs(MMOBegin, MMOEnd);
19147 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19149 MI->eraseFromParent();
19153 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19154 // accumulator loops. Writing back to the accumulator allows the coalescer
19155 // to remove extra copies in the loop.
19156 MachineBasicBlock *
19157 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19158 MachineBasicBlock *MBB) const {
19159 MachineOperand &AddendOp = MI->getOperand(3);
19161 // Bail out early if the addend isn't a register - we can't switch these.
19162 if (!AddendOp.isReg())
19165 MachineFunction &MF = *MBB->getParent();
19166 MachineRegisterInfo &MRI = MF.getRegInfo();
19168 // Check whether the addend is defined by a PHI:
19169 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19170 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19171 if (!AddendDef.isPHI())
19174 // Look for the following pattern:
19176 // %addend = phi [%entry, 0], [%loop, %result]
19178 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19182 // %addend = phi [%entry, 0], [%loop, %result]
19184 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19186 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19187 assert(AddendDef.getOperand(i).isReg());
19188 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19189 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19190 if (&PHISrcInst == MI) {
19191 // Found a matching instruction.
19192 unsigned NewFMAOpc = 0;
19193 switch (MI->getOpcode()) {
19194 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19195 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19196 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19197 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19198 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19199 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19200 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19201 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19202 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19203 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19204 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19205 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19206 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19207 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19208 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19209 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19210 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
19211 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
19212 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
19213 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
19215 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19216 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19217 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19218 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19219 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19220 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19221 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19222 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19223 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
19224 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
19225 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
19226 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
19227 default: llvm_unreachable("Unrecognized FMA variant.");
19230 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
19231 MachineInstrBuilder MIB =
19232 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19233 .addOperand(MI->getOperand(0))
19234 .addOperand(MI->getOperand(3))
19235 .addOperand(MI->getOperand(2))
19236 .addOperand(MI->getOperand(1));
19237 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19238 MI->eraseFromParent();
19245 MachineBasicBlock *
19246 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19247 MachineBasicBlock *BB) const {
19248 switch (MI->getOpcode()) {
19249 default: llvm_unreachable("Unexpected instr type to insert");
19250 case X86::TAILJMPd64:
19251 case X86::TAILJMPr64:
19252 case X86::TAILJMPm64:
19253 case X86::TAILJMPd64_REX:
19254 case X86::TAILJMPr64_REX:
19255 case X86::TAILJMPm64_REX:
19256 llvm_unreachable("TAILJMP64 would not be touched here.");
19257 case X86::TCRETURNdi64:
19258 case X86::TCRETURNri64:
19259 case X86::TCRETURNmi64:
19261 case X86::WIN_ALLOCA:
19262 return EmitLoweredWinAlloca(MI, BB);
19263 case X86::SEG_ALLOCA_32:
19264 case X86::SEG_ALLOCA_64:
19265 return EmitLoweredSegAlloca(MI, BB);
19266 case X86::TLSCall_32:
19267 case X86::TLSCall_64:
19268 return EmitLoweredTLSCall(MI, BB);
19269 case X86::CMOV_GR8:
19270 case X86::CMOV_FR32:
19271 case X86::CMOV_FR64:
19272 case X86::CMOV_V4F32:
19273 case X86::CMOV_V2F64:
19274 case X86::CMOV_V2I64:
19275 case X86::CMOV_V8F32:
19276 case X86::CMOV_V4F64:
19277 case X86::CMOV_V4I64:
19278 case X86::CMOV_V16F32:
19279 case X86::CMOV_V8F64:
19280 case X86::CMOV_V8I64:
19281 case X86::CMOV_GR16:
19282 case X86::CMOV_GR32:
19283 case X86::CMOV_RFP32:
19284 case X86::CMOV_RFP64:
19285 case X86::CMOV_RFP80:
19286 return EmitLoweredSelect(MI, BB);
19288 case X86::FP32_TO_INT16_IN_MEM:
19289 case X86::FP32_TO_INT32_IN_MEM:
19290 case X86::FP32_TO_INT64_IN_MEM:
19291 case X86::FP64_TO_INT16_IN_MEM:
19292 case X86::FP64_TO_INT32_IN_MEM:
19293 case X86::FP64_TO_INT64_IN_MEM:
19294 case X86::FP80_TO_INT16_IN_MEM:
19295 case X86::FP80_TO_INT32_IN_MEM:
19296 case X86::FP80_TO_INT64_IN_MEM: {
19297 MachineFunction *F = BB->getParent();
19298 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19299 DebugLoc DL = MI->getDebugLoc();
19301 // Change the floating point control register to use "round towards zero"
19302 // mode when truncating to an integer value.
19303 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19304 addFrameReference(BuildMI(*BB, MI, DL,
19305 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19307 // Load the old value of the high byte of the control word...
19309 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
19310 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
19313 // Set the high part to be round to zero...
19314 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
19317 // Reload the modified control word now...
19318 addFrameReference(BuildMI(*BB, MI, DL,
19319 TII->get(X86::FLDCW16m)), CWFrameIdx);
19321 // Restore the memory image of control word to original value
19322 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
19325 // Get the X86 opcode to use.
19327 switch (MI->getOpcode()) {
19328 default: llvm_unreachable("illegal opcode!");
19329 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
19330 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
19331 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
19332 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
19333 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
19334 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
19335 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
19336 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
19337 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
19341 MachineOperand &Op = MI->getOperand(0);
19343 AM.BaseType = X86AddressMode::RegBase;
19344 AM.Base.Reg = Op.getReg();
19346 AM.BaseType = X86AddressMode::FrameIndexBase;
19347 AM.Base.FrameIndex = Op.getIndex();
19349 Op = MI->getOperand(1);
19351 AM.Scale = Op.getImm();
19352 Op = MI->getOperand(2);
19354 AM.IndexReg = Op.getImm();
19355 Op = MI->getOperand(3);
19356 if (Op.isGlobal()) {
19357 AM.GV = Op.getGlobal();
19359 AM.Disp = Op.getImm();
19361 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
19362 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
19364 // Reload the original control word now.
19365 addFrameReference(BuildMI(*BB, MI, DL,
19366 TII->get(X86::FLDCW16m)), CWFrameIdx);
19368 MI->eraseFromParent(); // The pseudo instruction is gone now.
19371 // String/text processing lowering.
19372 case X86::PCMPISTRM128REG:
19373 case X86::VPCMPISTRM128REG:
19374 case X86::PCMPISTRM128MEM:
19375 case X86::VPCMPISTRM128MEM:
19376 case X86::PCMPESTRM128REG:
19377 case X86::VPCMPESTRM128REG:
19378 case X86::PCMPESTRM128MEM:
19379 case X86::VPCMPESTRM128MEM:
19380 assert(Subtarget->hasSSE42() &&
19381 "Target must have SSE4.2 or AVX features enabled");
19382 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
19384 // String/text processing lowering.
19385 case X86::PCMPISTRIREG:
19386 case X86::VPCMPISTRIREG:
19387 case X86::PCMPISTRIMEM:
19388 case X86::VPCMPISTRIMEM:
19389 case X86::PCMPESTRIREG:
19390 case X86::VPCMPESTRIREG:
19391 case X86::PCMPESTRIMEM:
19392 case X86::VPCMPESTRIMEM:
19393 assert(Subtarget->hasSSE42() &&
19394 "Target must have SSE4.2 or AVX features enabled");
19395 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
19397 // Thread synchronization.
19399 return EmitMonitor(MI, BB, Subtarget);
19403 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
19405 case X86::VASTART_SAVE_XMM_REGS:
19406 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
19408 case X86::VAARG_64:
19409 return EmitVAARG64WithCustomInserter(MI, BB);
19411 case X86::EH_SjLj_SetJmp32:
19412 case X86::EH_SjLj_SetJmp64:
19413 return emitEHSjLjSetJmp(MI, BB);
19415 case X86::EH_SjLj_LongJmp32:
19416 case X86::EH_SjLj_LongJmp64:
19417 return emitEHSjLjLongJmp(MI, BB);
19419 case TargetOpcode::STATEPOINT:
19420 // As an implementation detail, STATEPOINT shares the STACKMAP format at
19421 // this point in the process. We diverge later.
19422 return emitPatchPoint(MI, BB);
19424 case TargetOpcode::STACKMAP:
19425 case TargetOpcode::PATCHPOINT:
19426 return emitPatchPoint(MI, BB);
19428 case X86::VFMADDPDr213r:
19429 case X86::VFMADDPSr213r:
19430 case X86::VFMADDSDr213r:
19431 case X86::VFMADDSSr213r:
19432 case X86::VFMSUBPDr213r:
19433 case X86::VFMSUBPSr213r:
19434 case X86::VFMSUBSDr213r:
19435 case X86::VFMSUBSSr213r:
19436 case X86::VFNMADDPDr213r:
19437 case X86::VFNMADDPSr213r:
19438 case X86::VFNMADDSDr213r:
19439 case X86::VFNMADDSSr213r:
19440 case X86::VFNMSUBPDr213r:
19441 case X86::VFNMSUBPSr213r:
19442 case X86::VFNMSUBSDr213r:
19443 case X86::VFNMSUBSSr213r:
19444 case X86::VFMADDSUBPDr213r:
19445 case X86::VFMADDSUBPSr213r:
19446 case X86::VFMSUBADDPDr213r:
19447 case X86::VFMSUBADDPSr213r:
19448 case X86::VFMADDPDr213rY:
19449 case X86::VFMADDPSr213rY:
19450 case X86::VFMSUBPDr213rY:
19451 case X86::VFMSUBPSr213rY:
19452 case X86::VFNMADDPDr213rY:
19453 case X86::VFNMADDPSr213rY:
19454 case X86::VFNMSUBPDr213rY:
19455 case X86::VFNMSUBPSr213rY:
19456 case X86::VFMADDSUBPDr213rY:
19457 case X86::VFMADDSUBPSr213rY:
19458 case X86::VFMSUBADDPDr213rY:
19459 case X86::VFMSUBADDPSr213rY:
19460 return emitFMA3Instr(MI, BB);
19464 //===----------------------------------------------------------------------===//
19465 // X86 Optimization Hooks
19466 //===----------------------------------------------------------------------===//
19468 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
19471 const SelectionDAG &DAG,
19472 unsigned Depth) const {
19473 unsigned BitWidth = KnownZero.getBitWidth();
19474 unsigned Opc = Op.getOpcode();
19475 assert((Opc >= ISD::BUILTIN_OP_END ||
19476 Opc == ISD::INTRINSIC_WO_CHAIN ||
19477 Opc == ISD::INTRINSIC_W_CHAIN ||
19478 Opc == ISD::INTRINSIC_VOID) &&
19479 "Should use MaskedValueIsZero if you don't know whether Op"
19480 " is a target node!");
19482 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
19496 // These nodes' second result is a boolean.
19497 if (Op.getResNo() == 0)
19500 case X86ISD::SETCC:
19501 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
19503 case ISD::INTRINSIC_WO_CHAIN: {
19504 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
19505 unsigned NumLoBits = 0;
19508 case Intrinsic::x86_sse_movmsk_ps:
19509 case Intrinsic::x86_avx_movmsk_ps_256:
19510 case Intrinsic::x86_sse2_movmsk_pd:
19511 case Intrinsic::x86_avx_movmsk_pd_256:
19512 case Intrinsic::x86_mmx_pmovmskb:
19513 case Intrinsic::x86_sse2_pmovmskb_128:
19514 case Intrinsic::x86_avx2_pmovmskb: {
19515 // High bits of movmskp{s|d}, pmovmskb are known zero.
19517 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
19518 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
19519 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
19520 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
19521 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
19522 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
19523 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
19524 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
19526 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
19535 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
19537 const SelectionDAG &,
19538 unsigned Depth) const {
19539 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
19540 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
19541 return Op.getValueType().getScalarType().getSizeInBits();
19547 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
19548 /// node is a GlobalAddress + offset.
19549 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
19550 const GlobalValue* &GA,
19551 int64_t &Offset) const {
19552 if (N->getOpcode() == X86ISD::Wrapper) {
19553 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
19554 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
19555 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
19559 return TargetLowering::isGAPlusOffset(N, GA, Offset);
19562 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
19563 /// same as extracting the high 128-bit part of 256-bit vector and then
19564 /// inserting the result into the low part of a new 256-bit vector
19565 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
19566 EVT VT = SVOp->getValueType(0);
19567 unsigned NumElems = VT.getVectorNumElements();
19569 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19570 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
19571 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19572 SVOp->getMaskElt(j) >= 0)
19578 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
19579 /// same as extracting the low 128-bit part of 256-bit vector and then
19580 /// inserting the result into the high part of a new 256-bit vector
19581 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
19582 EVT VT = SVOp->getValueType(0);
19583 unsigned NumElems = VT.getVectorNumElements();
19585 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19586 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
19587 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19588 SVOp->getMaskElt(j) >= 0)
19594 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
19595 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
19596 TargetLowering::DAGCombinerInfo &DCI,
19597 const X86Subtarget* Subtarget) {
19599 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19600 SDValue V1 = SVOp->getOperand(0);
19601 SDValue V2 = SVOp->getOperand(1);
19602 EVT VT = SVOp->getValueType(0);
19603 unsigned NumElems = VT.getVectorNumElements();
19605 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19606 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19610 // V UNDEF BUILD_VECTOR UNDEF
19612 // CONCAT_VECTOR CONCAT_VECTOR
19615 // RESULT: V + zero extended
19617 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19618 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19619 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19622 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19625 // To match the shuffle mask, the first half of the mask should
19626 // be exactly the first vector, and all the rest a splat with the
19627 // first element of the second one.
19628 for (unsigned i = 0; i != NumElems/2; ++i)
19629 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19630 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19633 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19634 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19635 if (Ld->hasNUsesOfValue(1, 0)) {
19636 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19637 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19639 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19641 Ld->getPointerInfo(),
19642 Ld->getAlignment(),
19643 false/*isVolatile*/, true/*ReadMem*/,
19644 false/*WriteMem*/);
19646 // Make sure the newly-created LOAD is in the same position as Ld in
19647 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19648 // and update uses of Ld's output chain to use the TokenFactor.
19649 if (Ld->hasAnyUseOfValue(1)) {
19650 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19651 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19652 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19653 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19654 SDValue(ResNode.getNode(), 1));
19657 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19661 // Emit a zeroed vector and insert the desired subvector on its
19663 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19664 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19665 return DCI.CombineTo(N, InsV);
19668 //===--------------------------------------------------------------------===//
19669 // Combine some shuffles into subvector extracts and inserts:
19672 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19673 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19674 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19675 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19676 return DCI.CombineTo(N, InsV);
19679 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19680 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19681 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19682 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19683 return DCI.CombineTo(N, InsV);
19689 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19692 /// This is the leaf of the recursive combinine below. When we have found some
19693 /// chain of single-use x86 shuffle instructions and accumulated the combined
19694 /// shuffle mask represented by them, this will try to pattern match that mask
19695 /// into either a single instruction if there is a special purpose instruction
19696 /// for this operation, or into a PSHUFB instruction which is a fully general
19697 /// instruction but should only be used to replace chains over a certain depth.
19698 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19699 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19700 TargetLowering::DAGCombinerInfo &DCI,
19701 const X86Subtarget *Subtarget) {
19702 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19704 // Find the operand that enters the chain. Note that multiple uses are OK
19705 // here, we're not going to remove the operand we find.
19706 SDValue Input = Op.getOperand(0);
19707 while (Input.getOpcode() == ISD::BITCAST)
19708 Input = Input.getOperand(0);
19710 MVT VT = Input.getSimpleValueType();
19711 MVT RootVT = Root.getSimpleValueType();
19714 // Just remove no-op shuffle masks.
19715 if (Mask.size() == 1) {
19716 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19721 // Use the float domain if the operand type is a floating point type.
19722 bool FloatDomain = VT.isFloatingPoint();
19724 // For floating point shuffles, we don't have free copies in the shuffle
19725 // instructions or the ability to load as part of the instruction, so
19726 // canonicalize their shuffles to UNPCK or MOV variants.
19728 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19729 // vectors because it can have a load folded into it that UNPCK cannot. This
19730 // doesn't preclude something switching to the shorter encoding post-RA.
19732 // FIXME: Should teach these routines about AVX vector widths.
19733 if (FloatDomain && VT.getSizeInBits() == 128) {
19734 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
19735 bool Lo = Mask.equals({0, 0});
19738 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19739 // is no slower than UNPCKLPD but has the option to fold the input operand
19740 // into even an unaligned memory load.
19741 if (Lo && Subtarget->hasSSE3()) {
19742 Shuffle = X86ISD::MOVDDUP;
19743 ShuffleVT = MVT::v2f64;
19745 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
19746 // than the UNPCK variants.
19747 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
19748 ShuffleVT = MVT::v4f32;
19750 if (Depth == 1 && Root->getOpcode() == Shuffle)
19751 return false; // Nothing to do!
19752 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19753 DCI.AddToWorklist(Op.getNode());
19754 if (Shuffle == X86ISD::MOVDDUP)
19755 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19757 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19758 DCI.AddToWorklist(Op.getNode());
19759 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19763 if (Subtarget->hasSSE3() &&
19764 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
19765 bool Lo = Mask.equals({0, 0, 2, 2});
19766 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
19767 MVT ShuffleVT = MVT::v4f32;
19768 if (Depth == 1 && Root->getOpcode() == Shuffle)
19769 return false; // Nothing to do!
19770 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19771 DCI.AddToWorklist(Op.getNode());
19772 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19773 DCI.AddToWorklist(Op.getNode());
19774 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19778 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
19779 bool Lo = Mask.equals({0, 0, 1, 1});
19780 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19781 MVT ShuffleVT = MVT::v4f32;
19782 if (Depth == 1 && Root->getOpcode() == Shuffle)
19783 return false; // Nothing to do!
19784 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19785 DCI.AddToWorklist(Op.getNode());
19786 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19787 DCI.AddToWorklist(Op.getNode());
19788 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19794 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
19795 // variants as none of these have single-instruction variants that are
19796 // superior to the UNPCK formulation.
19797 if (!FloatDomain && VT.getSizeInBits() == 128 &&
19798 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
19799 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
19800 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
19802 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
19803 bool Lo = Mask[0] == 0;
19804 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19805 if (Depth == 1 && Root->getOpcode() == Shuffle)
19806 return false; // Nothing to do!
19808 switch (Mask.size()) {
19810 ShuffleVT = MVT::v8i16;
19813 ShuffleVT = MVT::v16i8;
19816 llvm_unreachable("Impossible mask size!");
19818 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19819 DCI.AddToWorklist(Op.getNode());
19820 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19821 DCI.AddToWorklist(Op.getNode());
19822 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19827 // Don't try to re-form single instruction chains under any circumstances now
19828 // that we've done encoding canonicalization for them.
