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 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
71 const X86Subtarget &STI)
72 : TargetLowering(TM), Subtarget(&STI) {
73 X86ScalarSSEf64 = Subtarget->hasSSE2();
74 X86ScalarSSEf32 = Subtarget->hasSSE1();
75 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
77 // Set up the TargetLowering object.
78 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
118 // The _ftol2 runtime function has an unusual calling conv, which
119 // is modeled by a special pseudo-instruction.
120 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
121 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
122 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
123 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
126 if (Subtarget->isTargetDarwin()) {
127 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
128 setUseUnderscoreSetJmp(false);
129 setUseUnderscoreLongJmp(false);
130 } else if (Subtarget->isTargetWindowsGNU()) {
131 // MS runtime is weird: it exports _setjmp, but longjmp!
132 setUseUnderscoreSetJmp(true);
133 setUseUnderscoreLongJmp(false);
135 setUseUnderscoreSetJmp(true);
136 setUseUnderscoreLongJmp(true);
139 // Set up the register classes.
140 addRegisterClass(MVT::i8, &X86::GR8RegClass);
141 addRegisterClass(MVT::i16, &X86::GR16RegClass);
142 addRegisterClass(MVT::i32, &X86::GR32RegClass);
143 if (Subtarget->is64Bit())
144 addRegisterClass(MVT::i64, &X86::GR64RegClass);
146 for (MVT VT : MVT::integer_valuetypes())
147 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
149 // We don't accept any truncstore of integer registers.
150 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
151 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
152 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
153 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
154 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
155 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
157 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
159 // SETOEQ and SETUNE require checking two conditions.
160 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
161 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
162 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
163 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
164 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
165 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
167 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
169 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
170 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
173 if (Subtarget->is64Bit()) {
174 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
175 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
176 } else if (!Subtarget->useSoftFloat()) {
177 // We have an algorithm for SSE2->double, and we turn this into a
178 // 64-bit FILD followed by conditional FADD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
180 // We have an algorithm for SSE2, and we turn this into a 64-bit
181 // FILD for other targets.
182 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
185 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
187 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
190 if (!Subtarget->useSoftFloat()) {
191 // SSE has no i16 to fp conversion, only i32
192 if (X86ScalarSSEf32) {
193 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
194 // f32 and f64 cases are Legal, f80 case is not
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
197 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
201 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
202 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
205 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
206 // are Legal, f80 is custom lowered.
207 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
208 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
210 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
212 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
213 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
215 if (X86ScalarSSEf32) {
216 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
217 // f32 and f64 cases are Legal, f80 case is not
218 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
220 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
221 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
224 // Handle FP_TO_UINT by promoting the destination to a larger signed
226 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
227 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
228 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
230 if (Subtarget->is64Bit()) {
231 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
233 } else if (!Subtarget->useSoftFloat()) {
234 // Since AVX is a superset of SSE3, only check for SSE here.
235 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
236 // Expand FP_TO_UINT into a select.
237 // FIXME: We would like to use a Custom expander here eventually to do
238 // the optimal thing for SSE vs. the default expansion in the legalizer.
239 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
241 // With SSE3 we can use fisttpll to convert to a signed i64; without
242 // SSE, we're stuck with a fistpll.
243 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
246 if (isTargetFTOL()) {
247 // Use the _ftol2 runtime function, which has a pseudo-instruction
248 // to handle its weird calling convention.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
252 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
253 if (!X86ScalarSSEf64) {
254 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
255 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
258 // Without SSE, i64->f64 goes through memory.
259 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
263 // Scalar integer divide and remainder are lowered to use operations that
264 // produce two results, to match the available instructions. This exposes
265 // the two-result form to trivial CSE, which is able to combine x/y and x%y
266 // into a single instruction.
268 // Scalar integer multiply-high is also lowered to use two-result
269 // operations, to match the available instructions. However, plain multiply
270 // (low) operations are left as Legal, as there are single-result
271 // instructions for this in x86. Using the two-result multiply instructions
272 // when both high and low results are needed must be arranged by dagcombine.
273 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
275 setOperationAction(ISD::MULHS, VT, Expand);
276 setOperationAction(ISD::MULHU, VT, Expand);
277 setOperationAction(ISD::SDIV, VT, Expand);
278 setOperationAction(ISD::UDIV, VT, Expand);
279 setOperationAction(ISD::SREM, VT, Expand);
280 setOperationAction(ISD::UREM, VT, Expand);
282 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
283 setOperationAction(ISD::ADDC, VT, Custom);
284 setOperationAction(ISD::ADDE, VT, Custom);
285 setOperationAction(ISD::SUBC, VT, Custom);
286 setOperationAction(ISD::SUBE, VT, Custom);
289 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
290 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
291 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
292 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
293 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
294 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
295 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
296 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
301 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
302 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
303 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
305 if (Subtarget->is64Bit())
306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
308 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
309 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
310 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
311 setOperationAction(ISD::FREM , MVT::f32 , Expand);
312 setOperationAction(ISD::FREM , MVT::f64 , Expand);
313 setOperationAction(ISD::FREM , MVT::f80 , Expand);
314 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
316 // Promote the i8 variants and force them on up to i32 which has a shorter
318 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
319 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
320 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
321 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
322 if (Subtarget->hasBMI()) {
323 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
324 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
325 if (Subtarget->is64Bit())
326 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
328 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
329 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
330 if (Subtarget->is64Bit())
331 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
334 if (Subtarget->hasLZCNT()) {
335 // When promoting the i8 variants, force them to i32 for a shorter
337 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
338 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
339 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
340 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
341 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
342 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
343 if (Subtarget->is64Bit())
344 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
347 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
348 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
352 if (Subtarget->is64Bit()) {
353 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
358 // Special handling for half-precision floating point conversions.
359 // If we don't have F16C support, then lower half float conversions
360 // into library calls.
361 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
362 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
363 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
366 // There's never any support for operations beyond MVT::f32.
367 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
368 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
369 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
370 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
372 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
373 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
374 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
375 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
376 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
377 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
379 if (Subtarget->hasPOPCNT()) {
380 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
382 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
383 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
384 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
385 if (Subtarget->is64Bit())
386 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
389 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
391 if (!Subtarget->hasMOVBE())
392 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
394 // These should be promoted to a larger select which is supported.
395 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
396 // X86 wants to expand cmov itself.
397 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
398 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
399 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
400 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
401 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
402 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
403 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
404 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
405 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
406 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
407 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
408 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
409 if (Subtarget->is64Bit()) {
410 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
413 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
414 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
415 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
416 // support continuation, user-level threading, and etc.. As a result, no
417 // other SjLj exception interfaces are implemented and please don't build
418 // your own exception handling based on them.
419 // LLVM/Clang supports zero-cost DWARF exception handling.
420 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
421 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
424 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
425 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
426 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
430 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
431 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
434 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
435 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
436 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
437 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
439 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
440 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
441 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
443 if (Subtarget->is64Bit()) {
444 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
445 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
449 if (Subtarget->hasSSE1())
450 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
452 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
454 // Expand certain atomics
455 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
457 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
458 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
459 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
462 if (Subtarget->hasCmpxchg16b()) {
463 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
466 // FIXME - use subtarget debug flags
467 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
468 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
469 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
472 if (Subtarget->is64Bit()) {
473 setExceptionPointerRegister(X86::RAX);
474 setExceptionSelectorRegister(X86::RDX);
476 setExceptionPointerRegister(X86::EAX);
477 setExceptionSelectorRegister(X86::EDX);
479 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
480 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
482 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
483 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
485 setOperationAction(ISD::TRAP, MVT::Other, Legal);
486 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
488 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
489 setOperationAction(ISD::VASTART , MVT::Other, Custom);
490 setOperationAction(ISD::VAEND , MVT::Other, Expand);
491 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
492 // TargetInfo::X86_64ABIBuiltinVaList
493 setOperationAction(ISD::VAARG , MVT::Other, Custom);
494 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
496 // TargetInfo::CharPtrBuiltinVaList
497 setOperationAction(ISD::VAARG , MVT::Other, Expand);
498 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
501 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
502 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
504 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
506 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
507 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
508 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
510 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
511 // f32 and f64 use SSE.
512 // Set up the FP register classes.
513 addRegisterClass(MVT::f32, &X86::FR32RegClass);
514 addRegisterClass(MVT::f64, &X86::FR64RegClass);
516 // Use ANDPD to simulate FABS.
517 setOperationAction(ISD::FABS , MVT::f64, Custom);
518 setOperationAction(ISD::FABS , MVT::f32, Custom);
520 // Use XORP to simulate FNEG.
521 setOperationAction(ISD::FNEG , MVT::f64, Custom);
522 setOperationAction(ISD::FNEG , MVT::f32, Custom);
524 // Use ANDPD and ORPD to simulate FCOPYSIGN.
525 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
526 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
528 // Lower this to FGETSIGNx86 plus an AND.
529 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
530 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
532 // We don't support sin/cos/fmod
533 setOperationAction(ISD::FSIN , MVT::f64, Expand);
534 setOperationAction(ISD::FCOS , MVT::f64, Expand);
535 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
536 setOperationAction(ISD::FSIN , MVT::f32, Expand);
537 setOperationAction(ISD::FCOS , MVT::f32, Expand);
538 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
540 // Expand FP immediates into loads from the stack, except for the special
542 addLegalFPImmediate(APFloat(+0.0)); // xorpd
543 addLegalFPImmediate(APFloat(+0.0f)); // xorps
544 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
545 // Use SSE for f32, x87 for f64.
546 // Set up the FP register classes.
547 addRegisterClass(MVT::f32, &X86::FR32RegClass);
548 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
550 // Use ANDPS to simulate FABS.
551 setOperationAction(ISD::FABS , MVT::f32, Custom);
553 // Use XORP to simulate FNEG.
554 setOperationAction(ISD::FNEG , MVT::f32, Custom);
556 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
558 // Use ANDPS and ORPS to simulate FCOPYSIGN.
559 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
560 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
562 // We don't support sin/cos/fmod
563 setOperationAction(ISD::FSIN , MVT::f32, Expand);
564 setOperationAction(ISD::FCOS , MVT::f32, Expand);
565 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
567 // Special cases we handle for FP constants.
568 addLegalFPImmediate(APFloat(+0.0f)); // xorps
569 addLegalFPImmediate(APFloat(+0.0)); // FLD0
570 addLegalFPImmediate(APFloat(+1.0)); // FLD1
571 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
572 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
574 if (!TM.Options.UnsafeFPMath) {
575 setOperationAction(ISD::FSIN , MVT::f64, Expand);
576 setOperationAction(ISD::FCOS , MVT::f64, Expand);
577 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
579 } else if (!Subtarget->useSoftFloat()) {
580 // f32 and f64 in x87.
581 // Set up the FP register classes.
582 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
583 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
585 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
586 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
587 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
588 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
590 if (!TM.Options.UnsafeFPMath) {
591 setOperationAction(ISD::FSIN , MVT::f64, Expand);
592 setOperationAction(ISD::FSIN , MVT::f32, Expand);
593 setOperationAction(ISD::FCOS , MVT::f64, Expand);
594 setOperationAction(ISD::FCOS , MVT::f32, Expand);
595 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
596 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
598 addLegalFPImmediate(APFloat(+0.0)); // FLD0
599 addLegalFPImmediate(APFloat(+1.0)); // FLD1
600 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
601 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
602 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
603 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
604 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
605 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
608 // We don't support FMA.
609 setOperationAction(ISD::FMA, MVT::f64, Expand);
610 setOperationAction(ISD::FMA, MVT::f32, Expand);
612 // Long double always uses X87.
613 if (!Subtarget->useSoftFloat()) {
614 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
615 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
616 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
618 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
619 addLegalFPImmediate(TmpFlt); // FLD0
621 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
624 APFloat TmpFlt2(+1.0);
625 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
627 addLegalFPImmediate(TmpFlt2); // FLD1
628 TmpFlt2.changeSign();
629 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
632 if (!TM.Options.UnsafeFPMath) {
633 setOperationAction(ISD::FSIN , MVT::f80, Expand);
634 setOperationAction(ISD::FCOS , MVT::f80, Expand);
635 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
638 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
639 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
640 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
641 setOperationAction(ISD::FRINT, MVT::f80, Expand);
642 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
643 setOperationAction(ISD::FMA, MVT::f80, Expand);
646 // Always use a library call for pow.
647 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
648 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
649 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
651 setOperationAction(ISD::FLOG, MVT::f80, Expand);
652 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
653 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
654 setOperationAction(ISD::FEXP, MVT::f80, Expand);
655 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
656 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
657 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
659 // First set operation action for all vector types to either promote
660 // (for widening) or expand (for scalarization). Then we will selectively
661 // turn on ones that can be effectively codegen'd.
662 for (MVT VT : MVT::vector_valuetypes()) {
663 setOperationAction(ISD::ADD , VT, Expand);
664 setOperationAction(ISD::SUB , VT, Expand);
665 setOperationAction(ISD::FADD, VT, Expand);
666 setOperationAction(ISD::FNEG, VT, Expand);
667 setOperationAction(ISD::FSUB, VT, Expand);
668 setOperationAction(ISD::MUL , VT, Expand);
669 setOperationAction(ISD::FMUL, VT, Expand);
670 setOperationAction(ISD::SDIV, VT, Expand);
671 setOperationAction(ISD::UDIV, VT, Expand);
672 setOperationAction(ISD::FDIV, VT, Expand);
673 setOperationAction(ISD::SREM, VT, Expand);
674 setOperationAction(ISD::UREM, VT, Expand);
675 setOperationAction(ISD::LOAD, VT, Expand);
676 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
677 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
678 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
679 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
680 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
681 setOperationAction(ISD::FABS, VT, Expand);
682 setOperationAction(ISD::FSIN, VT, Expand);
683 setOperationAction(ISD::FSINCOS, VT, Expand);
684 setOperationAction(ISD::FCOS, VT, Expand);
685 setOperationAction(ISD::FSINCOS, VT, Expand);
686 setOperationAction(ISD::FREM, VT, Expand);
687 setOperationAction(ISD::FMA, VT, Expand);
688 setOperationAction(ISD::FPOWI, VT, Expand);
689 setOperationAction(ISD::FSQRT, VT, Expand);
690 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
691 setOperationAction(ISD::FFLOOR, VT, Expand);
692 setOperationAction(ISD::FCEIL, VT, Expand);
693 setOperationAction(ISD::FTRUNC, VT, Expand);
694 setOperationAction(ISD::FRINT, VT, Expand);
695 setOperationAction(ISD::FNEARBYINT, VT, Expand);
696 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
697 setOperationAction(ISD::MULHS, VT, Expand);
698 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
699 setOperationAction(ISD::MULHU, VT, Expand);
700 setOperationAction(ISD::SDIVREM, VT, Expand);
701 setOperationAction(ISD::UDIVREM, VT, Expand);
702 setOperationAction(ISD::FPOW, VT, Expand);
703 setOperationAction(ISD::CTPOP, VT, Expand);
704 setOperationAction(ISD::CTTZ, VT, Expand);
705 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
706 setOperationAction(ISD::CTLZ, VT, Expand);
707 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
708 setOperationAction(ISD::SHL, VT, Expand);
709 setOperationAction(ISD::SRA, VT, Expand);
710 setOperationAction(ISD::SRL, VT, Expand);
711 setOperationAction(ISD::ROTL, VT, Expand);
712 setOperationAction(ISD::ROTR, VT, Expand);
713 setOperationAction(ISD::BSWAP, VT, Expand);
714 setOperationAction(ISD::SETCC, VT, Expand);
715 setOperationAction(ISD::FLOG, VT, Expand);
716 setOperationAction(ISD::FLOG2, VT, Expand);
717 setOperationAction(ISD::FLOG10, VT, Expand);
718 setOperationAction(ISD::FEXP, VT, Expand);
719 setOperationAction(ISD::FEXP2, VT, Expand);
720 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
721 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
722 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
723 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
725 setOperationAction(ISD::TRUNCATE, VT, Expand);
726 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
727 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
728 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
729 setOperationAction(ISD::VSELECT, VT, Expand);
730 setOperationAction(ISD::SELECT_CC, VT, Expand);
731 for (MVT InnerVT : MVT::vector_valuetypes()) {
732 setTruncStoreAction(InnerVT, VT, Expand);
734 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
735 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
737 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
738 // types, we have to deal with them whether we ask for Expansion or not.
739 // Setting Expand causes its own optimisation problems though, so leave
741 if (VT.getVectorElementType() == MVT::i1)
742 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
744 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
745 // split/scalarized right now.
746 if (VT.getVectorElementType() == MVT::f16)
747 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
751 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
752 // with -msoft-float, disable use of MMX as well.
753 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
754 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
755 // No operations on x86mmx supported, everything uses intrinsics.
758 // MMX-sized vectors (other than x86mmx) are expected to be expanded
759 // into smaller operations.
760 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
761 setOperationAction(ISD::MULHS, MMXTy, Expand);
762 setOperationAction(ISD::AND, MMXTy, Expand);
763 setOperationAction(ISD::OR, MMXTy, Expand);
764 setOperationAction(ISD::XOR, MMXTy, Expand);
765 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
766 setOperationAction(ISD::SELECT, MMXTy, Expand);
767 setOperationAction(ISD::BITCAST, MMXTy, Expand);
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
771 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
772 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
774 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
775 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
776 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
777 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
778 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
779 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
780 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
781 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
784 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
785 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
786 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
787 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
790 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
791 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
793 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
794 // registers cannot be used even for integer operations.
795 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
796 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
797 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
798 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
800 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
801 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
802 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
803 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
804 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
805 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
806 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
807 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
808 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
809 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
810 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
811 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
812 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
813 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
814 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
815 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
816 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
817 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
818 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
819 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
820 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
821 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
822 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
824 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
825 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
826 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
827 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
829 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
830 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
831 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
832 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
834 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
835 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
837 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
840 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
841 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
842 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
843 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
845 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
846 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
847 MVT VT = (MVT::SimpleValueType)i;
848 // Do not attempt to custom lower non-power-of-2 vectors
849 if (!isPowerOf2_32(VT.getVectorNumElements()))
851 // Do not attempt to custom lower non-128-bit vectors
852 if (!VT.is128BitVector())
854 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
855 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
856 setOperationAction(ISD::VSELECT, VT, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
860 // We support custom legalizing of sext and anyext loads for specific
861 // memory vector types which we can load as a scalar (or sequence of
862 // scalars) and extend in-register to a legal 128-bit vector type. For sext
863 // loads these must work with a single scalar load.
864 for (MVT VT : MVT::integer_vector_valuetypes()) {
865 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
866 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
867 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
868 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
869 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
870 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
871 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
872 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
873 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
876 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
877 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
878 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
880 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
881 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
882 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
883 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
885 if (Subtarget->is64Bit()) {
886 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
887 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
890 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
891 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
892 MVT VT = (MVT::SimpleValueType)i;
894 // Do not attempt to promote non-128-bit vectors
895 if (!VT.is128BitVector())
898 setOperationAction(ISD::AND, VT, Promote);
899 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
900 setOperationAction(ISD::OR, VT, Promote);
901 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
902 setOperationAction(ISD::XOR, VT, Promote);
903 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
904 setOperationAction(ISD::LOAD, VT, Promote);
905 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
906 setOperationAction(ISD::SELECT, VT, Promote);
907 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
910 // Custom lower v2i64 and v2f64 selects.
911 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
912 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
913 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
914 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
916 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
917 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
919 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
921 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
922 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
923 // As there is no 64-bit GPR available, we need build a special custom
924 // sequence to convert from v2i32 to v2f32.
925 if (!Subtarget->is64Bit())
926 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
928 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
929 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
931 for (MVT VT : MVT::fp_vector_valuetypes())
932 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
934 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
935 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
936 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
939 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
940 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
941 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
942 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
943 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
944 setOperationAction(ISD::FRINT, RoundedTy, Legal);
945 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
948 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
949 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
950 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
951 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
952 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
953 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
954 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
955 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
957 // FIXME: Do we need to handle scalar-to-vector here?
958 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
960 // We directly match byte blends in the backend as they match the VSELECT
962 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
964 // SSE41 brings specific instructions for doing vector sign extend even in
965 // cases where we don't have SRA.
966 for (MVT VT : MVT::integer_vector_valuetypes()) {
967 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
968 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
969 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
972 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
973 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
974 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
976 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
977 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
978 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
980 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
981 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
983 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
984 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
985 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
987 // i8 and i16 vectors are custom because the source register and source
988 // source memory operand types are not the same width. f32 vectors are
989 // custom since the immediate controlling the insert encodes additional
991 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
992 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
993 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
994 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
997 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
998 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1001 // FIXME: these should be Legal, but that's only for the case where
1002 // the index is constant. For now custom expand to deal with that.
1003 if (Subtarget->is64Bit()) {
1004 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1009 if (Subtarget->hasSSE2()) {
1010 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1011 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1012 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1014 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1015 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1017 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1018 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1020 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1021 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1023 // In the customized shift lowering, the legal cases in AVX2 will be
1025 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1026 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1028 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1031 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1032 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1035 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1036 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1037 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1038 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1039 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1040 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1041 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1043 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1044 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1045 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1047 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1048 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1049 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1050 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1051 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1052 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1053 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1054 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1055 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1056 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1057 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1058 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1060 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1063 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1064 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1065 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1066 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1067 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1068 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1069 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1070 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1071 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1073 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1074 // even though v8i16 is a legal type.
1075 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1076 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1077 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1079 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1080 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1081 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1083 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1084 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1086 for (MVT VT : MVT::fp_vector_valuetypes())
1087 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1089 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1090 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1092 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1093 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1095 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1096 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1098 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1099 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1100 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1101 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1103 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1104 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1105 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1107 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1108 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1109 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1110 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1111 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1112 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1113 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1114 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1115 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1116 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1117 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1118 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1120 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1121 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1122 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1123 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1125 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1126 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1127 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1128 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1129 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1130 setOperationAction(ISD::FMA, MVT::f32, Legal);
1131 setOperationAction(ISD::FMA, MVT::f64, Legal);
1134 if (Subtarget->hasInt256()) {
1135 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1136 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1137 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1138 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1140 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1141 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1142 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1143 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1145 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1146 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1147 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1148 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1150 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1151 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1152 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1153 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1155 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1156 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1157 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1158 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1159 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1160 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1161 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1162 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1163 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1164 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1165 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1166 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1168 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1169 // when we have a 256bit-wide blend with immediate.
1170 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1172 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1173 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1174 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1175 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1176 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1177 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1178 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1180 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1181 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1182 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1183 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1184 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1185 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1187 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1188 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1189 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1190 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1192 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1193 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1194 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1195 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1197 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1198 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1199 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1200 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1202 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1203 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1204 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1205 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1206 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1207 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1208 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1209 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1210 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1211 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1212 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1213 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1216 // In the customized shift lowering, the legal cases in AVX2 will be
1218 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1219 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1221 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1222 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1224 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1225 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1227 // Custom lower several nodes for 256-bit types.
1228 for (MVT VT : MVT::vector_valuetypes()) {
1229 if (VT.getScalarSizeInBits() >= 32) {
1230 setOperationAction(ISD::MLOAD, VT, Legal);
1231 setOperationAction(ISD::MSTORE, VT, Legal);
1233 // Extract subvector is special because the value type
1234 // (result) is 128-bit but the source is 256-bit wide.
1235 if (VT.is128BitVector()) {
1236 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1238 // Do not attempt to custom lower other non-256-bit vectors
1239 if (!VT.is256BitVector())
1242 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1243 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1244 setOperationAction(ISD::VSELECT, VT, Custom);
1245 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1246 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1247 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1248 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1249 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1252 if (Subtarget->hasInt256())
1253 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1256 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1257 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1258 MVT VT = (MVT::SimpleValueType)i;
1260 // Do not attempt to promote non-256-bit vectors
1261 if (!VT.is256BitVector())
1264 setOperationAction(ISD::AND, VT, Promote);
1265 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1266 setOperationAction(ISD::OR, VT, Promote);
1267 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1268 setOperationAction(ISD::XOR, VT, Promote);
1269 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1270 setOperationAction(ISD::LOAD, VT, Promote);
1271 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1272 setOperationAction(ISD::SELECT, VT, Promote);
1273 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1277 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1278 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1279 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1280 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1281 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1283 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1284 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1285 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1287 for (MVT VT : MVT::fp_vector_valuetypes())
1288 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1290 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1291 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1292 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1293 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1294 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1295 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1296 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1297 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1298 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1299 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1300 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1301 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1303 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1304 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1305 setOperationAction(ISD::XOR, MVT::i1, Legal);
1306 setOperationAction(ISD::OR, MVT::i1, Legal);
1307 setOperationAction(ISD::AND, MVT::i1, Legal);
1308 setOperationAction(ISD::SUB, MVT::i1, Custom);
1309 setOperationAction(ISD::ADD, MVT::i1, Custom);
1310 setOperationAction(ISD::MUL, MVT::i1, Custom);
1311 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1312 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1313 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1314 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1315 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1317 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1318 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1319 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1320 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1321 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1322 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1324 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1325 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1326 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1327 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1328 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1329 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1330 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1331 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1333 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1334 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1335 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1336 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1337 if (Subtarget->is64Bit()) {
1338 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1339 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1340 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1341 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1343 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1344 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1345 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1346 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1347 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1348 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1349 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1350 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1351 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1352 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1353 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1354 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1355 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1356 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1357 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1358 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1360 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1361 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1362 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1363 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1364 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1365 if (Subtarget->hasVLX()){
1366 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1367 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1368 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1369 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1370 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1372 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1373 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1374 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1375 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1376 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1378 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1379 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1380 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1381 if (Subtarget->hasDQI()) {
1382 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1383 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1385 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1386 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1387 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1388 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1389 if (Subtarget->hasVLX()) {
1390 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1391 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1392 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1393 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1394 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1395 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1396 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1397 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1400 if (Subtarget->hasVLX()) {
1401 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1402 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1403 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1404 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1405 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1406 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1407 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1408 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1410 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1411 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1412 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1413 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1414 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1415 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1416 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1417 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1418 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1419 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1420 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1421 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1422 if (Subtarget->hasDQI()) {
1423 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1424 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1426 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1427 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1428 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1429 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1430 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1431 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1432 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1433 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1434 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1437 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1438 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1439 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1440 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1441 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1443 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1444 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1446 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1448 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1449 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1450 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1451 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1452 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1453 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1454 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1455 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1456 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1457 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1458 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1460 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1461 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1462 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1463 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1464 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1465 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1466 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1467 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1469 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1470 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1472 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1473 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1475 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1477 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1478 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1480 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1481 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1483 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1484 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1486 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1487 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1488 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1489 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1490 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1491 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1493 if (Subtarget->hasCDI()) {
1494 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1495 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1497 if (Subtarget->hasDQI()) {
1498 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1499 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1500 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1502 // Custom lower several nodes.