19832 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
19833 // can replace them with a single PSHUFB instruction profitably. Intel's
19834 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
19835 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
19836 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
19837 SmallVector<SDValue, 16> PSHUFBMask;
19838 int NumBytes = VT.getSizeInBits() / 8;
19839 int Ratio = NumBytes / Mask.size();
19840 for (int i = 0; i < NumBytes; ++i) {
19841 if (Mask[i / Ratio] == SM_SentinelUndef) {
19842 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
19845 int M = Mask[i / Ratio] != SM_SentinelZero
19846 ? Ratio * Mask[i / Ratio] + i % Ratio
19848 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
19850 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
19851 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Input);
19852 DCI.AddToWorklist(Op.getNode());
19853 SDValue PSHUFBMaskOp =
19854 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
19855 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
19856 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
19857 DCI.AddToWorklist(Op.getNode());
19858 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19863 // Failed to find any combines.
19867 /// \brief Fully generic combining of x86 shuffle instructions.
19869 /// This should be the last combine run over the x86 shuffle instructions. Once
19870 /// they have been fully optimized, this will recursively consider all chains
19871 /// of single-use shuffle instructions, build a generic model of the cumulative
19872 /// shuffle operation, and check for simpler instructions which implement this
19873 /// operation. We use this primarily for two purposes:
19875 /// 1) Collapse generic shuffles to specialized single instructions when
19876 /// equivalent. In most cases, this is just an encoding size win, but
19877 /// sometimes we will collapse multiple generic shuffles into a single
19878 /// special-purpose shuffle.
19879 /// 2) Look for sequences of shuffle instructions with 3 or more total
19880 /// instructions, and replace them with the slightly more expensive SSSE3
19881 /// PSHUFB instruction if available. We do this as the last combining step
19882 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
19883 /// a suitable short sequence of other instructions. The PHUFB will either
19884 /// use a register or have to read from memory and so is slightly (but only
19885 /// slightly) more expensive than the other shuffle instructions.
19887 /// Because this is inherently a quadratic operation (for each shuffle in
19888 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
19889 /// This should never be an issue in practice as the shuffle lowering doesn't
19890 /// produce sequences of more than 8 instructions.
19892 /// FIXME: We will currently miss some cases where the redundant shuffling
19893 /// would simplify under the threshold for PSHUFB formation because of
19894 /// combine-ordering. To fix this, we should do the redundant instruction
19895 /// combining in this recursive walk.
19896 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
19897 ArrayRef<int> RootMask,
19898 int Depth, bool HasPSHUFB,
19900 TargetLowering::DAGCombinerInfo &DCI,
19901 const X86Subtarget *Subtarget) {
19902 // Bound the depth of our recursive combine because this is ultimately
19903 // quadratic in nature.
19907 // Directly rip through bitcasts to find the underlying operand.
19908 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
19909 Op = Op.getOperand(0);
19911 MVT VT = Op.getSimpleValueType();
19912 if (!VT.isVector())
19913 return false; // Bail if we hit a non-vector.
19915 assert(Root.getSimpleValueType().isVector() &&
19916 "Shuffles operate on vector types!");
19917 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
19918 "Can only combine shuffles of the same vector register size.");
19920 if (!isTargetShuffle(Op.getOpcode()))
19922 SmallVector<int, 16> OpMask;
19924 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
19925 // We only can combine unary shuffles which we can decode the mask for.
19926 if (!HaveMask || !IsUnary)
19929 assert(VT.getVectorNumElements() == OpMask.size() &&
19930 "Different mask size from vector size!");
19931 assert(((RootMask.size() > OpMask.size() &&
19932 RootMask.size() % OpMask.size() == 0) ||
19933 (OpMask.size() > RootMask.size() &&
19934 OpMask.size() % RootMask.size() == 0) ||
19935 OpMask.size() == RootMask.size()) &&
19936 "The smaller number of elements must divide the larger.");
19937 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
19938 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
19939 assert(((RootRatio == 1 && OpRatio == 1) ||
19940 (RootRatio == 1) != (OpRatio == 1)) &&
19941 "Must not have a ratio for both incoming and op masks!");
19943 SmallVector<int, 16> Mask;
19944 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
19946 // Merge this shuffle operation's mask into our accumulated mask. Note that
19947 // this shuffle's mask will be the first applied to the input, followed by the
19948 // root mask to get us all the way to the root value arrangement. The reason
19949 // for this order is that we are recursing up the operation chain.
19950 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
19951 int RootIdx = i / RootRatio;
19952 if (RootMask[RootIdx] < 0) {
19953 // This is a zero or undef lane, we're done.
19954 Mask.push_back(RootMask[RootIdx]);
19958 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
19959 int OpIdx = RootMaskedIdx / OpRatio;
19960 if (OpMask[OpIdx] < 0) {
19961 // The incoming lanes are zero or undef, it doesn't matter which ones we
19963 Mask.push_back(OpMask[OpIdx]);
19967 // Ok, we have non-zero lanes, map them through.
19968 Mask.push_back(OpMask[OpIdx] * OpRatio +
19969 RootMaskedIdx % OpRatio);
19972 // See if we can recurse into the operand to combine more things.
19973 switch (Op.getOpcode()) {
19974 case X86ISD::PSHUFB:
19976 case X86ISD::PSHUFD:
19977 case X86ISD::PSHUFHW:
19978 case X86ISD::PSHUFLW:
19979 if (Op.getOperand(0).hasOneUse() &&
19980 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19981 HasPSHUFB, DAG, DCI, Subtarget))
19985 case X86ISD::UNPCKL:
19986 case X86ISD::UNPCKH:
19987 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
19988 // We can't check for single use, we have to check that this shuffle is the only user.
19989 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
19990 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19991 HasPSHUFB, DAG, DCI, Subtarget))
19996 // Minor canonicalization of the accumulated shuffle mask to make it easier
19997 // to match below. All this does is detect masks with squential pairs of
19998 // elements, and shrink them to the half-width mask. It does this in a loop
19999 // so it will reduce the size of the mask to the minimal width mask which
20000 // performs an equivalent shuffle.
20001 SmallVector<int, 16> WidenedMask;
20002 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
20003 Mask = std::move(WidenedMask);
20004 WidenedMask.clear();
20007 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20011 /// \brief Get the PSHUF-style mask from PSHUF node.
20013 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20014 /// PSHUF-style masks that can be reused with such instructions.
20015 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20016 MVT VT = N.getSimpleValueType();
20017 SmallVector<int, 4> Mask;
20019 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
20023 // If we have more than 128-bits, only the low 128-bits of shuffle mask
20024 // matter. Check that the upper masks are repeats and remove them.
20025 if (VT.getSizeInBits() > 128) {
20026 int LaneElts = 128 / VT.getScalarSizeInBits();
20028 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
20029 for (int j = 0; j < LaneElts; ++j)
20030 assert(Mask[j] == Mask[i * LaneElts + j] - LaneElts &&
20031 "Mask doesn't repeat in high 128-bit lanes!");
20033 Mask.resize(LaneElts);
20036 switch (N.getOpcode()) {
20037 case X86ISD::PSHUFD:
20039 case X86ISD::PSHUFLW:
20042 case X86ISD::PSHUFHW:
20043 Mask.erase(Mask.begin(), Mask.begin() + 4);
20044 for (int &M : Mask)
20048 llvm_unreachable("No valid shuffle instruction found!");
20052 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20054 /// We walk up the chain and look for a combinable shuffle, skipping over
20055 /// shuffles that we could hoist this shuffle's transformation past without
20056 /// altering anything.
20058 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20060 TargetLowering::DAGCombinerInfo &DCI) {
20061 assert(N.getOpcode() == X86ISD::PSHUFD &&
20062 "Called with something other than an x86 128-bit half shuffle!");
20065 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20066 // of the shuffles in the chain so that we can form a fresh chain to replace
20068 SmallVector<SDValue, 8> Chain;
20069 SDValue V = N.getOperand(0);
20070 for (; V.hasOneUse(); V = V.getOperand(0)) {
20071 switch (V.getOpcode()) {
20073 return SDValue(); // Nothing combined!
20076 // Skip bitcasts as we always know the type for the target specific
20080 case X86ISD::PSHUFD:
20081 // Found another dword shuffle.
20084 case X86ISD::PSHUFLW:
20085 // Check that the low words (being shuffled) are the identity in the
20086 // dword shuffle, and the high words are self-contained.
20087 if (Mask[0] != 0 || Mask[1] != 1 ||
20088 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20091 Chain.push_back(V);
20094 case X86ISD::PSHUFHW:
20095 // Check that the high words (being shuffled) are the identity in the
20096 // dword shuffle, and the low words are self-contained.
20097 if (Mask[2] != 2 || Mask[3] != 3 ||
20098 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20101 Chain.push_back(V);
20104 case X86ISD::UNPCKL:
20105 case X86ISD::UNPCKH:
20106 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20107 // shuffle into a preceding word shuffle.
20108 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
20109 V.getSimpleValueType().getScalarType() != MVT::i16)
20112 // Search for a half-shuffle which we can combine with.
20113 unsigned CombineOp =
20114 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20115 if (V.getOperand(0) != V.getOperand(1) ||
20116 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20118 Chain.push_back(V);
20119 V = V.getOperand(0);
20121 switch (V.getOpcode()) {
20123 return SDValue(); // Nothing to combine.
20125 case X86ISD::PSHUFLW:
20126 case X86ISD::PSHUFHW:
20127 if (V.getOpcode() == CombineOp)
20130 Chain.push_back(V);
20134 V = V.getOperand(0);
20138 } while (V.hasOneUse());
20141 // Break out of the loop if we break out of the switch.
20145 if (!V.hasOneUse())
20146 // We fell out of the loop without finding a viable combining instruction.
20149 // Merge this node's mask and our incoming mask.
20150 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20151 for (int &M : Mask)
20153 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20154 getV4X86ShuffleImm8ForMask(Mask, DAG));
20156 // Rebuild the chain around this new shuffle.
20157 while (!Chain.empty()) {
20158 SDValue W = Chain.pop_back_val();
20160 if (V.getValueType() != W.getOperand(0).getValueType())
20161 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
20163 switch (W.getOpcode()) {
20165 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20167 case X86ISD::UNPCKL:
20168 case X86ISD::UNPCKH:
20169 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20172 case X86ISD::PSHUFD:
20173 case X86ISD::PSHUFLW:
20174 case X86ISD::PSHUFHW:
20175 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20179 if (V.getValueType() != N.getValueType())
20180 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
20182 // Return the new chain to replace N.
20186 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20188 /// We walk up the chain, skipping shuffles of the other half and looking
20189 /// through shuffles which switch halves trying to find a shuffle of the same
20190 /// pair of dwords.
20191 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20193 TargetLowering::DAGCombinerInfo &DCI) {
20195 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20196 "Called with something other than an x86 128-bit half shuffle!");
20198 unsigned CombineOpcode = N.getOpcode();
20200 // Walk up a single-use chain looking for a combinable shuffle.
20201 SDValue V = N.getOperand(0);
20202 for (; V.hasOneUse(); V = V.getOperand(0)) {
20203 switch (V.getOpcode()) {
20205 return false; // Nothing combined!
20208 // Skip bitcasts as we always know the type for the target specific
20212 case X86ISD::PSHUFLW:
20213 case X86ISD::PSHUFHW:
20214 if (V.getOpcode() == CombineOpcode)
20217 // Other-half shuffles are no-ops.
20220 // Break out of the loop if we break out of the switch.
20224 if (!V.hasOneUse())
20225 // We fell out of the loop without finding a viable combining instruction.
20228 // Combine away the bottom node as its shuffle will be accumulated into
20229 // a preceding shuffle.
20230 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20232 // Record the old value.
20235 // Merge this node's mask and our incoming mask (adjusted to account for all
20236 // the pshufd instructions encountered).
20237 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20238 for (int &M : Mask)
20240 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20241 getV4X86ShuffleImm8ForMask(Mask, DAG));
20243 // Check that the shuffles didn't cancel each other out. If not, we need to
20244 // combine to the new one.
20246 // Replace the combinable shuffle with the combined one, updating all users
20247 // so that we re-evaluate the chain here.
20248 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20253 /// \brief Try to combine x86 target specific shuffles.
20254 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20255 TargetLowering::DAGCombinerInfo &DCI,
20256 const X86Subtarget *Subtarget) {
20258 MVT VT = N.getSimpleValueType();
20259 SmallVector<int, 4> Mask;
20261 switch (N.getOpcode()) {
20262 case X86ISD::PSHUFD:
20263 case X86ISD::PSHUFLW:
20264 case X86ISD::PSHUFHW:
20265 Mask = getPSHUFShuffleMask(N);
20266 assert(Mask.size() == 4);
20272 // Nuke no-op shuffles that show up after combining.
20273 if (isNoopShuffleMask(Mask))
20274 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20276 // Look for simplifications involving one or two shuffle instructions.
20277 SDValue V = N.getOperand(0);
20278 switch (N.getOpcode()) {
20281 case X86ISD::PSHUFLW:
20282 case X86ISD::PSHUFHW:
20283 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
20285 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20286 return SDValue(); // We combined away this shuffle, so we're done.
20288 // See if this reduces to a PSHUFD which is no more expensive and can
20289 // combine with more operations. Note that it has to at least flip the
20290 // dwords as otherwise it would have been removed as a no-op.
20291 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
20292 int DMask[] = {0, 1, 2, 3};
20293 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20294 DMask[DOffset + 0] = DOffset + 1;
20295 DMask[DOffset + 1] = DOffset + 0;
20296 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
20297 V = DAG.getNode(ISD::BITCAST, DL, DVT, V);
20298 DCI.AddToWorklist(V.getNode());
20299 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
20300 getV4X86ShuffleImm8ForMask(DMask, DAG));
20301 DCI.AddToWorklist(V.getNode());
20302 return DAG.getNode(ISD::BITCAST, DL, VT, V);
20305 // Look for shuffle patterns which can be implemented as a single unpack.
20306 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20307 // only works when we have a PSHUFD followed by two half-shuffles.
20308 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20309 (V.getOpcode() == X86ISD::PSHUFLW ||
20310 V.getOpcode() == X86ISD::PSHUFHW) &&
20311 V.getOpcode() != N.getOpcode() &&
20313 SDValue D = V.getOperand(0);
20314 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20315 D = D.getOperand(0);
20316 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20317 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20318 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20319 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20320 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20322 for (int i = 0; i < 4; ++i) {
20323 WordMask[i + NOffset] = Mask[i] + NOffset;
20324 WordMask[i + VOffset] = VMask[i] + VOffset;
20326 // Map the word mask through the DWord mask.
20328 for (int i = 0; i < 8; ++i)
20329 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20330 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20331 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
20332 // We can replace all three shuffles with an unpack.
20333 V = DAG.getNode(ISD::BITCAST, DL, VT, D.getOperand(0));
20334 DCI.AddToWorklist(V.getNode());
20335 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
20344 case X86ISD::PSHUFD:
20345 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
20354 /// \brief Try to combine a shuffle into a target-specific add-sub node.
20356 /// We combine this directly on the abstract vector shuffle nodes so it is
20357 /// easier to generically match. We also insert dummy vector shuffle nodes for
20358 /// the operands which explicitly discard the lanes which are unused by this
20359 /// operation to try to flow through the rest of the combiner the fact that
20360 /// they're unused.
20361 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
20363 EVT VT = N->getValueType(0);
20365 // We only handle target-independent shuffles.
20366 // FIXME: It would be easy and harmless to use the target shuffle mask
20367 // extraction tool to support more.
20368 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
20371 auto *SVN = cast<ShuffleVectorSDNode>(N);
20372 ArrayRef<int> Mask = SVN->getMask();
20373 SDValue V1 = N->getOperand(0);
20374 SDValue V2 = N->getOperand(1);
20376 // We require the first shuffle operand to be the SUB node, and the second to
20377 // be the ADD node.
20378 // FIXME: We should support the commuted patterns.
20379 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
20382 // If there are other uses of these operations we can't fold them.
20383 if (!V1->hasOneUse() || !V2->hasOneUse())
20386 // Ensure that both operations have the same operands. Note that we can
20387 // commute the FADD operands.
20388 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
20389 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
20390 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
20393 // We're looking for blends between FADD and FSUB nodes. We insist on these
20394 // nodes being lined up in a specific expected pattern.
20395 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
20396 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
20397 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
20400 // Only specific types are legal at this point, assert so we notice if and
20401 // when these change.
20402 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
20403 VT == MVT::v4f64) &&
20404 "Unknown vector type encountered!");
20406 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
20409 /// PerformShuffleCombine - Performs several different shuffle combines.
20410 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
20411 TargetLowering::DAGCombinerInfo &DCI,
20412 const X86Subtarget *Subtarget) {
20414 SDValue N0 = N->getOperand(0);
20415 SDValue N1 = N->getOperand(1);
20416 EVT VT = N->getValueType(0);
20418 // Don't create instructions with illegal types after legalize types has run.
20419 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20420 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
20423 // If we have legalized the vector types, look for blends of FADD and FSUB
20424 // nodes that we can fuse into an ADDSUB node.
20425 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
20426 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
20429 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
20430 if (Subtarget->hasFp256() && VT.is256BitVector() &&
20431 N->getOpcode() == ISD::VECTOR_SHUFFLE)
20432 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
20434 // During Type Legalization, when promoting illegal vector types,
20435 // the backend might introduce new shuffle dag nodes and bitcasts.
20437 // This code performs the following transformation:
20438 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
20439 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
20441 // We do this only if both the bitcast and the BINOP dag nodes have
20442 // one use. Also, perform this transformation only if the new binary
20443 // operation is legal. This is to avoid introducing dag nodes that
20444 // potentially need to be further expanded (or custom lowered) into a
20445 // less optimal sequence of dag nodes.
20446 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
20447 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
20448 N0.getOpcode() == ISD::BITCAST) {
20449 SDValue BC0 = N0.getOperand(0);
20450 EVT SVT = BC0.getValueType();
20451 unsigned Opcode = BC0.getOpcode();
20452 unsigned NumElts = VT.getVectorNumElements();
20454 if (BC0.hasOneUse() && SVT.isVector() &&
20455 SVT.getVectorNumElements() * 2 == NumElts &&
20456 TLI.isOperationLegal(Opcode, VT)) {
20457 bool CanFold = false;
20469 unsigned SVTNumElts = SVT.getVectorNumElements();
20470 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20471 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
20472 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
20473 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
20474 CanFold = SVOp->getMaskElt(i) < 0;
20477 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
20478 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
20479 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
20480 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
20485 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
20486 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
20487 // consecutive, non-overlapping, and in the right order.
20488 SmallVector<SDValue, 16> Elts;
20489 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
20490 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
20492 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
20496 if (isTargetShuffle(N->getOpcode())) {
20498 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
20499 if (Shuffle.getNode())
20502 // Try recursively combining arbitrary sequences of x86 shuffle
20503 // instructions into higher-order shuffles. We do this after combining
20504 // specific PSHUF instruction sequences into their minimal form so that we
20505 // can evaluate how many specialized shuffle instructions are involved in
20506 // a particular chain.
20507 SmallVector<int, 1> NonceMask; // Just a placeholder.
20508 NonceMask.push_back(0);
20509 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
20510 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
20512 return SDValue(); // This routine will use CombineTo to replace N.
20518 /// PerformTruncateCombine - Converts truncate operation to
20519 /// a sequence of vector shuffle operations.
20520 /// It is possible when we truncate 256-bit vector to 128-bit vector
20521 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
20522 TargetLowering::DAGCombinerInfo &DCI,
20523 const X86Subtarget *Subtarget) {
20527 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
20528 /// specific shuffle of a load can be folded into a single element load.
20529 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
20530 /// shuffles have been custom lowered so we need to handle those here.
20531 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
20532 TargetLowering::DAGCombinerInfo &DCI) {
20533 if (DCI.isBeforeLegalizeOps())
20536 SDValue InVec = N->getOperand(0);
20537 SDValue EltNo = N->getOperand(1);
20539 if (!isa<ConstantSDNode>(EltNo))
20542 EVT OriginalVT = InVec.getValueType();
20544 if (InVec.getOpcode() == ISD::BITCAST) {
20545 // Don't duplicate a load with other uses.
20546 if (!InVec.hasOneUse())
20548 EVT BCVT = InVec.getOperand(0).getValueType();
20549 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
20551 InVec = InVec.getOperand(0);
20554 EVT CurrentVT = InVec.getValueType();
20556 if (!isTargetShuffle(InVec.getOpcode()))
20559 // Don't duplicate a load with other uses.
20560 if (!InVec.hasOneUse())
20563 SmallVector<int, 16> ShuffleMask;
20565 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
20566 ShuffleMask, UnaryShuffle))
20569 // Select the input vector, guarding against out of range extract vector.
20570 unsigned NumElems = CurrentVT.getVectorNumElements();
20571 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
20572 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
20573 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
20574 : InVec.getOperand(1);
20576 // If inputs to shuffle are the same for both ops, then allow 2 uses
20577 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
20578 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
20580 if (LdNode.getOpcode() == ISD::BITCAST) {
20581 // Don't duplicate a load with other uses.
20582 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
20585 AllowedUses = 1; // only allow 1 load use if we have a bitcast
20586 LdNode = LdNode.getOperand(0);
20589 if (!ISD::isNormalLoad(LdNode.getNode()))
20592 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
20594 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
20597 EVT EltVT = N->getValueType(0);
20598 // If there's a bitcast before the shuffle, check if the load type and
20599 // alignment is valid.
20600 unsigned Align = LN0->getAlignment();
20601 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20602 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
20603 EltVT.getTypeForEVT(*DAG.getContext()));
20605 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
20608 // All checks match so transform back to vector_shuffle so that DAG combiner
20609 // can finish the job
20612 // Create shuffle node taking into account the case that its a unary shuffle
20613 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20614 : InVec.getOperand(1);
20615 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20616 InVec.getOperand(0), Shuffle,
20618 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20619 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20623 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20624 /// special and don't usually play with other vector types, it's better to
20625 /// handle them early to be sure we emit efficient code by avoiding
20626 /// store-load conversions.
20627 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20628 if (N->getValueType(0) != MVT::x86mmx ||
20629 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20630 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20633 SDValue V = N->getOperand(0);
20634 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20635 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20636 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20637 N->getValueType(0), V.getOperand(0));
20642 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20643 /// generation and convert it from being a bunch of shuffles and extracts
20644 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20645 /// storing the value and loading scalars back, while for x64 we should
20646 /// use 64-bit extracts and shifts.
20647 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20648 TargetLowering::DAGCombinerInfo &DCI) {
20649 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20650 if (NewOp.getNode())
20653 SDValue InputVector = N->getOperand(0);
20655 // Detect mmx to i32 conversion through a v2i32 elt extract.
20656 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20657 N->getValueType(0) == MVT::i32 &&
20658 InputVector.getValueType() == MVT::v2i32) {
20660 // The bitcast source is a direct mmx result.
20661 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20662 if (MMXSrc.getValueType() == MVT::x86mmx)
20663 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20664 N->getValueType(0),
20665 InputVector.getNode()->getOperand(0));
20667 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20668 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20669 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20670 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20671 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20672 MMXSrcOp.getValueType() == MVT::v1i64 &&
20673 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20674 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20675 N->getValueType(0),
20676 MMXSrcOp.getOperand(0));
20679 // Only operate on vectors of 4 elements, where the alternative shuffling
20680 // gets to be more expensive.
20681 if (InputVector.getValueType() != MVT::v4i32)
20684 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20685 // single use which is a sign-extend or zero-extend, and all elements are
20687 SmallVector<SDNode *, 4> Uses;
20688 unsigned ExtractedElements = 0;
20689 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20690 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20691 if (UI.getUse().getResNo() != InputVector.getResNo())
20694 SDNode *Extract = *UI;
20695 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20698 if (Extract->getValueType(0) != MVT::i32)
20700 if (!Extract->hasOneUse())
20702 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20703 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20705 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20708 // Record which element was extracted.
20709 ExtractedElements |=
20710 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20712 Uses.push_back(Extract);
20715 // If not all the elements were used, this may not be worthwhile.
20716 if (ExtractedElements != 15)
20719 // Ok, we've now decided to do the transformation.
20720 // If 64-bit shifts are legal, use the extract-shift sequence,
20721 // otherwise bounce the vector off the cache.
20722 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20724 SDLoc dl(InputVector);
20726 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20727 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20728 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20729 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20730 DAG.getConstant(0, VecIdxTy));
20731 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20732 DAG.getConstant(1, VecIdxTy));
20734 SDValue ShAmt = DAG.getConstant(32,
20735 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20736 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20737 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20738 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20739 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20740 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20741 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
20743 // Store the value to a temporary stack slot.
20744 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
20745 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
20746 MachinePointerInfo(), false, false, 0);
20748 EVT ElementType = InputVector.getValueType().getVectorElementType();
20749 unsigned EltSize = ElementType.getSizeInBits() / 8;
20751 // Replace each use (extract) with a load of the appropriate element.
20752 for (unsigned i = 0; i < 4; ++i) {
20753 uint64_t Offset = EltSize * i;
20754 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
20756 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
20757 StackPtr, OffsetVal);
20759 // Load the scalar.
20760 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
20761 ScalarAddr, MachinePointerInfo(),
20762 false, false, false, 0);
20767 // Replace the extracts
20768 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
20769 UE = Uses.end(); UI != UE; ++UI) {
20770 SDNode *Extract = *UI;
20772 SDValue Idx = Extract->getOperand(1);
20773 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
20774 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
20777 // The replacement was made in place; don't return anything.
20781 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
20782 static std::pair<unsigned, bool>
20783 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
20784 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
20785 if (!VT.isVector())
20786 return std::make_pair(0, false);
20788 bool NeedSplit = false;
20789 switch (VT.getSimpleVT().SimpleTy) {
20790 default: return std::make_pair(0, false);
20793 if (!Subtarget->hasVLX())
20794 return std::make_pair(0, false);
20798 if (!Subtarget->hasBWI())
20799 return std::make_pair(0, false);
20803 if (!Subtarget->hasAVX512())
20804 return std::make_pair(0, false);
20809 if (!Subtarget->hasAVX2())
20811 if (!Subtarget->hasAVX())
20812 return std::make_pair(0, false);
20817 if (!Subtarget->hasSSE2())
20818 return std::make_pair(0, false);
20821 // SSE2 has only a small subset of the operations.
20822 bool hasUnsigned = Subtarget->hasSSE41() ||
20823 (Subtarget->hasSSE2() && VT == MVT::v16i8);
20824 bool hasSigned = Subtarget->hasSSE41() ||
20825 (Subtarget->hasSSE2() && VT == MVT::v8i16);
20827 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20830 // Check for x CC y ? x : y.
20831 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20832 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20837 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20840 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20843 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20846 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20848 // Check for x CC y ? y : x -- a min/max with reversed arms.
20849 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20850 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20855 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20858 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20861 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20864 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20868 return std::make_pair(Opc, NeedSplit);
20872 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
20873 const X86Subtarget *Subtarget) {
20875 SDValue Cond = N->getOperand(0);
20876 SDValue LHS = N->getOperand(1);
20877 SDValue RHS = N->getOperand(2);
20879 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
20880 SDValue CondSrc = Cond->getOperand(0);
20881 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
20882 Cond = CondSrc->getOperand(0);
20885 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
20888 // A vselect where all conditions and data are constants can be optimized into
20889 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
20890 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
20891 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
20894 unsigned MaskValue = 0;
20895 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
20898 MVT VT = N->getSimpleValueType(0);
20899 unsigned NumElems = VT.getVectorNumElements();
20900 SmallVector<int, 8> ShuffleMask(NumElems, -1);
20901 for (unsigned i = 0; i < NumElems; ++i) {
20902 // Be sure we emit undef where we can.
20903 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
20904 ShuffleMask[i] = -1;
20906 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
20909 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20910 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
20912 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
20915 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
20917 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
20918 TargetLowering::DAGCombinerInfo &DCI,
20919 const X86Subtarget *Subtarget) {
20921 SDValue Cond = N->getOperand(0);
20922 // Get the LHS/RHS of the select.
20923 SDValue LHS = N->getOperand(1);
20924 SDValue RHS = N->getOperand(2);
20925 EVT VT = LHS.getValueType();
20926 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20928 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
20929 // instructions match the semantics of the common C idiom x<y?x:y but not
20930 // x<=y?x:y, because of how they handle negative zero (which can be
20931 // ignored in unsafe-math mode).
20932 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
20933 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
20934 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
20935 (Subtarget->hasSSE2() ||
20936 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
20937 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20939 unsigned Opcode = 0;
20940 // Check for x CC y ? x : y.
20941 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20942 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20946 // Converting this to a min would handle NaNs incorrectly, and swapping
20947 // the operands would cause it to handle comparisons between positive
20948 // and negative zero incorrectly.
20949 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20950 if (!DAG.getTarget().Options.UnsafeFPMath &&
20951 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20953 std::swap(LHS, RHS);
20955 Opcode = X86ISD::FMIN;
20958 // Converting this to a min would handle comparisons between positive
20959 // and negative zero incorrectly.
20960 if (!DAG.getTarget().Options.UnsafeFPMath &&
20961 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20963 Opcode = X86ISD::FMIN;
20966 // Converting this to a min would handle both negative zeros and NaNs
20967 // incorrectly, but we can swap the operands to fix both.
20968 std::swap(LHS, RHS);
20972 Opcode = X86ISD::FMIN;
20976 // Converting this to a max would handle comparisons between positive
20977 // and negative zero incorrectly.
20978 if (!DAG.getTarget().Options.UnsafeFPMath &&
20979 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20981 Opcode = X86ISD::FMAX;
20984 // Converting this to a max would handle NaNs incorrectly, and swapping
20985 // the operands would cause it to handle comparisons between positive
20986 // and negative zero incorrectly.
20987 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20988 if (!DAG.getTarget().Options.UnsafeFPMath &&
20989 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20991 std::swap(LHS, RHS);
20993 Opcode = X86ISD::FMAX;
20996 // Converting this to a max would handle both negative zeros and NaNs
20997 // incorrectly, but we can swap the operands to fix both.
20998 std::swap(LHS, RHS);
21002 Opcode = X86ISD::FMAX;
21005 // Check for x CC y ? y : x -- a min/max with reversed arms.
21006 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21007 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21011 // Converting this to a min would handle comparisons between positive
21012 // and negative zero incorrectly, and swapping the operands would
21013 // cause it to handle NaNs incorrectly.
21014 if (!DAG.getTarget().Options.UnsafeFPMath &&
21015 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21016 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21018 std::swap(LHS, RHS);
21020 Opcode = X86ISD::FMIN;
21023 // Converting this to a min would handle NaNs incorrectly.
21024 if (!DAG.getTarget().Options.UnsafeFPMath &&
21025 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21027 Opcode = X86ISD::FMIN;
21030 // Converting this to a min would handle both negative zeros and NaNs
21031 // incorrectly, but we can swap the operands to fix both.
21032 std::swap(LHS, RHS);
21036 Opcode = X86ISD::FMIN;
21040 // Converting this to a max would handle NaNs incorrectly.
21041 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21043 Opcode = X86ISD::FMAX;
21046 // Converting this to a max would handle comparisons between positive
21047 // and negative zero incorrectly, and swapping the operands would
21048 // cause it to handle NaNs incorrectly.
21049 if (!DAG.getTarget().Options.UnsafeFPMath &&
21050 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21051 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21053 std::swap(LHS, RHS);
21055 Opcode = X86ISD::FMAX;
21058 // Converting this to a max would handle both negative zeros and NaNs
21059 // incorrectly, but we can swap the operands to fix both.
21060 std::swap(LHS, RHS);
21064 Opcode = X86ISD::FMAX;
21070 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21073 EVT CondVT = Cond.getValueType();
21074 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21075 CondVT.getVectorElementType() == MVT::i1) {
21076 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21077 // lowering on KNL. In this case we convert it to
21078 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21079 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21080 // Since SKX these selects have a proper lowering.
21081 EVT OpVT = LHS.getValueType();
21082 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21083 (OpVT.getVectorElementType() == MVT::i8 ||
21084 OpVT.getVectorElementType() == MVT::i16) &&
21085 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21086 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21087 DCI.AddToWorklist(Cond.getNode());
21088 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21091 // If this is a select between two integer constants, try to do some
21093 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21094 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21095 // Don't do this for crazy integer types.
21096 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21097 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21098 // so that TrueC (the true value) is larger than FalseC.
21099 bool NeedsCondInvert = false;
21101 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21102 // Efficiently invertible.
21103 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21104 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21105 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21106 NeedsCondInvert = true;
21107 std::swap(TrueC, FalseC);
21110 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21111 if (FalseC->getAPIntValue() == 0 &&
21112 TrueC->getAPIntValue().isPowerOf2()) {
21113 if (NeedsCondInvert) // Invert the condition if needed.
21114 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21115 DAG.getConstant(1, Cond.getValueType()));
21117 // Zero extend the condition if needed.
21118 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21120 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21121 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21122 DAG.getConstant(ShAmt, MVT::i8));
21125 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21126 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21127 if (NeedsCondInvert) // Invert the condition if needed.
21128 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21129 DAG.getConstant(1, Cond.getValueType()));
21131 // Zero extend the condition if needed.
21132 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21133 FalseC->getValueType(0), Cond);
21134 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21135 SDValue(FalseC, 0));
21138 // Optimize cases that will turn into an LEA instruction. This requires
21139 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21140 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21141 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21142 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21144 bool isFastMultiplier = false;
21146 switch ((unsigned char)Diff) {
21148 case 1: // result = add base, cond
21149 case 2: // result = lea base( , cond*2)
21150 case 3: // result = lea base(cond, cond*2)
21151 case 4: // result = lea base( , cond*4)
21152 case 5: // result = lea base(cond, cond*4)
21153 case 8: // result = lea base( , cond*8)
21154 case 9: // result = lea base(cond, cond*8)
21155 isFastMultiplier = true;
21160 if (isFastMultiplier) {
21161 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21162 if (NeedsCondInvert) // Invert the condition if needed.
21163 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21164 DAG.getConstant(1, Cond.getValueType()));
21166 // Zero extend the condition if needed.
21167 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21169 // Scale the condition by the difference.
21171 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21172 DAG.getConstant(Diff, Cond.getValueType()));
21174 // Add the base if non-zero.
21175 if (FalseC->getAPIntValue() != 0)
21176 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21177 SDValue(FalseC, 0));
21184 // Canonicalize max and min:
21185 // (x > y) ? x : y -> (x >= y) ? x : y
21186 // (x < y) ? x : y -> (x <= y) ? x : y
21187 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21188 // the need for an extra compare
21189 // against zero. e.g.
21190 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21192 // testl %edi, %edi
21194 // cmovgl %edi, %eax
21198 // cmovsl %eax, %edi
21199 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21200 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21201 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21202 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21207 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21208 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21209 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21210 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21215 // Early exit check
21216 if (!TLI.isTypeLegal(VT))
21219 // Match VSELECTs into subs with unsigned saturation.
21220 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21221 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21222 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21223 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21224 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21226 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21227 // left side invert the predicate to simplify logic below.
21229 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21231 CC = ISD::getSetCCInverse(CC, true);
21232 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21236 if (Other.getNode() && Other->getNumOperands() == 2 &&
21237 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21238 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21239 SDValue CondRHS = Cond->getOperand(1);
21241 // Look for a general sub with unsigned saturation first.
21242 // x >= y ? x-y : 0 --> subus x, y
21243 // x > y ? x-y : 0 --> subus x, y
21244 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21245 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21246 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21248 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21249 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21250 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21251 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21252 // If the RHS is a constant we have to reverse the const
21253 // canonicalization.
21254 // x > C-1 ? x+-C : 0 --> subus x, C
21255 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21256 CondRHSConst->getAPIntValue() ==
21257 (-OpRHSConst->getAPIntValue() - 1))
21258 return DAG.getNode(
21259 X86ISD::SUBUS, DL, VT, OpLHS,
21260 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
21262 // Another special case: If C was a sign bit, the sub has been
21263 // canonicalized into a xor.
21264 // FIXME: Would it be better to use computeKnownBits to determine
21265 // whether it's safe to decanonicalize the xor?
21266 // x s< 0 ? x^C : 0 --> subus x, C
21267 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21268 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21269 OpRHSConst->getAPIntValue().isSignBit())
21270 // Note that we have to rebuild the RHS constant here to ensure we
21271 // don't rely on particular values of undef lanes.
21272 return DAG.getNode(
21273 X86ISD::SUBUS, DL, VT, OpLHS,
21274 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
21279 // Try to match a min/max vector operation.