1503 for (MVT VT : MVT::vector_valuetypes()) {
1504 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1506 setOperationAction(ISD::AND, VT, Legal);
1507 setOperationAction(ISD::OR, VT, Legal);
1508 setOperationAction(ISD::XOR, VT, Legal);
1510 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1511 setOperationAction(ISD::MGATHER, VT, Custom);
1512 setOperationAction(ISD::MSCATTER, VT, Custom);
1514 // Extract subvector is special because the value type
1515 // (result) is 256/128-bit but the source is 512-bit wide.
1516 if (VT.is128BitVector() || VT.is256BitVector()) {
1517 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1519 if (VT.getVectorElementType() == MVT::i1)
1520 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1522 // Do not attempt to custom lower other non-512-bit vectors
1523 if (!VT.is512BitVector())
1526 if (EltSize >= 32) {
1527 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1528 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1529 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1530 setOperationAction(ISD::VSELECT, VT, Legal);
1531 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1532 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1533 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1534 setOperationAction(ISD::MLOAD, VT, Legal);
1535 setOperationAction(ISD::MSTORE, VT, Legal);
1538 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1539 MVT VT = (MVT::SimpleValueType)i;
1541 // Do not attempt to promote non-512-bit vectors.
1542 if (!VT.is512BitVector())
1545 setOperationAction(ISD::SELECT, VT, Promote);
1546 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1550 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1551 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1552 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1554 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1555 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1557 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1558 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1559 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1560 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1561 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1562 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1563 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1564 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1565 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1566 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1567 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1568 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1569 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1570 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1571 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1572 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1573 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1574 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1575 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1576 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1577 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1578 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1579 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1580 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1581 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1582 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1583 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1584 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1585 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1586 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1588 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1589 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1590 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1591 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1592 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1593 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1594 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1595 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1597 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1598 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1599 if (Subtarget->hasVLX())
1600 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1602 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1603 const MVT VT = (MVT::SimpleValueType)i;
1605 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1607 // Do not attempt to promote non-512-bit vectors.
1608 if (!VT.is512BitVector())
1612 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1613 setOperationAction(ISD::VSELECT, VT, Legal);
1618 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1619 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1620 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1622 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1623 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1624 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1625 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1626 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1627 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1628 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1629 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1630 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1631 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1633 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1634 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1635 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1636 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1637 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1638 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1639 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1640 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1642 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1643 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1644 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1645 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1646 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1647 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1648 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1649 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1652 // We want to custom lower some of our intrinsics.
1653 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1654 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1655 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1656 if (!Subtarget->is64Bit())
1657 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1659 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1660 // handle type legalization for these operations here.
1662 // FIXME: We really should do custom legalization for addition and
1663 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1664 // than generic legalization for 64-bit multiplication-with-overflow, though.
1665 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1666 // Add/Sub/Mul with overflow operations are custom lowered.
1668 setOperationAction(ISD::SADDO, VT, Custom);
1669 setOperationAction(ISD::UADDO, VT, Custom);
1670 setOperationAction(ISD::SSUBO, VT, Custom);
1671 setOperationAction(ISD::USUBO, VT, Custom);
1672 setOperationAction(ISD::SMULO, VT, Custom);
1673 setOperationAction(ISD::UMULO, VT, Custom);
1677 if (!Subtarget->is64Bit()) {
1678 // These libcalls are not available in 32-bit.
1679 setLibcallName(RTLIB::SHL_I128, nullptr);
1680 setLibcallName(RTLIB::SRL_I128, nullptr);
1681 setLibcallName(RTLIB::SRA_I128, nullptr);
1684 // Combine sin / cos into one node or libcall if possible.
1685 if (Subtarget->hasSinCos()) {
1686 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1687 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1688 if (Subtarget->isTargetDarwin()) {
1689 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1690 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1691 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1692 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1696 if (Subtarget->isTargetWin64()) {
1697 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1698 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1699 setOperationAction(ISD::SREM, MVT::i128, Custom);
1700 setOperationAction(ISD::UREM, MVT::i128, Custom);
1701 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1702 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1705 // We have target-specific dag combine patterns for the following nodes:
1706 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1707 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1708 setTargetDAGCombine(ISD::BITCAST);
1709 setTargetDAGCombine(ISD::VSELECT);
1710 setTargetDAGCombine(ISD::SELECT);
1711 setTargetDAGCombine(ISD::SHL);
1712 setTargetDAGCombine(ISD::SRA);
1713 setTargetDAGCombine(ISD::SRL);
1714 setTargetDAGCombine(ISD::OR);
1715 setTargetDAGCombine(ISD::AND);
1716 setTargetDAGCombine(ISD::ADD);
1717 setTargetDAGCombine(ISD::FADD);
1718 setTargetDAGCombine(ISD::FSUB);
1719 setTargetDAGCombine(ISD::FMA);
1720 setTargetDAGCombine(ISD::SUB);
1721 setTargetDAGCombine(ISD::LOAD);
1722 setTargetDAGCombine(ISD::MLOAD);
1723 setTargetDAGCombine(ISD::STORE);
1724 setTargetDAGCombine(ISD::MSTORE);
1725 setTargetDAGCombine(ISD::ZERO_EXTEND);
1726 setTargetDAGCombine(ISD::ANY_EXTEND);
1727 setTargetDAGCombine(ISD::SIGN_EXTEND);
1728 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1729 setTargetDAGCombine(ISD::SINT_TO_FP);
1730 setTargetDAGCombine(ISD::UINT_TO_FP);
1731 setTargetDAGCombine(ISD::SETCC);
1732 setTargetDAGCombine(ISD::BUILD_VECTOR);
1733 setTargetDAGCombine(ISD::MUL);
1734 setTargetDAGCombine(ISD::XOR);
1736 computeRegisterProperties(Subtarget->getRegisterInfo());
1738 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1739 MaxStoresPerMemsetOptSize = 8;
1740 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1741 MaxStoresPerMemcpyOptSize = 4;
1742 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1743 MaxStoresPerMemmoveOptSize = 4;
1744 setPrefLoopAlignment(4); // 2^4 bytes.
1746 // Predictable cmov don't hurt on atom because it's in-order.
1747 PredictableSelectIsExpensive = !Subtarget->isAtom();
1748 EnableExtLdPromotion = true;
1749 setPrefFunctionAlignment(4); // 2^4 bytes.
1751 verifyIntrinsicTables();
1754 // This has so far only been implemented for 64-bit MachO.
1755 bool X86TargetLowering::useLoadStackGuardNode() const {
1756 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1759 TargetLoweringBase::LegalizeTypeAction
1760 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1761 if (ExperimentalVectorWideningLegalization &&
1762 VT.getVectorNumElements() != 1 &&
1763 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1764 return TypeWidenVector;
1766 return TargetLoweringBase::getPreferredVectorAction(VT);
1769 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1772 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1774 const unsigned NumElts = VT.getVectorNumElements();
1775 const EVT EltVT = VT.getVectorElementType();
1776 if (VT.is512BitVector()) {
1777 if (Subtarget->hasAVX512())
1778 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1779 EltVT == MVT::f32 || EltVT == MVT::f64)
1781 case 8: return MVT::v8i1;
1782 case 16: return MVT::v16i1;
1784 if (Subtarget->hasBWI())
1785 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1787 case 32: return MVT::v32i1;
1788 case 64: return MVT::v64i1;
1792 if (VT.is256BitVector() || VT.is128BitVector()) {
1793 if (Subtarget->hasVLX())
1794 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1795 EltVT == MVT::f32 || EltVT == MVT::f64)
1797 case 2: return MVT::v2i1;
1798 case 4: return MVT::v4i1;
1799 case 8: return MVT::v8i1;
1801 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1802 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1804 case 8: return MVT::v8i1;
1805 case 16: return MVT::v16i1;
1806 case 32: return MVT::v32i1;
1810 return VT.changeVectorElementTypeToInteger();
1813 /// Helper for getByValTypeAlignment to determine
1814 /// the desired ByVal argument alignment.
1815 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1818 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1819 if (VTy->getBitWidth() == 128)
1821 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1822 unsigned EltAlign = 0;
1823 getMaxByValAlign(ATy->getElementType(), EltAlign);
1824 if (EltAlign > MaxAlign)
1825 MaxAlign = EltAlign;
1826 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1827 for (auto *EltTy : STy->elements()) {
1828 unsigned EltAlign = 0;
1829 getMaxByValAlign(EltTy, EltAlign);
1830 if (EltAlign > MaxAlign)
1831 MaxAlign = EltAlign;
1838 /// Return the desired alignment for ByVal aggregate
1839 /// function arguments in the caller parameter area. For X86, aggregates
1840 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1841 /// are at 4-byte boundaries.
1842 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1843 const DataLayout &DL) const {
1844 if (Subtarget->is64Bit()) {
1845 // Max of 8 and alignment of type.
1846 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1853 if (Subtarget->hasSSE1())
1854 getMaxByValAlign(Ty, Align);
1858 /// Returns the target specific optimal type for load
1859 /// and store operations as a result of memset, memcpy, and memmove
1860 /// lowering. If DstAlign is zero that means it's safe to destination
1861 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1862 /// means there isn't a need to check it against alignment requirement,
1863 /// probably because the source does not need to be loaded. If 'IsMemset' is
1864 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1865 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1866 /// source is constant so it does not need to be loaded.
1867 /// It returns EVT::Other if the type should be determined using generic
1868 /// target-independent logic.
1870 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1871 unsigned DstAlign, unsigned SrcAlign,
1872 bool IsMemset, bool ZeroMemset,
1874 MachineFunction &MF) const {
1875 const Function *F = MF.getFunction();
1876 if ((!IsMemset || ZeroMemset) &&
1877 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1879 (Subtarget->isUnalignedMemAccessFast() ||
1880 ((DstAlign == 0 || DstAlign >= 16) &&
1881 (SrcAlign == 0 || SrcAlign >= 16)))) {
1883 if (Subtarget->hasInt256())
1885 if (Subtarget->hasFp256())
1888 if (Subtarget->hasSSE2())
1890 if (Subtarget->hasSSE1())
1892 } else if (!MemcpyStrSrc && Size >= 8 &&
1893 !Subtarget->is64Bit() &&
1894 Subtarget->hasSSE2()) {
1895 // Do not use f64 to lower memcpy if source is string constant. It's
1896 // better to use i32 to avoid the loads.
1900 if (Subtarget->is64Bit() && Size >= 8)
1905 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1907 return X86ScalarSSEf32;
1908 else if (VT == MVT::f64)
1909 return X86ScalarSSEf64;
1914 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1919 *Fast = Subtarget->isUnalignedMemAccessFast();
1923 /// Return the entry encoding for a jump table in the
1924 /// current function. The returned value is a member of the
1925 /// MachineJumpTableInfo::JTEntryKind enum.
1926 unsigned X86TargetLowering::getJumpTableEncoding() const {
1927 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1929 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1930 Subtarget->isPICStyleGOT())
1931 return MachineJumpTableInfo::EK_Custom32;
1933 // Otherwise, use the normal jump table encoding heuristics.
1934 return TargetLowering::getJumpTableEncoding();
1937 bool X86TargetLowering::useSoftFloat() const {
1938 return Subtarget->useSoftFloat();
1942 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1943 const MachineBasicBlock *MBB,
1944 unsigned uid,MCContext &Ctx) const{
1945 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1946 Subtarget->isPICStyleGOT());
1947 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1949 return MCSymbolRefExpr::create(MBB->getSymbol(),
1950 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1953 /// Returns relocation base for the given PIC jumptable.
1954 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1955 SelectionDAG &DAG) const {
1956 if (!Subtarget->is64Bit())
1957 // This doesn't have SDLoc associated with it, but is not really the
1958 // same as a Register.
1959 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
1960 getPointerTy(DAG.getDataLayout()));
1964 /// This returns the relocation base for the given PIC jumptable,
1965 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1966 const MCExpr *X86TargetLowering::
1967 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1968 MCContext &Ctx) const {
1969 // X86-64 uses RIP relative addressing based on the jump table label.
1970 if (Subtarget->isPICStyleRIPRel())
1971 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1973 // Otherwise, the reference is relative to the PIC base.
1974 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
1977 std::pair<const TargetRegisterClass *, uint8_t>
1978 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1980 const TargetRegisterClass *RRC = nullptr;
1982 switch (VT.SimpleTy) {
1984 return TargetLowering::findRepresentativeClass(TRI, VT);
1985 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1986 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1989 RRC = &X86::VR64RegClass;
1991 case MVT::f32: case MVT::f64:
1992 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1993 case MVT::v4f32: case MVT::v2f64:
1994 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1996 RRC = &X86::VR128RegClass;
1999 return std::make_pair(RRC, Cost);
2002 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2003 unsigned &Offset) const {
2004 if (!Subtarget->isTargetLinux())
2007 if (Subtarget->is64Bit()) {
2008 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2010 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2022 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2023 unsigned DestAS) const {
2024 assert(SrcAS != DestAS && "Expected different address spaces!");
2026 return SrcAS < 256 && DestAS < 256;
2029 //===----------------------------------------------------------------------===//
2030 // Return Value Calling Convention Implementation
2031 //===----------------------------------------------------------------------===//
2033 #include "X86GenCallingConv.inc"
2036 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2037 MachineFunction &MF, bool isVarArg,
2038 const SmallVectorImpl<ISD::OutputArg> &Outs,
2039 LLVMContext &Context) const {
2040 SmallVector<CCValAssign, 16> RVLocs;
2041 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2042 return CCInfo.CheckReturn(Outs, RetCC_X86);
2045 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2046 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2051 X86TargetLowering::LowerReturn(SDValue Chain,
2052 CallingConv::ID CallConv, bool isVarArg,
2053 const SmallVectorImpl<ISD::OutputArg> &Outs,
2054 const SmallVectorImpl<SDValue> &OutVals,
2055 SDLoc dl, SelectionDAG &DAG) const {
2056 MachineFunction &MF = DAG.getMachineFunction();
2057 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2059 SmallVector<CCValAssign, 16> RVLocs;
2060 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2061 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2064 SmallVector<SDValue, 6> RetOps;
2065 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2066 // Operand #1 = Bytes To Pop
2067 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2070 // Copy the result values into the output registers.
2071 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2072 CCValAssign &VA = RVLocs[i];
2073 assert(VA.isRegLoc() && "Can only return in registers!");
2074 SDValue ValToCopy = OutVals[i];
2075 EVT ValVT = ValToCopy.getValueType();
2077 // Promote values to the appropriate types.
2078 if (VA.getLocInfo() == CCValAssign::SExt)
2079 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2080 else if (VA.getLocInfo() == CCValAssign::ZExt)
2081 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2082 else if (VA.getLocInfo() == CCValAssign::AExt) {
2083 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2084 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2086 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2088 else if (VA.getLocInfo() == CCValAssign::BCvt)
2089 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2091 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2092 "Unexpected FP-extend for return value.");
2094 // If this is x86-64, and we disabled SSE, we can't return FP values,
2095 // or SSE or MMX vectors.
2096 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2097 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2098 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2099 report_fatal_error("SSE register return with SSE disabled");
2101 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2102 // llvm-gcc has never done it right and no one has noticed, so this
2103 // should be OK for now.
2104 if (ValVT == MVT::f64 &&
2105 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2106 report_fatal_error("SSE2 register return with SSE2 disabled");
2108 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2109 // the RET instruction and handled by the FP Stackifier.
2110 if (VA.getLocReg() == X86::FP0 ||
2111 VA.getLocReg() == X86::FP1) {
2112 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2113 // change the value to the FP stack register class.
2114 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2115 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2116 RetOps.push_back(ValToCopy);
2117 // Don't emit a copytoreg.
2121 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2122 // which is returned in RAX / RDX.
2123 if (Subtarget->is64Bit()) {
2124 if (ValVT == MVT::x86mmx) {
2125 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2126 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2127 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2129 // If we don't have SSE2 available, convert to v4f32 so the generated
2130 // register is legal.
2131 if (!Subtarget->hasSSE2())
2132 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2137 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2138 Flag = Chain.getValue(1);
2139 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2142 // All x86 ABIs require that for returning structs by value we copy
2143 // the sret argument into %rax/%eax (depending on ABI) for the return.
2144 // We saved the argument into a virtual register in the entry block,
2145 // so now we copy the value out and into %rax/%eax.
2147 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2148 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2149 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2150 // either case FuncInfo->setSRetReturnReg() will have been called.
2151 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2152 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2153 getPointerTy(MF.getDataLayout()));
2156 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2157 X86::RAX : X86::EAX;
2158 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2159 Flag = Chain.getValue(1);
2161 // RAX/EAX now acts like a return value.
2163 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2166 RetOps[0] = Chain; // Update chain.
2168 // Add the flag if we have it.
2170 RetOps.push_back(Flag);
2172 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2175 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2176 if (N->getNumValues() != 1)
2178 if (!N->hasNUsesOfValue(1, 0))
2181 SDValue TCChain = Chain;
2182 SDNode *Copy = *N->use_begin();
2183 if (Copy->getOpcode() == ISD::CopyToReg) {
2184 // If the copy has a glue operand, we conservatively assume it isn't safe to
2185 // perform a tail call.
2186 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2188 TCChain = Copy->getOperand(0);
2189 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2192 bool HasRet = false;
2193 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2195 if (UI->getOpcode() != X86ISD::RET_FLAG)
2197 // If we are returning more than one value, we can definitely
2198 // not make a tail call see PR19530
2199 if (UI->getNumOperands() > 4)
2201 if (UI->getNumOperands() == 4 &&
2202 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2215 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2216 ISD::NodeType ExtendKind) const {
2218 // TODO: Is this also valid on 32-bit?
2219 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2220 ReturnMVT = MVT::i8;
2222 ReturnMVT = MVT::i32;
2224 EVT MinVT = getRegisterType(Context, ReturnMVT);
2225 return VT.bitsLT(MinVT) ? MinVT : VT;
2228 /// Lower the result values of a call into the
2229 /// appropriate copies out of appropriate physical registers.
2232 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2233 CallingConv::ID CallConv, bool isVarArg,
2234 const SmallVectorImpl<ISD::InputArg> &Ins,
2235 SDLoc dl, SelectionDAG &DAG,
2236 SmallVectorImpl<SDValue> &InVals) const {
2238 // Assign locations to each value returned by this call.
2239 SmallVector<CCValAssign, 16> RVLocs;
2240 bool Is64Bit = Subtarget->is64Bit();
2241 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2243 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2245 // Copy all of the result registers out of their specified physreg.
2246 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2247 CCValAssign &VA = RVLocs[i];
2248 EVT CopyVT = VA.getLocVT();
2250 // If this is x86-64, and we disabled SSE, we can't return FP values
2251 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2252 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2253 report_fatal_error("SSE register return with SSE disabled");
2256 // If we prefer to use the value in xmm registers, copy it out as f80 and
2257 // use a truncate to move it from fp stack reg to xmm reg.
2258 bool RoundAfterCopy = false;
2259 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2260 isScalarFPTypeInSSEReg(VA.getValVT())) {
2262 RoundAfterCopy = (CopyVT != VA.getLocVT());
2265 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2266 CopyVT, InFlag).getValue(1);
2267 SDValue Val = Chain.getValue(0);
2270 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2271 // This truncation won't change the value.
2272 DAG.getIntPtrConstant(1, dl));
2274 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2275 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2277 InFlag = Chain.getValue(2);
2278 InVals.push_back(Val);
2284 //===----------------------------------------------------------------------===//
2285 // C & StdCall & Fast Calling Convention implementation
2286 //===----------------------------------------------------------------------===//
2287 // StdCall calling convention seems to be standard for many Windows' API
2288 // routines and around. It differs from C calling convention just a little:
2289 // callee should clean up the stack, not caller. Symbols should be also
2290 // decorated in some fancy way :) It doesn't support any vector arguments.
2291 // For info on fast calling convention see Fast Calling Convention (tail call)
2292 // implementation LowerX86_32FastCCCallTo.
2294 /// CallIsStructReturn - Determines whether a call uses struct return
2296 enum StructReturnType {
2301 static StructReturnType
2302 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2304 return NotStructReturn;
2306 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2307 if (!Flags.isSRet())
2308 return NotStructReturn;
2309 if (Flags.isInReg())
2310 return RegStructReturn;
2311 return StackStructReturn;
2314 /// Determines whether a function uses struct return semantics.
2315 static StructReturnType
2316 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2318 return NotStructReturn;
2320 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2321 if (!Flags.isSRet())
2322 return NotStructReturn;
2323 if (Flags.isInReg())
2324 return RegStructReturn;
2325 return StackStructReturn;
2328 /// Make a copy of an aggregate at address specified by "Src" to address
2329 /// "Dst" with size and alignment information specified by the specific
2330 /// parameter attribute. The copy will be passed as a byval function parameter.
2332 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2333 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2335 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2337 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2338 /*isVolatile*/false, /*AlwaysInline=*/true,
2339 /*isTailCall*/false,
2340 MachinePointerInfo(), MachinePointerInfo());
2343 /// Return true if the calling convention is one that
2344 /// supports tail call optimization.
2345 static bool IsTailCallConvention(CallingConv::ID CC) {
2346 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2347 CC == CallingConv::HiPE);
2350 /// \brief Return true if the calling convention is a C calling convention.
2351 static bool IsCCallConvention(CallingConv::ID CC) {
2352 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2353 CC == CallingConv::X86_64_SysV);
2356 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2358 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2359 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2363 CallingConv::ID CalleeCC = CS.getCallingConv();
2364 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2370 /// Return true if the function is being made into
2371 /// a tailcall target by changing its ABI.
2372 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2373 bool GuaranteedTailCallOpt) {
2374 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2378 X86TargetLowering::LowerMemArgument(SDValue Chain,
2379 CallingConv::ID CallConv,
2380 const SmallVectorImpl<ISD::InputArg> &Ins,
2381 SDLoc dl, SelectionDAG &DAG,
2382 const CCValAssign &VA,
2383 MachineFrameInfo *MFI,
2385 // Create the nodes corresponding to a load from this parameter slot.
2386 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2387 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2388 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2389 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2392 // If value is passed by pointer we have address passed instead of the value
2394 bool ExtendedInMem = VA.isExtInLoc() &&
2395 VA.getValVT().getScalarType() == MVT::i1;
2397 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2398 ValVT = VA.getLocVT();
2400 ValVT = VA.getValVT();
2402 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2403 // changed with more analysis.
2404 // In case of tail call optimization mark all arguments mutable. Since they
2405 // could be overwritten by lowering of arguments in case of a tail call.
2406 if (Flags.isByVal()) {
2407 unsigned Bytes = Flags.getByValSize();
2408 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2409 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2410 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2412 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2413 VA.getLocMemOffset(), isImmutable);
2414 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2415 SDValue Val = DAG.getLoad(
2416 ValVT, dl, Chain, FIN,
2417 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2419 return ExtendedInMem ?
2420 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2424 // FIXME: Get this from tablegen.
2425 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2426 const X86Subtarget *Subtarget) {
2427 assert(Subtarget->is64Bit());
2429 if (Subtarget->isCallingConvWin64(CallConv)) {
2430 static const MCPhysReg GPR64ArgRegsWin64[] = {
2431 X86::RCX, X86::RDX, X86::R8, X86::R9
2433 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2436 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2437 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2439 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2442 // FIXME: Get this from tablegen.
2443 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2444 CallingConv::ID CallConv,
2445 const X86Subtarget *Subtarget) {
2446 assert(Subtarget->is64Bit());
2447 if (Subtarget->isCallingConvWin64(CallConv)) {
2448 // The XMM registers which might contain var arg parameters are shadowed
2449 // in their paired GPR. So we only need to save the GPR to their home
2451 // TODO: __vectorcall will change this.
2455 const Function *Fn = MF.getFunction();
2456 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2457 bool isSoftFloat = Subtarget->useSoftFloat();
2458 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2459 "SSE register cannot be used when SSE is disabled!");
2460 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2461 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2465 static const MCPhysReg XMMArgRegs64Bit[] = {
2466 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2467 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2469 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2473 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2474 CallingConv::ID CallConv,
2476 const SmallVectorImpl<ISD::InputArg> &Ins,
2479 SmallVectorImpl<SDValue> &InVals)
2481 MachineFunction &MF = DAG.getMachineFunction();
2482 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2483 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2485 const Function* Fn = MF.getFunction();
2486 if (Fn->hasExternalLinkage() &&
2487 Subtarget->isTargetCygMing() &&
2488 Fn->getName() == "main")
2489 FuncInfo->setForceFramePointer(true);
2491 MachineFrameInfo *MFI = MF.getFrameInfo();
2492 bool Is64Bit = Subtarget->is64Bit();
2493 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2495 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2496 "Var args not supported with calling convention fastcc, ghc or hipe");
2498 // Assign locations to all of the incoming arguments.
2499 SmallVector<CCValAssign, 16> ArgLocs;
2500 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2502 // Allocate shadow area for Win64
2504 CCInfo.AllocateStack(32, 8);
2506 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2508 unsigned LastVal = ~0U;
2510 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2511 CCValAssign &VA = ArgLocs[i];
2512 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2514 assert(VA.getValNo() != LastVal &&
2515 "Don't support value assigned to multiple locs yet");
2517 LastVal = VA.getValNo();
2519 if (VA.isRegLoc()) {
2520 EVT RegVT = VA.getLocVT();
2521 const TargetRegisterClass *RC;
2522 if (RegVT == MVT::i32)
2523 RC = &X86::GR32RegClass;
2524 else if (Is64Bit && RegVT == MVT::i64)
2525 RC = &X86::GR64RegClass;
2526 else if (RegVT == MVT::f32)
2527 RC = &X86::FR32RegClass;
2528 else if (RegVT == MVT::f64)
2529 RC = &X86::FR64RegClass;
2530 else if (RegVT.is512BitVector())
2531 RC = &X86::VR512RegClass;
2532 else if (RegVT.is256BitVector())
2533 RC = &X86::VR256RegClass;
2534 else if (RegVT.is128BitVector())
2535 RC = &X86::VR128RegClass;
2536 else if (RegVT == MVT::x86mmx)
2537 RC = &X86::VR64RegClass;
2538 else if (RegVT == MVT::i1)
2539 RC = &X86::VK1RegClass;
2540 else if (RegVT == MVT::v8i1)
2541 RC = &X86::VK8RegClass;
2542 else if (RegVT == MVT::v16i1)
2543 RC = &X86::VK16RegClass;
2544 else if (RegVT == MVT::v32i1)
2545 RC = &X86::VK32RegClass;
2546 else if (RegVT == MVT::v64i1)
2547 RC = &X86::VK64RegClass;
2549 llvm_unreachable("Unknown argument type!");
2551 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2552 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2554 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2555 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2557 if (VA.getLocInfo() == CCValAssign::SExt)
2558 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2559 DAG.getValueType(VA.getValVT()));
2560 else if (VA.getLocInfo() == CCValAssign::ZExt)
2561 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2562 DAG.getValueType(VA.getValVT()));
2563 else if (VA.getLocInfo() == CCValAssign::BCvt)
2564 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2566 if (VA.isExtInLoc()) {
2567 // Handle MMX values passed in XMM regs.