21280 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21281 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21282 unsigned Opc = ret.first;
21283 bool NeedSplit = ret.second;
21285 if (Opc && NeedSplit) {
21286 unsigned NumElems = VT.getVectorNumElements();
21287 // Extract the LHS vectors
21288 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21289 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21291 // Extract the RHS vectors
21292 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21293 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21295 // Create min/max for each subvector
21296 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21297 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21299 // Merge the result
21300 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21302 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21305 // Simplify vector selection if condition value type matches vselect
21307 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
21308 assert(Cond.getValueType().isVector() &&
21309 "vector select expects a vector selector!");
21311 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21312 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21314 // Try invert the condition if true value is not all 1s and false value
21316 if (!TValIsAllOnes && !FValIsAllZeros &&
21317 // Check if the selector will be produced by CMPP*/PCMP*
21318 Cond.getOpcode() == ISD::SETCC &&
21319 // Check if SETCC has already been promoted
21320 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
21321 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21322 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21324 if (TValIsAllZeros || FValIsAllOnes) {
21325 SDValue CC = Cond.getOperand(2);
21326 ISD::CondCode NewCC =
21327 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21328 Cond.getOperand(0).getValueType().isInteger());
21329 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21330 std::swap(LHS, RHS);
21331 TValIsAllOnes = FValIsAllOnes;
21332 FValIsAllZeros = TValIsAllZeros;
21336 if (TValIsAllOnes || FValIsAllZeros) {
21339 if (TValIsAllOnes && FValIsAllZeros)
21341 else if (TValIsAllOnes)
21342 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
21343 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
21344 else if (FValIsAllZeros)
21345 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21346 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
21348 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
21352 // We should generate an X86ISD::BLENDI from a vselect if its argument
21353 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
21354 // constants. This specific pattern gets generated when we split a
21355 // selector for a 512 bit vector in a machine without AVX512 (but with
21356 // 256-bit vectors), during legalization:
21358 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
21360 // Iff we find this pattern and the build_vectors are built from
21361 // constants, we translate the vselect into a shuffle_vector that we
21362 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
21363 if ((N->getOpcode() == ISD::VSELECT ||
21364 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
21365 !DCI.isBeforeLegalize()) {
21366 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
21367 if (Shuffle.getNode())
21371 // If this is a *dynamic* select (non-constant condition) and we can match
21372 // this node with one of the variable blend instructions, restructure the
21373 // condition so that the blends can use the high bit of each element and use
21374 // SimplifyDemandedBits to simplify the condition operand.
21375 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
21376 !DCI.isBeforeLegalize() &&
21377 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
21378 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
21380 // Don't optimize vector selects that map to mask-registers.
21384 // We can only handle the cases where VSELECT is directly legal on the
21385 // subtarget. We custom lower VSELECT nodes with constant conditions and
21386 // this makes it hard to see whether a dynamic VSELECT will correctly
21387 // lower, so we both check the operation's status and explicitly handle the
21388 // cases where a *dynamic* blend will fail even though a constant-condition
21389 // blend could be custom lowered.
21390 // FIXME: We should find a better way to handle this class of problems.
21391 // Potentially, we should combine constant-condition vselect nodes
21392 // pre-legalization into shuffles and not mark as many types as custom
21394 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
21396 // FIXME: We don't support i16-element blends currently. We could and
21397 // should support them by making *all* the bits in the condition be set
21398 // rather than just the high bit and using an i8-element blend.
21399 if (VT.getScalarType() == MVT::i16)
21401 // Dynamic blending was only available from SSE4.1 onward.
21402 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
21404 // Byte blends are only available in AVX2
21405 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
21406 !Subtarget->hasAVX2())
21409 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
21410 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
21412 APInt KnownZero, KnownOne;
21413 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
21414 DCI.isBeforeLegalizeOps());
21415 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
21416 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
21418 // If we changed the computation somewhere in the DAG, this change
21419 // will affect all users of Cond.
21420 // Make sure it is fine and update all the nodes so that we do not
21421 // use the generic VSELECT anymore. Otherwise, we may perform
21422 // wrong optimizations as we messed up with the actual expectation
21423 // for the vector boolean values.
21424 if (Cond != TLO.Old) {
21425 // Check all uses of that condition operand to check whether it will be
21426 // consumed by non-BLEND instructions, which may depend on all bits are
21428 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21430 if (I->getOpcode() != ISD::VSELECT)
21431 // TODO: Add other opcodes eventually lowered into BLEND.
21434 // Update all the users of the condition, before committing the change,
21435 // so that the VSELECT optimizations that expect the correct vector
21436 // boolean value will not be triggered.
21437 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21439 DAG.ReplaceAllUsesOfValueWith(
21441 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
21442 Cond, I->getOperand(1), I->getOperand(2)));
21443 DCI.CommitTargetLoweringOpt(TLO);
21446 // At this point, only Cond is changed. Change the condition
21447 // just for N to keep the opportunity to optimize all other
21448 // users their own way.
21449 DAG.ReplaceAllUsesOfValueWith(
21451 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
21452 TLO.New, N->getOperand(1), N->getOperand(2)));
21460 // Check whether a boolean test is testing a boolean value generated by
21461 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
21464 // Simplify the following patterns:
21465 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
21466 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
21467 // to (Op EFLAGS Cond)
21469 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
21470 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
21471 // to (Op EFLAGS !Cond)
21473 // where Op could be BRCOND or CMOV.
21475 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
21476 // Quit if not CMP and SUB with its value result used.
21477 if (Cmp.getOpcode() != X86ISD::CMP &&
21478 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
21481 // Quit if not used as a boolean value.
21482 if (CC != X86::COND_E && CC != X86::COND_NE)
21485 // Check CMP operands. One of them should be 0 or 1 and the other should be
21486 // an SetCC or extended from it.
21487 SDValue Op1 = Cmp.getOperand(0);
21488 SDValue Op2 = Cmp.getOperand(1);
21491 const ConstantSDNode* C = nullptr;
21492 bool needOppositeCond = (CC == X86::COND_E);
21493 bool checkAgainstTrue = false; // Is it a comparison against 1?
21495 if ((C = dyn_cast<ConstantSDNode>(Op1)))
21497 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
21499 else // Quit if all operands are not constants.
21502 if (C->getZExtValue() == 1) {
21503 needOppositeCond = !needOppositeCond;
21504 checkAgainstTrue = true;
21505 } else if (C->getZExtValue() != 0)
21506 // Quit if the constant is neither 0 or 1.
21509 bool truncatedToBoolWithAnd = false;
21510 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
21511 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
21512 SetCC.getOpcode() == ISD::TRUNCATE ||
21513 SetCC.getOpcode() == ISD::AND) {
21514 if (SetCC.getOpcode() == ISD::AND) {
21516 ConstantSDNode *CS;
21517 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
21518 CS->getZExtValue() == 1)
21520 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
21521 CS->getZExtValue() == 1)
21525 SetCC = SetCC.getOperand(OpIdx);
21526 truncatedToBoolWithAnd = true;
21528 SetCC = SetCC.getOperand(0);
21531 switch (SetCC.getOpcode()) {
21532 case X86ISD::SETCC_CARRY:
21533 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
21534 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
21535 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
21536 // truncated to i1 using 'and'.
21537 if (checkAgainstTrue && !truncatedToBoolWithAnd)
21539 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
21540 "Invalid use of SETCC_CARRY!");
21542 case X86ISD::SETCC:
21543 // Set the condition code or opposite one if necessary.
21544 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
21545 if (needOppositeCond)
21546 CC = X86::GetOppositeBranchCondition(CC);
21547 return SetCC.getOperand(1);
21548 case X86ISD::CMOV: {
21549 // Check whether false/true value has canonical one, i.e. 0 or 1.
21550 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
21551 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
21552 // Quit if true value is not a constant.
21555 // Quit if false value is not a constant.
21557 SDValue Op = SetCC.getOperand(0);
21558 // Skip 'zext' or 'trunc' node.
21559 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
21560 Op.getOpcode() == ISD::TRUNCATE)
21561 Op = Op.getOperand(0);
21562 // A special case for rdrand/rdseed, where 0 is set if false cond is
21564 if ((Op.getOpcode() != X86ISD::RDRAND &&
21565 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
21568 // Quit if false value is not the constant 0 or 1.
21569 bool FValIsFalse = true;
21570 if (FVal && FVal->getZExtValue() != 0) {
21571 if (FVal->getZExtValue() != 1)
21573 // If FVal is 1, opposite cond is needed.
21574 needOppositeCond = !needOppositeCond;
21575 FValIsFalse = false;
21577 // Quit if TVal is not the constant opposite of FVal.
21578 if (FValIsFalse && TVal->getZExtValue() != 1)
21580 if (!FValIsFalse && TVal->getZExtValue() != 0)
21582 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
21583 if (needOppositeCond)
21584 CC = X86::GetOppositeBranchCondition(CC);
21585 return SetCC.getOperand(3);
21592 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
21594 /// (X86or (X86setcc) (X86setcc))
21595 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
21596 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
21597 X86::CondCode &CC1, SDValue &Flags,
21599 if (Cond->getOpcode() == X86ISD::CMP) {
21600 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
21601 if (!CondOp1C || !CondOp1C->isNullValue())
21604 Cond = Cond->getOperand(0);
21609 SDValue SetCC0, SetCC1;
21610 switch (Cond->getOpcode()) {
21611 default: return false;
21618 SetCC0 = Cond->getOperand(0);
21619 SetCC1 = Cond->getOperand(1);
21623 // Make sure we have SETCC nodes, using the same flags value.
21624 if (SetCC0.getOpcode() != X86ISD::SETCC ||
21625 SetCC1.getOpcode() != X86ISD::SETCC ||
21626 SetCC0->getOperand(1) != SetCC1->getOperand(1))
21629 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
21630 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
21631 Flags = SetCC0->getOperand(1);
21635 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
21636 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
21637 TargetLowering::DAGCombinerInfo &DCI,
21638 const X86Subtarget *Subtarget) {
21641 // If the flag operand isn't dead, don't touch this CMOV.
21642 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
21645 SDValue FalseOp = N->getOperand(0);
21646 SDValue TrueOp = N->getOperand(1);
21647 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
21648 SDValue Cond = N->getOperand(3);
21650 if (CC == X86::COND_E || CC == X86::COND_NE) {
21651 switch (Cond.getOpcode()) {
21655 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
21656 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
21657 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21663 Flags = checkBoolTestSetCCCombine(Cond, CC);
21664 if (Flags.getNode() &&
21665 // Extra check as FCMOV only supports a subset of X86 cond.
21666 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21667 SDValue Ops[] = { FalseOp, TrueOp,
21668 DAG.getConstant(CC, MVT::i8), Flags };
21669 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21672 // If this is a select between two integer constants, try to do some
21673 // optimizations. Note that the operands are ordered the opposite of SELECT
21675 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21676 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21677 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21678 // larger than FalseC (the false value).
21679 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21680 CC = X86::GetOppositeBranchCondition(CC);
21681 std::swap(TrueC, FalseC);
21682 std::swap(TrueOp, FalseOp);
21685 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21686 // This is efficient for any integer data type (including i8/i16) and
21688 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21689 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21690 DAG.getConstant(CC, MVT::i8), Cond);
21692 // Zero extend the condition if needed.
21693 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21695 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21696 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21697 DAG.getConstant(ShAmt, MVT::i8));
21698 if (N->getNumValues() == 2) // Dead flag value?
21699 return DCI.CombineTo(N, Cond, SDValue());
21703 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21704 // for any integer data type, including i8/i16.
21705 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21706 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21707 DAG.getConstant(CC, MVT::i8), Cond);
21709 // Zero extend the condition if needed.
21710 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21711 FalseC->getValueType(0), Cond);
21712 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21713 SDValue(FalseC, 0));
21715 if (N->getNumValues() == 2) // Dead flag value?
21716 return DCI.CombineTo(N, Cond, SDValue());
21720 // Optimize cases that will turn into an LEA instruction. This requires
21721 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21722 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21723 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21724 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21726 bool isFastMultiplier = false;
21728 switch ((unsigned char)Diff) {
21730 case 1: // result = add base, cond
21731 case 2: // result = lea base( , cond*2)
21732 case 3: // result = lea base(cond, cond*2)
21733 case 4: // result = lea base( , cond*4)
21734 case 5: // result = lea base(cond, cond*4)
21735 case 8: // result = lea base( , cond*8)
21736 case 9: // result = lea base(cond, cond*8)
21737 isFastMultiplier = true;
21742 if (isFastMultiplier) {
21743 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21744 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21745 DAG.getConstant(CC, MVT::i8), Cond);
21746 // Zero extend the condition if needed.
21747 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21749 // Scale the condition by the difference.
21751 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21752 DAG.getConstant(Diff, Cond.getValueType()));
21754 // Add the base if non-zero.
21755 if (FalseC->getAPIntValue() != 0)
21756 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21757 SDValue(FalseC, 0));
21758 if (N->getNumValues() == 2) // Dead flag value?
21759 return DCI.CombineTo(N, Cond, SDValue());
21766 // Handle these cases:
21767 // (select (x != c), e, c) -> select (x != c), e, x),
21768 // (select (x == c), c, e) -> select (x == c), x, e)
21769 // where the c is an integer constant, and the "select" is the combination
21770 // of CMOV and CMP.
21772 // The rationale for this change is that the conditional-move from a constant
21773 // needs two instructions, however, conditional-move from a register needs
21774 // only one instruction.
21776 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
21777 // some instruction-combining opportunities. This opt needs to be
21778 // postponed as late as possible.
21780 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
21781 // the DCI.xxxx conditions are provided to postpone the optimization as
21782 // late as possible.
21784 ConstantSDNode *CmpAgainst = nullptr;
21785 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
21786 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
21787 !isa<ConstantSDNode>(Cond.getOperand(0))) {
21789 if (CC == X86::COND_NE &&
21790 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
21791 CC = X86::GetOppositeBranchCondition(CC);
21792 std::swap(TrueOp, FalseOp);
21795 if (CC == X86::COND_E &&
21796 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
21797 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
21798 DAG.getConstant(CC, MVT::i8), Cond };
21799 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
21804 // Fold and/or of setcc's to double CMOV:
21805 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
21806 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
21808 // This combine lets us generate:
21809 // cmovcc1 (jcc1 if we don't have CMOV)
21815 // cmovne (jne if we don't have CMOV)
21816 // When we can't use the CMOV instruction, it might increase branch
21818 // When we can use CMOV, or when there is no mispredict, this improves
21819 // throughput and reduces register pressure.
21821 if (CC == X86::COND_NE) {
21823 X86::CondCode CC0, CC1;
21825 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
21827 std::swap(FalseOp, TrueOp);
21828 CC0 = X86::GetOppositeBranchCondition(CC0);
21829 CC1 = X86::GetOppositeBranchCondition(CC1);
21832 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, MVT::i8),
21834 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
21835 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, MVT::i8), Flags};
21836 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21837 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
21845 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
21846 const X86Subtarget *Subtarget) {
21847 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
21849 default: return SDValue();
21850 // SSE/AVX/AVX2 blend intrinsics.
21851 case Intrinsic::x86_avx2_pblendvb:
21852 // Don't try to simplify this intrinsic if we don't have AVX2.
21853 if (!Subtarget->hasAVX2())
21856 case Intrinsic::x86_avx_blendv_pd_256:
21857 case Intrinsic::x86_avx_blendv_ps_256:
21858 // Don't try to simplify this intrinsic if we don't have AVX.
21859 if (!Subtarget->hasAVX())
21862 case Intrinsic::x86_sse41_blendvps:
21863 case Intrinsic::x86_sse41_blendvpd:
21864 case Intrinsic::x86_sse41_pblendvb: {
21865 SDValue Op0 = N->getOperand(1);
21866 SDValue Op1 = N->getOperand(2);
21867 SDValue Mask = N->getOperand(3);
21869 // Don't try to simplify this intrinsic if we don't have SSE4.1.
21870 if (!Subtarget->hasSSE41())
21873 // fold (blend A, A, Mask) -> A
21876 // fold (blend A, B, allZeros) -> A
21877 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
21879 // fold (blend A, B, allOnes) -> B
21880 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
21883 // Simplify the case where the mask is a constant i32 value.
21884 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
21885 if (C->isNullValue())
21887 if (C->isAllOnesValue())
21894 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
21895 case Intrinsic::x86_sse2_psrai_w:
21896 case Intrinsic::x86_sse2_psrai_d:
21897 case Intrinsic::x86_avx2_psrai_w:
21898 case Intrinsic::x86_avx2_psrai_d:
21899 case Intrinsic::x86_sse2_psra_w:
21900 case Intrinsic::x86_sse2_psra_d:
21901 case Intrinsic::x86_avx2_psra_w:
21902 case Intrinsic::x86_avx2_psra_d: {
21903 SDValue Op0 = N->getOperand(1);
21904 SDValue Op1 = N->getOperand(2);
21905 EVT VT = Op0.getValueType();
21906 assert(VT.isVector() && "Expected a vector type!");
21908 if (isa<BuildVectorSDNode>(Op1))
21909 Op1 = Op1.getOperand(0);
21911 if (!isa<ConstantSDNode>(Op1))
21914 EVT SVT = VT.getVectorElementType();
21915 unsigned SVTBits = SVT.getSizeInBits();
21917 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
21918 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
21919 uint64_t ShAmt = C.getZExtValue();
21921 // Don't try to convert this shift into a ISD::SRA if the shift
21922 // count is bigger than or equal to the element size.
21923 if (ShAmt >= SVTBits)
21926 // Trivial case: if the shift count is zero, then fold this
21927 // into the first operand.
21931 // Replace this packed shift intrinsic with a target independent
21933 SDValue Splat = DAG.getConstant(C, VT);
21934 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
21939 /// PerformMulCombine - Optimize a single multiply with constant into two
21940 /// in order to implement it with two cheaper instructions, e.g.
21941 /// LEA + SHL, LEA + LEA.
21942 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
21943 TargetLowering::DAGCombinerInfo &DCI) {
21944 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
21947 EVT VT = N->getValueType(0);
21948 if (VT != MVT::i64 && VT != MVT::i32)
21951 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
21954 uint64_t MulAmt = C->getZExtValue();
21955 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
21958 uint64_t MulAmt1 = 0;
21959 uint64_t MulAmt2 = 0;
21960 if ((MulAmt % 9) == 0) {
21962 MulAmt2 = MulAmt / 9;
21963 } else if ((MulAmt % 5) == 0) {
21965 MulAmt2 = MulAmt / 5;
21966 } else if ((MulAmt % 3) == 0) {
21968 MulAmt2 = MulAmt / 3;
21971 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
21974 if (isPowerOf2_64(MulAmt2) &&
21975 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
21976 // If second multiplifer is pow2, issue it first. We want the multiply by
21977 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
21979 std::swap(MulAmt1, MulAmt2);
21982 if (isPowerOf2_64(MulAmt1))
21983 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
21984 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
21986 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
21987 DAG.getConstant(MulAmt1, VT));
21989 if (isPowerOf2_64(MulAmt2))
21990 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
21991 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
21993 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
21994 DAG.getConstant(MulAmt2, VT));
21996 // Do not add new nodes to DAG combiner worklist.