2568 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2569 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2571 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2574 assert(VA.isMemLoc());
2575 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2578 // If value is passed via pointer - do a load.
2579 if (VA.getLocInfo() == CCValAssign::Indirect)
2580 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2581 MachinePointerInfo(), false, false, false, 0);
2583 InVals.push_back(ArgValue);
2586 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2587 // All x86 ABIs require that for returning structs by value we copy the
2588 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2589 // the argument into a virtual register so that we can access it from the
2591 if (Ins[i].Flags.isSRet()) {
2592 unsigned Reg = FuncInfo->getSRetReturnReg();
2594 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2595 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2596 FuncInfo->setSRetReturnReg(Reg);
2598 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2599 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2604 unsigned StackSize = CCInfo.getNextStackOffset();
2605 // Align stack specially for tail calls.
2606 if (FuncIsMadeTailCallSafe(CallConv,
2607 MF.getTarget().Options.GuaranteedTailCallOpt))
2608 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2610 // If the function takes variable number of arguments, make a frame index for
2611 // the start of the first vararg value... for expansion of llvm.va_start. We
2612 // can skip this if there are no va_start calls.
2613 if (MFI->hasVAStart() &&
2614 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2615 CallConv != CallingConv::X86_ThisCall))) {
2616 FuncInfo->setVarArgsFrameIndex(
2617 MFI->CreateFixedObject(1, StackSize, true));
2620 MachineModuleInfo &MMI = MF.getMMI();
2621 const Function *WinEHParent = nullptr;
2622 if (MMI.hasWinEHFuncInfo(Fn))
2623 WinEHParent = MMI.getWinEHParent(Fn);
2624 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2625 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2627 // Figure out if XMM registers are in use.
2628 assert(!(Subtarget->useSoftFloat() &&
2629 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2630 "SSE register cannot be used when SSE is disabled!");
2632 // 64-bit calling conventions support varargs and register parameters, so we
2633 // have to do extra work to spill them in the prologue.
2634 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2635 // Find the first unallocated argument registers.
2636 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2637 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2638 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2639 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2640 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2641 "SSE register cannot be used when SSE is disabled!");
2643 // Gather all the live in physical registers.
2644 SmallVector<SDValue, 6> LiveGPRs;
2645 SmallVector<SDValue, 8> LiveXMMRegs;
2647 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2648 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2650 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2652 if (!ArgXMMs.empty()) {
2653 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2654 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2655 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2656 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2657 LiveXMMRegs.push_back(
2658 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2663 // Get to the caller-allocated home save location. Add 8 to account
2664 // for the return address.
2665 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2666 FuncInfo->setRegSaveFrameIndex(
2667 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2668 // Fixup to set vararg frame on shadow area (4 x i64).
2670 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2672 // For X86-64, if there are vararg parameters that are passed via
2673 // registers, then we must store them to their spots on the stack so
2674 // they may be loaded by deferencing the result of va_next.
2675 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2676 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2677 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2678 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2681 // Store the integer parameter registers.
2682 SmallVector<SDValue, 8> MemOps;
2683 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2684 getPointerTy(DAG.getDataLayout()));
2685 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2686 for (SDValue Val : LiveGPRs) {
2687 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2688 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2690 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2691 MachinePointerInfo::getFixedStack(
2692 DAG.getMachineFunction(),
2693 FuncInfo->getRegSaveFrameIndex(), Offset),
2695 MemOps.push_back(Store);
2699 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2700 // Now store the XMM (fp + vector) parameter registers.
2701 SmallVector<SDValue, 12> SaveXMMOps;
2702 SaveXMMOps.push_back(Chain);
2703 SaveXMMOps.push_back(ALVal);
2704 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2705 FuncInfo->getRegSaveFrameIndex(), dl));
2706 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2707 FuncInfo->getVarArgsFPOffset(), dl));
2708 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2710 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2711 MVT::Other, SaveXMMOps));
2714 if (!MemOps.empty())
2715 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2716 } else if (IsWin64 && IsWinEHOutlined) {
2717 // Get to the caller-allocated home save location. Add 8 to account
2718 // for the return address.
2719 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2720 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2721 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2723 MMI.getWinEHFuncInfo(Fn)
2724 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2725 FuncInfo->getRegSaveFrameIndex();
2727 // Store the second integer parameter (rdx) into rsp+16 relative to the
2728 // stack pointer at the entry of the function.
2729 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2730 getPointerTy(DAG.getDataLayout()));
2731 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2732 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2733 Chain = DAG.getStore(
2734 Val.getValue(1), dl, Val, RSFIN,
2735 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
2736 FuncInfo->getRegSaveFrameIndex()),
2737 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2740 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2741 // Find the largest legal vector type.
2742 MVT VecVT = MVT::Other;
2743 // FIXME: Only some x86_32 calling conventions support AVX512.
2744 if (Subtarget->hasAVX512() &&
2745 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2746 CallConv == CallingConv::Intel_OCL_BI)))
2747 VecVT = MVT::v16f32;
2748 else if (Subtarget->hasAVX())
2750 else if (Subtarget->hasSSE2())
2753 // We forward some GPRs and some vector types.
2754 SmallVector<MVT, 2> RegParmTypes;
2755 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2756 RegParmTypes.push_back(IntVT);
2757 if (VecVT != MVT::Other)
2758 RegParmTypes.push_back(VecVT);
2760 // Compute the set of forwarded registers. The rest are scratch.
2761 SmallVectorImpl<ForwardedRegister> &Forwards =
2762 FuncInfo->getForwardedMustTailRegParms();
2763 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2765 // Conservatively forward AL on x86_64, since it might be used for varargs.
2766 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2767 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2768 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2771 // Copy all forwards from physical to virtual registers.
2772 for (ForwardedRegister &F : Forwards) {
2773 // FIXME: Can we use a less constrained schedule?
2774 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2775 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2776 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2780 // Some CCs need callee pop.
2781 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2782 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2783 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2785 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2786 // If this is an sret function, the return should pop the hidden pointer.
2787 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2788 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2789 argsAreStructReturn(Ins) == StackStructReturn)
2790 FuncInfo->setBytesToPopOnReturn(4);
2794 // RegSaveFrameIndex is X86-64 only.
2795 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2796 if (CallConv == CallingConv::X86_FastCall ||
2797 CallConv == CallingConv::X86_ThisCall)
2798 // fastcc functions can't have varargs.
2799 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2802 FuncInfo->setArgumentStackSize(StackSize);
2804 if (IsWinEHParent) {
2806 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2807 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2808 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2809 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2810 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2811 MachinePointerInfo::getFixedStack(
2812 DAG.getMachineFunction(), UnwindHelpFI),
2813 /*isVolatile=*/true,
2814 /*isNonTemporal=*/false, /*Alignment=*/0);
2816 // Functions using Win32 EH are considered to have opaque SP adjustments
2817 // to force local variables to be addressed from the frame or base
2819 MFI->setHasOpaqueSPAdjustment(true);
2827 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2828 SDValue StackPtr, SDValue Arg,
2829 SDLoc dl, SelectionDAG &DAG,
2830 const CCValAssign &VA,
2831 ISD::ArgFlagsTy Flags) const {
2832 unsigned LocMemOffset = VA.getLocMemOffset();
2833 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2834 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2836 if (Flags.isByVal())
2837 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2839 return DAG.getStore(
2840 Chain, dl, Arg, PtrOff,
2841 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2845 /// Emit a load of return address if tail call
2846 /// optimization is performed and it is required.
2848 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2849 SDValue &OutRetAddr, SDValue Chain,
2850 bool IsTailCall, bool Is64Bit,
2851 int FPDiff, SDLoc dl) const {
2852 // Adjust the Return address stack slot.
2853 EVT VT = getPointerTy(DAG.getDataLayout());
2854 OutRetAddr = getReturnAddressFrameIndex(DAG);
2856 // Load the "old" Return address.
2857 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2858 false, false, false, 0);
2859 return SDValue(OutRetAddr.getNode(), 1);
2862 /// Emit a store of the return address if tail call
2863 /// optimization is performed and it is required (FPDiff!=0).
2864 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2865 SDValue Chain, SDValue RetAddrFrIdx,
2866 EVT PtrVT, unsigned SlotSize,
2867 int FPDiff, SDLoc dl) {
2868 // Store the return address to the appropriate stack slot.
2869 if (!FPDiff) return Chain;
2870 // Calculate the new stack slot for the return address.
2871 int NewReturnAddrFI =
2872 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2874 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2875 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2876 MachinePointerInfo::getFixedStack(
2877 DAG.getMachineFunction(), NewReturnAddrFI),
2882 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2883 /// operation of specified width.
2884 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2886 unsigned NumElems = VT.getVectorNumElements();
2887 SmallVector<int, 8> Mask;
2888 Mask.push_back(NumElems);
2889 for (unsigned i = 1; i != NumElems; ++i)
2891 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2895 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2896 SmallVectorImpl<SDValue> &InVals) const {
2897 SelectionDAG &DAG = CLI.DAG;
2899 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2900 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2901 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2902 SDValue Chain = CLI.Chain;
2903 SDValue Callee = CLI.Callee;
2904 CallingConv::ID CallConv = CLI.CallConv;
2905 bool &isTailCall = CLI.IsTailCall;
2906 bool isVarArg = CLI.IsVarArg;
2908 MachineFunction &MF = DAG.getMachineFunction();
2909 bool Is64Bit = Subtarget->is64Bit();
2910 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2911 StructReturnType SR = callIsStructReturn(Outs);
2912 bool IsSibcall = false;
2913 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2914 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2916 if (Attr.getValueAsString() == "true")
2919 if (Subtarget->isPICStyleGOT() &&
2920 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2921 // If we are using a GOT, disable tail calls to external symbols with
2922 // default visibility. Tail calling such a symbol requires using a GOT
2923 // relocation, which forces early binding of the symbol. This breaks code
2924 // that require lazy function symbol resolution. Using musttail or
2925 // GuaranteedTailCallOpt will override this.
2926 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2927 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2928 G->getGlobal()->hasDefaultVisibility()))
2932 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2934 // Force this to be a tail call. The verifier rules are enough to ensure
2935 // that we can lower this successfully without moving the return address
2938 } else if (isTailCall) {
2939 // Check if it's really possible to do a tail call.
2940 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2941 isVarArg, SR != NotStructReturn,
2942 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2943 Outs, OutVals, Ins, DAG);
2945 // Sibcalls are automatically detected tailcalls which do not require
2947 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2954 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2955 "Var args not supported with calling convention fastcc, ghc or hipe");
2957 // Analyze operands of the call, assigning locations to each operand.
2958 SmallVector<CCValAssign, 16> ArgLocs;
2959 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2961 // Allocate shadow area for Win64
2963 CCInfo.AllocateStack(32, 8);
2965 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2967 // Get a count of how many bytes are to be pushed on the stack.
2968 unsigned NumBytes = CCInfo.getNextStackOffset();
2970 // This is a sibcall. The memory operands are available in caller's
2971 // own caller's stack.
2973 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2974 IsTailCallConvention(CallConv))
2975 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2978 if (isTailCall && !IsSibcall && !IsMustTail) {
2979 // Lower arguments at fp - stackoffset + fpdiff.
2980 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2982 FPDiff = NumBytesCallerPushed - NumBytes;
2984 // Set the delta of movement of the returnaddr stackslot.
2985 // But only set if delta is greater than previous delta.
2986 if (FPDiff < X86Info->getTCReturnAddrDelta())
2987 X86Info->setTCReturnAddrDelta(FPDiff);
2990 unsigned NumBytesToPush = NumBytes;
2991 unsigned NumBytesToPop = NumBytes;
2993 // If we have an inalloca argument, all stack space has already been allocated
2994 // for us and be right at the top of the stack. We don't support multiple
2995 // arguments passed in memory when using inalloca.
2996 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2998 if (!ArgLocs.back().isMemLoc())
2999 report_fatal_error("cannot use inalloca attribute on a register "
3001 if (ArgLocs.back().getLocMemOffset() != 0)
3002 report_fatal_error("any parameter with the inalloca attribute must be "
3003 "the only memory argument");
3007 Chain = DAG.getCALLSEQ_START(
3008 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3010 SDValue RetAddrFrIdx;
3011 // Load return address for tail calls.
3012 if (isTailCall && FPDiff)
3013 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3014 Is64Bit, FPDiff, dl);
3016 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3017 SmallVector<SDValue, 8> MemOpChains;
3020 // Walk the register/memloc assignments, inserting copies/loads. In the case
3021 // of tail call optimization arguments are handle later.
3022 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3023 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3024 // Skip inalloca arguments, they have already been written.
3025 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3026 if (Flags.isInAlloca())
3029 CCValAssign &VA = ArgLocs[i];
3030 EVT RegVT = VA.getLocVT();
3031 SDValue Arg = OutVals[i];
3032 bool isByVal = Flags.isByVal();
3034 // Promote the value if needed.
3035 switch (VA.getLocInfo()) {
3036 default: llvm_unreachable("Unknown loc info!");
3037 case CCValAssign::Full: break;
3038 case CCValAssign::SExt:
3039 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3041 case CCValAssign::ZExt:
3042 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3044 case CCValAssign::AExt:
3045 if (Arg.getValueType().isVector() &&
3046 Arg.getValueType().getScalarType() == MVT::i1)
3047 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3048 else if (RegVT.is128BitVector()) {
3049 // Special case: passing MMX values in XMM registers.
3050 Arg = DAG.getBitcast(MVT::i64, Arg);
3051 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3052 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3054 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3056 case CCValAssign::BCvt:
3057 Arg = DAG.getBitcast(RegVT, Arg);
3059 case CCValAssign::Indirect: {
3060 // Store the argument.
3061 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3062 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3063 Chain = DAG.getStore(
3064 Chain, dl, Arg, SpillSlot,
3065 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3072 if (VA.isRegLoc()) {
3073 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3074 if (isVarArg && IsWin64) {
3075 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3076 // shadow reg if callee is a varargs function.
3077 unsigned ShadowReg = 0;
3078 switch (VA.getLocReg()) {
3079 case X86::XMM0: ShadowReg = X86::RCX; break;
3080 case X86::XMM1: ShadowReg = X86::RDX; break;
3081 case X86::XMM2: ShadowReg = X86::R8; break;
3082 case X86::XMM3: ShadowReg = X86::R9; break;
3085 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3087 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3088 assert(VA.isMemLoc());
3089 if (!StackPtr.getNode())
3090 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3091 getPointerTy(DAG.getDataLayout()));
3092 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3093 dl, DAG, VA, Flags));
3097 if (!MemOpChains.empty())
3098 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3100 if (Subtarget->isPICStyleGOT()) {
3101 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3104 RegsToPass.push_back(std::make_pair(
3105 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3106 getPointerTy(DAG.getDataLayout()))));
3108 // If we are tail calling and generating PIC/GOT style code load the
3109 // address of the callee into ECX. The value in ecx is used as target of
3110 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3111 // for tail calls on PIC/GOT architectures. Normally we would just put the
3112 // address of GOT into ebx and then call target@PLT. But for tail calls
3113 // ebx would be restored (since ebx is callee saved) before jumping to the
3116 // Note: The actual moving to ECX is done further down.
3117 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3118 if (G && !G->getGlobal()->hasLocalLinkage() &&
3119 G->getGlobal()->hasDefaultVisibility())
3120 Callee = LowerGlobalAddress(Callee, DAG);
3121 else if (isa<ExternalSymbolSDNode>(Callee))
3122 Callee = LowerExternalSymbol(Callee, DAG);
3126 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3127 // From AMD64 ABI document:
3128 // For calls that may call functions that use varargs or stdargs
3129 // (prototype-less calls or calls to functions containing ellipsis (...) in
3130 // the declaration) %al is used as hidden argument to specify the number
3131 // of SSE registers used. The contents of %al do not need to match exactly
3132 // the number of registers, but must be an ubound on the number of SSE
3133 // registers used and is in the range 0 - 8 inclusive.
3135 // Count the number of XMM registers allocated.
3136 static const MCPhysReg XMMArgRegs[] = {
3137 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3138 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3140 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3141 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3142 && "SSE registers cannot be used when SSE is disabled");
3144 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3145 DAG.getConstant(NumXMMRegs, dl,
3149 if (isVarArg && IsMustTail) {
3150 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3151 for (const auto &F : Forwards) {
3152 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3153 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3157 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3158 // don't need this because the eligibility check rejects calls that require
3159 // shuffling arguments passed in memory.
3160 if (!IsSibcall && isTailCall) {
3161 // Force all the incoming stack arguments to be loaded from the stack
3162 // before any new outgoing arguments are stored to the stack, because the
3163 // outgoing stack slots may alias the incoming argument stack slots, and
3164 // the alias isn't otherwise explicit. This is slightly more conservative
3165 // than necessary, because it means that each store effectively depends
3166 // on every argument instead of just those arguments it would clobber.
3167 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3169 SmallVector<SDValue, 8> MemOpChains2;
3172 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3173 CCValAssign &VA = ArgLocs[i];
3176 assert(VA.isMemLoc());
3177 SDValue Arg = OutVals[i];
3178 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3179 // Skip inalloca arguments. They don't require any work.
3180 if (Flags.isInAlloca())
3182 // Create frame index.
3183 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3184 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3185 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3186 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3188 if (Flags.isByVal()) {
3189 // Copy relative to framepointer.
3190 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3191 if (!StackPtr.getNode())
3192 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3193 getPointerTy(DAG.getDataLayout()));
3194 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3197 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3201 // Store relative to framepointer.
3202 MemOpChains2.push_back(DAG.getStore(
3203 ArgChain, dl, Arg, FIN,
3204 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3209 if (!MemOpChains2.empty())
3210 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3212 // Store the return address to the appropriate stack slot.
3213 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3214 getPointerTy(DAG.getDataLayout()),
3215 RegInfo->getSlotSize(), FPDiff, dl);
3218 // Build a sequence of copy-to-reg nodes chained together with token chain
3219 // and flag operands which copy the outgoing args into registers.
3221 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3222 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3223 RegsToPass[i].second, InFlag);
3224 InFlag = Chain.getValue(1);
3227 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3228 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3229 // In the 64-bit large code model, we have to make all calls
3230 // through a register, since the call instruction's 32-bit
3231 // pc-relative offset may not be large enough to hold the whole
3233 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3234 // If the callee is a GlobalAddress node (quite common, every direct call
3235 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3237 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3239 // We should use extra load for direct calls to dllimported functions in
3241 const GlobalValue *GV = G->getGlobal();
3242 if (!GV->hasDLLImportStorageClass()) {
3243 unsigned char OpFlags = 0;
3244 bool ExtraLoad = false;
3245 unsigned WrapperKind = ISD::DELETED_NODE;
3247 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3248 // external symbols most go through the PLT in PIC mode. If the symbol
3249 // has hidden or protected visibility, or if it is static or local, then
3250 // we don't need to use the PLT - we can directly call it.
3251 if (Subtarget->isTargetELF() &&
3252 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3253 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3254 OpFlags = X86II::MO_PLT;
3255 } else if (Subtarget->isPICStyleStubAny() &&
3256 !GV->isStrongDefinitionForLinker() &&
3257 (!Subtarget->getTargetTriple().isMacOSX() ||
3258 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3259 // PC-relative references to external symbols should go through $stub,
3260 // unless we're building with the leopard linker or later, which
3261 // automatically synthesizes these stubs.
3262 OpFlags = X86II::MO_DARWIN_STUB;
3263 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3264 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3265 // If the function is marked as non-lazy, generate an indirect call
3266 // which loads from the GOT directly. This avoids runtime overhead
3267 // at the cost of eager binding (and one extra byte of encoding).
3268 OpFlags = X86II::MO_GOTPCREL;
3269 WrapperKind = X86ISD::WrapperRIP;
3273 Callee = DAG.getTargetGlobalAddress(
3274 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3276 // Add a wrapper if needed.
3277 if (WrapperKind != ISD::DELETED_NODE)
3278 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3279 getPointerTy(DAG.getDataLayout()), Callee);
3280 // Add extra indirection if needed.
3282 Callee = DAG.getLoad(
3283 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3284 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3287 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3288 unsigned char OpFlags = 0;
3290 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3291 // external symbols should go through the PLT.
3292 if (Subtarget->isTargetELF() &&
3293 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3294 OpFlags = X86II::MO_PLT;
3295 } else if (Subtarget->isPICStyleStubAny() &&
3296 (!Subtarget->getTargetTriple().isMacOSX() ||
3297 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3298 // PC-relative references to external symbols should go through $stub,
3299 // unless we're building with the leopard linker or later, which
3300 // automatically synthesizes these stubs.
3301 OpFlags = X86II::MO_DARWIN_STUB;
3304 Callee = DAG.getTargetExternalSymbol(
3305 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3306 } else if (Subtarget->isTarget64BitILP32() &&
3307 Callee->getValueType(0) == MVT::i32) {
3308 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3309 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3312 // Returns a chain & a flag for retval copy to use.
3313 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3314 SmallVector<SDValue, 8> Ops;
3316 if (!IsSibcall && isTailCall) {
3317 Chain = DAG.getCALLSEQ_END(Chain,
3318 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3319 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3320 InFlag = Chain.getValue(1);
3323 Ops.push_back(Chain);
3324 Ops.push_back(Callee);
3327 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3329 // Add argument registers to the end of the list so that they are known live
3331 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3332 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3333 RegsToPass[i].second.getValueType()));
3335 // Add a register mask operand representing the call-preserved registers.
3336 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3337 assert(Mask && "Missing call preserved mask for calling convention");
3339 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3340 // the function clobbers all registers. If an exception is thrown, the runtime
3341 // will not restore CSRs.
3342 // FIXME: Model this more precisely so that we can register allocate across
3343 // the normal edge and spill and fill across the exceptional edge.
3344 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3345 const Function *CallerFn = MF.getFunction();
3346 EHPersonality Pers =
3347 CallerFn->hasPersonalityFn()
3348 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3349 : EHPersonality::Unknown;
3350 if (isMSVCEHPersonality(Pers))
3351 Mask = RegInfo->getNoPreservedMask();
3354 Ops.push_back(DAG.getRegisterMask(Mask));
3356 if (InFlag.getNode())
3357 Ops.push_back(InFlag);
3361 //// If this is the first return lowered for this function, add the regs
3362 //// to the liveout set for the function.
3363 // This isn't right, although it's probably harmless on x86; liveouts
3364 // should be computed from returns not tail calls. Consider a void
3365 // function making a tail call to a function returning int.
3366 MF.getFrameInfo()->setHasTailCall();
3367 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3370 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3371 InFlag = Chain.getValue(1);
3373 // Create the CALLSEQ_END node.
3374 unsigned NumBytesForCalleeToPop;
3375 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3376 DAG.getTarget().Options.GuaranteedTailCallOpt))
3377 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3378 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3379 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3380 SR == StackStructReturn)
3381 // If this is a call to a struct-return function, the callee
3382 // pops the hidden struct pointer, so we have to push it back.
3383 // This is common for Darwin/X86, Linux & Mingw32 targets.
3384 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3385 NumBytesForCalleeToPop = 4;
3387 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3389 // Returns a flag for retval copy to use.
3391 Chain = DAG.getCALLSEQ_END(Chain,
3392 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3393 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3396 InFlag = Chain.getValue(1);
3399 // Handle result values, copying them out of physregs into vregs that we
3401 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3402 Ins, dl, DAG, InVals);
3405 //===----------------------------------------------------------------------===//
3406 // Fast Calling Convention (tail call) implementation
3407 //===----------------------------------------------------------------------===//
3409 // Like std call, callee cleans arguments, convention except that ECX is
3410 // reserved for storing the tail called function address. Only 2 registers are
3411 // free for argument passing (inreg). Tail call optimization is performed
3413 // * tailcallopt is enabled
3414 // * caller/callee are fastcc
3415 // On X86_64 architecture with GOT-style position independent code only local
3416 // (within module) calls are supported at the moment.
3417 // To keep the stack aligned according to platform abi the function
3418 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3419 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3420 // If a tail called function callee has more arguments than the caller the
3421 // caller needs to make sure that there is room to move the RETADDR to. This is
3422 // achieved by reserving an area the size of the argument delta right after the
3423 // original RETADDR, but before the saved framepointer or the spilled registers
3424 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3436 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3439 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3440 SelectionDAG& DAG) const {
3441 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3442 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3443 unsigned StackAlignment = TFI.getStackAlignment();
3444 uint64_t AlignMask = StackAlignment - 1;
3445 int64_t Offset = StackSize;
3446 unsigned SlotSize = RegInfo->getSlotSize();
3447 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3448 // Number smaller than 12 so just add the difference.
3449 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3451 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3452 Offset = ((~AlignMask) & Offset) + StackAlignment +
3453 (StackAlignment-SlotSize);
3458 /// Return true if the given stack call argument is already available in the
3459 /// same position (relatively) of the caller's incoming argument stack.
3461 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3462 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3463 const X86InstrInfo *TII) {
3464 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3466 if (Arg.getOpcode() == ISD::CopyFromReg) {
3467 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3468 if (!TargetRegisterInfo::isVirtualRegister(VR))
3470 MachineInstr *Def = MRI->getVRegDef(VR);
3473 if (!Flags.isByVal()) {
3474 if (!TII->isLoadFromStackSlot(Def, FI))
3477 unsigned Opcode = Def->getOpcode();
3478 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3479 Opcode == X86::LEA64_32r) &&
3480 Def->getOperand(1).isFI()) {
3481 FI = Def->getOperand(1).getIndex();
3482 Bytes = Flags.getByValSize();
3486 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3487 if (Flags.isByVal())
3488 // ByVal argument is passed in as a pointer but it's now being
3489 // dereferenced. e.g.
3490 // define @foo(%struct.X* %A) {
3491 // tail call @bar(%struct.X* byval %A)
3494 SDValue Ptr = Ld->getBasePtr();
3495 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3498 FI = FINode->getIndex();
3499 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3500 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3501 FI = FINode->getIndex();
3502 Bytes = Flags.getByValSize();
3506 assert(FI != INT_MAX);
3507 if (!MFI->isFixedObjectIndex(FI))
3509 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3512 /// Check whether the call is eligible for tail call optimization. Targets
3513 /// that want to do tail call optimization should implement this function.