21997 DCI.CombineTo(N, NewMul, false);
22002 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22003 SDValue N0 = N->getOperand(0);
22004 SDValue N1 = N->getOperand(1);
22005 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22006 EVT VT = N0.getValueType();
22008 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22009 // since the result of setcc_c is all zero's or all ones.
22010 if (VT.isInteger() && !VT.isVector() &&
22011 N1C && N0.getOpcode() == ISD::AND &&
22012 N0.getOperand(1).getOpcode() == ISD::Constant) {
22013 SDValue N00 = N0.getOperand(0);
22014 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22015 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22016 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22017 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22018 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22019 APInt ShAmt = N1C->getAPIntValue();
22020 Mask = Mask.shl(ShAmt);
22022 return DAG.getNode(ISD::AND, SDLoc(N), VT,
22023 N00, DAG.getConstant(Mask, VT));
22027 // Hardware support for vector shifts is sparse which makes us scalarize the
22028 // vector operations in many cases. Also, on sandybridge ADD is faster than
22030 // (shl V, 1) -> add V,V
22031 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22032 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22033 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22034 // We shift all of the values by one. In many cases we do not have
22035 // hardware support for this operation. This is better expressed as an ADD
22037 if (N1SplatC->getZExtValue() == 1)
22038 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22044 /// \brief Returns a vector of 0s if the node in input is a vector logical
22045 /// shift by a constant amount which is known to be bigger than or equal
22046 /// to the vector element size in bits.
22047 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22048 const X86Subtarget *Subtarget) {
22049 EVT VT = N->getValueType(0);
22051 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22052 (!Subtarget->hasInt256() ||
22053 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22056 SDValue Amt = N->getOperand(1);
22058 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22059 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22060 APInt ShiftAmt = AmtSplat->getAPIntValue();
22061 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22063 // SSE2/AVX2 logical shifts always return a vector of 0s
22064 // if the shift amount is bigger than or equal to
22065 // the element size. The constant shift amount will be
22066 // encoded as a 8-bit immediate.
22067 if (ShiftAmt.trunc(8).uge(MaxAmount))
22068 return getZeroVector(VT, Subtarget, DAG, DL);
22074 /// PerformShiftCombine - Combine shifts.
22075 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22076 TargetLowering::DAGCombinerInfo &DCI,
22077 const X86Subtarget *Subtarget) {
22078 if (N->getOpcode() == ISD::SHL) {
22079 SDValue V = PerformSHLCombine(N, DAG);
22080 if (V.getNode()) return V;
22083 if (N->getOpcode() != ISD::SRA) {
22084 // Try to fold this logical shift into a zero vector.
22085 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22086 if (V.getNode()) return V;
22092 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22093 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22094 // and friends. Likewise for OR -> CMPNEQSS.
22095 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22096 TargetLowering::DAGCombinerInfo &DCI,
22097 const X86Subtarget *Subtarget) {
22100 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22101 // we're requiring SSE2 for both.
22102 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22103 SDValue N0 = N->getOperand(0);
22104 SDValue N1 = N->getOperand(1);
22105 SDValue CMP0 = N0->getOperand(1);
22106 SDValue CMP1 = N1->getOperand(1);
22109 // The SETCCs should both refer to the same CMP.
22110 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22113 SDValue CMP00 = CMP0->getOperand(0);
22114 SDValue CMP01 = CMP0->getOperand(1);
22115 EVT VT = CMP00.getValueType();
22117 if (VT == MVT::f32 || VT == MVT::f64) {
22118 bool ExpectingFlags = false;
22119 // Check for any users that want flags:
22120 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22121 !ExpectingFlags && UI != UE; ++UI)
22122 switch (UI->getOpcode()) {
22127 ExpectingFlags = true;
22129 case ISD::CopyToReg:
22130 case ISD::SIGN_EXTEND:
22131 case ISD::ZERO_EXTEND:
22132 case ISD::ANY_EXTEND:
22136 if (!ExpectingFlags) {
22137 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22138 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22140 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22141 X86::CondCode tmp = cc0;
22146 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22147 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22148 // FIXME: need symbolic constants for these magic numbers.
22149 // See X86ATTInstPrinter.cpp:printSSECC().
22150 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22151 if (Subtarget->hasAVX512()) {
22152 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22153 CMP01, DAG.getConstant(x86cc, MVT::i8));
22154 if (N->getValueType(0) != MVT::i1)
22155 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22159 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22160 CMP00.getValueType(), CMP00, CMP01,
22161 DAG.getConstant(x86cc, MVT::i8));
22163 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22164 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22166 if (is64BitFP && !Subtarget->is64Bit()) {
22167 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22168 // 64-bit integer, since that's not a legal type. Since
22169 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22170 // bits, but can do this little dance to extract the lowest 32 bits
22171 // and work with those going forward.
22172 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22174 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
22176 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22177 Vector32, DAG.getIntPtrConstant(0));
22181 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
22182 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22183 DAG.getConstant(1, IntVT));
22184 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
22185 return OneBitOfTruth;
22193 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22194 /// so it can be folded inside ANDNP.
22195 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22196 EVT VT = N->getValueType(0);
22198 // Match direct AllOnes for 128 and 256-bit vectors
22199 if (ISD::isBuildVectorAllOnes(N))
22202 // Look through a bit convert.
22203 if (N->getOpcode() == ISD::BITCAST)
22204 N = N->getOperand(0).getNode();
22206 // Sometimes the operand may come from a insert_subvector building a 256-bit
22208 if (VT.is256BitVector() &&
22209 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22210 SDValue V1 = N->getOperand(0);
22211 SDValue V2 = N->getOperand(1);
22213 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22214 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22215 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22216 ISD::isBuildVectorAllOnes(V2.getNode()))
22223 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22224 // register. In most cases we actually compare or select YMM-sized registers
22225 // and mixing the two types creates horrible code. This method optimizes
22226 // some of the transition sequences.
22227 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22228 TargetLowering::DAGCombinerInfo &DCI,
22229 const X86Subtarget *Subtarget) {
22230 EVT VT = N->getValueType(0);
22231 if (!VT.is256BitVector())
22234 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22235 N->getOpcode() == ISD::ZERO_EXTEND ||
22236 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22238 SDValue Narrow = N->getOperand(0);
22239 EVT NarrowVT = Narrow->getValueType(0);
22240 if (!NarrowVT.is128BitVector())
22243 if (Narrow->getOpcode() != ISD::XOR &&
22244 Narrow->getOpcode() != ISD::AND &&
22245 Narrow->getOpcode() != ISD::OR)
22248 SDValue N0 = Narrow->getOperand(0);
22249 SDValue N1 = Narrow->getOperand(1);
22252 // The Left side has to be a trunc.
22253 if (N0.getOpcode() != ISD::TRUNCATE)
22256 // The type of the truncated inputs.
22257 EVT WideVT = N0->getOperand(0)->getValueType(0);
22261 // The right side has to be a 'trunc' or a constant vector.
22262 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22263 ConstantSDNode *RHSConstSplat = nullptr;
22264 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22265 RHSConstSplat = RHSBV->getConstantSplatNode();
22266 if (!RHSTrunc && !RHSConstSplat)
22269 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22271 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22274 // Set N0 and N1 to hold the inputs to the new wide operation.
22275 N0 = N0->getOperand(0);
22276 if (RHSConstSplat) {
22277 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22278 SDValue(RHSConstSplat, 0));
22279 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22280 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22281 } else if (RHSTrunc) {
22282 N1 = N1->getOperand(0);
22285 // Generate the wide operation.
22286 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22287 unsigned Opcode = N->getOpcode();
22289 case ISD::ANY_EXTEND:
22291 case ISD::ZERO_EXTEND: {
22292 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22293 APInt Mask = APInt::getAllOnesValue(InBits);
22294 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22295 return DAG.getNode(ISD::AND, DL, VT,
22296 Op, DAG.getConstant(Mask, VT));
22298 case ISD::SIGN_EXTEND:
22299 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22300 Op, DAG.getValueType(NarrowVT));
22302 llvm_unreachable("Unexpected opcode");
22306 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
22307 TargetLowering::DAGCombinerInfo &DCI,
22308 const X86Subtarget *Subtarget) {
22309 SDValue N0 = N->getOperand(0);
22310 SDValue N1 = N->getOperand(1);
22313 // A vector zext_in_reg may be represented as a shuffle,
22314 // feeding into a bitcast (this represents anyext) feeding into
22315 // an and with a mask.
22316 // We'd like to try to combine that into a shuffle with zero
22317 // plus a bitcast, removing the and.
22318 if (N0.getOpcode() != ISD::BITCAST ||
22319 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
22322 // The other side of the AND should be a splat of 2^C, where C
22323 // is the number of bits in the source type.
22324 if (N1.getOpcode() == ISD::BITCAST)
22325 N1 = N1.getOperand(0);
22326 if (N1.getOpcode() != ISD::BUILD_VECTOR)
22328 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
22330 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
22331 EVT SrcType = Shuffle->getValueType(0);
22333 // We expect a single-source shuffle
22334 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
22337 unsigned SrcSize = SrcType.getScalarSizeInBits();
22339 APInt SplatValue, SplatUndef;
22340 unsigned SplatBitSize;
22342 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
22343 SplatBitSize, HasAnyUndefs))
22346 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
22347 // Make sure the splat matches the mask we expect
22348 if (SplatBitSize > ResSize ||
22349 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
22352 // Make sure the input and output size make sense
22353 if (SrcSize >= ResSize || ResSize % SrcSize)
22356 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
22357 // The number of u's between each two values depends on the ratio between
22358 // the source and dest type.
22359 unsigned ZextRatio = ResSize / SrcSize;
22360 bool IsZext = true;
22361 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
22362 if (i % ZextRatio) {
22363 if (Shuffle->getMaskElt(i) > 0) {
22369 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
22370 // Expected element number
22380 // Ok, perform the transformation - replace the shuffle with
22381 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
22382 // (instead of undef) where the k elements come from the zero vector.
22383 SmallVector<int, 8> Mask;
22384 unsigned NumElems = SrcType.getVectorNumElements();
22385 for (unsigned i = 0; i < NumElems; ++i)
22387 Mask.push_back(NumElems);
22389 Mask.push_back(i / ZextRatio);
22391 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
22392 Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
22393 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
22396 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22397 TargetLowering::DAGCombinerInfo &DCI,
22398 const X86Subtarget *Subtarget) {
22399 if (DCI.isBeforeLegalizeOps())
22402 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
22405 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
22408 EVT VT = N->getValueType(0);
22409 SDValue N0 = N->getOperand(0);
22410 SDValue N1 = N->getOperand(1);
22413 // Create BEXTR instructions
22414 // BEXTR is ((X >> imm) & (2**size-1))
22415 if (VT == MVT::i32 || VT == MVT::i64) {
22416 // Check for BEXTR.
22417 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22418 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22419 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22420 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22421 if (MaskNode && ShiftNode) {
22422 uint64_t Mask = MaskNode->getZExtValue();
22423 uint64_t Shift = ShiftNode->getZExtValue();
22424 if (isMask_64(Mask)) {
22425 uint64_t MaskSize = countPopulation(Mask);
22426 if (Shift + MaskSize <= VT.getSizeInBits())
22427 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22428 DAG.getConstant(Shift | (MaskSize << 8), VT));
22436 // Want to form ANDNP nodes:
22437 // 1) In the hopes of then easily combining them with OR and AND nodes
22438 // to form PBLEND/PSIGN.
22439 // 2) To match ANDN packed intrinsics
22440 if (VT != MVT::v2i64 && VT != MVT::v4i64)
22443 // Check LHS for vnot
22444 if (N0.getOpcode() == ISD::XOR &&
22445 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
22446 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
22447 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
22449 // Check RHS for vnot
22450 if (N1.getOpcode() == ISD::XOR &&
22451 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
22452 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
22453 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
22458 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
22459 TargetLowering::DAGCombinerInfo &DCI,
22460 const X86Subtarget *Subtarget) {
22461 if (DCI.isBeforeLegalizeOps())
22464 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22468 SDValue N0 = N->getOperand(0);
22469 SDValue N1 = N->getOperand(1);
22470 EVT VT = N->getValueType(0);
22472 // look for psign/blend
22473 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
22474 if (!Subtarget->hasSSSE3() ||
22475 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
22478 // Canonicalize pandn to RHS
22479 if (N0.getOpcode() == X86ISD::ANDNP)
22481 // or (and (m, y), (pandn m, x))
22482 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
22483 SDValue Mask = N1.getOperand(0);
22484 SDValue X = N1.getOperand(1);
22486 if (N0.getOperand(0) == Mask)
22487 Y = N0.getOperand(1);
22488 if (N0.getOperand(1) == Mask)
22489 Y = N0.getOperand(0);
22491 // Check to see if the mask appeared in both the AND and ANDNP and
22495 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
22496 // Look through mask bitcast.
22497 if (Mask.getOpcode() == ISD::BITCAST)
22498 Mask = Mask.getOperand(0);
22499 if (X.getOpcode() == ISD::BITCAST)
22500 X = X.getOperand(0);
22501 if (Y.getOpcode() == ISD::BITCAST)
22502 Y = Y.getOperand(0);
22504 EVT MaskVT = Mask.getValueType();
22506 // Validate that the Mask operand is a vector sra node.
22507 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
22508 // there is no psrai.b
22509 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
22510 unsigned SraAmt = ~0;
22511 if (Mask.getOpcode() == ISD::SRA) {
22512 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
22513 if (auto *AmtConst = AmtBV->getConstantSplatNode())
22514 SraAmt = AmtConst->getZExtValue();
22515 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
22516 SDValue SraC = Mask.getOperand(1);
22517 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
22519 if ((SraAmt + 1) != EltBits)
22524 // Now we know we at least have a plendvb with the mask val. See if
22525 // we can form a psignb/w/d.
22526 // psign = x.type == y.type == mask.type && y = sub(0, x);
22527 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
22528 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
22529 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
22530 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
22531 "Unsupported VT for PSIGN");
22532 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
22533 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22535 // PBLENDVB only available on SSE 4.1
22536 if (!Subtarget->hasSSE41())
22539 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
22541 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
22542 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
22543 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
22544 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
22545 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22549 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
22552 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
22553 MachineFunction &MF = DAG.getMachineFunction();
22555 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
22557 // SHLD/SHRD instructions have lower register pressure, but on some
22558 // platforms they have higher latency than the equivalent
22559 // series of shifts/or that would otherwise be generated.
22560 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
22561 // have higher latencies and we are not optimizing for size.
22562 if (!OptForSize && Subtarget->isSHLDSlow())
22565 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
22567 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
22569 if (!N0.hasOneUse() || !N1.hasOneUse())
22572 SDValue ShAmt0 = N0.getOperand(1);
22573 if (ShAmt0.getValueType() != MVT::i8)
22575 SDValue ShAmt1 = N1.getOperand(1);
22576 if (ShAmt1.getValueType() != MVT::i8)
22578 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
22579 ShAmt0 = ShAmt0.getOperand(0);
22580 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
22581 ShAmt1 = ShAmt1.getOperand(0);
22584 unsigned Opc = X86ISD::SHLD;
22585 SDValue Op0 = N0.getOperand(0);
22586 SDValue Op1 = N1.getOperand(0);
22587 if (ShAmt0.getOpcode() == ISD::SUB) {
22588 Opc = X86ISD::SHRD;
22589 std::swap(Op0, Op1);
22590 std::swap(ShAmt0, ShAmt1);
22593 unsigned Bits = VT.getSizeInBits();
22594 if (ShAmt1.getOpcode() == ISD::SUB) {
22595 SDValue Sum = ShAmt1.getOperand(0);
22596 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
22597 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
22598 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
22599 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
22600 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
22601 return DAG.getNode(Opc, DL, VT,
22603 DAG.getNode(ISD::TRUNCATE, DL,
22606 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
22607 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
22609 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
22610 return DAG.getNode(Opc, DL, VT,
22611 N0.getOperand(0), N1.getOperand(0),
22612 DAG.getNode(ISD::TRUNCATE, DL,
22619 // Generate NEG and CMOV for integer abs.
22620 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
22621 EVT VT = N->getValueType(0);
22623 // Since X86 does not have CMOV for 8-bit integer, we don't convert
22624 // 8-bit integer abs to NEG and CMOV.
22625 if (VT.isInteger() && VT.getSizeInBits() == 8)
22628 SDValue N0 = N->getOperand(0);
22629 SDValue N1 = N->getOperand(1);
22632 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
22633 // and change it to SUB and CMOV.
22634 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
22635 N0.getOpcode() == ISD::ADD &&
22636 N0.getOperand(1) == N1 &&
22637 N1.getOpcode() == ISD::SRA &&
22638 N1.getOperand(0) == N0.getOperand(0))
22639 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
22640 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
22641 // Generate SUB & CMOV.
22642 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
22643 DAG.getConstant(0, VT), N0.getOperand(0));
22645 SDValue Ops[] = { N0.getOperand(0), Neg,
22646 DAG.getConstant(X86::COND_GE, MVT::i8),
22647 SDValue(Neg.getNode(), 1) };
22648 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
22653 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
22654 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
22655 TargetLowering::DAGCombinerInfo &DCI,
22656 const X86Subtarget *Subtarget) {
22657 if (DCI.isBeforeLegalizeOps())
22660 if (Subtarget->hasCMov()) {
22661 SDValue RV = performIntegerAbsCombine(N, DAG);
22669 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
22670 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
22671 TargetLowering::DAGCombinerInfo &DCI,
22672 const X86Subtarget *Subtarget) {
22673 LoadSDNode *Ld = cast<LoadSDNode>(N);
22674 EVT RegVT = Ld->getValueType(0);
22675 EVT MemVT = Ld->getMemoryVT();
22677 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22679 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
22680 // into two 16-byte operations.