3515 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3516 CallingConv::ID CalleeCC,
3518 bool isCalleeStructRet,
3519 bool isCallerStructRet,
3521 const SmallVectorImpl<ISD::OutputArg> &Outs,
3522 const SmallVectorImpl<SDValue> &OutVals,
3523 const SmallVectorImpl<ISD::InputArg> &Ins,
3524 SelectionDAG &DAG) const {
3525 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3528 // If -tailcallopt is specified, make fastcc functions tail-callable.
3529 const MachineFunction &MF = DAG.getMachineFunction();
3530 const Function *CallerF = MF.getFunction();
3532 // If the function return type is x86_fp80 and the callee return type is not,
3533 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3534 // perform a tailcall optimization here.
3535 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3538 CallingConv::ID CallerCC = CallerF->getCallingConv();
3539 bool CCMatch = CallerCC == CalleeCC;
3540 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3541 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3543 // Win64 functions have extra shadow space for argument homing. Don't do the
3544 // sibcall if the caller and callee have mismatched expectations for this
3546 if (IsCalleeWin64 != IsCallerWin64)
3549 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3550 if (IsTailCallConvention(CalleeCC) && CCMatch)
3555 // Look for obvious safe cases to perform tail call optimization that do not
3556 // require ABI changes. This is what gcc calls sibcall.
3558 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3559 // emit a special epilogue.
3560 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3561 if (RegInfo->needsStackRealignment(MF))
3564 // Also avoid sibcall optimization if either caller or callee uses struct
3565 // return semantics.
3566 if (isCalleeStructRet || isCallerStructRet)
3569 // An stdcall/thiscall caller is expected to clean up its arguments; the
3570 // callee isn't going to do that.
3571 // FIXME: this is more restrictive than needed. We could produce a tailcall
3572 // when the stack adjustment matches. For example, with a thiscall that takes
3573 // only one argument.
3574 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3575 CallerCC == CallingConv::X86_ThisCall))
3578 // Do not sibcall optimize vararg calls unless all arguments are passed via
3580 if (isVarArg && !Outs.empty()) {
3582 // Optimizing for varargs on Win64 is unlikely to be safe without
3583 // additional testing.
3584 if (IsCalleeWin64 || IsCallerWin64)
3587 SmallVector<CCValAssign, 16> ArgLocs;
3588 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3591 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3592 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3593 if (!ArgLocs[i].isRegLoc())
3597 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3598 // stack. Therefore, if it's not used by the call it is not safe to optimize
3599 // this into a sibcall.
3600 bool Unused = false;
3601 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3608 SmallVector<CCValAssign, 16> RVLocs;
3609 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3611 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3612 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3613 CCValAssign &VA = RVLocs[i];
3614 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3619 // If the calling conventions do not match, then we'd better make sure the
3620 // results are returned in the same way as what the caller expects.
3622 SmallVector<CCValAssign, 16> RVLocs1;
3623 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3625 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3627 SmallVector<CCValAssign, 16> RVLocs2;
3628 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3630 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3632 if (RVLocs1.size() != RVLocs2.size())
3634 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3635 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3637 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3639 if (RVLocs1[i].isRegLoc()) {
3640 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3643 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3649 // If the callee takes no arguments then go on to check the results of the
3651 if (!Outs.empty()) {
3652 // Check if stack adjustment is needed. For now, do not do this if any
3653 // argument is passed on the stack.
3654 SmallVector<CCValAssign, 16> ArgLocs;
3655 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3658 // Allocate shadow area for Win64
3660 CCInfo.AllocateStack(32, 8);
3662 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3663 if (CCInfo.getNextStackOffset()) {
3664 MachineFunction &MF = DAG.getMachineFunction();
3665 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3668 // Check if the arguments are already laid out in the right way as
3669 // the caller's fixed stack objects.
3670 MachineFrameInfo *MFI = MF.getFrameInfo();
3671 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3672 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3673 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3674 CCValAssign &VA = ArgLocs[i];
3675 SDValue Arg = OutVals[i];
3676 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3677 if (VA.getLocInfo() == CCValAssign::Indirect)
3679 if (!VA.isRegLoc()) {
3680 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3687 // If the tailcall address may be in a register, then make sure it's
3688 // possible to register allocate for it. In 32-bit, the call address can
3689 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3690 // callee-saved registers are restored. These happen to be the same
3691 // registers used to pass 'inreg' arguments so watch out for those.
3692 if (!Subtarget->is64Bit() &&
3693 ((!isa<GlobalAddressSDNode>(Callee) &&
3694 !isa<ExternalSymbolSDNode>(Callee)) ||
3695 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3696 unsigned NumInRegs = 0;
3697 // In PIC we need an extra register to formulate the address computation
3699 unsigned MaxInRegs =
3700 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3702 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3703 CCValAssign &VA = ArgLocs[i];
3706 unsigned Reg = VA.getLocReg();
3709 case X86::EAX: case X86::EDX: case X86::ECX:
3710 if (++NumInRegs == MaxInRegs)
3722 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3723 const TargetLibraryInfo *libInfo) const {
3724 return X86::createFastISel(funcInfo, libInfo);
3727 //===----------------------------------------------------------------------===//
3728 // Other Lowering Hooks
3729 //===----------------------------------------------------------------------===//
3731 static bool MayFoldLoad(SDValue Op) {
3732 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3735 static bool MayFoldIntoStore(SDValue Op) {
3736 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3739 static bool isTargetShuffle(unsigned Opcode) {
3741 default: return false;
3742 case X86ISD::BLENDI:
3743 case X86ISD::PSHUFB:
3744 case X86ISD::PSHUFD:
3745 case X86ISD::PSHUFHW:
3746 case X86ISD::PSHUFLW:
3748 case X86ISD::PALIGNR:
3749 case X86ISD::MOVLHPS:
3750 case X86ISD::MOVLHPD:
3751 case X86ISD::MOVHLPS:
3752 case X86ISD::MOVLPS:
3753 case X86ISD::MOVLPD:
3754 case X86ISD::MOVSHDUP:
3755 case X86ISD::MOVSLDUP:
3756 case X86ISD::MOVDDUP:
3759 case X86ISD::UNPCKL:
3760 case X86ISD::UNPCKH:
3761 case X86ISD::VPERMILPI:
3762 case X86ISD::VPERM2X128:
3763 case X86ISD::VPERMI:
3768 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3769 SDValue V1, unsigned TargetMask,
3770 SelectionDAG &DAG) {
3772 default: llvm_unreachable("Unknown x86 shuffle node");
3773 case X86ISD::PSHUFD:
3774 case X86ISD::PSHUFHW:
3775 case X86ISD::PSHUFLW:
3776 case X86ISD::VPERMILPI:
3777 case X86ISD::VPERMI:
3778 return DAG.getNode(Opc, dl, VT, V1,
3779 DAG.getConstant(TargetMask, dl, MVT::i8));
3783 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3784 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3786 default: llvm_unreachable("Unknown x86 shuffle node");
3787 case X86ISD::MOVLHPS:
3788 case X86ISD::MOVLHPD:
3789 case X86ISD::MOVHLPS:
3790 case X86ISD::MOVLPS:
3791 case X86ISD::MOVLPD:
3794 case X86ISD::UNPCKL:
3795 case X86ISD::UNPCKH:
3796 return DAG.getNode(Opc, dl, VT, V1, V2);
3800 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3801 MachineFunction &MF = DAG.getMachineFunction();
3802 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3803 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3804 int ReturnAddrIndex = FuncInfo->getRAIndex();
3806 if (ReturnAddrIndex == 0) {
3807 // Set up a frame object for the return address.
3808 unsigned SlotSize = RegInfo->getSlotSize();
3809 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3812 FuncInfo->setRAIndex(ReturnAddrIndex);
3815 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3818 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3819 bool hasSymbolicDisplacement) {
3820 // Offset should fit into 32 bit immediate field.
3821 if (!isInt<32>(Offset))
3824 // If we don't have a symbolic displacement - we don't have any extra
3826 if (!hasSymbolicDisplacement)
3829 // FIXME: Some tweaks might be needed for medium code model.
3830 if (M != CodeModel::Small && M != CodeModel::Kernel)
3833 // For small code model we assume that latest object is 16MB before end of 31
3834 // bits boundary. We may also accept pretty large negative constants knowing
3835 // that all objects are in the positive half of address space.
3836 if (M == CodeModel::Small && Offset < 16*1024*1024)
3839 // For kernel code model we know that all object resist in the negative half
3840 // of 32bits address space. We may not accept negative offsets, since they may
3841 // be just off and we may accept pretty large positive ones.
3842 if (M == CodeModel::Kernel && Offset >= 0)
3848 /// Determines whether the callee is required to pop its own arguments.
3849 /// Callee pop is necessary to support tail calls.
3850 bool X86::isCalleePop(CallingConv::ID CallingConv,
3851 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3852 switch (CallingConv) {
3855 case CallingConv::X86_StdCall:
3856 case CallingConv::X86_FastCall:
3857 case CallingConv::X86_ThisCall:
3859 case CallingConv::Fast:
3860 case CallingConv::GHC:
3861 case CallingConv::HiPE:
3868 /// \brief Return true if the condition is an unsigned comparison operation.
3869 static bool isX86CCUnsigned(unsigned X86CC) {
3871 default: llvm_unreachable("Invalid integer condition!");
3872 case X86::COND_E: return true;
3873 case X86::COND_G: return false;
3874 case X86::COND_GE: return false;
3875 case X86::COND_L: return false;
3876 case X86::COND_LE: return false;
3877 case X86::COND_NE: return true;
3878 case X86::COND_B: return true;
3879 case X86::COND_A: return true;
3880 case X86::COND_BE: return true;
3881 case X86::COND_AE: return true;
3883 llvm_unreachable("covered switch fell through?!");
3886 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3887 /// condition code, returning the condition code and the LHS/RHS of the
3888 /// comparison to make.
3889 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3890 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3892 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3893 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3894 // X > -1 -> X == 0, jump !sign.
3895 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3896 return X86::COND_NS;
3898 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3899 // X < 0 -> X == 0, jump on sign.
3902 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3904 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3905 return X86::COND_LE;
3909 switch (SetCCOpcode) {
3910 default: llvm_unreachable("Invalid integer condition!");
3911 case ISD::SETEQ: return X86::COND_E;
3912 case ISD::SETGT: return X86::COND_G;
3913 case ISD::SETGE: return X86::COND_GE;
3914 case ISD::SETLT: return X86::COND_L;
3915 case ISD::SETLE: return X86::COND_LE;
3916 case ISD::SETNE: return X86::COND_NE;
3917 case ISD::SETULT: return X86::COND_B;
3918 case ISD::SETUGT: return X86::COND_A;
3919 case ISD::SETULE: return X86::COND_BE;
3920 case ISD::SETUGE: return X86::COND_AE;
3924 // First determine if it is required or is profitable to flip the operands.
3926 // If LHS is a foldable load, but RHS is not, flip the condition.
3927 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3928 !ISD::isNON_EXTLoad(RHS.getNode())) {
3929 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3930 std::swap(LHS, RHS);
3933 switch (SetCCOpcode) {
3939 std::swap(LHS, RHS);
3943 // On a floating point condition, the flags are set as follows:
3945 // 0 | 0 | 0 | X > Y
3946 // 0 | 0 | 1 | X < Y
3947 // 1 | 0 | 0 | X == Y
3948 // 1 | 1 | 1 | unordered
3949 switch (SetCCOpcode) {
3950 default: llvm_unreachable("Condcode should be pre-legalized away");
3952 case ISD::SETEQ: return X86::COND_E;
3953 case ISD::SETOLT: // flipped
3955 case ISD::SETGT: return X86::COND_A;
3956 case ISD::SETOLE: // flipped
3958 case ISD::SETGE: return X86::COND_AE;
3959 case ISD::SETUGT: // flipped
3961 case ISD::SETLT: return X86::COND_B;
3962 case ISD::SETUGE: // flipped
3964 case ISD::SETLE: return X86::COND_BE;
3966 case ISD::SETNE: return X86::COND_NE;
3967 case ISD::SETUO: return X86::COND_P;
3968 case ISD::SETO: return X86::COND_NP;
3970 case ISD::SETUNE: return X86::COND_INVALID;
3974 /// Is there a floating point cmov for the specific X86 condition code?
3975 /// Current x86 isa includes the following FP cmov instructions:
3976 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3977 static bool hasFPCMov(unsigned X86CC) {
3993 /// Returns true if the target can instruction select the
3994 /// specified FP immediate natively. If false, the legalizer will
3995 /// materialize the FP immediate as a load from a constant pool.
3996 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3997 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3998 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4004 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4005 ISD::LoadExtType ExtTy,
4007 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4008 // relocation target a movq or addq instruction: don't let the load shrink.
4009 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4010 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4011 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4012 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4016 /// \brief Returns true if it is beneficial to convert a load of a constant
4017 /// to just the constant itself.
4018 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4020 assert(Ty->isIntegerTy());
4022 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4023 if (BitSize == 0 || BitSize > 64)
4028 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4029 unsigned Index) const {
4030 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4033 return (Index == 0 || Index == ResVT.getVectorNumElements());
4036 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4037 // Speculate cttz only if we can directly use TZCNT.
4038 return Subtarget->hasBMI();
4041 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4042 // Speculate ctlz only if we can directly use LZCNT.
4043 return Subtarget->hasLZCNT();
4046 /// Return true if every element in Mask, beginning
4047 /// from position Pos and ending in Pos+Size is undef.
4048 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4049 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4055 /// Return true if Val is undef or if its value falls within the
4056 /// specified range (L, H].
4057 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4058 return (Val < 0) || (Val >= Low && Val < Hi);
4061 /// Val is either less than zero (undef) or equal to the specified value.
4062 static bool isUndefOrEqual(int Val, int CmpVal) {
4063 return (Val < 0 || Val == CmpVal);
4066 /// Return true if every element in Mask, beginning
4067 /// from position Pos and ending in Pos+Size, falls within the specified
4068 /// sequential range (Low, Low+Size]. or is undef.
4069 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4070 unsigned Pos, unsigned Size, int Low) {
4071 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4072 if (!isUndefOrEqual(Mask[i], Low))
4077 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4078 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4079 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4080 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4081 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4084 // The index should be aligned on a vecWidth-bit boundary.
4086 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4088 MVT VT = N->getSimpleValueType(0);
4089 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4090 bool Result = (Index * ElSize) % vecWidth == 0;
4095 /// Return true if the specified INSERT_SUBVECTOR
4096 /// operand specifies a subvector insert that is suitable for input to
4097 /// insertion of 128 or 256-bit subvectors
4098 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4099 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4100 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4102 // The index should be aligned on a vecWidth-bit boundary.
4104 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4106 MVT VT = N->getSimpleValueType(0);
4107 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4108 bool Result = (Index * ElSize) % vecWidth == 0;
4113 bool X86::isVINSERT128Index(SDNode *N) {
4114 return isVINSERTIndex(N, 128);
4117 bool X86::isVINSERT256Index(SDNode *N) {
4118 return isVINSERTIndex(N, 256);
4121 bool X86::isVEXTRACT128Index(SDNode *N) {
4122 return isVEXTRACTIndex(N, 128);
4125 bool X86::isVEXTRACT256Index(SDNode *N) {
4126 return isVEXTRACTIndex(N, 256);
4129 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4130 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4131 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4132 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4135 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4137 MVT VecVT = N->getOperand(0).getSimpleValueType();
4138 MVT ElVT = VecVT.getVectorElementType();
4140 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4141 return Index / NumElemsPerChunk;
4144 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4145 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4146 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4147 llvm_unreachable("Illegal insert subvector for VINSERT");
4150 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4152 MVT VecVT = N->getSimpleValueType(0);
4153 MVT ElVT = VecVT.getVectorElementType();
4155 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4156 return Index / NumElemsPerChunk;
4159 /// Return the appropriate immediate to extract the specified
4160 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4161 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4162 return getExtractVEXTRACTImmediate(N, 128);
4165 /// Return the appropriate immediate to extract the specified
4166 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4167 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4168 return getExtractVEXTRACTImmediate(N, 256);
4171 /// Return the appropriate immediate to insert at the specified
4172 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4173 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4174 return getInsertVINSERTImmediate(N, 128);
4177 /// Return the appropriate immediate to insert at the specified
4178 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4179 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4180 return getInsertVINSERTImmediate(N, 256);
4183 /// Returns true if Elt is a constant integer zero
4184 static bool isZero(SDValue V) {
4185 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4186 return C && C->isNullValue();
4189 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4190 bool X86::isZeroNode(SDValue Elt) {
4193 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4194 return CFP->getValueAPF().isPosZero();
4198 /// Returns a vector of specified type with all zero elements.
4199 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4200 SelectionDAG &DAG, SDLoc dl) {
4201 assert(VT.isVector() && "Expected a vector type");
4203 // Always build SSE zero vectors as <4 x i32> bitcasted
4204 // to their dest type. This ensures they get CSE'd.
4206 if (VT.is128BitVector()) { // SSE
4207 if (Subtarget->hasSSE2()) { // SSE2
4208 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4209 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4211 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4212 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4214 } else if (VT.is256BitVector()) { // AVX
4215 if (Subtarget->hasInt256()) { // AVX2
4216 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4217 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4218 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4220 // 256-bit logic and arithmetic instructions in AVX are all
4221 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4222 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4223 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4224 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4226 } else if (VT.is512BitVector()) { // AVX-512
4227 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4228 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4229 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4230 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4231 } else if (VT.getScalarType() == MVT::i1) {
4233 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4234 && "Unexpected vector type");
4235 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4236 && "Unexpected vector type");
4237 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4238 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4239 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4241 llvm_unreachable("Unexpected vector type");
4243 return DAG.getBitcast(VT, Vec);
4246 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4247 SelectionDAG &DAG, SDLoc dl,
4248 unsigned vectorWidth) {
4249 assert((vectorWidth == 128 || vectorWidth == 256) &&
4250 "Unsupported vector width");
4251 EVT VT = Vec.getValueType();
4252 EVT ElVT = VT.getVectorElementType();
4253 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4254 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4255 VT.getVectorNumElements()/Factor);
4257 // Extract from UNDEF is UNDEF.
4258 if (Vec.getOpcode() == ISD::UNDEF)
4259 return DAG.getUNDEF(ResultVT);
4261 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4262 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4264 // This is the index of the first element of the vectorWidth-bit chunk
4266 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4269 // If the input is a buildvector just emit a smaller one.
4270 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4271 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4272 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4275 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4276 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4279 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4280 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4281 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4282 /// instructions or a simple subregister reference. Idx is an index in the
4283 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4284 /// lowering EXTRACT_VECTOR_ELT operations easier.
4285 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4286 SelectionDAG &DAG, SDLoc dl) {
4287 assert((Vec.getValueType().is256BitVector() ||
4288 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4289 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4292 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4293 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4294 SelectionDAG &DAG, SDLoc dl) {
4295 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4296 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4299 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4300 unsigned IdxVal, SelectionDAG &DAG,
4301 SDLoc dl, unsigned vectorWidth) {
4302 assert((vectorWidth == 128 || vectorWidth == 256) &&
4303 "Unsupported vector width");
4304 // Inserting UNDEF is Result
4305 if (Vec.getOpcode() == ISD::UNDEF)
4307 EVT VT = Vec.getValueType();
4308 EVT ElVT = VT.getVectorElementType();
4309 EVT ResultVT = Result.getValueType();
4311 // Insert the relevant vectorWidth bits.
4312 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4314 // This is the index of the first element of the vectorWidth-bit chunk
4316 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4319 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4320 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4323 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4324 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4325 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4326 /// simple superregister reference. Idx is an index in the 128 bits
4327 /// we want. It need not be aligned to a 128-bit boundary. That makes
4328 /// lowering INSERT_VECTOR_ELT operations easier.
4329 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4330 SelectionDAG &DAG, SDLoc dl) {
4331 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4333 // For insertion into the zero index (low half) of a 256-bit vector, it is
4334 // more efficient to generate a blend with immediate instead of an insert*128.
4335 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4336 // extend the subvector to the size of the result vector. Make sure that
4337 // we are not recursing on that node by checking for undef here.
4338 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4339 Result.getOpcode() != ISD::UNDEF) {
4340 EVT ResultVT = Result.getValueType();
4341 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4342 SDValue Undef = DAG.getUNDEF(ResultVT);
4343 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4346 // The blend instruction, and therefore its mask, depend on the data type.
4347 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4348 if (ScalarType.isFloatingPoint()) {
4349 // Choose either vblendps (float) or vblendpd (double).
4350 unsigned ScalarSize = ScalarType.getSizeInBits();
4351 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4352 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4353 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4354 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4357 const X86Subtarget &Subtarget =
4358 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4360 // AVX2 is needed for 256-bit integer blend support.
4361 // Integers must be cast to 32-bit because there is only vpblendd;
4362 // vpblendw can't be used for this because it has a handicapped mask.
4364 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4365 // is still more efficient than using the wrong domain vinsertf128 that
4366 // will be created by InsertSubVector().
4367 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4369 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4370 Vec256 = DAG.getBitcast(CastVT, Vec256);
4371 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4372 return DAG.getBitcast(ResultVT, Vec256);
4375 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4378 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4379 SelectionDAG &DAG, SDLoc dl) {
4380 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4381 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4384 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4385 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4386 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4387 /// large BUILD_VECTORS.
4388 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4389 unsigned NumElems, SelectionDAG &DAG,
4391 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4392 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4395 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4396 unsigned NumElems, SelectionDAG &DAG,
4398 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4399 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4402 /// Returns a vector of specified type with all bits set.
4403 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4404 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4405 /// Then bitcast to their original type, ensuring they get CSE'd.
4406 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4408 assert(VT.isVector() && "Expected a vector type");
4410 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4412 if (VT.is256BitVector()) {
4413 if (HasInt256) { // AVX2
4414 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4415 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4417 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4418 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4420 } else if (VT.is128BitVector()) {
4421 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4423 llvm_unreachable("Unexpected vector type");
4425 return DAG.getBitcast(VT, Vec);
4428 /// Returns a vector_shuffle node for an unpackl operation.
4429 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4431 unsigned NumElems = VT.getVectorNumElements();
4432 SmallVector<int, 8> Mask;
4433 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4435 Mask.push_back(i + NumElems);
4437 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4440 /// Returns a vector_shuffle node for an unpackh operation.
4441 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4443 unsigned NumElems = VT.getVectorNumElements();
4444 SmallVector<int, 8> Mask;
4445 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4446 Mask.push_back(i + Half);
4447 Mask.push_back(i + NumElems + Half);
4449 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4452 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4453 /// This produces a shuffle where the low element of V2 is swizzled into the
4454 /// zero/undef vector, landing at element Idx.
4455 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4456 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4458 const X86Subtarget *Subtarget,
4459 SelectionDAG &DAG) {
4460 MVT VT = V2.getSimpleValueType();
4462 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4463 unsigned NumElems = VT.getVectorNumElements();
4464 SmallVector<int, 16> MaskVec;
4465 for (unsigned i = 0; i != NumElems; ++i)
4466 // If this is the insertion idx, put the low elt of V2 here.
4467 MaskVec.push_back(i == Idx ? NumElems : i);
4468 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4471 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4472 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4473 /// uses one source. Note that this will set IsUnary for shuffles which use a
4474 /// single input multiple times, and in those cases it will
4475 /// adjust the mask to only have indices within that single input.
4476 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4477 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4478 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4479 unsigned NumElems = VT.getVectorNumElements();
4483 bool IsFakeUnary = false;
4484 switch(N->getOpcode()) {
4485 case X86ISD::BLENDI:
4486 ImmN = N->getOperand(N->getNumOperands()-1);
4487 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4490 ImmN = N->getOperand(N->getNumOperands()-1);
4491 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4492 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4494 case X86ISD::UNPCKH:
4495 DecodeUNPCKHMask(VT, Mask);
4496 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4498 case X86ISD::UNPCKL:
4499 DecodeUNPCKLMask(VT, Mask);
4500 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4502 case X86ISD::MOVHLPS:
4503 DecodeMOVHLPSMask(NumElems, Mask);
4504 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4506 case X86ISD::MOVLHPS:
4507 DecodeMOVLHPSMask(NumElems, Mask);
4508 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4510 case X86ISD::PALIGNR:
4511 ImmN = N->getOperand(N->getNumOperands()-1);
4512 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4514 case X86ISD::PSHUFD:
4515 case X86ISD::VPERMILPI:
4516 ImmN = N->getOperand(N->getNumOperands()-1);
4517 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4520 case X86ISD::PSHUFHW:
4521 ImmN = N->getOperand(N->getNumOperands()-1);
4522 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4525 case X86ISD::PSHUFLW:
4526 ImmN = N->getOperand(N->getNumOperands()-1);
4527 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4530 case X86ISD::PSHUFB: {
4532 SDValue MaskNode = N->getOperand(1);
4533 while (MaskNode->getOpcode() == ISD::BITCAST)
4534 MaskNode = MaskNode->getOperand(0);
4536 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4537 // If we have a build-vector, then things are easy.
4538 EVT VT = MaskNode.getValueType();
4539 assert(VT.isVector() &&
4540 "Can't produce a non-vector with a build_vector!");
4541 if (!VT.isInteger())
4544 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4546 SmallVector<uint64_t, 32> RawMask;
4547 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4548 SDValue Op = MaskNode->getOperand(i);
4549 if (Op->getOpcode() == ISD::UNDEF) {
4550 RawMask.push_back((uint64_t)SM_SentinelUndef);
4553 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4556 APInt MaskElement = CN->getAPIntValue();
4558 // We now have to decode the element which could be any integer size and
4559 // extract each byte of it.
4560 for (int j = 0; j < NumBytesPerElement; ++j) {
4561 // Note that this is x86 and so always little endian: the low byte is
4562 // the first byte of the mask.
4563 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4564 MaskElement = MaskElement.lshr(8);
4567 DecodePSHUFBMask(RawMask, Mask);
4571 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4575 SDValue Ptr = MaskLoad->getBasePtr();
4576 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4577 Ptr->getOpcode() == X86ISD::WrapperRIP)
4578 Ptr = Ptr->getOperand(0);
4580 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4581 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4584 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4585 DecodePSHUFBMask(C, Mask);
4593 case X86ISD::VPERMI:
4594 ImmN = N->getOperand(N->getNumOperands()-1);
4595 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4600 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4602 case X86ISD::VPERM2X128:
4603 ImmN = N->getOperand(N->getNumOperands()-1);
4604 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4605 if (Mask.empty()) return false;
4606 // Mask only contains negative index if an element is zero.
4607 if (std::any_of(Mask.begin(), Mask.end(),
4608 [](int M){ return M == SM_SentinelZero; }))
4611 case X86ISD::MOVSLDUP:
4612 DecodeMOVSLDUPMask(VT, Mask);
4615 case X86ISD::MOVSHDUP:
4616 DecodeMOVSHDUPMask(VT, Mask);
4619 case X86ISD::MOVDDUP:
4620 DecodeMOVDDUPMask(VT, Mask);
4623 case X86ISD::MOVLHPD:
4624 case X86ISD::MOVLPD:
4625 case X86ISD::MOVLPS:
4626 // Not yet implemented
4628 default: llvm_unreachable("unknown target shuffle node");
4631 // If we have a fake unary shuffle, the shuffle mask is spread across two
4632 // inputs that are actually the same node. Re-map the mask to always point
4633 // into the first input.
4636 if (M >= (int)Mask.size())
4642 /// Returns the scalar element that will make up the ith
4643 /// element of the result of the vector shuffle.
4644 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4647 return SDValue(); // Limit search depth.
4649 SDValue V = SDValue(N, 0);
4650 EVT VT = V.getValueType();
4651 unsigned Opcode = V.getOpcode();
4653 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4654 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4655 int Elt = SV->getMaskElt(Index);
4658 return DAG.getUNDEF(VT.getVectorElementType());
4660 unsigned NumElems = VT.getVectorNumElements();
4661 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4662 : SV->getOperand(1);
4663 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4666 // Recurse into target specific vector shuffles to find scalars.
4667 if (isTargetShuffle(Opcode)) {
4668 MVT ShufVT = V.getSimpleValueType();
4669 unsigned NumElems = ShufVT.getVectorNumElements();
4670 SmallVector<int, 16> ShuffleMask;
4673 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4676 int Elt = ShuffleMask[Index];
4678 return DAG.getUNDEF(ShufVT.getVectorElementType());
4680 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4682 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4686 // Actual nodes that may contain scalar elements
4687 if (Opcode == ISD::BITCAST) {
4688 V = V.getOperand(0);
4689 EVT SrcVT = V.getValueType();
4690 unsigned NumElems = VT.getVectorNumElements();
4692 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4696 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4697 return (Index == 0) ? V.getOperand(0)
4698 : DAG.getUNDEF(VT.getVectorElementType());
4700 if (V.getOpcode() == ISD::BUILD_VECTOR)
4701 return V.getOperand(Index);
4706 /// Custom lower build_vector of v16i8.
4707 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4708 unsigned NumNonZero, unsigned NumZero,
4710 const X86Subtarget* Subtarget,
4711 const TargetLowering &TLI) {
4719 // SSE4.1 - use PINSRB to insert each byte directly.
4720 if (Subtarget->hasSSE41()) {
4721 for (unsigned i = 0; i < 16; ++i) {
4722 bool isNonZero = (NonZeros & (1 << i)) != 0;
4726 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4728 V = DAG.getUNDEF(MVT::v16i8);
4731 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4732 MVT::v16i8, V, Op.getOperand(i),
4733 DAG.getIntPtrConstant(i, dl));
4740 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4741 for (unsigned i = 0; i < 16; ++i) {
4742 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4743 if (ThisIsNonZero && First) {
4745 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4747 V = DAG.getUNDEF(MVT::v8i16);
4752 SDValue ThisElt, LastElt;
4753 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4754 if (LastIsNonZero) {
4755 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4756 MVT::i16, Op.getOperand(i-1));
4758 if (ThisIsNonZero) {
4759 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4760 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4761 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4763 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4767 if (ThisElt.getNode())
4768 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4769 DAG.getIntPtrConstant(i/2, dl));
4773 return DAG.getBitcast(MVT::v16i8, V);
4776 /// Custom lower build_vector of v8i16.
4777 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4778 unsigned NumNonZero, unsigned NumZero,
4780 const X86Subtarget* Subtarget,
4781 const TargetLowering &TLI) {
4788 for (unsigned i = 0; i < 8; ++i) {
4789 bool isNonZero = (NonZeros & (1 << i)) != 0;
4793 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4795 V = DAG.getUNDEF(MVT::v8i16);
4798 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4799 MVT::v8i16, V, Op.getOperand(i),
4800 DAG.getIntPtrConstant(i, dl));
4807 /// Custom lower build_vector of v4i32 or v4f32.
4808 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4809 const X86Subtarget *Subtarget,
4810 const TargetLowering &TLI) {
4811 // Find all zeroable elements.
4812 std::bitset<4> Zeroable;
4813 for (int i=0; i < 4; ++i) {
4814 SDValue Elt = Op->getOperand(i);
4815 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4817 assert(Zeroable.size() - Zeroable.count() > 1 &&
4818 "We expect at least two non-zero elements!");
4820 // We only know how to deal with build_vector nodes where elements are either
4821 // zeroable or extract_vector_elt with constant index.
4822 SDValue FirstNonZero;
4823 unsigned FirstNonZeroIdx;
4824 for (unsigned i=0; i < 4; ++i) {
4827 SDValue Elt = Op->getOperand(i);
4828 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4829 !isa<ConstantSDNode>(Elt.getOperand(1)))
4831 // Make sure that this node is extracting from a 128-bit vector.
4832 MVT VT = Elt.getOperand(0).getSimpleValueType();
4833 if (!VT.is128BitVector())
4835 if (!FirstNonZero.getNode()) {
4837 FirstNonZeroIdx = i;
4841 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4842 SDValue V1 = FirstNonZero.getOperand(0);
4843 MVT VT = V1.getSimpleValueType();
4845 // See if this build_vector can be lowered as a blend with zero.
4847 unsigned EltMaskIdx, EltIdx;
4849 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4850 if (Zeroable[EltIdx]) {
4851 // The zero vector will be on the right hand side.
4852 Mask[EltIdx] = EltIdx+4;
4856 Elt = Op->getOperand(EltIdx);
4857 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4858 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4859 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4861 Mask[EltIdx] = EltIdx;
4865 // Let the shuffle legalizer deal with blend operations.
4866 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4867 if (V1.getSimpleValueType() != VT)
4868 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4869 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4872 // See if we can lower this build_vector to a INSERTPS.
4873 if (!Subtarget->hasSSE41())
4876 SDValue V2 = Elt.getOperand(0);
4877 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4880 bool CanFold = true;
4881 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4885 SDValue Current = Op->getOperand(i);
4886 SDValue SrcVector = Current->getOperand(0);
4889 CanFold = SrcVector == V1 &&
4890 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4896 assert(V1.getNode() && "Expected at least two non-zero elements!");
4897 if (V1.getSimpleValueType() != MVT::v4f32)
4898 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4899 if (V2.getSimpleValueType() != MVT::v4f32)
4900 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4902 // Ok, we can emit an INSERTPS instruction.
4903 unsigned ZMask = Zeroable.to_ulong();
4905 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4906 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4908 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4909 DAG.getIntPtrConstant(InsertPSMask, DL));
4910 return DAG.getBitcast(VT, Result);
4913 /// Return a vector logical shift node.
4914 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4915 unsigned NumBits, SelectionDAG &DAG,
4916 const TargetLowering &TLI, SDLoc dl) {
4917 assert(VT.is128BitVector() && "Unknown type for VShift");
4918 MVT ShVT = MVT::v2i64;
4919 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4920 SrcOp = DAG.getBitcast(ShVT, SrcOp);
4921 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
4922 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4923 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4924 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4928 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4930 // Check if the scalar load can be widened into a vector load. And if
4931 // the address is "base + cst" see if the cst can be "absorbed" into
4932 // the shuffle mask.
4933 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4934 SDValue Ptr = LD->getBasePtr();
4935 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4937 EVT PVT = LD->getValueType(0);
4938 if (PVT != MVT::i32 && PVT != MVT::f32)
4943 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4944 FI = FINode->getIndex();
4946 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4947 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4948 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4949 Offset = Ptr.getConstantOperandVal(1);
4950 Ptr = Ptr.getOperand(0);
4955 // FIXME: 256-bit vector instructions don't require a strict alignment,
4956 // improve this code to support it better.
4957 unsigned RequiredAlign = VT.getSizeInBits()/8;
4958 SDValue Chain = LD->getChain();
4959 // Make sure the stack object alignment is at least 16 or 32.
4960 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4961 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4962 if (MFI->isFixedObjectIndex(FI)) {
4963 // Can't change the alignment. FIXME: It's possible to compute
4964 // the exact stack offset and reference FI + adjust offset instead.
4965 // If someone *really* cares about this. That's the way to implement it.
4968 MFI->setObjectAlignment(FI, RequiredAlign);
4972 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4973 // Ptr + (Offset & ~15).
4976 if ((Offset % RequiredAlign) & 3)
4978 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4981 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4982 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4985 int EltNo = (Offset - StartOffset) >> 2;
4986 unsigned NumElems = VT.getVectorNumElements();
4988 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4989 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4990 LD->getPointerInfo().getWithOffset(StartOffset),
4991 false, false, false, 0);
4993 SmallVector<int, 8> Mask(NumElems, EltNo);
4995 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5001 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5002 /// elements can be replaced by a single large load which has the same value as
5003 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5005 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5007 /// FIXME: we'd also like to handle the case where the last elements are zero
5008 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5009 /// There's even a handy isZeroNode for that purpose.
5010 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5011 SDLoc &DL, SelectionDAG &DAG,
5012 bool isAfterLegalize) {
5013 unsigned NumElems = Elts.size();
5015 LoadSDNode *LDBase = nullptr;
5016 unsigned LastLoadedElt = -1U;
5018 // For each element in the initializer, see if we've found a load or an undef.
5019 // If we don't find an initial load element, or later load elements are
5020 // non-consecutive, bail out.
5021 for (unsigned i = 0; i < NumElems; ++i) {
5022 SDValue Elt = Elts[i];
5023 // Look through a bitcast.
5024 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5025 Elt = Elt.getOperand(0);
5026 if (!Elt.getNode() ||
5027 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5030 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5032 LDBase = cast<LoadSDNode>(Elt.getNode());
5036 if (Elt.getOpcode() == ISD::UNDEF)
5039 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5040 EVT LdVT = Elt.getValueType();
5041 // Each loaded element must be the correct fractional portion of the
5042 // requested vector load.
5043 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5045 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5050 // If we have found an entire vector of loads and undefs, then return a large
5051 // load of the entire vector width starting at the base pointer. If we found
5052 // consecutive loads for the low half, generate a vzext_load node.
5053 if (LastLoadedElt == NumElems - 1) {
5054 assert(LDBase && "Did not find base load for merging consecutive loads");
5055 EVT EltVT = LDBase->getValueType(0);
5056 // Ensure that the input vector size for the merged loads matches the
5057 // cumulative size of the input elements.
5058 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5061 if (isAfterLegalize &&
5062 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5065 SDValue NewLd = SDValue();
5067 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5068 LDBase->getPointerInfo(), LDBase->isVolatile(),
5069 LDBase->isNonTemporal(), LDBase->isInvariant(),
5070 LDBase->getAlignment());
5072 if (LDBase->hasAnyUseOfValue(1)) {
5073 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5075 SDValue(NewLd.getNode(), 1));
5076 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5077 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5078 SDValue(NewLd.getNode(), 1));
5084 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5085 //of a v4i32 / v4f32. It's probably worth generalizing.
5086 EVT EltVT = VT.getVectorElementType();
5087 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5088 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5089 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5090 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5092 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5093 LDBase->getPointerInfo(),
5094 LDBase->getAlignment(),
5095 false/*isVolatile*/, true/*ReadMem*/,
5098 // Make sure the newly-created LOAD is in the same position as LDBase in
5099 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5100 // update uses of LDBase's output chain to use the TokenFactor.
5101 if (LDBase->hasAnyUseOfValue(1)) {
5102 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5103 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5104 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5105 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5106 SDValue(ResNode.getNode(), 1));
5109 return DAG.getBitcast(VT, ResNode);
5114 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5115 /// to generate a splat value for the following cases:
5116 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5117 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5118 /// a scalar load, or a constant.
5119 /// The VBROADCAST node is returned when a pattern is found,
5120 /// or SDValue() otherwise.
5121 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5122 SelectionDAG &DAG) {
5123 // VBROADCAST requires AVX.
5124 // TODO: Splats could be generated for non-AVX CPUs using SSE
5125 // instructions, but there's less potential gain for only 128-bit vectors.
5126 if (!Subtarget->hasAVX())
5129 MVT VT = Op.getSimpleValueType();
5132 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5133 "Unsupported vector type for broadcast.");
5138 switch (Op.getOpcode()) {
5140 // Unknown pattern found.
5143 case ISD::BUILD_VECTOR: {
5144 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5145 BitVector UndefElements;
5146 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5148 // We need a splat of a single value to use broadcast, and it doesn't
5149 // make any sense if the value is only in one element of the vector.
5150 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5154 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5155 Ld.getOpcode() == ISD::ConstantFP);
5157 // Make sure that all of the users of a non-constant load are from the
5158 // BUILD_VECTOR node.
5159 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5164 case ISD::VECTOR_SHUFFLE: {
5165 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5167 // Shuffles must have a splat mask where the first element is
5169 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5172 SDValue Sc = Op.getOperand(0);
5173 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5174 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5176 if (!Subtarget->hasInt256())
5179 // Use the register form of the broadcast instruction available on AVX2.
5180 if (VT.getSizeInBits() >= 256)
5181 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5182 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5185 Ld = Sc.getOperand(0);
5186 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5187 Ld.getOpcode() == ISD::ConstantFP);
5189 // The scalar_to_vector node and the suspected
5190 // load node must have exactly one user.
5191 // Constants may have multiple users.
5193 // AVX-512 has register version of the broadcast
5194 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5195 Ld.getValueType().getSizeInBits() >= 32;
5196 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5203 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5204 bool IsGE256 = (VT.getSizeInBits() >= 256);
5206 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5207 // instruction to save 8 or more bytes of constant pool data.
5208 // TODO: If multiple splats are generated to load the same constant,
5209 // it may be detrimental to overall size. There needs to be a way to detect
5210 // that condition to know if this is truly a size win.
5211 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5213 // Handle broadcasting a single constant scalar from the constant pool
5215 // On Sandybridge (no AVX2), it is still better to load a constant vector
5216 // from the constant pool and not to broadcast it from a scalar.
5217 // But override that restriction when optimizing for size.
5218 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5219 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5220 EVT CVT = Ld.getValueType();
5221 assert(!CVT.isVector() && "Must not broadcast a vector type");
5223 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5224 // For size optimization, also splat v2f64 and v2i64, and for size opt
5225 // with AVX2, also splat i8 and i16.
5226 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5227 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5228 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5229 const Constant *C = nullptr;
5230 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5231 C = CI->getConstantIntValue();
5232 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5233 C = CF->getConstantFPValue();
5235 assert(C && "Invalid constant type");
5237 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5239 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5240 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5242 CVT, dl, DAG.getEntryNode(), CP,
5243 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5244 false, false, Alignment);
5246 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5250 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5252 // Handle AVX2 in-register broadcasts.
5253 if (!IsLoad && Subtarget->hasInt256() &&
5254 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5255 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5257 // The scalar source must be a normal load.
5261 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5262 (Subtarget->hasVLX() && ScalarSize == 64))
5263 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5265 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5266 // double since there is no vbroadcastsd xmm
5267 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5268 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5269 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5272 // Unsupported broadcast.
5276 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5277 /// underlying vector and index.
5279 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5281 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5283 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5284 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5287 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5289 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5291 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5292 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5295 // In this case the vector is the extract_subvector expression and the index
5296 // is 2, as specified by the shuffle.
5297 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5298 SDValue ShuffleVec = SVOp->getOperand(0);
5299 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5300 assert(ShuffleVecVT.getVectorElementType() ==
5301 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5303 int ShuffleIdx = SVOp->getMaskElt(Idx);
5304 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5305 ExtractedFromVec = ShuffleVec;
5311 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5312 MVT VT = Op.getSimpleValueType();
5314 // Skip if insert_vec_elt is not supported.
5315 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5316 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5320 unsigned NumElems = Op.getNumOperands();
5324 SmallVector<unsigned, 4> InsertIndices;
5325 SmallVector<int, 8> Mask(NumElems, -1);
5327 for (unsigned i = 0; i != NumElems; ++i) {
5328 unsigned Opc = Op.getOperand(i).getOpcode();
5330 if (Opc == ISD::UNDEF)
5333 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5334 // Quit if more than 1 elements need inserting.
5335 if (InsertIndices.size() > 1)
5338 InsertIndices.push_back(i);
5342 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5343 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5344 // Quit if non-constant index.
5345 if (!isa<ConstantSDNode>(ExtIdx))
5347 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5349 // Quit if extracted from vector of different type.
5350 if (ExtractedFromVec.getValueType() != VT)
5353 if (!VecIn1.getNode())
5354 VecIn1 = ExtractedFromVec;
5355 else if (VecIn1 != ExtractedFromVec) {
5356 if (!VecIn2.getNode())
5357 VecIn2 = ExtractedFromVec;
5358 else if (VecIn2 != ExtractedFromVec)
5359 // Quit if more than 2 vectors to shuffle
5363 if (ExtractedFromVec == VecIn1)
5365 else if (ExtractedFromVec == VecIn2)
5366 Mask[i] = Idx + NumElems;
5369 if (!VecIn1.getNode())
5372 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5373 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5374 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5375 unsigned Idx = InsertIndices[i];
5376 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5377 DAG.getIntPtrConstant(Idx, DL));
5383 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5384 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5385 Op.getScalarValueSizeInBits() == 1 &&
5386 "Can not convert non-constant vector");
5387 uint64_t Immediate = 0;
5388 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5389 SDValue In = Op.getOperand(idx);
5390 if (In.getOpcode() != ISD::UNDEF)
5391 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5395 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5396 return DAG.getConstant(Immediate, dl, VT);
5398 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5400 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5402 MVT VT = Op.getSimpleValueType();
5403 assert((VT.getVectorElementType() == MVT::i1) &&
5404 "Unexpected type in LowerBUILD_VECTORvXi1!");
5407 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5408 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5409 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5410 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5413 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5414 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5415 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5416 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5419 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5420 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5421 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5422 return DAG.getBitcast(VT, Imm);
5423 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5424 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5425 DAG.getIntPtrConstant(0, dl));
5428 // Vector has one or more non-const elements
5429 uint64_t Immediate = 0;
5430 SmallVector<unsigned, 16> NonConstIdx;
5431 bool IsSplat = true;
5432 bool HasConstElts = false;
5434 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5435 SDValue In = Op.getOperand(idx);
5436 if (In.getOpcode() == ISD::UNDEF)
5438 if (!isa<ConstantSDNode>(In))
5439 NonConstIdx.push_back(idx);
5441 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5442 HasConstElts = true;
5446 else if (In != Op.getOperand(SplatIdx))
5450 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5452 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5453 DAG.getConstant(1, dl, VT),
5454 DAG.getConstant(0, dl, VT));
5456 // insert elements one by one
5460 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5461 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5463 else if (HasConstElts)
5464 Imm = DAG.getConstant(0, dl, VT);
5466 Imm = DAG.getUNDEF(VT);
5467 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5468 DstVec = DAG.getBitcast(VT, Imm);
5470 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5471 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5472 DAG.getIntPtrConstant(0, dl));
5475 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5476 unsigned InsertIdx = NonConstIdx[i];
5477 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5478 Op.getOperand(InsertIdx),
5479 DAG.getIntPtrConstant(InsertIdx, dl));
5484 /// \brief Return true if \p N implements a horizontal binop and return the
5485 /// operands for the horizontal binop into V0 and V1.
5487 /// This is a helper function of LowerToHorizontalOp().
5488 /// This function checks that the build_vector \p N in input implements a
5489 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5490 /// operation to match.
5491 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5492 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5493 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5496 /// This function only analyzes elements of \p N whose indices are
5497 /// in range [BaseIdx, LastIdx).
5498 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5500 unsigned BaseIdx, unsigned LastIdx,
5501 SDValue &V0, SDValue &V1) {
5502 EVT VT = N->getValueType(0);
5504 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5505 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5506 "Invalid Vector in input!");
5508 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5509 bool CanFold = true;
5510 unsigned ExpectedVExtractIdx = BaseIdx;
5511 unsigned NumElts = LastIdx - BaseIdx;
5512 V0 = DAG.getUNDEF(VT);
5513 V1 = DAG.getUNDEF(VT);
5515 // Check if N implements a horizontal binop.
5516 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5517 SDValue Op = N->getOperand(i + BaseIdx);
5520 if (Op->getOpcode() == ISD::UNDEF) {
5521 // Update the expected vector extract index.
5522 if (i * 2 == NumElts)
5523 ExpectedVExtractIdx = BaseIdx;
5524 ExpectedVExtractIdx += 2;
5528 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5533 SDValue Op0 = Op.getOperand(0);
5534 SDValue Op1 = Op.getOperand(1);
5536 // Try to match the following pattern:
5537 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5538 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5539 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5540 Op0.getOperand(0) == Op1.getOperand(0) &&
5541 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5542 isa<ConstantSDNode>(Op1.getOperand(1)));
5546 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5547 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5549 if (i * 2 < NumElts) {
5550 if (V0.getOpcode() == ISD::UNDEF) {
5551 V0 = Op0.getOperand(0);
5552 if (V0.getValueType() != VT)
5556 if (V1.getOpcode() == ISD::UNDEF) {
5557 V1 = Op0.getOperand(0);
5558 if (V1.getValueType() != VT)
5561 if (i * 2 == NumElts)
5562 ExpectedVExtractIdx = BaseIdx;
5565 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5566 if (I0 == ExpectedVExtractIdx)
5567 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5568 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5569 // Try to match the following dag sequence:
5570 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5571 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5575 ExpectedVExtractIdx += 2;
5581 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5582 /// a concat_vector.
5584 /// This is a helper function of LowerToHorizontalOp().
5585 /// This function expects two 256-bit vectors called V0 and V1.
5586 /// At first, each vector is split into two separate 128-bit vectors.
5587 /// Then, the resulting 128-bit vectors are used to implement two
5588 /// horizontal binary operations.
5590 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5592 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5593 /// the two new horizontal binop.
5594 /// When Mode is set, the first horizontal binop dag node would take as input
5595 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5596 /// horizontal binop dag node would take as input the lower 128-bit of V1
5597 /// and the upper 128-bit of V1.
5599 /// HADD V0_LO, V0_HI
5600 /// HADD V1_LO, V1_HI
5602 /// Otherwise, the first horizontal binop dag node takes as input the lower
5603 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5604 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5606 /// HADD V0_LO, V1_LO
5607 /// HADD V0_HI, V1_HI
5609 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5610 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5611 /// the upper 128-bits of the result.
5612 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5613 SDLoc DL, SelectionDAG &DAG,
5614 unsigned X86Opcode, bool Mode,
5615 bool isUndefLO, bool isUndefHI) {
5616 EVT VT = V0.getValueType();
5617 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5618 "Invalid nodes in input!");
5620 unsigned NumElts = VT.getVectorNumElements();
5621 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5622 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5623 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5624 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5625 EVT NewVT = V0_LO.getValueType();
5627 SDValue LO = DAG.getUNDEF(NewVT);
5628 SDValue HI = DAG.getUNDEF(NewVT);
5631 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5632 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5633 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5634 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5635 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5637 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5638 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5639 V1_LO->getOpcode() != ISD::UNDEF))
5640 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5642 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5643 V1_HI->getOpcode() != ISD::UNDEF))
5644 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5647 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5650 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5652 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5653 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5654 EVT VT = BV->getValueType(0);
5655 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5656 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5660 unsigned NumElts = VT.getVectorNumElements();
5661 SDValue InVec0 = DAG.getUNDEF(VT);
5662 SDValue InVec1 = DAG.getUNDEF(VT);
5664 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5665 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5667 // Odd-numbered elements in the input build vector are obtained from
5668 // adding two integer/float elements.
5669 // Even-numbered elements in the input build vector are obtained from
5670 // subtracting two integer/float elements.
5671 unsigned ExpectedOpcode = ISD::FSUB;
5672 unsigned NextExpectedOpcode = ISD::FADD;
5673 bool AddFound = false;
5674 bool SubFound = false;
5676 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5677 SDValue Op = BV->getOperand(i);
5679 // Skip 'undef' values.
5680 unsigned Opcode = Op.getOpcode();
5681 if (Opcode == ISD::UNDEF) {
5682 std::swap(ExpectedOpcode, NextExpectedOpcode);
5686 // Early exit if we found an unexpected opcode.
5687 if (Opcode != ExpectedOpcode)
5690 SDValue Op0 = Op.getOperand(0);
5691 SDValue Op1 = Op.getOperand(1);
5693 // Try to match the following pattern:
5694 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5695 // Early exit if we cannot match that sequence.
5696 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5697 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5698 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5699 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5700 Op0.getOperand(1) != Op1.getOperand(1))
5703 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5707 // We found a valid add/sub node. Update the information accordingly.
5713 // Update InVec0 and InVec1.
5714 if (InVec0.getOpcode() == ISD::UNDEF) {
5715 InVec0 = Op0.getOperand(0);
5716 if (InVec0.getValueType() != VT)
5719 if (InVec1.getOpcode() == ISD::UNDEF) {
5720 InVec1 = Op1.getOperand(0);
5721 if (InVec1.getValueType() != VT)
5725 // Make sure that operands in input to each add/sub node always
5726 // come from a same pair of vectors.
5727 if (InVec0 != Op0.getOperand(0)) {
5728 if (ExpectedOpcode == ISD::FSUB)
5731 // FADD is commutable. Try to commute the operands
5732 // and then test again.
5733 std::swap(Op0, Op1);
5734 if (InVec0 != Op0.getOperand(0))
5738 if (InVec1 != Op1.getOperand(0))
5741 // Update the pair of expected opcodes.
5742 std::swap(ExpectedOpcode, NextExpectedOpcode);
5745 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5746 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5747 InVec1.getOpcode() != ISD::UNDEF)
5748 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5753 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5754 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5755 const X86Subtarget *Subtarget,
5756 SelectionDAG &DAG) {
5757 EVT VT = BV->getValueType(0);
5758 unsigned NumElts = VT.getVectorNumElements();
5759 unsigned NumUndefsLO = 0;
5760 unsigned NumUndefsHI = 0;
5761 unsigned Half = NumElts/2;
5763 // Count the number of UNDEF operands in the build_vector in input.