22681 ISD::LoadExtType Ext = Ld->getExtensionType();
22682 unsigned Alignment = Ld->getAlignment();
22683 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
22684 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22685 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
22686 unsigned NumElems = RegVT.getVectorNumElements();
22690 SDValue Ptr = Ld->getBasePtr();
22691 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
22693 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
22695 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22696 Ld->getPointerInfo(), Ld->isVolatile(),
22697 Ld->isNonTemporal(), Ld->isInvariant(),
22699 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22700 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22701 Ld->getPointerInfo(), Ld->isVolatile(),
22702 Ld->isNonTemporal(), Ld->isInvariant(),
22703 std::min(16U, Alignment));
22704 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22706 Load2.getValue(1));
22708 SDValue NewVec = DAG.getUNDEF(RegVT);
22709 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22710 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22711 return DCI.CombineTo(N, NewVec, TF, true);
22717 /// PerformMLOADCombine - Resolve extending loads
22718 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22719 TargetLowering::DAGCombinerInfo &DCI,
22720 const X86Subtarget *Subtarget) {
22721 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22722 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22725 EVT VT = Mld->getValueType(0);
22726 unsigned NumElems = VT.getVectorNumElements();
22727 EVT LdVT = Mld->getMemoryVT();
22730 assert(LdVT != VT && "Cannot extend to the same type");
22731 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
22732 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
22733 // From, To sizes and ElemCount must be pow of two
22734 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22735 "Unexpected size for extending masked load");
22737 unsigned SizeRatio = ToSz / FromSz;
22738 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
22740 // Create a type on which we perform the shuffle
22741 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22742 LdVT.getScalarType(), NumElems*SizeRatio);
22743 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22745 // Convert Src0 value
22746 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
22747 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
22748 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22749 for (unsigned i = 0; i != NumElems; ++i)
22750 ShuffleVec[i] = i * SizeRatio;
22752 // Can't shuffle using an illegal type.
22753 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22754 && "WideVecVT should be legal");
22755 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
22756 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
22758 // Prepare the new mask
22760 SDValue Mask = Mld->getMask();
22761 if (Mask.getValueType() == VT) {
22762 // Mask and original value have the same type
22763 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22764 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22765 for (unsigned i = 0; i != NumElems; ++i)
22766 ShuffleVec[i] = i * SizeRatio;
22767 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22768 ShuffleVec[i] = NumElems*SizeRatio;
22769 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22770 DAG.getConstant(0, WideVecVT),
22774 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22775 unsigned WidenNumElts = NumElems*SizeRatio;
22776 unsigned MaskNumElts = VT.getVectorNumElements();
22777 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22780 unsigned NumConcat = WidenNumElts / MaskNumElts;
22781 SmallVector<SDValue, 16> Ops(NumConcat);
22782 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22784 for (unsigned i = 1; i != NumConcat; ++i)
22787 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22790 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
22791 Mld->getBasePtr(), NewMask, WideSrc0,
22792 Mld->getMemoryVT(), Mld->getMemOperand(),
22794 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
22795 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
22798 /// PerformMSTORECombine - Resolve truncating stores
22799 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
22800 const X86Subtarget *Subtarget) {
22801 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
22802 if (!Mst->isTruncatingStore())
22805 EVT VT = Mst->getValue().getValueType();
22806 unsigned NumElems = VT.getVectorNumElements();
22807 EVT StVT = Mst->getMemoryVT();
22810 assert(StVT != VT && "Cannot truncate to the same type");
22811 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22812 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22814 // From, To sizes and ElemCount must be pow of two
22815 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22816 "Unexpected size for truncating masked store");
22817 // We are going to use the original vector elt for storing.
22818 // Accumulated smaller vector elements must be a multiple of the store size.
22819 assert (((NumElems * FromSz) % ToSz) == 0 &&
22820 "Unexpected ratio for truncating masked store");
22822 unsigned SizeRatio = FromSz / ToSz;
22823 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22825 // Create a type on which we perform the shuffle
22826 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22827 StVT.getScalarType(), NumElems*SizeRatio);
22829 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22831 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
22832 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22833 for (unsigned i = 0; i != NumElems; ++i)
22834 ShuffleVec[i] = i * SizeRatio;
22836 // Can't shuffle using an illegal type.
22837 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22838 && "WideVecVT should be legal");
22840 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22841 DAG.getUNDEF(WideVecVT),
22845 SDValue Mask = Mst->getMask();
22846 if (Mask.getValueType() == VT) {
22847 // Mask and original value have the same type
22848 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22849 for (unsigned i = 0; i != NumElems; ++i)
22850 ShuffleVec[i] = i * SizeRatio;
22851 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22852 ShuffleVec[i] = NumElems*SizeRatio;
22853 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22854 DAG.getConstant(0, WideVecVT),
22858 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22859 unsigned WidenNumElts = NumElems*SizeRatio;
22860 unsigned MaskNumElts = VT.getVectorNumElements();
22861 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22864 unsigned NumConcat = WidenNumElts / MaskNumElts;
22865 SmallVector<SDValue, 16> Ops(NumConcat);
22866 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22868 for (unsigned i = 1; i != NumConcat; ++i)
22871 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22874 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
22875 NewMask, StVT, Mst->getMemOperand(), false);
22877 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
22878 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
22879 const X86Subtarget *Subtarget) {
22880 StoreSDNode *St = cast<StoreSDNode>(N);
22881 EVT VT = St->getValue().getValueType();
22882 EVT StVT = St->getMemoryVT();
22884 SDValue StoredVal = St->getOperand(1);
22885 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22887 // If we are saving a concatenation of two XMM registers and 32-byte stores
22888 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
22889 unsigned Alignment = St->getAlignment();
22890 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
22891 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22892 StVT == VT && !IsAligned) {
22893 unsigned NumElems = VT.getVectorNumElements();
22897 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
22898 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
22900 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
22901 SDValue Ptr0 = St->getBasePtr();
22902 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
22904 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
22905 St->getPointerInfo(), St->isVolatile(),
22906 St->isNonTemporal(), Alignment);
22907 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
22908 St->getPointerInfo(), St->isVolatile(),
22909 St->isNonTemporal(),
22910 std::min(16U, Alignment));
22911 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
22914 // Optimize trunc store (of multiple scalars) to shuffle and store.
22915 // First, pack all of the elements in one place. Next, store to memory
22916 // in fewer chunks.
22917 if (St->isTruncatingStore() && VT.isVector()) {
22918 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22919 unsigned NumElems = VT.getVectorNumElements();
22920 assert(StVT != VT && "Cannot truncate to the same type");
22921 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22922 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22924 // From, To sizes and ElemCount must be pow of two
22925 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
22926 // We are going to use the original vector elt for storing.
22927 // Accumulated smaller vector elements must be a multiple of the store size.
22928 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
22930 unsigned SizeRatio = FromSz / ToSz;
22932 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22934 // Create a type on which we perform the shuffle
22935 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22936 StVT.getScalarType(), NumElems*SizeRatio);
22938 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22940 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
22941 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
22942 for (unsigned i = 0; i != NumElems; ++i)
22943 ShuffleVec[i] = i * SizeRatio;
22945 // Can't shuffle using an illegal type.
22946 if (!TLI.isTypeLegal(WideVecVT))
22949 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22950 DAG.getUNDEF(WideVecVT),
22952 // At this point all of the data is stored at the bottom of the
22953 // register. We now need to save it to mem.
22955 // Find the largest store unit
22956 MVT StoreType = MVT::i8;
22957 for (MVT Tp : MVT::integer_valuetypes()) {
22958 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
22962 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
22963 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
22964 (64 <= NumElems * ToSz))
22965 StoreType = MVT::f64;
22967 // Bitcast the original vector into a vector of store-size units
22968 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
22969 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
22970 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
22971 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
22972 SmallVector<SDValue, 8> Chains;
22973 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
22974 TLI.getPointerTy());
22975 SDValue Ptr = St->getBasePtr();
22977 // Perform one or more big stores into memory.
22978 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
22979 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
22980 StoreType, ShuffWide,
22981 DAG.getIntPtrConstant(i));
22982 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
22983 St->getPointerInfo(), St->isVolatile(),
22984 St->isNonTemporal(), St->getAlignment());
22985 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22986 Chains.push_back(Ch);
22989 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
22992 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
22993 // the FP state in cases where an emms may be missing.
22994 // A preferable solution to the general problem is to figure out the right
22995 // places to insert EMMS. This qualifies as a quick hack.
22997 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
22998 if (VT.getSizeInBits() != 64)
23001 const Function *F = DAG.getMachineFunction().getFunction();
23002 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
23003 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
23004 && Subtarget->hasSSE2();
23005 if ((VT.isVector() ||
23006 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23007 isa<LoadSDNode>(St->getValue()) &&
23008 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23009 St->getChain().hasOneUse() && !St->isVolatile()) {
23010 SDNode* LdVal = St->getValue().getNode();
23011 LoadSDNode *Ld = nullptr;
23012 int TokenFactorIndex = -1;
23013 SmallVector<SDValue, 8> Ops;
23014 SDNode* ChainVal = St->getChain().getNode();
23015 // Must be a store of a load. We currently handle two cases: the load
23016 // is a direct child, and it's under an intervening TokenFactor. It is
23017 // possible to dig deeper under nested TokenFactors.
23018 if (ChainVal == LdVal)
23019 Ld = cast<LoadSDNode>(St->getChain());
23020 else if (St->getValue().hasOneUse() &&
23021 ChainVal->getOpcode() == ISD::TokenFactor) {
23022 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23023 if (ChainVal->getOperand(i).getNode() == LdVal) {
23024 TokenFactorIndex = i;
23025 Ld = cast<LoadSDNode>(St->getValue());
23027 Ops.push_back(ChainVal->getOperand(i));
23031 if (!Ld || !ISD::isNormalLoad(Ld))
23034 // If this is not the MMX case, i.e. we are just turning i64 load/store
23035 // into f64 load/store, avoid the transformation if there are multiple
23036 // uses of the loaded value.
23037 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23042 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23043 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23045 if (Subtarget->is64Bit() || F64IsLegal) {
23046 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23047 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23048 Ld->getPointerInfo(), Ld->isVolatile(),
23049 Ld->isNonTemporal(), Ld->isInvariant(),
23050 Ld->getAlignment());
23051 SDValue NewChain = NewLd.getValue(1);
23052 if (TokenFactorIndex != -1) {
23053 Ops.push_back(NewChain);
23054 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23056 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23057 St->getPointerInfo(),
23058 St->isVolatile(), St->isNonTemporal(),
23059 St->getAlignment());
23062 // Otherwise, lower to two pairs of 32-bit loads / stores.
23063 SDValue LoAddr = Ld->getBasePtr();
23064 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23065 DAG.getConstant(4, MVT::i32));
23067 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23068 Ld->getPointerInfo(),
23069 Ld->isVolatile(), Ld->isNonTemporal(),
23070 Ld->isInvariant(), Ld->getAlignment());
23071 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23072 Ld->getPointerInfo().getWithOffset(4),
23073 Ld->isVolatile(), Ld->isNonTemporal(),
23075 MinAlign(Ld->getAlignment(), 4));
23077 SDValue NewChain = LoLd.getValue(1);
23078 if (TokenFactorIndex != -1) {
23079 Ops.push_back(LoLd);
23080 Ops.push_back(HiLd);
23081 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23084 LoAddr = St->getBasePtr();
23085 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23086 DAG.getConstant(4, MVT::i32));
23088 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23089 St->getPointerInfo(),
23090 St->isVolatile(), St->isNonTemporal(),
23091 St->getAlignment());
23092 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23093 St->getPointerInfo().getWithOffset(4),
23095 St->isNonTemporal(),
23096 MinAlign(St->getAlignment(), 4));
23097 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23100 // This is similar to the above case, but here we handle a scalar 64-bit
23101 // integer store that is extracted from a vector on a 32-bit target.
23102 // If we have SSE2, then we can treat it like a floating-point double
23103 // to get past legalization. The execution dependencies fixup pass will
23104 // choose the optimal machine instruction for the store if this really is
23105 // an integer or v2f32 rather than an f64.
23106 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
23107 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
23108 SDValue OldExtract = St->getOperand(1);
23109 SDValue ExtOp0 = OldExtract.getOperand(0);
23110 unsigned VecSize = ExtOp0.getValueSizeInBits();
23111 MVT VecVT = MVT::getVectorVT(MVT::f64, VecSize / 64);
23112 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtOp0);
23113 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
23114 BitCast, OldExtract.getOperand(1));
23115 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
23116 St->getPointerInfo(), St->isVolatile(),
23117 St->isNonTemporal(), St->getAlignment());
23123 /// Return 'true' if this vector operation is "horizontal"
23124 /// and return the operands for the horizontal operation in LHS and RHS. A
23125 /// horizontal operation performs the binary operation on successive elements
23126 /// of its first operand, then on successive elements of its second operand,
23127 /// returning the resulting values in a vector. For example, if
23128 /// A = < float a0, float a1, float a2, float a3 >
23130 /// B = < float b0, float b1, float b2, float b3 >
23131 /// then the result of doing a horizontal operation on A and B is
23132 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23133 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23134 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23135 /// set to A, RHS to B, and the routine returns 'true'.
23136 /// Note that the binary operation should have the property that if one of the
23137 /// operands is UNDEF then the result is UNDEF.
23138 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23139 // Look for the following pattern: if
23140 // A = < float a0, float a1, float a2, float a3 >
23141 // B = < float b0, float b1, float b2, float b3 >
23143 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23144 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23145 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23146 // which is A horizontal-op B.
23148 // At least one of the operands should be a vector shuffle.
23149 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23150 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23153 MVT VT = LHS.getSimpleValueType();
23155 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23156 "Unsupported vector type for horizontal add/sub");
23158 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23159 // operate independently on 128-bit lanes.
23160 unsigned NumElts = VT.getVectorNumElements();
23161 unsigned NumLanes = VT.getSizeInBits()/128;
23162 unsigned NumLaneElts = NumElts / NumLanes;
23163 assert((NumLaneElts % 2 == 0) &&
23164 "Vector type should have an even number of elements in each lane");
23165 unsigned HalfLaneElts = NumLaneElts/2;
23167 // View LHS in the form
23168 // LHS = VECTOR_SHUFFLE A, B, LMask
23169 // If LHS is not a shuffle then pretend it is the shuffle
23170 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23171 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23174 SmallVector<int, 16> LMask(NumElts);
23175 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23176 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23177 A = LHS.getOperand(0);
23178 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23179 B = LHS.getOperand(1);
23180 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23181 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23183 if (LHS.getOpcode() != ISD::UNDEF)
23185 for (unsigned i = 0; i != NumElts; ++i)
23189 // Likewise, view RHS in the form
23190 // RHS = VECTOR_SHUFFLE C, D, RMask
23192 SmallVector<int, 16> RMask(NumElts);
23193 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23194 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23195 C = RHS.getOperand(0);
23196 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23197 D = RHS.getOperand(1);
23198 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23199 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23201 if (RHS.getOpcode() != ISD::UNDEF)
23203 for (unsigned i = 0; i != NumElts; ++i)
23207 // Check that the shuffles are both shuffling the same vectors.
23208 if (!(A == C && B == D) && !(A == D && B == C))
23211 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23212 if (!A.getNode() && !B.getNode())
23215 // If A and B occur in reverse order in RHS, then "swap" them (which means
23216 // rewriting the mask).
23218 ShuffleVectorSDNode::commuteMask(RMask);
23220 // At this point LHS and RHS are equivalent to
23221 // LHS = VECTOR_SHUFFLE A, B, LMask
23222 // RHS = VECTOR_SHUFFLE A, B, RMask
23223 // Check that the masks correspond to performing a horizontal operation.
23224 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23225 for (unsigned i = 0; i != NumLaneElts; ++i) {
23226 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23228 // Ignore any UNDEF components.
23229 if (LIdx < 0 || RIdx < 0 ||
23230 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23231 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23234 // Check that successive elements are being operated on. If not, this is
23235 // not a horizontal operation.
23236 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23237 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23238 if (!(LIdx == Index && RIdx == Index + 1) &&
23239 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23244 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23245 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23249 /// Do target-specific dag combines on floating point adds.
23250 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23251 const X86Subtarget *Subtarget) {
23252 EVT VT = N->getValueType(0);
23253 SDValue LHS = N->getOperand(0);
23254 SDValue RHS = N->getOperand(1);
23256 // Try to synthesize horizontal adds from adds of shuffles.
23257 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23258 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23259 isHorizontalBinOp(LHS, RHS, true))
23260 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23264 /// Do target-specific dag combines on floating point subs.
23265 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23266 const X86Subtarget *Subtarget) {
23267 EVT VT = N->getValueType(0);
23268 SDValue LHS = N->getOperand(0);
23269 SDValue RHS = N->getOperand(1);
23271 // Try to synthesize horizontal subs from subs of shuffles.
23272 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23273 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23274 isHorizontalBinOp(LHS, RHS, false))
23275 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23279 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
23280 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23281 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23283 // F[X]OR(0.0, x) -> x
23284 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23285 if (C->getValueAPF().isPosZero())
23286 return N->getOperand(1);
23288 // F[X]OR(x, 0.0) -> x
23289 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23290 if (C->getValueAPF().isPosZero())
23291 return N->getOperand(0);
23295 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
23296 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23297 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23299 // Only perform optimizations if UnsafeMath is used.
23300 if (!DAG.getTarget().Options.UnsafeFPMath)
23303 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23304 // into FMINC and FMAXC, which are Commutative operations.
23305 unsigned NewOp = 0;
23306 switch (N->getOpcode()) {
23307 default: llvm_unreachable("unknown opcode");
23308 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23309 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23312 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23313 N->getOperand(0), N->getOperand(1));
23316 /// Do target-specific dag combines on X86ISD::FAND nodes.
23317 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23318 // FAND(0.0, x) -> 0.0
23319 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23320 if (C->getValueAPF().isPosZero())
23321 return N->getOperand(0);
23323 // FAND(x, 0.0) -> 0.0
23324 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23325 if (C->getValueAPF().isPosZero())
23326 return N->getOperand(1);
23331 /// Do target-specific dag combines on X86ISD::FANDN nodes
23332 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23333 // FANDN(0.0, x) -> x
23334 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23335 if (C->getValueAPF().isPosZero())
23336 return N->getOperand(1);
23338 // FANDN(x, 0.0) -> 0.0
23339 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23340 if (C->getValueAPF().isPosZero())
23341 return N->getOperand(1);
23346 static SDValue PerformBTCombine(SDNode *N,
23348 TargetLowering::DAGCombinerInfo &DCI) {
23349 // BT ignores high bits in the bit index operand.
23350 SDValue Op1 = N->getOperand(1);
23351 if (Op1.hasOneUse()) {
23352 unsigned BitWidth = Op1.getValueSizeInBits();
23353 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23354 APInt KnownZero, KnownOne;
23355 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23356 !DCI.isBeforeLegalizeOps());
23357 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23358 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23359 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23360 DCI.CommitTargetLoweringOpt(TLO);
23365 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23366 SDValue Op = N->getOperand(0);
23367 if (Op.getOpcode() == ISD::BITCAST)
23368 Op = Op.getOperand(0);
23369 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23370 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23371 VT.getVectorElementType().getSizeInBits() ==
23372 OpVT.getVectorElementType().getSizeInBits()) {
23373 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23378 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23379 const X86Subtarget *Subtarget) {
23380 EVT VT = N->getValueType(0);
23381 if (!VT.isVector())
23384 SDValue N0 = N->getOperand(0);
23385 SDValue N1 = N->getOperand(1);
23386 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23389 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23390 // both SSE and AVX2 since there is no sign-extended shift right
23391 // operation on a vector with 64-bit elements.
23392 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23393 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23394 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23395 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23396 SDValue N00 = N0.getOperand(0);
23398 // EXTLOAD has a better solution on AVX2,
23399 // it may be replaced with X86ISD::VSEXT node.
23400 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23401 if (!ISD::isNormalLoad(N00.getNode()))
23404 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23405 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23407 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23413 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23414 TargetLowering::DAGCombinerInfo &DCI,
23415 const X86Subtarget *Subtarget) {
23416 SDValue N0 = N->getOperand(0);
23417 EVT VT = N->getValueType(0);
23419 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
23420 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
23421 // This exposes the sext to the sdivrem lowering, so that it directly extends
23422 // from AH (which we otherwise need to do contortions to access).
23423 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
23424 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
23426 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23427 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
23428 N0.getOperand(0), N0.getOperand(1));
23429 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23430 return R.getValue(1);
23433 if (!DCI.isBeforeLegalizeOps())
23436 if (!Subtarget->hasFp256())
23439 if (VT.isVector() && VT.getSizeInBits() == 256) {
23440 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23448 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
23449 const X86Subtarget* Subtarget) {
23451 EVT VT = N->getValueType(0);
23453 // Let legalize expand this if it isn't a legal type yet.
23454 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
23457 EVT ScalarVT = VT.getScalarType();
23458 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
23459 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
23462 SDValue A = N->getOperand(0);
23463 SDValue B = N->getOperand(1);
23464 SDValue C = N->getOperand(2);
23466 bool NegA = (A.getOpcode() == ISD::FNEG);
23467 bool NegB = (B.getOpcode() == ISD::FNEG);
23468 bool NegC = (C.getOpcode() == ISD::FNEG);
23470 // Negative multiplication when NegA xor NegB
23471 bool NegMul = (NegA != NegB);
23473 A = A.getOperand(0);
23475 B = B.getOperand(0);
23477 C = C.getOperand(0);
23481 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
23483 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
23485 return DAG.getNode(Opcode, dl, VT, A, B, C);
23488 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
23489 TargetLowering::DAGCombinerInfo &DCI,
23490 const X86Subtarget *Subtarget) {
23491 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
23492 // (and (i32 x86isd::setcc_carry), 1)
23493 // This eliminates the zext. This transformation is necessary because
23494 // ISD::SETCC is always legalized to i8.
23496 SDValue N0 = N->getOperand(0);
23497 EVT VT = N->getValueType(0);
23499 if (N0.getOpcode() == ISD::AND &&
23501 N0.getOperand(0).hasOneUse()) {
23502 SDValue N00 = N0.getOperand(0);
23503 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23504 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23505 if (!C || C->getZExtValue() != 1)
23507 return DAG.getNode(ISD::AND, dl, VT,
23508 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23509 N00.getOperand(0), N00.getOperand(1)),
23510 DAG.getConstant(1, VT));
23514 if (N0.getOpcode() == ISD::TRUNCATE &&
23516 N0.getOperand(0).hasOneUse()) {
23517 SDValue N00 = N0.getOperand(0);
23518 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23519 return DAG.getNode(ISD::AND, dl, VT,
23520 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23521 N00.getOperand(0), N00.getOperand(1)),
23522 DAG.getConstant(1, VT));
23525 if (VT.is256BitVector()) {
23526 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23531 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
23532 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
23533 // This exposes the zext to the udivrem lowering, so that it directly extends
23534 // from AH (which we otherwise need to do contortions to access).
23535 if (N0.getOpcode() == ISD::UDIVREM &&
23536 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
23537 (VT == MVT::i32 || VT == MVT::i64)) {
23538 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23539 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
23540 N0.getOperand(0), N0.getOperand(1));
23541 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23542 return R.getValue(1);
23548 // Optimize x == -y --> x+y == 0
23549 // x != -y --> x+y != 0
23550 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
23551 const X86Subtarget* Subtarget) {
23552 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
23553 SDValue LHS = N->getOperand(0);
23554 SDValue RHS = N->getOperand(1);
23555 EVT VT = N->getValueType(0);
23558 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
23559 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
23560 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
23561 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N), LHS.getValueType(), RHS,
23562 LHS.getOperand(1));
23563 return DAG.getSetCC(SDLoc(N), N->getValueType(0), addV,
23564 DAG.getConstant(0, addV.getValueType()), CC);
23566 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
23567 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
23568 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
23569 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N), RHS.getValueType(), LHS,
23570 RHS.getOperand(1));
23571 return DAG.getSetCC(SDLoc(N), N->getValueType(0), addV,
23572 DAG.getConstant(0, addV.getValueType()), CC);
23575 if (VT.getScalarType() == MVT::i1 &&
23576 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
23578 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23579 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23580 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23582 if (!IsSEXT0 || !IsVZero1) {
23583 // Swap the operands and update the condition code.
23584 std::swap(LHS, RHS);
23585 CC = ISD::getSetCCSwappedOperands(CC);
23587 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23588 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23589 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23592 if (IsSEXT0 && IsVZero1) {
23593 assert(VT == LHS.getOperand(0).getValueType() &&
23594 "Uexpected operand type");
23595 if (CC == ISD::SETGT)
23596 return DAG.getConstant(0, VT);
23597 if (CC == ISD::SETLE)
23598 return DAG.getConstant(1, VT);
23599 if (CC == ISD::SETEQ || CC == ISD::SETGE)
23600 return DAG.getNOT(DL, LHS.getOperand(0), VT);
23602 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
23603 "Unexpected condition code!");
23604 return LHS.getOperand(0);
23611 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
23612 SelectionDAG &DAG) {
23614 MVT VT = Load->getSimpleValueType(0);
23615 MVT EVT = VT.getVectorElementType();
23616 SDValue Addr = Load->getOperand(1);
23617 SDValue NewAddr = DAG.getNode(
23618 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
23619 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
23622 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
23623 DAG.getMachineFunction().getMachineMemOperand(
23624 Load->getMemOperand(), 0, EVT.getStoreSize()));
23628 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
23629 const X86Subtarget *Subtarget) {
23631 MVT VT = N->getOperand(1)->getSimpleValueType(0);
23632 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
23633 "X86insertps is only defined for v4x32");
23635 SDValue Ld = N->getOperand(1);
23636 if (MayFoldLoad(Ld)) {
23637 // Extract the countS bits from the immediate so we can get the proper
23638 // address when narrowing the vector load to a specific element.
23639 // When the second source op is a memory address, insertps doesn't use
23640 // countS and just gets an f32 from that address.
23641 unsigned DestIndex =
23642 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
23644 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
23646 // Create this as a scalar to vector to match the instruction pattern.
23647 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
23648 // countS bits are ignored when loading from memory on insertps, which
23649 // means we don't need to explicitly set them to 0.
23650 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
23651 LoadScalarToVector, N->getOperand(2));
23656 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
23657 SDValue V0 = N->getOperand(0);
23658 SDValue V1 = N->getOperand(1);
23660 EVT VT = N->getValueType(0);
23662 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23663 // operands and changing the mask to 1. This saves us a bunch of
23664 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23665 // x86InstrInfo knows how to commute this back after instruction selection
23666 // if it would help register allocation.
23668 // TODO: If optimizing for size or a processor that doesn't suffer from
23669 // partial register update stalls, this should be transformed into a MOVSD
23670 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23672 if (VT == MVT::v2f64)
23673 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23674 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23675 SDValue NewMask = DAG.getConstant(1, MVT::i8);
23676 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23682 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
23683 // as "sbb reg,reg", since it can be extended without zext and produces
23684 // an all-ones bit which is more useful than 0/1 in some cases.
23685 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
23688 return DAG.getNode(ISD::AND, DL, VT,
23689 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23690 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
23691 DAG.getConstant(1, VT));
23692 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
23693 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
23694 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23695 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
23698 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
23699 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
23700 TargetLowering::DAGCombinerInfo &DCI,
23701 const X86Subtarget *Subtarget) {
23703 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
23704 SDValue EFLAGS = N->getOperand(1);
23706 if (CC == X86::COND_A) {
23707 // Try to convert COND_A into COND_B in an attempt to facilitate
23708 // materializing "setb reg".
23710 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
23711 // cannot take an immediate as its first operand.
23713 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
23714 EFLAGS.getValueType().isInteger() &&
23715 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
23716 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23717 EFLAGS.getNode()->getVTList(),
23718 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23719 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23720 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
23724 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
23725 // a zext and produces an all-ones bit which is more useful than 0/1 in some
23727 if (CC == X86::COND_B)
23728 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23732 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23733 if (Flags.getNode()) {
23734 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23735 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23741 // Optimize branch condition evaluation.
23743 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23744 TargetLowering::DAGCombinerInfo &DCI,
23745 const X86Subtarget *Subtarget) {
23747 SDValue Chain = N->getOperand(0);
23748 SDValue Dest = N->getOperand(1);
23749 SDValue EFLAGS = N->getOperand(3);
23750 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23754 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23755 if (Flags.getNode()) {
23756 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23757 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23764 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23765 SelectionDAG &DAG) {
23766 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23767 // optimize away operation when it's from a constant.
23769 // The general transformation is:
23770 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23771 // AND(VECTOR_CMP(x,y), constant2)
23772 // constant2 = UNARYOP(constant)
23774 // Early exit if this isn't a vector operation, the operand of the
23775 // unary operation isn't a bitwise AND, or if the sizes of the operations
23776 // aren't the same.
23777 EVT VT = N->getValueType(0);
23778 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23779 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23780 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23783 // Now check that the other operand of the AND is a constant. We could
23784 // make the transformation for non-constant splats as well, but it's unclear
23785 // that would be a benefit as it would not eliminate any operations, just
23786 // perform one more step in scalar code before moving to the vector unit.
23787 if (BuildVectorSDNode *BV =
23788 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23789 // Bail out if the vector isn't a constant.
23790 if (!BV->isConstant())
23793 // Everything checks out. Build up the new and improved node.
23795 EVT IntVT = BV->getValueType(0);
23796 // Create a new constant of the appropriate type for the transformed
23798 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23799 // The AND node needs bitcasts to/from an integer vector type around it.
23800 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
23801 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
23802 N->getOperand(0)->getOperand(0), MaskConst);
23803 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
23810 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
23811 const X86Subtarget *Subtarget) {
23812 // First try to optimize away the conversion entirely when it's
23813 // conditionally from a constant. Vectors only.
23814 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
23815 if (Res != SDValue())
23818 // Now move on to more general possibilities.
23819 SDValue Op0 = N->getOperand(0);
23820 EVT InVT = Op0->getValueType(0);
23822 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
23823 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
23825 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
23826 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
23827 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
23830 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
23831 // a 32-bit target where SSE doesn't support i64->FP operations.
23832 if (Op0.getOpcode() == ISD::LOAD) {
23833 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
23834 EVT VT = Ld->getValueType(0);
23835 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
23836 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
23837 !Subtarget->is64Bit() && VT == MVT::i64) {
23838 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
23839 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
23840 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
23847 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
23848 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
23849 X86TargetLowering::DAGCombinerInfo &DCI) {
23850 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
23851 // the result is either zero or one (depending on the input carry bit).
23852 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
23853 if (X86::isZeroNode(N->getOperand(0)) &&
23854 X86::isZeroNode(N->getOperand(1)) &&
23855 // We don't have a good way to replace an EFLAGS use, so only do this when
23857 SDValue(N, 1).use_empty()) {
23859 EVT VT = N->getValueType(0);
23860 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
23861 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
23862 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
23863 DAG.getConstant(X86::COND_B,MVT::i8),
23865 DAG.getConstant(1, VT));
23866 return DCI.CombineTo(N, Res1, CarryOut);
23872 // fold (add Y, (sete X, 0)) -> adc 0, Y
23873 // (add Y, (setne X, 0)) -> sbb -1, Y
23874 // (sub (sete X, 0), Y) -> sbb 0, Y
23875 // (sub (setne X, 0), Y) -> adc -1, Y
23876 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
23879 // Look through ZExts.
23880 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
23881 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
23884 SDValue SetCC = Ext.getOperand(0);
23885 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
23888 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
23889 if (CC != X86::COND_E && CC != X86::COND_NE)
23892 SDValue Cmp = SetCC.getOperand(1);
23893 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
23894 !X86::isZeroNode(Cmp.getOperand(1)) ||
23895 !Cmp.getOperand(0).getValueType().isInteger())
23898 SDValue CmpOp0 = Cmp.getOperand(0);
23899 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
23900 DAG.getConstant(1, CmpOp0.getValueType()));
23902 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
23903 if (CC == X86::COND_NE)
23904 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
23905 DL, OtherVal.getValueType(), OtherVal,
23906 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
23907 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
23908 DL, OtherVal.getValueType(), OtherVal,
23909 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
23912 /// PerformADDCombine - Do target-specific dag combines on integer adds.
23913 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
23914 const X86Subtarget *Subtarget) {
23915 EVT VT = N->getValueType(0);
23916 SDValue Op0 = N->getOperand(0);
23917 SDValue Op1 = N->getOperand(1);
23919 // Try to synthesize horizontal adds from adds of shuffles.
23920 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23921 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23922 isHorizontalBinOp(Op0, Op1, true))
23923 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
23925 return OptimizeConditionalInDecrement(N, DAG);
23928 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
23929 const X86Subtarget *Subtarget) {
23930 SDValue Op0 = N->getOperand(0);
23931 SDValue Op1 = N->getOperand(1);
23933 // X86 can't encode an immediate LHS of a sub. See if we can push the
23934 // negation into a preceding instruction.
23935 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
23936 // If the RHS of the sub is a XOR with one use and a constant, invert the
23937 // immediate. Then add one to the LHS of the sub so we can turn
23938 // X-Y -> X+~Y+1, saving one register.
23939 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
23940 isa<ConstantSDNode>(Op1.getOperand(1))) {
23941 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
23942 EVT VT = Op0.getValueType();
23943 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
23945 DAG.getConstant(~XorC, VT));
23946 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
23947 DAG.getConstant(C->getAPIntValue()+1, VT));
23951 // Try to synthesize horizontal adds from adds of shuffles.
23952 EVT VT = N->getValueType(0);
23953 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23954 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23955 isHorizontalBinOp(Op0, Op1, true))
23956 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
23958 return OptimizeConditionalInDecrement(N, DAG);
23961 /// performVZEXTCombine - Performs build vector combines
23962 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
23963 TargetLowering::DAGCombinerInfo &DCI,
23964 const X86Subtarget *Subtarget) {
23966 MVT VT = N->getSimpleValueType(0);
23967 SDValue Op = N->getOperand(0);
23968 MVT OpVT = Op.getSimpleValueType();
23969 MVT OpEltVT = OpVT.getVectorElementType();
23970 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
23972 // (vzext (bitcast (vzext (x)) -> (vzext x)
23974 while (V.getOpcode() == ISD::BITCAST)
23975 V = V.getOperand(0);
23977 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
23978 MVT InnerVT = V.getSimpleValueType();
23979 MVT InnerEltVT = InnerVT.getVectorElementType();
23981 // If the element sizes match exactly, we can just do one larger vzext. This
23982 // is always an exact type match as vzext operates on integer types.
23983 if (OpEltVT == InnerEltVT) {
23984 assert(OpVT == InnerVT && "Types must match for vzext!");
23985 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
23988 // The only other way we can combine them is if only a single element of the
23989 // inner vzext is used in the input to the outer vzext.
23990 if (InnerEltVT.getSizeInBits() < InputBits)
23993 // In this case, the inner vzext is completely dead because we're going to
23994 // only look at bits inside of the low element. Just do the outer vzext on
23995 // a bitcast of the input to the inner.
23996 return DAG.getNode(X86ISD::VZEXT, DL, VT,
23997 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
24000 // Check if we can bypass extracting and re-inserting an element of an input
24001 // vector. Essentialy:
24002 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
24003 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
24004 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
24005 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
24006 SDValue ExtractedV = V.getOperand(0);
24007 SDValue OrigV = ExtractedV.getOperand(0);
24008 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
24009 if (ExtractIdx->getZExtValue() == 0) {
24010 MVT OrigVT = OrigV.getSimpleValueType();
24011 // Extract a subvector if necessary...
24012 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
24013 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
24014 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
24015 OrigVT.getVectorNumElements() / Ratio);
24016 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
24017 DAG.getIntPtrConstant(0));
24019 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
24020 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
24027 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24028 DAGCombinerInfo &DCI) const {
24029 SelectionDAG &DAG = DCI.DAG;
24030 switch (N->getOpcode()) {
24032 case ISD::EXTRACT_VECTOR_ELT:
24033 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24036 case X86ISD::SHRUNKBLEND:
24037 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24038 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
24039 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24040 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24041 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24042 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24043 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24046 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24047 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24048 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24049 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24050 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24051 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
24052 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24053 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
24054 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
24055 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24056 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24058 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24060 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24061 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24062 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24063 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24064 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24065 case ISD::ANY_EXTEND:
24066 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24067 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24068 case ISD::SIGN_EXTEND_INREG:
24069 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24070 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
24071 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24072 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24073 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24074 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24075 case X86ISD::SHUFP: // Handle all target specific shuffles
24076 case X86ISD::PALIGNR:
24077 case X86ISD::UNPCKH:
24078 case X86ISD::UNPCKL:
24079 case X86ISD::MOVHLPS:
24080 case X86ISD::MOVLHPS:
24081 case X86ISD::PSHUFB:
24082 case X86ISD::PSHUFD:
24083 case X86ISD::PSHUFHW:
24084 case X86ISD::PSHUFLW:
24085 case X86ISD::MOVSS:
24086 case X86ISD::MOVSD:
24087 case X86ISD::VPERMILPI:
24088 case X86ISD::VPERM2X128:
24089 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24090 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24091 case ISD::INTRINSIC_WO_CHAIN:
24092 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24093 case X86ISD::INSERTPS: {
24094 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
24095 return PerformINSERTPSCombine(N, DAG, Subtarget);
24098 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
24104 /// isTypeDesirableForOp - Return true if the target has native support for
24105 /// the specified value type and it is 'desirable' to use the type for the
24106 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24107 /// instruction encodings are longer and some i16 instructions are slow.