5764 for (unsigned i = 0, e = Half; i != e; ++i)
5765 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5768 for (unsigned i = Half, e = NumElts; i != e; ++i)
5769 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5772 // Early exit if this is either a build_vector of all UNDEFs or all the
5773 // operands but one are UNDEF.
5774 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5778 SDValue InVec0, InVec1;
5779 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5780 // Try to match an SSE3 float HADD/HSUB.
5781 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5782 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5784 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5785 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5786 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5787 // Try to match an SSSE3 integer HADD/HSUB.
5788 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5789 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5791 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5792 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5795 if (!Subtarget->hasAVX())
5798 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5799 // Try to match an AVX horizontal add/sub of packed single/double
5800 // precision floating point values from 256-bit vectors.
5801 SDValue InVec2, InVec3;
5802 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5803 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5804 ((InVec0.getOpcode() == ISD::UNDEF ||
5805 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5806 ((InVec1.getOpcode() == ISD::UNDEF ||
5807 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5808 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5810 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5811 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5812 ((InVec0.getOpcode() == ISD::UNDEF ||
5813 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5814 ((InVec1.getOpcode() == ISD::UNDEF ||
5815 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5816 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5817 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5818 // Try to match an AVX2 horizontal add/sub of signed integers.
5819 SDValue InVec2, InVec3;
5821 bool CanFold = true;
5823 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5824 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5825 ((InVec0.getOpcode() == ISD::UNDEF ||
5826 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5827 ((InVec1.getOpcode() == ISD::UNDEF ||
5828 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5829 X86Opcode = X86ISD::HADD;
5830 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5831 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5832 ((InVec0.getOpcode() == ISD::UNDEF ||
5833 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5834 ((InVec1.getOpcode() == ISD::UNDEF ||
5835 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5836 X86Opcode = X86ISD::HSUB;
5841 // Fold this build_vector into a single horizontal add/sub.
5842 // Do this only if the target has AVX2.
5843 if (Subtarget->hasAVX2())
5844 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5846 // Do not try to expand this build_vector into a pair of horizontal
5847 // add/sub if we can emit a pair of scalar add/sub.
5848 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5851 // Convert this build_vector into a pair of horizontal binop followed by
5853 bool isUndefLO = NumUndefsLO == Half;
5854 bool isUndefHI = NumUndefsHI == Half;
5855 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5856 isUndefLO, isUndefHI);
5860 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5861 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5863 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5864 X86Opcode = X86ISD::HADD;
5865 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5866 X86Opcode = X86ISD::HSUB;
5867 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5868 X86Opcode = X86ISD::FHADD;
5869 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5870 X86Opcode = X86ISD::FHSUB;
5874 // Don't try to expand this build_vector into a pair of horizontal add/sub
5875 // if we can simply emit a pair of scalar add/sub.
5876 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5879 // Convert this build_vector into two horizontal add/sub followed by
5881 bool isUndefLO = NumUndefsLO == Half;
5882 bool isUndefHI = NumUndefsHI == Half;
5883 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5884 isUndefLO, isUndefHI);
5891 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5894 MVT VT = Op.getSimpleValueType();
5895 MVT ExtVT = VT.getVectorElementType();
5896 unsigned NumElems = Op.getNumOperands();
5898 // Generate vectors for predicate vectors.
5899 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5900 return LowerBUILD_VECTORvXi1(Op, DAG);
5902 // Vectors containing all zeros can be matched by pxor and xorps later
5903 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5904 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5905 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5906 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5909 return getZeroVector(VT, Subtarget, DAG, dl);
5912 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5913 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5914 // vpcmpeqd on 256-bit vectors.
5915 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5916 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5919 if (!VT.is512BitVector())
5920 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5923 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5924 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5926 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5927 return HorizontalOp;
5928 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5931 unsigned EVTBits = ExtVT.getSizeInBits();
5933 unsigned NumZero = 0;
5934 unsigned NumNonZero = 0;
5935 unsigned NonZeros = 0;
5936 bool IsAllConstants = true;
5937 SmallSet<SDValue, 8> Values;
5938 for (unsigned i = 0; i < NumElems; ++i) {
5939 SDValue Elt = Op.getOperand(i);
5940 if (Elt.getOpcode() == ISD::UNDEF)
5943 if (Elt.getOpcode() != ISD::Constant &&
5944 Elt.getOpcode() != ISD::ConstantFP)
5945 IsAllConstants = false;
5946 if (X86::isZeroNode(Elt))
5949 NonZeros |= (1 << i);
5954 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5955 if (NumNonZero == 0)
5956 return DAG.getUNDEF(VT);
5958 // Special case for single non-zero, non-undef, element.
5959 if (NumNonZero == 1) {
5960 unsigned Idx = countTrailingZeros(NonZeros);
5961 SDValue Item = Op.getOperand(Idx);
5963 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5964 // the value are obviously zero, truncate the value to i32 and do the
5965 // insertion that way. Only do this if the value is non-constant or if the
5966 // value is a constant being inserted into element 0. It is cheaper to do
5967 // a constant pool load than it is to do a movd + shuffle.
5968 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5969 (!IsAllConstants || Idx == 0)) {
5970 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5972 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5973 EVT VecVT = MVT::v4i32;
5975 // Truncate the value (which may itself be a constant) to i32, and
5976 // convert it to a vector with movd (S2V+shuffle to zero extend).
5977 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5978 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5979 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
5980 Item, Idx * 2, true, Subtarget, DAG));
5984 // If we have a constant or non-constant insertion into the low element of
5985 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5986 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5987 // depending on what the source datatype is.
5990 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5992 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5993 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5994 if (VT.is512BitVector()) {
5995 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5996 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5997 Item, DAG.getIntPtrConstant(0, dl));
5999 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6000 "Expected an SSE value type!");
6001 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6002 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6003 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6006 // We can't directly insert an i8 or i16 into a vector, so zero extend
6008 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6009 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6010 if (VT.is256BitVector()) {
6011 if (Subtarget->hasAVX()) {
6012 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6013 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6015 // Without AVX, we need to extend to a 128-bit vector and then
6016 // insert into the 256-bit vector.
6017 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6018 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6019 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6022 assert(VT.is128BitVector() && "Expected an SSE value type!");
6023 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6024 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6026 return DAG.getBitcast(VT, Item);
6030 // Is it a vector logical left shift?
6031 if (NumElems == 2 && Idx == 1 &&
6032 X86::isZeroNode(Op.getOperand(0)) &&
6033 !X86::isZeroNode(Op.getOperand(1))) {
6034 unsigned NumBits = VT.getSizeInBits();
6035 return getVShift(true, VT,
6036 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6037 VT, Op.getOperand(1)),
6038 NumBits/2, DAG, *this, dl);
6041 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6044 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6045 // is a non-constant being inserted into an element other than the low one,
6046 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6047 // movd/movss) to move this into the low element, then shuffle it into
6049 if (EVTBits == 32) {
6050 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6051 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6055 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6056 if (Values.size() == 1) {
6057 if (EVTBits == 32) {
6058 // Instead of a shuffle like this:
6059 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6060 // Check if it's possible to issue this instead.
6061 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6062 unsigned Idx = countTrailingZeros(NonZeros);
6063 SDValue Item = Op.getOperand(Idx);
6064 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6065 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6070 // A vector full of immediates; various special cases are already
6071 // handled, so this is best done with a single constant-pool load.
6075 // For AVX-length vectors, see if we can use a vector load to get all of the
6076 // elements, otherwise build the individual 128-bit pieces and use
6077 // shuffles to put them in place.
6078 if (VT.is256BitVector() || VT.is512BitVector()) {
6079 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6081 // Check for a build vector of consecutive loads.
6082 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6085 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6087 // Build both the lower and upper subvector.
6088 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6089 makeArrayRef(&V[0], NumElems/2));
6090 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6091 makeArrayRef(&V[NumElems / 2], NumElems/2));
6093 // Recreate the wider vector with the lower and upper part.
6094 if (VT.is256BitVector())
6095 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6096 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6099 // Let legalizer expand 2-wide build_vectors.
6100 if (EVTBits == 64) {
6101 if (NumNonZero == 1) {
6102 // One half is zero or undef.
6103 unsigned Idx = countTrailingZeros(NonZeros);
6104 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6105 Op.getOperand(Idx));
6106 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6111 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6112 if (EVTBits == 8 && NumElems == 16)
6113 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6117 if (EVTBits == 16 && NumElems == 8)
6118 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6122 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6123 if (EVTBits == 32 && NumElems == 4)
6124 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6127 // If element VT is == 32 bits, turn it into a number of shuffles.
6128 SmallVector<SDValue, 8> V(NumElems);
6129 if (NumElems == 4 && NumZero > 0) {
6130 for (unsigned i = 0; i < 4; ++i) {
6131 bool isZero = !(NonZeros & (1 << i));
6133 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6135 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6138 for (unsigned i = 0; i < 2; ++i) {
6139 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6142 V[i] = V[i*2]; // Must be a zero vector.
6145 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6148 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6151 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6156 bool Reverse1 = (NonZeros & 0x3) == 2;
6157 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6161 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6162 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6164 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6167 if (Values.size() > 1 && VT.is128BitVector()) {
6168 // Check for a build vector of consecutive loads.
6169 for (unsigned i = 0; i < NumElems; ++i)
6170 V[i] = Op.getOperand(i);
6172 // Check for elements which are consecutive loads.
6173 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6176 // Check for a build vector from mostly shuffle plus few inserting.
6177 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6180 // For SSE 4.1, use insertps to put the high elements into the low element.
6181 if (Subtarget->hasSSE41()) {
6183 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6184 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6186 Result = DAG.getUNDEF(VT);
6188 for (unsigned i = 1; i < NumElems; ++i) {
6189 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6190 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6191 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6196 // Otherwise, expand into a number of unpckl*, start by extending each of
6197 // our (non-undef) elements to the full vector width with the element in the
6198 // bottom slot of the vector (which generates no code for SSE).
6199 for (unsigned i = 0; i < NumElems; ++i) {
6200 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6201 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6203 V[i] = DAG.getUNDEF(VT);
6206 // Next, we iteratively mix elements, e.g. for v4f32:
6207 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6208 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6209 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6210 unsigned EltStride = NumElems >> 1;
6211 while (EltStride != 0) {
6212 for (unsigned i = 0; i < EltStride; ++i) {
6213 // If V[i+EltStride] is undef and this is the first round of mixing,
6214 // then it is safe to just drop this shuffle: V[i] is already in the
6215 // right place, the one element (since it's the first round) being
6216 // inserted as undef can be dropped. This isn't safe for successive
6217 // rounds because they will permute elements within both vectors.
6218 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6219 EltStride == NumElems/2)
6222 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6231 // 256-bit AVX can use the vinsertf128 instruction
6232 // to create 256-bit vectors from two other 128-bit ones.
6233 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6235 MVT ResVT = Op.getSimpleValueType();
6237 assert((ResVT.is256BitVector() ||
6238 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6240 SDValue V1 = Op.getOperand(0);
6241 SDValue V2 = Op.getOperand(1);
6242 unsigned NumElems = ResVT.getVectorNumElements();
6243 if (ResVT.is256BitVector())
6244 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6246 if (Op.getNumOperands() == 4) {
6247 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6248 ResVT.getVectorNumElements()/2);
6249 SDValue V3 = Op.getOperand(2);
6250 SDValue V4 = Op.getOperand(3);
6251 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6252 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6254 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6257 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6258 const X86Subtarget *Subtarget,
6259 SelectionDAG & DAG) {
6261 MVT ResVT = Op.getSimpleValueType();
6262 unsigned NumOfOperands = Op.getNumOperands();
6264 assert(isPowerOf2_32(NumOfOperands) &&
6265 "Unexpected number of operands in CONCAT_VECTORS");
6267 if (NumOfOperands > 2) {
6268 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6269 ResVT.getVectorNumElements()/2);
6270 SmallVector<SDValue, 2> Ops;
6271 for (unsigned i = 0; i < NumOfOperands/2; i++)
6272 Ops.push_back(Op.getOperand(i));
6273 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6275 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6276 Ops.push_back(Op.getOperand(i));
6277 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6278 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6281 SDValue V1 = Op.getOperand(0);
6282 SDValue V2 = Op.getOperand(1);
6283 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6284 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6286 if (IsZeroV1 && IsZeroV2)
6287 return getZeroVector(ResVT, Subtarget, DAG, dl);
6289 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6290 SDValue Undef = DAG.getUNDEF(ResVT);
6291 unsigned NumElems = ResVT.getVectorNumElements();
6292 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6294 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6295 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6299 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6300 // Zero the upper bits of V1
6301 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6302 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6305 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6308 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6309 const X86Subtarget *Subtarget,
6310 SelectionDAG &DAG) {
6311 MVT VT = Op.getSimpleValueType();
6312 if (VT.getVectorElementType() == MVT::i1)
6313 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6315 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6316 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6317 Op.getNumOperands() == 4)));
6319 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6320 // from two other 128-bit ones.
6322 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6323 return LowerAVXCONCAT_VECTORS(Op, DAG);
6327 //===----------------------------------------------------------------------===//
6328 // Vector shuffle lowering
6330 // This is an experimental code path for lowering vector shuffles on x86. It is
6331 // designed to handle arbitrary vector shuffles and blends, gracefully
6332 // degrading performance as necessary. It works hard to recognize idiomatic
6333 // shuffles and lower them to optimal instruction patterns without leaving
6334 // a framework that allows reasonably efficient handling of all vector shuffle
6336 //===----------------------------------------------------------------------===//
6338 /// \brief Tiny helper function to identify a no-op mask.
6340 /// This is a somewhat boring predicate function. It checks whether the mask
6341 /// array input, which is assumed to be a single-input shuffle mask of the kind
6342 /// used by the X86 shuffle instructions (not a fully general
6343 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6344 /// in-place shuffle are 'no-op's.
6345 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6346 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6347 if (Mask[i] != -1 && Mask[i] != i)
6352 /// \brief Helper function to classify a mask as a single-input mask.
6354 /// This isn't a generic single-input test because in the vector shuffle
6355 /// lowering we canonicalize single inputs to be the first input operand. This
6356 /// means we can more quickly test for a single input by only checking whether
6357 /// an input from the second operand exists. We also assume that the size of
6358 /// mask corresponds to the size of the input vectors which isn't true in the
6359 /// fully general case.
6360 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6362 if (M >= (int)Mask.size())
6367 /// \brief Test whether there are elements crossing 128-bit lanes in this
6370 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6371 /// and we routinely test for these.
6372 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6373 int LaneSize = 128 / VT.getScalarSizeInBits();
6374 int Size = Mask.size();
6375 for (int i = 0; i < Size; ++i)
6376 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6381 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6383 /// This checks a shuffle mask to see if it is performing the same
6384 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6385 /// that it is also not lane-crossing. It may however involve a blend from the
6386 /// same lane of a second vector.
6388 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6389 /// non-trivial to compute in the face of undef lanes. The representation is
6390 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6391 /// entries from both V1 and V2 inputs to the wider mask.
6393 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6394 SmallVectorImpl<int> &RepeatedMask) {
6395 int LaneSize = 128 / VT.getScalarSizeInBits();
6396 RepeatedMask.resize(LaneSize, -1);
6397 int Size = Mask.size();
6398 for (int i = 0; i < Size; ++i) {
6401 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6402 // This entry crosses lanes, so there is no way to model this shuffle.
6405 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6406 if (RepeatedMask[i % LaneSize] == -1)
6407 // This is the first non-undef entry in this slot of a 128-bit lane.
6408 RepeatedMask[i % LaneSize] =
6409 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6410 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6411 // Found a mismatch with the repeated mask.
6417 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6420 /// This is a fast way to test a shuffle mask against a fixed pattern:
6422 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6424 /// It returns true if the mask is exactly as wide as the argument list, and
6425 /// each element of the mask is either -1 (signifying undef) or the value given
6426 /// in the argument.
6427 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6428 ArrayRef<int> ExpectedMask) {
6429 if (Mask.size() != ExpectedMask.size())
6432 int Size = Mask.size();
6434 // If the values are build vectors, we can look through them to find
6435 // equivalent inputs that make the shuffles equivalent.
6436 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6437 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6439 for (int i = 0; i < Size; ++i)
6440 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6441 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6442 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6443 if (!MaskBV || !ExpectedBV ||
6444 MaskBV->getOperand(Mask[i] % Size) !=
6445 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6452 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6454 /// This helper function produces an 8-bit shuffle immediate corresponding to
6455 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6456 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6459 /// NB: We rely heavily on "undef" masks preserving the input lane.
6460 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6461 SelectionDAG &DAG) {
6462 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6463 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6464 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6465 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6466 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6469 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6470 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6471 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6472 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6473 return DAG.getConstant(Imm, DL, MVT::i8);
6476 /// \brief Compute whether each element of a shuffle is zeroable.
6478 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6479 /// Either it is an undef element in the shuffle mask, the element of the input
6480 /// referenced is undef, or the element of the input referenced is known to be
6481 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6482 /// as many lanes with this technique as possible to simplify the remaining
6484 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6485 SDValue V1, SDValue V2) {
6486 SmallBitVector Zeroable(Mask.size(), false);
6488 while (V1.getOpcode() == ISD::BITCAST)
6489 V1 = V1->getOperand(0);
6490 while (V2.getOpcode() == ISD::BITCAST)
6491 V2 = V2->getOperand(0);
6493 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6494 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6496 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6498 // Handle the easy cases.
6499 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6504 // If this is an index into a build_vector node (which has the same number
6505 // of elements), dig out the input value and use it.
6506 SDValue V = M < Size ? V1 : V2;
6507 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6510 SDValue Input = V.getOperand(M % Size);
6511 // The UNDEF opcode check really should be dead code here, but not quite
6512 // worth asserting on (it isn't invalid, just unexpected).
6513 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6520 /// \brief Try to emit a bitmask instruction for a shuffle.
6522 /// This handles cases where we can model a blend exactly as a bitmask due to
6523 /// one of the inputs being zeroable.
6524 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6525 SDValue V2, ArrayRef<int> Mask,
6526 SelectionDAG &DAG) {
6527 MVT EltVT = VT.getScalarType();
6528 int NumEltBits = EltVT.getSizeInBits();
6529 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6530 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6531 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6533 if (EltVT.isFloatingPoint()) {
6534 Zero = DAG.getBitcast(EltVT, Zero);
6535 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6537 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6538 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6540 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6543 if (Mask[i] % Size != i)
6544 return SDValue(); // Not a blend.
6546 V = Mask[i] < Size ? V1 : V2;
6547 else if (V != (Mask[i] < Size ? V1 : V2))
6548 return SDValue(); // Can only let one input through the mask.
6550 VMaskOps[i] = AllOnes;
6553 return SDValue(); // No non-zeroable elements!
6555 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6556 V = DAG.getNode(VT.isFloatingPoint()
6557 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6562 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6564 /// This is used as a fallback approach when first class blend instructions are
6565 /// unavailable. Currently it is only suitable for integer vectors, but could
6566 /// be generalized for floating point vectors if desirable.
6567 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6568 SDValue V2, ArrayRef<int> Mask,
6569 SelectionDAG &DAG) {
6570 assert(VT.isInteger() && "Only supports integer vector types!");
6571 MVT EltVT = VT.getScalarType();
6572 int NumEltBits = EltVT.getSizeInBits();
6573 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6574 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6576 SmallVector<SDValue, 16> MaskOps;
6577 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6578 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6579 return SDValue(); // Shuffled input!
6580 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6583 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6584 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6585 // We have to cast V2 around.
6586 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6587 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6588 DAG.getBitcast(MaskVT, V1Mask),
6589 DAG.getBitcast(MaskVT, V2)));
6590 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6593 /// \brief Try to emit a blend instruction for a shuffle.
6595 /// This doesn't do any checks for the availability of instructions for blending
6596 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6597 /// be matched in the backend with the type given. What it does check for is
6598 /// that the shuffle mask is in fact a blend.
6599 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6600 SDValue V2, ArrayRef<int> Mask,
6601 const X86Subtarget *Subtarget,
6602 SelectionDAG &DAG) {
6603 unsigned BlendMask = 0;
6604 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6605 if (Mask[i] >= Size) {
6606 if (Mask[i] != i + Size)
6607 return SDValue(); // Shuffled V2 input!
6608 BlendMask |= 1u << i;
6611 if (Mask[i] >= 0 && Mask[i] != i)
6612 return SDValue(); // Shuffled V1 input!
6614 switch (VT.SimpleTy) {
6619 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6620 DAG.getConstant(BlendMask, DL, MVT::i8));
6624 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6628 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6629 // that instruction.
6630 if (Subtarget->hasAVX2()) {
6631 // Scale the blend by the number of 32-bit dwords per element.
6632 int Scale = VT.getScalarSizeInBits() / 32;
6634 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6635 if (Mask[i] >= Size)
6636 for (int j = 0; j < Scale; ++j)
6637 BlendMask |= 1u << (i * Scale + j);
6639 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6640 V1 = DAG.getBitcast(BlendVT, V1);
6641 V2 = DAG.getBitcast(BlendVT, V2);
6642 return DAG.getBitcast(
6643 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6644 DAG.getConstant(BlendMask, DL, MVT::i8)));
6648 // For integer shuffles we need to expand the mask and cast the inputs to
6649 // v8i16s prior to blending.
6650 int Scale = 8 / VT.getVectorNumElements();
6652 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6653 if (Mask[i] >= Size)
6654 for (int j = 0; j < Scale; ++j)
6655 BlendMask |= 1u << (i * Scale + j);
6657 V1 = DAG.getBitcast(MVT::v8i16, V1);
6658 V2 = DAG.getBitcast(MVT::v8i16, V2);
6659 return DAG.getBitcast(VT,
6660 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6661 DAG.getConstant(BlendMask, DL, MVT::i8)));
6665 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6666 SmallVector<int, 8> RepeatedMask;
6667 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6668 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6669 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6671 for (int i = 0; i < 8; ++i)
6672 if (RepeatedMask[i] >= 16)
6673 BlendMask |= 1u << i;
6674 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6675 DAG.getConstant(BlendMask, DL, MVT::i8));
6681 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6682 "256-bit byte-blends require AVX2 support!");
6684 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6685 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6688 // Scale the blend by the number of bytes per element.
6689 int Scale = VT.getScalarSizeInBits() / 8;
6691 // This form of blend is always done on bytes. Compute the byte vector
6693 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6695 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6696 // mix of LLVM's code generator and the x86 backend. We tell the code
6697 // generator that boolean values in the elements of an x86 vector register
6698 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6699 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6700 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6701 // of the element (the remaining are ignored) and 0 in that high bit would
6702 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6703 // the LLVM model for boolean values in vector elements gets the relevant
6704 // bit set, it is set backwards and over constrained relative to x86's
6706 SmallVector<SDValue, 32> VSELECTMask;
6707 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6708 for (int j = 0; j < Scale; ++j)
6709 VSELECTMask.push_back(
6710 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6711 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6714 V1 = DAG.getBitcast(BlendVT, V1);
6715 V2 = DAG.getBitcast(BlendVT, V2);
6716 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6717 DAG.getNode(ISD::BUILD_VECTOR, DL,
6718 BlendVT, VSELECTMask),
6723 llvm_unreachable("Not a supported integer vector type!");
6727 /// \brief Try to lower as a blend of elements from two inputs followed by
6728 /// a single-input permutation.
6730 /// This matches the pattern where we can blend elements from two inputs and
6731 /// then reduce the shuffle to a single-input permutation.
6732 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6735 SelectionDAG &DAG) {
6736 // We build up the blend mask while checking whether a blend is a viable way
6737 // to reduce the shuffle.
6738 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6739 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6741 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6745 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6747 if (BlendMask[Mask[i] % Size] == -1)
6748 BlendMask[Mask[i] % Size] = Mask[i];
6749 else if (BlendMask[Mask[i] % Size] != Mask[i])
6750 return SDValue(); // Can't blend in the needed input!
6752 PermuteMask[i] = Mask[i] % Size;
6755 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6756 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6759 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6760 /// blends and permutes.
6762 /// This matches the extremely common pattern for handling combined
6763 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6764 /// operations. It will try to pick the best arrangement of shuffles and
6766 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6770 SelectionDAG &DAG) {
6771 // Shuffle the input elements into the desired positions in V1 and V2 and
6772 // blend them together.
6773 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6774 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6775 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6776 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6777 if (Mask[i] >= 0 && Mask[i] < Size) {
6778 V1Mask[i] = Mask[i];
6780 } else if (Mask[i] >= Size) {
6781 V2Mask[i] = Mask[i] - Size;
6782 BlendMask[i] = i + Size;
6785 // Try to lower with the simpler initial blend strategy unless one of the
6786 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6787 // shuffle may be able to fold with a load or other benefit. However, when
6788 // we'll have to do 2x as many shuffles in order to achieve this, blending
6789 // first is a better strategy.
6790 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6791 if (SDValue BlendPerm =
6792 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6795 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6796 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6797 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6800 /// \brief Try to lower a vector shuffle as a byte rotation.
6802 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6803 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6804 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6805 /// try to generically lower a vector shuffle through such an pattern. It
6806 /// does not check for the profitability of lowering either as PALIGNR or
6807 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6808 /// This matches shuffle vectors that look like:
6810 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6812 /// Essentially it concatenates V1 and V2, shifts right by some number of
6813 /// elements, and takes the low elements as the result. Note that while this is
6814 /// specified as a *right shift* because x86 is little-endian, it is a *left
6815 /// rotate* of the vector lanes.
6816 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6819 const X86Subtarget *Subtarget,
6820 SelectionDAG &DAG) {
6821 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6823 int NumElts = Mask.size();
6824 int NumLanes = VT.getSizeInBits() / 128;
6825 int NumLaneElts = NumElts / NumLanes;
6827 // We need to detect various ways of spelling a rotation:
6828 // [11, 12, 13, 14, 15, 0, 1, 2]
6829 // [-1, 12, 13, 14, -1, -1, 1, -1]
6830 // [-1, -1, -1, -1, -1, -1, 1, 2]
6831 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6832 // [-1, 4, 5, 6, -1, -1, 9, -1]
6833 // [-1, 4, 5, 6, -1, -1, -1, -1]
6836 for (int l = 0; l < NumElts; l += NumLaneElts) {
6837 for (int i = 0; i < NumLaneElts; ++i) {
6838 if (Mask[l + i] == -1)
6840 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6842 // Get the mod-Size index and lane correct it.
6843 int LaneIdx = (Mask[l + i] % NumElts) - l;
6844 // Make sure it was in this lane.
6845 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6848 // Determine where a rotated vector would have started.
6849 int StartIdx = i - LaneIdx;
6851 // The identity rotation isn't interesting, stop.