24108 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24109 if (!isTypeLegal(VT))
24111 if (VT != MVT::i16)
24118 case ISD::SIGN_EXTEND:
24119 case ISD::ZERO_EXTEND:
24120 case ISD::ANY_EXTEND:
24133 /// IsDesirableToPromoteOp - This method query the target whether it is
24134 /// beneficial for dag combiner to promote the specified node. If true, it
24135 /// should return the desired promotion type by reference.
24136 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24137 EVT VT = Op.getValueType();
24138 if (VT != MVT::i16)
24141 bool Promote = false;
24142 bool Commute = false;
24143 switch (Op.getOpcode()) {
24146 LoadSDNode *LD = cast<LoadSDNode>(Op);
24147 // If the non-extending load has a single use and it's not live out, then it
24148 // might be folded.
24149 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24150 Op.hasOneUse()*/) {
24151 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24152 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24153 // The only case where we'd want to promote LOAD (rather then it being
24154 // promoted as an operand is when it's only use is liveout.
24155 if (UI->getOpcode() != ISD::CopyToReg)
24162 case ISD::SIGN_EXTEND:
24163 case ISD::ZERO_EXTEND:
24164 case ISD::ANY_EXTEND:
24169 SDValue N0 = Op.getOperand(0);
24170 // Look out for (store (shl (load), x)).
24171 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24184 SDValue N0 = Op.getOperand(0);
24185 SDValue N1 = Op.getOperand(1);
24186 if (!Commute && MayFoldLoad(N1))
24188 // Avoid disabling potential load folding opportunities.
24189 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24191 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24201 //===----------------------------------------------------------------------===//
24202 // X86 Inline Assembly Support
24203 //===----------------------------------------------------------------------===//
24205 // Helper to match a string separated by whitespace.
24206 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
24207 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
24209 for (StringRef Piece : Pieces) {
24210 if (!S.startswith(Piece)) // Check if the piece matches.
24213 S = S.substr(Piece.size());
24214 StringRef::size_type Pos = S.find_first_not_of(" \t");
24215 if (Pos == 0) // We matched a prefix.
24224 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24226 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24227 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24228 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24229 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24231 if (AsmPieces.size() == 3)
24233 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24240 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24241 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24243 std::string AsmStr = IA->getAsmString();
24245 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24246 if (!Ty || Ty->getBitWidth() % 16 != 0)
24249 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24250 SmallVector<StringRef, 4> AsmPieces;
24251 SplitString(AsmStr, AsmPieces, ";\n");
24253 switch (AsmPieces.size()) {
24254 default: return false;
24256 // FIXME: this should verify that we are targeting a 486 or better. If not,
24257 // we will turn this bswap into something that will be lowered to logical
24258 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24259 // lower so don't worry about this.
24261 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
24262 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
24263 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
24264 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
24265 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
24266 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
24267 // No need to check constraints, nothing other than the equivalent of
24268 // "=r,0" would be valid here.
24269 return IntrinsicLowering::LowerToByteSwap(CI);
24272 // rorw $$8, ${0:w} --> llvm.bswap.i16
24273 if (CI->getType()->isIntegerTy(16) &&
24274 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24275 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
24276 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
24278 const std::string &ConstraintsStr = IA->getConstraintString();
24279 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24280 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24281 if (clobbersFlagRegisters(AsmPieces))
24282 return IntrinsicLowering::LowerToByteSwap(CI);
24286 if (CI->getType()->isIntegerTy(32) &&
24287 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24288 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
24289 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
24290 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
24292 const std::string &ConstraintsStr = IA->getConstraintString();
24293 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24294 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24295 if (clobbersFlagRegisters(AsmPieces))
24296 return IntrinsicLowering::LowerToByteSwap(CI);
24299 if (CI->getType()->isIntegerTy(64)) {
24300 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24301 if (Constraints.size() >= 2 &&
24302 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24303 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24304 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24305 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
24306 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
24307 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
24308 return IntrinsicLowering::LowerToByteSwap(CI);
24316 /// getConstraintType - Given a constraint letter, return the type of
24317 /// constraint it is for this target.
24318 X86TargetLowering::ConstraintType
24319 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24320 if (Constraint.size() == 1) {
24321 switch (Constraint[0]) {
24332 return C_RegisterClass;
24356 return TargetLowering::getConstraintType(Constraint);
24359 /// Examine constraint type and operand type and determine a weight value.
24360 /// This object must already have been set up with the operand type
24361 /// and the current alternative constraint selected.
24362 TargetLowering::ConstraintWeight
24363 X86TargetLowering::getSingleConstraintMatchWeight(
24364 AsmOperandInfo &info, const char *constraint) const {
24365 ConstraintWeight weight = CW_Invalid;
24366 Value *CallOperandVal = info.CallOperandVal;
24367 // If we don't have a value, we can't do a match,
24368 // but allow it at the lowest weight.
24369 if (!CallOperandVal)
24371 Type *type = CallOperandVal->getType();
24372 // Look at the constraint type.
24373 switch (*constraint) {
24375 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
24386 if (CallOperandVal->getType()->isIntegerTy())
24387 weight = CW_SpecificReg;
24392 if (type->isFloatingPointTy())
24393 weight = CW_SpecificReg;
24396 if (type->isX86_MMXTy() && Subtarget->hasMMX())
24397 weight = CW_SpecificReg;
24401 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
24402 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
24403 weight = CW_Register;
24406 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
24407 if (C->getZExtValue() <= 31)
24408 weight = CW_Constant;
24412 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24413 if (C->getZExtValue() <= 63)
24414 weight = CW_Constant;
24418 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24419 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
24420 weight = CW_Constant;
24424 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24425 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
24426 weight = CW_Constant;
24430 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24431 if (C->getZExtValue() <= 3)
24432 weight = CW_Constant;
24436 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24437 if (C->getZExtValue() <= 0xff)
24438 weight = CW_Constant;
24443 if (isa<ConstantFP>(CallOperandVal)) {
24444 weight = CW_Constant;
24448 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24449 if ((C->getSExtValue() >= -0x80000000LL) &&
24450 (C->getSExtValue() <= 0x7fffffffLL))
24451 weight = CW_Constant;
24455 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24456 if (C->getZExtValue() <= 0xffffffff)
24457 weight = CW_Constant;
24464 /// LowerXConstraint - try to replace an X constraint, which matches anything,
24465 /// with another that has more specific requirements based on the type of the
24466 /// corresponding operand.
24467 const char *X86TargetLowering::
24468 LowerXConstraint(EVT ConstraintVT) const {
24469 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
24470 // 'f' like normal targets.
24471 if (ConstraintVT.isFloatingPoint()) {
24472 if (Subtarget->hasSSE2())
24474 if (Subtarget->hasSSE1())
24478 return TargetLowering::LowerXConstraint(ConstraintVT);
24481 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
24482 /// vector. If it is invalid, don't add anything to Ops.
24483 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
24484 std::string &Constraint,
24485 std::vector<SDValue>&Ops,
24486 SelectionDAG &DAG) const {
24489 // Only support length 1 constraints for now.
24490 if (Constraint.length() > 1) return;
24492 char ConstraintLetter = Constraint[0];
24493 switch (ConstraintLetter) {
24496 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24497 if (C->getZExtValue() <= 31) {
24498 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24504 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24505 if (C->getZExtValue() <= 63) {
24506 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24512 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24513 if (isInt<8>(C->getSExtValue())) {
24514 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24520 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24521 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
24522 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
24523 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
24529 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24530 if (C->getZExtValue() <= 3) {
24531 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24537 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24538 if (C->getZExtValue() <= 255) {
24539 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24545 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24546 if (C->getZExtValue() <= 127) {
24547 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24553 // 32-bit signed value
24554 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24555 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24556 C->getSExtValue())) {
24557 // Widen to 64 bits here to get it sign extended.
24558 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
24561 // FIXME gcc accepts some relocatable values here too, but only in certain
24562 // memory models; it's complicated.
24567 // 32-bit unsigned value
24568 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24569 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24570 C->getZExtValue())) {
24571 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
24575 // FIXME gcc accepts some relocatable values here too, but only in certain
24576 // memory models; it's complicated.
24580 // Literal immediates are always ok.
24581 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
24582 // Widen to 64 bits here to get it sign extended.
24583 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
24587 // In any sort of PIC mode addresses need to be computed at runtime by
24588 // adding in a register or some sort of table lookup. These can't
24589 // be used as immediates.
24590 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
24593 // If we are in non-pic codegen mode, we allow the address of a global (with
24594 // an optional displacement) to be used with 'i'.
24595 GlobalAddressSDNode *GA = nullptr;
24596 int64_t Offset = 0;
24598 // Match either (GA), (GA+C), (GA+C1+C2), etc.
24600 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
24601 Offset += GA->getOffset();
24603 } else if (Op.getOpcode() == ISD::ADD) {
24604 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24605 Offset += C->getZExtValue();
24606 Op = Op.getOperand(0);
24609 } else if (Op.getOpcode() == ISD::SUB) {
24610 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
24611 Offset += -C->getZExtValue();
24612 Op = Op.getOperand(0);
24617 // Otherwise, this isn't something we can handle, reject it.
24621 const GlobalValue *GV = GA->getGlobal();
24622 // If we require an extra load to get this address, as in PIC mode, we
24623 // can't accept it.
24624 if (isGlobalStubReference(
24625 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
24628 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
24629 GA->getValueType(0), Offset);
24634 if (Result.getNode()) {
24635 Ops.push_back(Result);
24638 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
24641 std::pair<unsigned, const TargetRegisterClass *>
24642 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
24643 const std::string &Constraint,
24645 // First, see if this is a constraint that directly corresponds to an LLVM
24647 if (Constraint.size() == 1) {
24648 // GCC Constraint Letters
24649 switch (Constraint[0]) {
24651 // TODO: Slight differences here in allocation order and leaving
24652 // RIP in the class. Do they matter any more here than they do
24653 // in the normal allocation?
24654 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
24655 if (Subtarget->is64Bit()) {
24656 if (VT == MVT::i32 || VT == MVT::f32)
24657 return std::make_pair(0U, &X86::GR32RegClass);
24658 if (VT == MVT::i16)
24659 return std::make_pair(0U, &X86::GR16RegClass);
24660 if (VT == MVT::i8 || VT == MVT::i1)
24661 return std::make_pair(0U, &X86::GR8RegClass);
24662 if (VT == MVT::i64 || VT == MVT::f64)
24663 return std::make_pair(0U, &X86::GR64RegClass);
24666 // 32-bit fallthrough
24667 case 'Q': // Q_REGS
24668 if (VT == MVT::i32 || VT == MVT::f32)
24669 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
24670 if (VT == MVT::i16)
24671 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
24672 if (VT == MVT::i8 || VT == MVT::i1)
24673 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
24674 if (VT == MVT::i64)
24675 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
24677 case 'r': // GENERAL_REGS
24678 case 'l': // INDEX_REGS
24679 if (VT == MVT::i8 || VT == MVT::i1)
24680 return std::make_pair(0U, &X86::GR8RegClass);
24681 if (VT == MVT::i16)
24682 return std::make_pair(0U, &X86::GR16RegClass);
24683 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
24684 return std::make_pair(0U, &X86::GR32RegClass);
24685 return std::make_pair(0U, &X86::GR64RegClass);
24686 case 'R': // LEGACY_REGS
24687 if (VT == MVT::i8 || VT == MVT::i1)
24688 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
24689 if (VT == MVT::i16)
24690 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
24691 if (VT == MVT::i32 || !Subtarget->is64Bit())
24692 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
24693 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
24694 case 'f': // FP Stack registers.
24695 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
24696 // value to the correct fpstack register class.
24697 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
24698 return std::make_pair(0U, &X86::RFP32RegClass);
24699 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
24700 return std::make_pair(0U, &X86::RFP64RegClass);
24701 return std::make_pair(0U, &X86::RFP80RegClass);
24702 case 'y': // MMX_REGS if MMX allowed.
24703 if (!Subtarget->hasMMX()) break;
24704 return std::make_pair(0U, &X86::VR64RegClass);
24705 case 'Y': // SSE_REGS if SSE2 allowed
24706 if (!Subtarget->hasSSE2()) break;
24708 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
24709 if (!Subtarget->hasSSE1()) break;
24711 switch (VT.SimpleTy) {
24713 // Scalar SSE types.
24716 return std::make_pair(0U, &X86::FR32RegClass);
24719 return std::make_pair(0U, &X86::FR64RegClass);
24727 return std::make_pair(0U, &X86::VR128RegClass);
24735 return std::make_pair(0U, &X86::VR256RegClass);
24740 return std::make_pair(0U, &X86::VR512RegClass);
24746 // Use the default implementation in TargetLowering to convert the register
24747 // constraint into a member of a register class.
24748 std::pair<unsigned, const TargetRegisterClass*> Res;
24749 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
24751 // Not found as a standard register?
24753 // Map st(0) -> st(7) -> ST0
24754 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24755 tolower(Constraint[1]) == 's' &&
24756 tolower(Constraint[2]) == 't' &&
24757 Constraint[3] == '(' &&
24758 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24759 Constraint[5] == ')' &&
24760 Constraint[6] == '}') {
24762 Res.first = X86::FP0+Constraint[4]-'0';
24763 Res.second = &X86::RFP80RegClass;
24767 // GCC allows "st(0)" to be called just plain "st".
24768 if (StringRef("{st}").equals_lower(Constraint)) {
24769 Res.first = X86::FP0;
24770 Res.second = &X86::RFP80RegClass;
24775 if (StringRef("{flags}").equals_lower(Constraint)) {
24776 Res.first = X86::EFLAGS;
24777 Res.second = &X86::CCRRegClass;
24781 // 'A' means EAX + EDX.
24782 if (Constraint == "A") {
24783 Res.first = X86::EAX;
24784 Res.second = &X86::GR32_ADRegClass;
24790 // Otherwise, check to see if this is a register class of the wrong value
24791 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24792 // turn into {ax},{dx}.
24793 if (Res.second->hasType(VT))
24794 return Res; // Correct type already, nothing to do.
24796 // All of the single-register GCC register classes map their values onto
24797 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24798 // really want an 8-bit or 32-bit register, map to the appropriate register
24799 // class and return the appropriate register.
24800 if (Res.second == &X86::GR16RegClass) {
24801 if (VT == MVT::i8 || VT == MVT::i1) {
24802 unsigned DestReg = 0;
24803 switch (Res.first) {
24805 case X86::AX: DestReg = X86::AL; break;
24806 case X86::DX: DestReg = X86::DL; break;
24807 case X86::CX: DestReg = X86::CL; break;
24808 case X86::BX: DestReg = X86::BL; break;
24811 Res.first = DestReg;
24812 Res.second = &X86::GR8RegClass;
24814 } else if (VT == MVT::i32 || VT == MVT::f32) {
24815 unsigned DestReg = 0;
24816 switch (Res.first) {
24818 case X86::AX: DestReg = X86::EAX; break;
24819 case X86::DX: DestReg = X86::EDX; break;
24820 case X86::CX: DestReg = X86::ECX; break;
24821 case X86::BX: DestReg = X86::EBX; break;
24822 case X86::SI: DestReg = X86::ESI; break;
24823 case X86::DI: DestReg = X86::EDI; break;
24824 case X86::BP: DestReg = X86::EBP; break;
24825 case X86::SP: DestReg = X86::ESP; break;
24828 Res.first = DestReg;
24829 Res.second = &X86::GR32RegClass;
24831 } else if (VT == MVT::i64 || VT == MVT::f64) {
24832 unsigned DestReg = 0;
24833 switch (Res.first) {
24835 case X86::AX: DestReg = X86::RAX; break;
24836 case X86::DX: DestReg = X86::RDX; break;
24837 case X86::CX: DestReg = X86::RCX; break;
24838 case X86::BX: DestReg = X86::RBX; break;
24839 case X86::SI: DestReg = X86::RSI; break;
24840 case X86::DI: DestReg = X86::RDI; break;
24841 case X86::BP: DestReg = X86::RBP; break;
24842 case X86::SP: DestReg = X86::RSP; break;
24845 Res.first = DestReg;
24846 Res.second = &X86::GR64RegClass;
24849 } else if (Res.second == &X86::FR32RegClass ||
24850 Res.second == &X86::FR64RegClass ||
24851 Res.second == &X86::VR128RegClass ||
24852 Res.second == &X86::VR256RegClass ||
24853 Res.second == &X86::FR32XRegClass ||
24854 Res.second == &X86::FR64XRegClass ||
24855 Res.second == &X86::VR128XRegClass ||
24856 Res.second == &X86::VR256XRegClass ||
24857 Res.second == &X86::VR512RegClass) {
24858 // Handle references to XMM physical registers that got mapped into the
24859 // wrong class. This can happen with constraints like {xmm0} where the
24860 // target independent register mapper will just pick the first match it can
24861 // find, ignoring the required type.
24863 if (VT == MVT::f32 || VT == MVT::i32)
24864 Res.second = &X86::FR32RegClass;
24865 else if (VT == MVT::f64 || VT == MVT::i64)
24866 Res.second = &X86::FR64RegClass;
24867 else if (X86::VR128RegClass.hasType(VT))
24868 Res.second = &X86::VR128RegClass;
24869 else if (X86::VR256RegClass.hasType(VT))
24870 Res.second = &X86::VR256RegClass;
24871 else if (X86::VR512RegClass.hasType(VT))
24872 Res.second = &X86::VR512RegClass;
24878 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
24880 // Scaling factors are not free at all.
24881 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
24882 // will take 2 allocations in the out of order engine instead of 1
24883 // for plain addressing mode, i.e. inst (reg1).
24885 // vaddps (%rsi,%drx), %ymm0, %ymm1
24886 // Requires two allocations (one for the load, one for the computation)
24888 // vaddps (%rsi), %ymm0, %ymm1
24889 // Requires just 1 allocation, i.e., freeing allocations for other operations
24890 // and having less micro operations to execute.
24892 // For some X86 architectures, this is even worse because for instance for
24893 // stores, the complex addressing mode forces the instruction to use the
24894 // "load" ports instead of the dedicated "store" port.
24895 // E.g., on Haswell:
24896 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
24897 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
24898 if (isLegalAddressingMode(AM, Ty))
24899 // Scale represents reg2 * scale, thus account for 1
24900 // as soon as we use a second register.
24901 return AM.Scale != 0;
24905 bool X86TargetLowering::isTargetFTOL() const {
24906 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();