6854 // If we found the tail of a vector the rotation must be the missing
6855 // front. If we found the head of a vector, it must be how much of the
6857 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6860 Rotation = CandidateRotation;
6861 else if (Rotation != CandidateRotation)
6862 // The rotations don't match, so we can't match this mask.
6865 // Compute which value this mask is pointing at.
6866 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6868 // Compute which of the two target values this index should be assigned
6869 // to. This reflects whether the high elements are remaining or the low
6870 // elements are remaining.
6871 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6873 // Either set up this value if we've not encountered it before, or check
6874 // that it remains consistent.
6877 else if (TargetV != MaskV)
6878 // This may be a rotation, but it pulls from the inputs in some
6879 // unsupported interleaving.
6884 // Check that we successfully analyzed the mask, and normalize the results.
6885 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6886 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6892 // The actual rotate instruction rotates bytes, so we need to scale the
6893 // rotation based on how many bytes are in the vector lane.
6894 int Scale = 16 / NumLaneElts;
6896 // SSSE3 targets can use the palignr instruction.
6897 if (Subtarget->hasSSSE3()) {
6898 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6899 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6900 Lo = DAG.getBitcast(AlignVT, Lo);
6901 Hi = DAG.getBitcast(AlignVT, Hi);
6903 return DAG.getBitcast(
6904 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6905 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6908 assert(VT.getSizeInBits() == 128 &&
6909 "Rotate-based lowering only supports 128-bit lowering!");
6910 assert(Mask.size() <= 16 &&
6911 "Can shuffle at most 16 bytes in a 128-bit vector!");
6913 // Default SSE2 implementation
6914 int LoByteShift = 16 - Rotation * Scale;
6915 int HiByteShift = Rotation * Scale;
6917 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6918 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6919 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6921 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6922 DAG.getConstant(LoByteShift, DL, MVT::i8));
6923 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6924 DAG.getConstant(HiByteShift, DL, MVT::i8));
6925 return DAG.getBitcast(VT,
6926 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6929 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6931 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6932 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6933 /// matches elements from one of the input vectors shuffled to the left or
6934 /// right with zeroable elements 'shifted in'. It handles both the strictly
6935 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6938 /// PSHL : (little-endian) left bit shift.
6939 /// [ zz, 0, zz, 2 ]
6940 /// [ -1, 4, zz, -1 ]
6941 /// PSRL : (little-endian) right bit shift.
6943 /// [ -1, -1, 7, zz]
6944 /// PSLLDQ : (little-endian) left byte shift
6945 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6946 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6947 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6948 /// PSRLDQ : (little-endian) right byte shift
6949 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6950 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6951 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6952 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6953 SDValue V2, ArrayRef<int> Mask,
6954 SelectionDAG &DAG) {
6955 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6957 int Size = Mask.size();
6958 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6960 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6961 for (int i = 0; i < Size; i += Scale)
6962 for (int j = 0; j < Shift; ++j)
6963 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6969 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6970 for (int i = 0; i != Size; i += Scale) {
6971 unsigned Pos = Left ? i + Shift : i;
6972 unsigned Low = Left ? i : i + Shift;
6973 unsigned Len = Scale - Shift;
6974 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6975 Low + (V == V1 ? 0 : Size)))
6979 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6980 bool ByteShift = ShiftEltBits > 64;
6981 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6982 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6983 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6985 // Normalize the scale for byte shifts to still produce an i64 element
6987 Scale = ByteShift ? Scale / 2 : Scale;
6989 // We need to round trip through the appropriate type for the shift.
6990 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6991 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6992 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6993 "Illegal integer vector type");
6994 V = DAG.getBitcast(ShiftVT, V);
6996 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6997 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6998 return DAG.getBitcast(VT, V);
7001 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7002 // keep doubling the size of the integer elements up to that. We can
7003 // then shift the elements of the integer vector by whole multiples of
7004 // their width within the elements of the larger integer vector. Test each
7005 // multiple to see if we can find a match with the moved element indices
7006 // and that the shifted in elements are all zeroable.
7007 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7008 for (int Shift = 1; Shift != Scale; ++Shift)
7009 for (bool Left : {true, false})
7010 if (CheckZeros(Shift, Scale, Left))
7011 for (SDValue V : {V1, V2})
7012 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7019 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7020 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7021 SDValue V2, ArrayRef<int> Mask,
7022 SelectionDAG &DAG) {
7023 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7024 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7026 int Size = Mask.size();
7027 int HalfSize = Size / 2;
7028 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7030 // Upper half must be undefined.
7031 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7034 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7035 // Remainder of lower half result is zero and upper half is all undef.
7036 auto LowerAsEXTRQ = [&]() {
7037 // Determine the extraction length from the part of the
7038 // lower half that isn't zeroable.
7040 for (; Len >= 0; --Len)
7041 if (!Zeroable[Len - 1])
7043 assert(Len > 0 && "Zeroable shuffle mask");
7045 // Attempt to match first Len sequential elements from the lower half.
7048 for (int i = 0; i != Len; ++i) {
7052 SDValue &V = (M < Size ? V1 : V2);
7055 // All mask elements must be in the lower half.
7059 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7070 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7071 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7072 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7073 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7074 DAG.getConstant(BitLen, DL, MVT::i8),
7075 DAG.getConstant(BitIdx, DL, MVT::i8));
7078 if (SDValue ExtrQ = LowerAsEXTRQ())
7081 // INSERTQ: Extract lowest Len elements from lower half of second source and
7082 // insert over first source, starting at Idx.
7083 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7084 auto LowerAsInsertQ = [&]() {
7085 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7088 // Attempt to match first source from mask before insertion point.
7089 if (isUndefInRange(Mask, 0, Idx)) {
7091 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7093 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7099 // Extend the extraction length looking to match both the insertion of
7100 // the second source and the remaining elements of the first.
7101 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7106 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7108 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7114 // Match the remaining elements of the lower half.
7115 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7117 } else if ((!Base || (Base == V1)) &&
7118 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7120 } else if ((!Base || (Base == V2)) &&
7121 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7128 // We may not have a base (first source) - this can safely be undefined.
7130 Base = DAG.getUNDEF(VT);
7132 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7133 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7134 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7135 DAG.getConstant(BitLen, DL, MVT::i8),
7136 DAG.getConstant(BitIdx, DL, MVT::i8));
7143 if (SDValue InsertQ = LowerAsInsertQ())
7149 /// \brief Lower a vector shuffle as a zero or any extension.
7151 /// Given a specific number of elements, element bit width, and extension
7152 /// stride, produce either a zero or any extension based on the available
7153 /// features of the subtarget.
7154 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7155 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7156 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7157 assert(Scale > 1 && "Need a scale to extend.");
7158 int NumElements = VT.getVectorNumElements();
7159 int EltBits = VT.getScalarSizeInBits();
7160 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7161 "Only 8, 16, and 32 bit elements can be extended.");
7162 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7164 // Found a valid zext mask! Try various lowering strategies based on the
7165 // input type and available ISA extensions.
7166 if (Subtarget->hasSSE41()) {
7167 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7168 NumElements / Scale);
7169 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7172 // For any extends we can cheat for larger element sizes and use shuffle
7173 // instructions that can fold with a load and/or copy.
7174 if (AnyExt && EltBits == 32) {
7175 int PSHUFDMask[4] = {0, -1, 1, -1};
7176 return DAG.getBitcast(
7177 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7178 DAG.getBitcast(MVT::v4i32, InputV),
7179 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7181 if (AnyExt && EltBits == 16 && Scale > 2) {
7182 int PSHUFDMask[4] = {0, -1, 0, -1};
7183 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7184 DAG.getBitcast(MVT::v4i32, InputV),
7185 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7186 int PSHUFHWMask[4] = {1, -1, -1, -1};
7187 return DAG.getBitcast(
7188 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7189 DAG.getBitcast(MVT::v8i16, InputV),
7190 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7193 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7195 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7196 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7197 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7199 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7200 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7201 DAG.getConstant(EltBits, DL, MVT::i8),
7202 DAG.getConstant(0, DL, MVT::i8)));
7203 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7204 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7207 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7208 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7209 DAG.getConstant(EltBits, DL, MVT::i8),
7210 DAG.getConstant(EltBits, DL, MVT::i8)));
7211 return DAG.getNode(ISD::BITCAST, DL, VT,
7212 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7215 // If this would require more than 2 unpack instructions to expand, use
7216 // pshufb when available. We can only use more than 2 unpack instructions
7217 // when zero extending i8 elements which also makes it easier to use pshufb.
7218 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7219 assert(NumElements == 16 && "Unexpected byte vector width!");
7220 SDValue PSHUFBMask[16];
7221 for (int i = 0; i < 16; ++i)
7223 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7224 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7225 return DAG.getBitcast(VT,
7226 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7227 DAG.getNode(ISD::BUILD_VECTOR, DL,
7228 MVT::v16i8, PSHUFBMask)));
7231 // Otherwise emit a sequence of unpacks.
7233 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7234 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7235 : getZeroVector(InputVT, Subtarget, DAG, DL);
7236 InputV = DAG.getBitcast(InputVT, InputV);
7237 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7241 } while (Scale > 1);
7242 return DAG.getBitcast(VT, InputV);
7245 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7247 /// This routine will try to do everything in its power to cleverly lower
7248 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7249 /// check for the profitability of this lowering, it tries to aggressively
7250 /// match this pattern. It will use all of the micro-architectural details it
7251 /// can to emit an efficient lowering. It handles both blends with all-zero
7252 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7253 /// masking out later).
7255 /// The reason we have dedicated lowering for zext-style shuffles is that they
7256 /// are both incredibly common and often quite performance sensitive.
7257 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7258 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7259 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7260 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7262 int Bits = VT.getSizeInBits();
7263 int NumElements = VT.getVectorNumElements();
7264 assert(VT.getScalarSizeInBits() <= 32 &&
7265 "Exceeds 32-bit integer zero extension limit");
7266 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7268 // Define a helper function to check a particular ext-scale and lower to it if
7270 auto Lower = [&](int Scale) -> SDValue {
7273 for (int i = 0; i < NumElements; ++i) {
7275 continue; // Valid anywhere but doesn't tell us anything.
7276 if (i % Scale != 0) {
7277 // Each of the extended elements need to be zeroable.
7281 // We no longer are in the anyext case.
7286 // Each of the base elements needs to be consecutive indices into the
7287 // same input vector.
7288 SDValue V = Mask[i] < NumElements ? V1 : V2;
7291 else if (InputV != V)
7292 return SDValue(); // Flip-flopping inputs.
7294 if (Mask[i] % NumElements != i / Scale)
7295 return SDValue(); // Non-consecutive strided elements.
7298 // If we fail to find an input, we have a zero-shuffle which should always
7299 // have already been handled.
7300 // FIXME: Maybe handle this here in case during blending we end up with one?
7304 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7305 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7308 // The widest scale possible for extending is to a 64-bit integer.
7309 assert(Bits % 64 == 0 &&
7310 "The number of bits in a vector must be divisible by 64 on x86!");
7311 int NumExtElements = Bits / 64;
7313 // Each iteration, try extending the elements half as much, but into twice as
7315 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7316 assert(NumElements % NumExtElements == 0 &&
7317 "The input vector size must be divisible by the extended size.");
7318 if (SDValue V = Lower(NumElements / NumExtElements))
7322 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7326 // Returns one of the source operands if the shuffle can be reduced to a
7327 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7328 auto CanZExtLowHalf = [&]() {
7329 for (int i = NumElements / 2; i != NumElements; ++i)
7332 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7334 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7339 if (SDValue V = CanZExtLowHalf()) {
7340 V = DAG.getBitcast(MVT::v2i64, V);
7341 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7342 return DAG.getBitcast(VT, V);
7345 // No viable ext lowering found.
7349 /// \brief Try to get a scalar value for a specific element of a vector.
7351 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7352 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7353 SelectionDAG &DAG) {
7354 MVT VT = V.getSimpleValueType();
7355 MVT EltVT = VT.getVectorElementType();
7356 while (V.getOpcode() == ISD::BITCAST)
7357 V = V.getOperand(0);
7358 // If the bitcasts shift the element size, we can't extract an equivalent
7360 MVT NewVT = V.getSimpleValueType();
7361 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7364 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7365 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7366 // Ensure the scalar operand is the same size as the destination.
7367 // FIXME: Add support for scalar truncation where possible.
7368 SDValue S = V.getOperand(Idx);
7369 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7370 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7376 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7378 /// This is particularly important because the set of instructions varies
7379 /// significantly based on whether the operand is a load or not.
7380 static bool isShuffleFoldableLoad(SDValue V) {
7381 while (V.getOpcode() == ISD::BITCAST)
7382 V = V.getOperand(0);
7384 return ISD::isNON_EXTLoad(V.getNode());
7387 /// \brief Try to lower insertion of a single element into a zero vector.
7389 /// This is a common pattern that we have especially efficient patterns to lower
7390 /// across all subtarget feature sets.
7391 static SDValue lowerVectorShuffleAsElementInsertion(
7392 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7393 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7394 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7396 MVT EltVT = VT.getVectorElementType();
7398 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7399 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7401 bool IsV1Zeroable = true;
7402 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7403 if (i != V2Index && !Zeroable[i]) {
7404 IsV1Zeroable = false;
7408 // Check for a single input from a SCALAR_TO_VECTOR node.
7409 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7410 // all the smarts here sunk into that routine. However, the current
7411 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7412 // vector shuffle lowering is dead.
7413 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7415 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7416 // We need to zext the scalar if it is smaller than an i32.
7417 V2S = DAG.getBitcast(EltVT, V2S);
7418 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7419 // Using zext to expand a narrow element won't work for non-zero
7424 // Zero-extend directly to i32.
7426 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7428 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7429 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7430 EltVT == MVT::i16) {
7431 // Either not inserting from the low element of the input or the input
7432 // element size is too small to use VZEXT_MOVL to clear the high bits.
7436 if (!IsV1Zeroable) {
7437 // If V1 can't be treated as a zero vector we have fewer options to lower
7438 // this. We can't support integer vectors or non-zero targets cheaply, and
7439 // the V1 elements can't be permuted in any way.
7440 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7441 if (!VT.isFloatingPoint() || V2Index != 0)
7443 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7444 V1Mask[V2Index] = -1;
7445 if (!isNoopShuffleMask(V1Mask))
7447 // This is essentially a special case blend operation, but if we have
7448 // general purpose blend operations, they are always faster. Bail and let
7449 // the rest of the lowering handle these as blends.
7450 if (Subtarget->hasSSE41())
7453 // Otherwise, use MOVSD or MOVSS.
7454 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7455 "Only two types of floating point element types to handle!");
7456 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7460 // This lowering only works for the low element with floating point vectors.
7461 if (VT.isFloatingPoint() && V2Index != 0)
7464 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7466 V2 = DAG.getBitcast(VT, V2);
7469 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7470 // the desired position. Otherwise it is more efficient to do a vector
7471 // shift left. We know that we can do a vector shift left because all
7472 // the inputs are zero.
7473 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7474 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7475 V2Shuffle[V2Index] = 0;
7476 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7478 V2 = DAG.getBitcast(MVT::v2i64, V2);
7480 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7481 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7482 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7483 DAG.getDataLayout(), VT)));
7484 V2 = DAG.getBitcast(VT, V2);
7490 /// \brief Try to lower broadcast of a single element.
7492 /// For convenience, this code also bundles all of the subtarget feature set
7493 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7494 /// a convenient way to factor it out.
7495 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7497 const X86Subtarget *Subtarget,
7498 SelectionDAG &DAG) {
7499 if (!Subtarget->hasAVX())
7501 if (VT.isInteger() && !Subtarget->hasAVX2())
7504 // Check that the mask is a broadcast.
7505 int BroadcastIdx = -1;
7507 if (M >= 0 && BroadcastIdx == -1)
7509 else if (M >= 0 && M != BroadcastIdx)
7512 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7513 "a sorted mask where the broadcast "
7516 // Go up the chain of (vector) values to find a scalar load that we can
7517 // combine with the broadcast.
7519 switch (V.getOpcode()) {
7520 case ISD::CONCAT_VECTORS: {
7521 int OperandSize = Mask.size() / V.getNumOperands();
7522 V = V.getOperand(BroadcastIdx / OperandSize);
7523 BroadcastIdx %= OperandSize;
7527 case ISD::INSERT_SUBVECTOR: {
7528 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7529 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7533 int BeginIdx = (int)ConstantIdx->getZExtValue();
7535 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7536 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7537 BroadcastIdx -= BeginIdx;
7548 // Check if this is a broadcast of a scalar. We special case lowering
7549 // for scalars so that we can more effectively fold with loads.
7550 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7551 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7552 V = V.getOperand(BroadcastIdx);
7554 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7555 // Only AVX2 has register broadcasts.
7556 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7558 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7559 // We can't broadcast from a vector register without AVX2, and we can only
7560 // broadcast from the zero-element of a vector register.
7564 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7567 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7568 // INSERTPS when the V1 elements are already in the correct locations
7569 // because otherwise we can just always use two SHUFPS instructions which
7570 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7571 // perform INSERTPS if a single V1 element is out of place and all V2
7572 // elements are zeroable.
7573 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7575 SelectionDAG &DAG) {
7576 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7577 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7578 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7579 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7581 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7584 int V1DstIndex = -1;
7585 int V2DstIndex = -1;
7586 bool V1UsedInPlace = false;
7588 for (int i = 0; i < 4; ++i) {
7589 // Synthesize a zero mask from the zeroable elements (includes undefs).
7595 // Flag if we use any V1 inputs in place.
7597 V1UsedInPlace = true;
7601 // We can only insert a single non-zeroable element.
7602 if (V1DstIndex != -1 || V2DstIndex != -1)
7606 // V1 input out of place for insertion.
7609 // V2 input for insertion.
7614 // Don't bother if we have no (non-zeroable) element for insertion.
7615 if (V1DstIndex == -1 && V2DstIndex == -1)
7618 // Determine element insertion src/dst indices. The src index is from the
7619 // start of the inserted vector, not the start of the concatenated vector.
7620 unsigned V2SrcIndex = 0;
7621 if (V1DstIndex != -1) {
7622 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7623 // and don't use the original V2 at all.
7624 V2SrcIndex = Mask[V1DstIndex];
7625 V2DstIndex = V1DstIndex;
7628 V2SrcIndex = Mask[V2DstIndex] - 4;
7631 // If no V1 inputs are used in place, then the result is created only from
7632 // the zero mask and the V2 insertion - so remove V1 dependency.
7634 V1 = DAG.getUNDEF(MVT::v4f32);
7636 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7637 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7639 // Insert the V2 element into the desired position.
7641 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7642 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7645 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7646 /// UNPCK instruction.
7648 /// This specifically targets cases where we end up with alternating between
7649 /// the two inputs, and so can permute them into something that feeds a single
7650 /// UNPCK instruction. Note that this routine only targets integer vectors
7651 /// because for floating point vectors we have a generalized SHUFPS lowering
7652 /// strategy that handles everything that doesn't *exactly* match an unpack,
7653 /// making this clever lowering unnecessary.
7654 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7655 SDValue V2, ArrayRef<int> Mask,
7656 SelectionDAG &DAG) {
7657 assert(!VT.isFloatingPoint() &&
7658 "This routine only supports integer vectors.");
7659 assert(!isSingleInputShuffleMask(Mask) &&
7660 "This routine should only be used when blending two inputs.");
7661 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7663 int Size = Mask.size();
7665 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7666 return M >= 0 && M % Size < Size / 2;
7668 int NumHiInputs = std::count_if(
7669 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7671 bool UnpackLo = NumLoInputs >= NumHiInputs;
7673 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7674 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7675 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7677 for (int i = 0; i < Size; ++i) {
7681 // Each element of the unpack contains Scale elements from this mask.
7682 int UnpackIdx = i / Scale;
7684 // We only handle the case where V1 feeds the first slots of the unpack.
7685 // We rely on canonicalization to ensure this is the case.
7686 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7689 // Setup the mask for this input. The indexing is tricky as we have to
7690 // handle the unpack stride.
7691 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7692 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7696 // If we will have to shuffle both inputs to use the unpack, check whether
7697 // we can just unpack first and shuffle the result. If so, skip this unpack.
7698 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7699 !isNoopShuffleMask(V2Mask))
7702 // Shuffle the inputs into place.
7703 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7704 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7706 // Cast the inputs to the type we will use to unpack them.
7707 V1 = DAG.getBitcast(UnpackVT, V1);
7708 V2 = DAG.getBitcast(UnpackVT, V2);
7710 // Unpack the inputs and cast the result back to the desired type.
7711 return DAG.getBitcast(
7712 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7716 // We try each unpack from the largest to the smallest to try and find one
7717 // that fits this mask.
7718 int OrigNumElements = VT.getVectorNumElements();
7719 int OrigScalarSize = VT.getScalarSizeInBits();
7720 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7721 int Scale = ScalarSize / OrigScalarSize;
7722 int NumElements = OrigNumElements / Scale;
7723 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7724 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7728 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7730 if (NumLoInputs == 0 || NumHiInputs == 0) {
7731 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7732 "We have to have *some* inputs!");
7733 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7735 // FIXME: We could consider the total complexity of the permute of each
7736 // possible unpacking. Or at the least we should consider how many
7737 // half-crossings are created.
7738 // FIXME: We could consider commuting the unpacks.
7740 SmallVector<int, 32> PermMask;
7741 PermMask.assign(Size, -1);
7742 for (int i = 0; i < Size; ++i) {
7746 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7749 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7751 return DAG.getVectorShuffle(
7752 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7754 DAG.getUNDEF(VT), PermMask);
7760 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7762 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7763 /// support for floating point shuffles but not integer shuffles. These
7764 /// instructions will incur a domain crossing penalty on some chips though so
7765 /// it is better to avoid lowering through this for integer vectors where
7767 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7768 const X86Subtarget *Subtarget,
7769 SelectionDAG &DAG) {
7771 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7772 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7773 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7774 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7775 ArrayRef<int> Mask = SVOp->getMask();
7776 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7778 if (isSingleInputShuffleMask(Mask)) {
7779 // Use low duplicate instructions for masks that match their pattern.
7780 if (Subtarget->hasSSE3())
7781 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7782 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7784 // Straight shuffle of a single input vector. Simulate this by using the
7785 // single input as both of the "inputs" to this instruction..
7786 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7788 if (Subtarget->hasAVX()) {
7789 // If we have AVX, we can use VPERMILPS which will allow folding a load
7790 // into the shuffle.
7791 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7792 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7795 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7796 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7798 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7799 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7801 // If we have a single input, insert that into V1 if we can do so cheaply.
7802 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7803 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7804 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7806 // Try inverting the insertion since for v2 masks it is easy to do and we
7807 // can't reliably sort the mask one way or the other.
7808 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7809 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7810 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7811 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7815 // Try to use one of the special instruction patterns to handle two common
7816 // blend patterns if a zero-blend above didn't work.
7817 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7818 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7819 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7820 // We can either use a special instruction to load over the low double or
7821 // to move just the low double.
7823 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7825 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7827 if (Subtarget->hasSSE41())
7828 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7832 // Use dedicated unpack instructions for masks that match their pattern.
7833 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7834 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7835 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7836 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7838 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7839 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7840 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7843 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7845 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7846 /// the integer unit to minimize domain crossing penalties. However, for blends
7847 /// it falls back to the floating point shuffle operation with appropriate bit
7849 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7850 const X86Subtarget *Subtarget,
7851 SelectionDAG &DAG) {
7853 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7854 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7855 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7856 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7857 ArrayRef<int> Mask = SVOp->getMask();
7858 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7860 if (isSingleInputShuffleMask(Mask)) {
7861 // Check for being able to broadcast a single element.
7862 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7863 Mask, Subtarget, DAG))
7866 // Straight shuffle of a single input vector. For everything from SSE2
7867 // onward this has a single fast instruction with no scary immediates.
7868 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7869 V1 = DAG.getBitcast(MVT::v4i32, V1);
7870 int WidenedMask[4] = {
7871 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7872 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7873 return DAG.getBitcast(
7875 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7876 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7878 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7879 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7880 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7881 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7883 // If we have a blend of two PACKUS operations an the blend aligns with the
7884 // low and half halves, we can just merge the PACKUS operations. This is
7885 // particularly important as it lets us merge shuffles that this routine itself
7887 auto GetPackNode = [](SDValue V) {
7888 while (V.getOpcode() == ISD::BITCAST)
7889 V = V.getOperand(0);
7891 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7893 if (SDValue V1Pack = GetPackNode(V1))
7894 if (SDValue V2Pack = GetPackNode(V2))
7895 return DAG.getBitcast(MVT::v2i64,
7896 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7897 Mask[0] == 0 ? V1Pack.getOperand(0)
7898 : V1Pack.getOperand(1),
7899 Mask[1] == 2 ? V2Pack.getOperand(0)
7900 : V2Pack.getOperand(1)));
7902 // Try to use shift instructions.
7904 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7907 // When loading a scalar and then shuffling it into a vector we can often do
7908 // the insertion cheaply.
7909 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7910 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7912 // Try inverting the insertion since for v2 masks it is easy to do and we
7913 // can't reliably sort the mask one way or the other.
7914 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7915 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7916 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7919 // We have different paths for blend lowering, but they all must use the
7920 // *exact* same predicate.
7921 bool IsBlendSupported = Subtarget->hasSSE41();
7922 if (IsBlendSupported)
7923 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7927 // Use dedicated unpack instructions for masks that match their pattern.
7928 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7929 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7930 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7931 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7933 // Try to use byte rotation instructions.
7934 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7935 if (Subtarget->hasSSSE3())
7936 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7937 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7940 // If we have direct support for blends, we should lower by decomposing into
7941 // a permute. That will be faster than the domain cross.
7942 if (IsBlendSupported)
7943 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7946 // We implement this with SHUFPD which is pretty lame because it will likely
7947 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7948 // However, all the alternatives are still more cycles and newer chips don't
7949 // have this problem. It would be really nice if x86 had better shuffles here.
7950 V1 = DAG.getBitcast(MVT::v2f64, V1);
7951 V2 = DAG.getBitcast(MVT::v2f64, V2);
7952 return DAG.getBitcast(MVT::v2i64,
7953 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7956 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7958 /// This is used to disable more specialized lowerings when the shufps lowering
7959 /// will happen to be efficient.
7960 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7961 // This routine only handles 128-bit shufps.
7962 assert(Mask.size() == 4 && "Unsupported mask size!");
7964 // To lower with a single SHUFPS we need to have the low half and high half
7965 // each requiring a single input.
7966 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7968 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7974 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7976 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7977 /// It makes no assumptions about whether this is the *best* lowering, it simply
7979 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7980 ArrayRef<int> Mask, SDValue V1,
7981 SDValue V2, SelectionDAG &DAG) {
7982 SDValue LowV = V1, HighV = V2;
7983 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7986 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7988 if (NumV2Elements == 1) {
7990 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7993 // Compute the index adjacent to V2Index and in the same half by toggling
7995 int V2AdjIndex = V2Index ^ 1;
7997 if (Mask[V2AdjIndex] == -1) {
7998 // Handles all the cases where we have a single V2 element and an undef.
7999 // This will only ever happen in the high lanes because we commute the
8000 // vector otherwise.
8002 std::swap(LowV, HighV);
8003 NewMask[V2Index] -= 4;
8005 // Handle the case where the V2 element ends up adjacent to a V1 element.
8006 // To make this work, blend them together as the first step.
8007 int V1Index = V2AdjIndex;
8008 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8009 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8010 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8012 // Now proceed to reconstruct the final blend as we have the necessary
8013 // high or low half formed.
8020 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8021 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8023 } else if (NumV2Elements == 2) {
8024 if (Mask[0] < 4 && Mask[1] < 4) {
8025 // Handle the easy case where we have V1 in the low lanes and V2 in the
8029 } else if (Mask[2] < 4 && Mask[3] < 4) {
8030 // We also handle the reversed case because this utility may get called
8031 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8032 // arrange things in the right direction.
8038 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8039 // trying to place elements directly, just blend them and set up the final
8040 // shuffle to place them.
8042 // The first two blend mask elements are for V1, the second two are for
8044 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8045 Mask[2] < 4 ? Mask[2] : Mask[3],
8046 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8047 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8048 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8049 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8051 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8054 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8055 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8056 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8057 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8060 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8061 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8064 /// \brief Lower 4-lane 32-bit floating point shuffles.
8066 /// Uses instructions exclusively from the floating point unit to minimize
8067 /// domain crossing penalties, as these are sufficient to implement all v4f32
8069 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8070 const X86Subtarget *Subtarget,
8071 SelectionDAG &DAG) {
8073 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8074 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8075 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8076 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8077 ArrayRef<int> Mask = SVOp->getMask();
8078 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8081 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8083 if (NumV2Elements == 0) {
8084 // Check for being able to broadcast a single element.
8085 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8086 Mask, Subtarget, DAG))
8089 // Use even/odd duplicate instructions for masks that match their pattern.
8090 if (Subtarget->hasSSE3()) {
8091 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8092 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8093 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8094 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8097 if (Subtarget->hasAVX()) {
8098 // If we have AVX, we can use VPERMILPS which will allow folding a load
8099 // into the shuffle.
8100 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8101 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8104 // Otherwise, use a straight shuffle of a single input vector. We pass the
8105 // input vector to both operands to simulate this with a SHUFPS.
8106 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8107 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8110 // There are special ways we can lower some single-element blends. However, we
8111 // have custom ways we can lower more complex single-element blends below that
8112 // we defer to if both this and BLENDPS fail to match, so restrict this to
8113 // when the V2 input is targeting element 0 of the mask -- that is the fast
8115 if (NumV2Elements == 1 && Mask[0] >= 4)
8116 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8117 Mask, Subtarget, DAG))
8120 if (Subtarget->hasSSE41()) {
8121 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8125 // Use INSERTPS if we can complete the shuffle efficiently.
8126 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8129 if (!isSingleSHUFPSMask(Mask))
8130 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8131 DL, MVT::v4f32, V1, V2, Mask, DAG))
8135 // Use dedicated unpack instructions for masks that match their pattern.
8136 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8137 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8138 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8139 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8140 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8141 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8142 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8143 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8145 // Otherwise fall back to a SHUFPS lowering strategy.
8146 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8149 /// \brief Lower 4-lane i32 vector shuffles.
8151 /// We try to handle these with integer-domain shuffles where we can, but for
8152 /// blends we use the floating point domain blend instructions.
8153 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8154 const X86Subtarget *Subtarget,
8155 SelectionDAG &DAG) {
8157 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8158 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8159 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8160 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8161 ArrayRef<int> Mask = SVOp->getMask();
8162 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8164 // Whenever we can lower this as a zext, that instruction is strictly faster
8165 // than any alternative. It also allows us to fold memory operands into the
8166 // shuffle in many cases.
8167 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8168 Mask, Subtarget, DAG))
8172 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8174 if (NumV2Elements == 0) {
8175 // Check for being able to broadcast a single element.
8176 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8177 Mask, Subtarget, DAG))
8180 // Straight shuffle of a single input vector. For everything from SSE2
8181 // onward this has a single fast instruction with no scary immediates.
8182 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8183 // but we aren't actually going to use the UNPCK instruction because doing
8184 // so prevents folding a load into this instruction or making a copy.
8185 const int UnpackLoMask[] = {0, 0, 1, 1};
8186 const int UnpackHiMask[] = {2, 2, 3, 3};
8187 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8188 Mask = UnpackLoMask;
8189 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8190 Mask = UnpackHiMask;
8192 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8193 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8196 // Try to use shift instructions.
8198 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8201 // There are special ways we can lower some single-element blends.
8202 if (NumV2Elements == 1)
8203 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8204 Mask, Subtarget, DAG))
8207 // We have different paths for blend lowering, but they all must use the
8208 // *exact* same predicate.
8209 bool IsBlendSupported = Subtarget->hasSSE41();
8210 if (IsBlendSupported)
8211 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8215 if (SDValue Masked =
8216 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8219 // Use dedicated unpack instructions for masks that match their pattern.
8220 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8221 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8222 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8223 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8224 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8225 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8226 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8227 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8229 // Try to use byte rotation instructions.
8230 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8231 if (Subtarget->hasSSSE3())
8232 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8233 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8236 // If we have direct support for blends, we should lower by decomposing into
8237 // a permute. That will be faster than the domain cross.
8238 if (IsBlendSupported)
8239 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8242 // Try to lower by permuting the inputs into an unpack instruction.
8243 if (SDValue Unpack =
8244 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
8247 // We implement this with SHUFPS because it can blend from two vectors.
8248 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8249 // up the inputs, bypassing domain shift penalties that we would encur if we
8250 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8252 return DAG.getBitcast(
8254 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8255 DAG.getBitcast(MVT::v4f32, V2), Mask));
8258 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8259 /// shuffle lowering, and the most complex part.
8261 /// The lowering strategy is to try to form pairs of input lanes which are
8262 /// targeted at the same half of the final vector, and then use a dword shuffle
8263 /// to place them onto the right half, and finally unpack the paired lanes into
8264 /// their final position.
8266 /// The exact breakdown of how to form these dword pairs and align them on the
8267 /// correct sides is really tricky. See the comments within the function for
8268 /// more of the details.
8270 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8271 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8272 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8273 /// vector, form the analogous 128-bit 8-element Mask.
8274 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8275 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8276 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8277 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8278 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8280 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8281 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8282 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8284 SmallVector<int, 4> LoInputs;
8285 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8286 [](int M) { return M >= 0; });
8287 std::sort(LoInputs.begin(), LoInputs.end());
8288 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8289 SmallVector<int, 4> HiInputs;
8290 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8291 [](int M) { return M >= 0; });
8292 std::sort(HiInputs.begin(), HiInputs.end());
8293 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8295 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8296 int NumHToL = LoInputs.size() - NumLToL;
8298 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8299 int NumHToH = HiInputs.size() - NumLToH;
8300 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8301 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8302 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8303 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8305 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8306 // such inputs we can swap two of the dwords across the half mark and end up
8307 // with <=2 inputs to each half in each half. Once there, we can fall through
8308 // to the generic code below. For example:
8310 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8311 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8313 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8314 // and an existing 2-into-2 on the other half. In this case we may have to
8315 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8316 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8317 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8318 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8319 // half than the one we target for fixing) will be fixed when we re-enter this
8320 // path. We will also combine away any sequence of PSHUFD instructions that
8321 // result into a single instruction. Here is an example of the tricky case:
8323 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8324 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8326 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8328 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8329 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8331 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8332 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8334 // The result is fine to be handled by the generic logic.
8335 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8336 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8337 int AOffset, int BOffset) {
8338 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8339 "Must call this with A having 3 or 1 inputs from the A half.");
8340 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8341 "Must call this with B having 1 or 3 inputs from the B half.");
8342 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8343 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8345 bool ThreeAInputs = AToAInputs.size() == 3;
8347 // Compute the index of dword with only one word among the three inputs in
8348 // a half by taking the sum of the half with three inputs and subtracting
8349 // the sum of the actual three inputs. The difference is the remaining
8352 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8353 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8354 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8355 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8356 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8357 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8358 int TripleNonInputIdx =
8359 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8360 TripleDWord = TripleNonInputIdx / 2;
8362 // We use xor with one to compute the adjacent DWord to whichever one the
8364 OneInputDWord = (OneInput / 2) ^ 1;
8366 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8367 // and BToA inputs. If there is also such a problem with the BToB and AToB
8368 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8369 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8370 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8371 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8372 // Compute how many inputs will be flipped by swapping these DWords. We
8374 // to balance this to ensure we don't form a 3-1 shuffle in the other
8376 int NumFlippedAToBInputs =
8377 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8378 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8379 int NumFlippedBToBInputs =
8380 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8381 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8382 if ((NumFlippedAToBInputs == 1 &&
8383 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8384 (NumFlippedBToBInputs == 1 &&
8385 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8386 // We choose whether to fix the A half or B half based on whether that
8387 // half has zero flipped inputs. At zero, we may not be able to fix it
8388 // with that half. We also bias towards fixing the B half because that
8389 // will more commonly be the high half, and we have to bias one way.
8390 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8391 ArrayRef<int> Inputs) {
8392 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8393 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8394 PinnedIdx ^ 1) != Inputs.end();
8395 // Determine whether the free index is in the flipped dword or the
8396 // unflipped dword based on where the pinned index is. We use this bit
8397 // in an xor to conditionally select the adjacent dword.
8398 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8399 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8400 FixFreeIdx) != Inputs.end();
8401 if (IsFixIdxInput == IsFixFreeIdxInput)
8403 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8404 FixFreeIdx) != Inputs.end();
8405 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8406 "We need to be changing the number of flipped inputs!");
8407 int PSHUFHalfMask[] = {0, 1, 2, 3};
8408 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8409 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8411 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8414 if (M != -1 && M == FixIdx)
8416 else if (M != -1 && M == FixFreeIdx)
8419 if (NumFlippedBToBInputs != 0) {
8421 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8422 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8424 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8425 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8426 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8431 int PSHUFDMask[] = {0, 1, 2, 3};
8432 PSHUFDMask[ADWord] = BDWord;
8433 PSHUFDMask[BDWord] = ADWord;
8436 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8437 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8439 // Adjust the mask to match the new locations of A and B.
8441 if (M != -1 && M/2 == ADWord)
8442 M = 2 * BDWord + M % 2;
8443 else if (M != -1 && M/2 == BDWord)
8444 M = 2 * ADWord + M % 2;
8446 // Recurse back into this routine to re-compute state now that this isn't
8447 // a 3 and 1 problem.
8448 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8451 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8452 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8453 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8454 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8456 // At this point there are at most two inputs to the low and high halves from
8457 // each half. That means the inputs can always be grouped into dwords and
8458 // those dwords can then be moved to the correct half with a dword shuffle.
8459 // We use at most one low and one high word shuffle to collect these paired
8460 // inputs into dwords, and finally a dword shuffle to place them.
8461 int PSHUFLMask[4] = {-1, -1, -1, -1};
8462 int PSHUFHMask[4] = {-1, -1, -1, -1};
8463 int PSHUFDMask[4] = {-1, -1, -1, -1};
8465 // First fix the masks for all the inputs that are staying in their
8466 // original halves. This will then dictate the targets of the cross-half
8468 auto fixInPlaceInputs =
8469 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8470 MutableArrayRef<int> SourceHalfMask,
8471 MutableArrayRef<int> HalfMask, int HalfOffset) {
8472 if (InPlaceInputs.empty())
8474 if (InPlaceInputs.size() == 1) {
8475 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8476 InPlaceInputs[0] - HalfOffset;
8477 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8480 if (IncomingInputs.empty()) {
8481 // Just fix all of the in place inputs.
8482 for (int Input : InPlaceInputs) {
8483 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8484 PSHUFDMask[Input / 2] = Input / 2;
8489 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8490 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8491 InPlaceInputs[0] - HalfOffset;
8492 // Put the second input next to the first so that they are packed into
8493 // a dword. We find the adjacent index by toggling the low bit.
8494 int AdjIndex = InPlaceInputs[0] ^ 1;
8495 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8496 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8497 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8499 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8500 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8502 // Now gather the cross-half inputs and place them into a free dword of
8503 // their target half.
8504 // FIXME: This operation could almost certainly be simplified dramatically to
8505 // look more like the 3-1 fixing operation.
8506 auto moveInputsToRightHalf = [&PSHUFDMask](
8507 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8508 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8509 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8511 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8512 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8514 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8516 int LowWord = Word & ~1;
8517 int HighWord = Word | 1;
8518 return isWordClobbered(SourceHalfMask, LowWord) ||
8519 isWordClobbered(SourceHalfMask, HighWord);
8522 if (IncomingInputs.empty())
8525 if (ExistingInputs.empty()) {
8526 // Map any dwords with inputs from them into the right half.
8527 for (int Input : IncomingInputs) {
8528 // If the source half mask maps over the inputs, turn those into
8529 // swaps and use the swapped lane.
8530 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8531 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8532 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8533 Input - SourceOffset;
8534 // We have to swap the uses in our half mask in one sweep.
8535 for (int &M : HalfMask)
8536 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8538 else if (M == Input)
8539 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8541 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8542 Input - SourceOffset &&
8543 "Previous placement doesn't match!");
8545 // Note that this correctly re-maps both when we do a swap and when
8546 // we observe the other side of the swap above. We rely on that to
8547 // avoid swapping the members of the input list directly.
8548 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8551 // Map the input's dword into the correct half.
8552 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8553 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8555 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8557 "Previous placement doesn't match!");
8560 // And just directly shift any other-half mask elements to be same-half
8561 // as we will have mirrored the dword containing the element into the
8562 // same position within that half.
8563 for (int &M : HalfMask)
8564 if (M >= SourceOffset && M < SourceOffset + 4) {
8565 M = M - SourceOffset + DestOffset;
8566 assert(M >= 0 && "This should never wrap below zero!");
8571 // Ensure we have the input in a viable dword of its current half. This
8572 // is particularly tricky because the original position may be clobbered
8573 // by inputs being moved and *staying* in that half.
8574 if (IncomingInputs.size() == 1) {
8575 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8576 int InputFixed = std::find(std::begin(SourceHalfMask),
8577 std::end(SourceHalfMask), -1) -
8578 std::begin(SourceHalfMask) + SourceOffset;
8579 SourceHalfMask[InputFixed - SourceOffset] =
8580 IncomingInputs[0] - SourceOffset;
8581 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8583 IncomingInputs[0] = InputFixed;
8585 } else if (IncomingInputs.size() == 2) {
8586 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8587 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8588 // We have two non-adjacent or clobbered inputs we need to extract from
8589 // the source half. To do this, we need to map them into some adjacent
8590 // dword slot in the source mask.
8591 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8592 IncomingInputs[1] - SourceOffset};
8594 // If there is a free slot in the source half mask adjacent to one of
8595 // the inputs, place the other input in it. We use (Index XOR 1) to
8596 // compute an adjacent index.
8597 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8598 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8599 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8600 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8601 InputsFixed[1] = InputsFixed[0] ^ 1;
8602 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8603 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8604 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8605 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8606 InputsFixed[0] = InputsFixed[1] ^ 1;
8607 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8608 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8609 // The two inputs are in the same DWord but it is clobbered and the
8610 // adjacent DWord isn't used at all. Move both inputs to the free
8612 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8613 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8614 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8615 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8617 // The only way we hit this point is if there is no clobbering
8618 // (because there are no off-half inputs to this half) and there is no
8619 // free slot adjacent to one of the inputs. In this case, we have to
8620 // swap an input with a non-input.
8621 for (int i = 0; i < 4; ++i)
8622 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8623 "We can't handle any clobbers here!");
8624 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8625 "Cannot have adjacent inputs here!");
8627 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8628 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8630 // We also have to update the final source mask in this case because
8631 // it may need to undo the above swap.
8632 for (int &M : FinalSourceHalfMask)
8633 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8634 M = InputsFixed[1] + SourceOffset;
8635 else if (M == InputsFixed[1] + SourceOffset)
8636 M = (InputsFixed[0] ^ 1) + SourceOffset;
8638 InputsFixed[1] = InputsFixed[0] ^ 1;
8641 // Point everything at the fixed inputs.
8642 for (int &M : HalfMask)
8643 if (M == IncomingInputs[0])
8644 M = InputsFixed[0] + SourceOffset;
8645 else if (M == IncomingInputs[1])
8646 M = InputsFixed[1] + SourceOffset;
8648 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8649 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8652 llvm_unreachable("Unhandled input size!");
8655 // Now hoist the DWord down to the right half.
8656 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8657 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8658 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8659 for (int &M : HalfMask)
8660 for (int Input : IncomingInputs)
8662 M = FreeDWord * 2 + Input % 2;
8664 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8665 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8666 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8667 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8669 // Now enact all the shuffles we've computed to move the inputs into their
8671 if (!isNoopShuffleMask(PSHUFLMask))
8672 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8673 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8674 if (!isNoopShuffleMask(PSHUFHMask))
8675 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8676 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8677 if (!isNoopShuffleMask(PSHUFDMask))
8680 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8681 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8683 // At this point, each half should contain all its inputs, and we can then
8684 // just shuffle them into their final position.
8685 assert(std::count_if(LoMask.begin(), LoMask.end(),
8686 [](int M) { return M >= 4; }) == 0 &&
8687 "Failed to lift all the high half inputs to the low mask!");
8688 assert(std::count_if(HiMask.begin(), HiMask.end(),
8689 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8690 "Failed to lift all the low half inputs to the high mask!");
8692 // Do a half shuffle for the low mask.
8693 if (!isNoopShuffleMask(LoMask))
8694 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8695 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8697 // Do a half shuffle with the high mask after shifting its values down.
8698 for (int &M : HiMask)
8701 if (!isNoopShuffleMask(HiMask))
8702 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8703 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8708 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8709 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8710 SDValue V2, ArrayRef<int> Mask,
8711 SelectionDAG &DAG, bool &V1InUse,
8713 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8719 int Size = Mask.size();
8720 int Scale = 16 / Size;
8721 for (int i = 0; i < 16; ++i) {
8722 if (Mask[i / Scale] == -1) {
8723 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8725 const int ZeroMask = 0x80;
8726 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8728 int V2Idx = Mask[i / Scale] < Size
8730 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8731 if (Zeroable[i / Scale])
8732 V1Idx = V2Idx = ZeroMask;
8733 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8734 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8735 V1InUse |= (ZeroMask != V1Idx);
8736 V2InUse |= (ZeroMask != V2Idx);
8741 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8742 DAG.getBitcast(MVT::v16i8, V1),
8743 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8745 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8746 DAG.getBitcast(MVT::v16i8, V2),
8747 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8749 // If we need shuffled inputs from both, blend the two.
8751 if (V1InUse && V2InUse)
8752 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8754 V = V1InUse ? V1 : V2;
8756 // Cast the result back to the correct type.
8757 return DAG.getBitcast(VT, V);
8760 /// \brief Generic lowering of 8-lane i16 shuffles.
8762 /// This handles both single-input shuffles and combined shuffle/blends with
8763 /// two inputs. The single input shuffles are immediately delegated to
8764 /// a dedicated lowering routine.
8766 /// The blends are lowered in one of three fundamental ways. If there are few
8767 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8768 /// of the input is significantly cheaper when lowered as an interleaving of
8769 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8770 /// halves of the inputs separately (making them have relatively few inputs)
8771 /// and then concatenate them.
8772 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8773 const X86Subtarget *Subtarget,
8774 SelectionDAG &DAG) {
8776 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8777 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8778 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8779 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8780 ArrayRef<int> OrigMask = SVOp->getMask();
8781 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8782 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8783 MutableArrayRef<int> Mask(MaskStorage);
8785 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8787 // Whenever we can lower this as a zext, that instruction is strictly faster
8788 // than any alternative.
8789 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8790 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8793 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8795 auto isV2 = [](int M) { return M >= 8; };
8797 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8799 if (NumV2Inputs == 0) {
8800 // Check for being able to broadcast a single element.
8801 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8802 Mask, Subtarget, DAG))
8805 // Try to use shift instructions.
8807 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8810 // Use dedicated unpack instructions for masks that match their pattern.
8811 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8812 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8813 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8814 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8816 // Try to use byte rotation instructions.
8817 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8818 Mask, Subtarget, DAG))
8821 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8825 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8826 "All single-input shuffles should be canonicalized to be V1-input "
8829 // Try to use shift instructions.
8831 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8834 // See if we can use SSE4A Extraction / Insertion.
8835 if (Subtarget->hasSSE4A())
8836 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
8839 // There are special ways we can lower some single-element blends.
8840 if (NumV2Inputs == 1)
8841 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8842 Mask, Subtarget, DAG))
8845 // We have different paths for blend lowering, but they all must use the
8846 // *exact* same predicate.
8847 bool IsBlendSupported = Subtarget->hasSSE41();
8848 if (IsBlendSupported)
8849 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8853 if (SDValue Masked =
8854 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8857 // Use dedicated unpack instructions for masks that match their pattern.
8858 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8859 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8860 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8861 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8863 // Try to use byte rotation instructions.
8864 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8865 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8868 if (SDValue BitBlend =
8869 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8872 if (SDValue Unpack =
8873 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8876 // If we can't directly blend but can use PSHUFB, that will be better as it
8877 // can both shuffle and set up the inefficient blend.
8878 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8879 bool V1InUse, V2InUse;
8880 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8884 // We can always bit-blend if we have to so the fallback strategy is to
8885 // decompose into single-input permutes and blends.
8886 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8890 /// \brief Check whether a compaction lowering can be done by dropping even
8891 /// elements and compute how many times even elements must be dropped.
8893 /// This handles shuffles which take every Nth element where N is a power of
8894 /// two. Example shuffle masks:
8896 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8897 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8898 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8899 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8900 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8901 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8903 /// Any of these lanes can of course be undef.
8905 /// This routine only supports N <= 3.
8906 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8909 /// \returns N above, or the number of times even elements must be dropped if
8910 /// there is such a number. Otherwise returns zero.
8911 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8912 // Figure out whether we're looping over two inputs or just one.
8913 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8915 // The modulus for the shuffle vector entries is based on whether this is
8916 // a single input or not.
8917 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8918 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8919 "We should only be called with masks with a power-of-2 size!");
8921 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8923 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8924 // and 2^3 simultaneously. This is because we may have ambiguity with
8925 // partially undef inputs.
8926 bool ViableForN[3] = {true, true, true};
8928 for (int i = 0, e = Mask.size(); i < e; ++i) {
8929 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8934 bool IsAnyViable = false;
8935 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8936 if (ViableForN[j]) {
8939 // The shuffle mask must be equal to (i * 2^N) % M.
8940 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8943 ViableForN[j] = false;
8945 // Early exit if we exhaust the possible powers of two.
8950 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8954 // Return 0 as there is no viable power of two.
8958 /// \brief Generic lowering of v16i8 shuffles.
8960 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8961 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8962 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8963 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8965 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8966 const X86Subtarget *Subtarget,
8967 SelectionDAG &DAG) {
8969 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8970 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8971 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8972 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8973 ArrayRef<int> Mask = SVOp->getMask();
8974 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8976 // Try to use shift instructions.
8978 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8981 // Try to use byte rotation instructions.
8982 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8983 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8986 // Try to use a zext lowering.
8987 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8988 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8991 // See if we can use SSE4A Extraction / Insertion.
8992 if (Subtarget->hasSSE4A())
8993 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
8997 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8999 // For single-input shuffles, there are some nicer lowering tricks we can use.
9000 if (NumV2Elements == 0) {
9001 // Check for being able to broadcast a single element.
9002 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9003 Mask, Subtarget, DAG))
9006 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9007 // Notably, this handles splat and partial-splat shuffles more efficiently.
9008 // However, it only makes sense if the pre-duplication shuffle simplifies
9009 // things significantly. Currently, this means we need to be able to
9010 // express the pre-duplication shuffle as an i16 shuffle.
9012 // FIXME: We should check for other patterns which can be widened into an
9013 // i16 shuffle as well.
9014 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9015 for (int i = 0; i < 16; i += 2)
9016 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9021 auto tryToWidenViaDuplication = [&]() -> SDValue {
9022 if (!canWidenViaDuplication(Mask))
9024 SmallVector<int, 4> LoInputs;
9025 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9026 [](int M) { return M >= 0 && M < 8; });
9027 std::sort(LoInputs.begin(), LoInputs.end());
9028 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9030 SmallVector<int, 4> HiInputs;
9031 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9032 [](int M) { return M >= 8; });
9033 std::sort(HiInputs.begin(), HiInputs.end());
9034 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9037 bool TargetLo = LoInputs.size() >= HiInputs.size();
9038 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9039 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9041 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9042 SmallDenseMap<int, int, 8> LaneMap;
9043 for (int I : InPlaceInputs) {
9044 PreDupI16Shuffle[I/2] = I/2;
9047 int j = TargetLo ? 0 : 4, je = j + 4;
9048 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9049 // Check if j is already a shuffle of this input. This happens when
9050 // there are two adjacent bytes after we move the low one.
9051 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9052 // If we haven't yet mapped the input, search for a slot into which
9054 while (j < je && PreDupI16Shuffle[j] != -1)
9058 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9061 // Map this input with the i16 shuffle.
9062 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9065 // Update the lane map based on the mapping we ended up with.
9066 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9068 V1 = DAG.getBitcast(
9070 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9071 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9073 // Unpack the bytes to form the i16s that will be shuffled into place.
9074 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9075 MVT::v16i8, V1, V1);
9077 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9078 for (int i = 0; i < 16; ++i)
9079 if (Mask[i] != -1) {
9080 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9081 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9082 if (PostDupI16Shuffle[i / 2] == -1)
9083 PostDupI16Shuffle[i / 2] = MappedMask;
9085 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9086 "Conflicting entrties in the original shuffle!");
9088 return DAG.getBitcast(
9090 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9091 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9093 if (SDValue V = tryToWidenViaDuplication())
9097 if (SDValue Masked =
9098 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9101 // Use dedicated unpack instructions for masks that match their pattern.
9102 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9103 0, 16, 1, 17, 2, 18, 3, 19,
9105 4, 20, 5, 21, 6, 22, 7, 23}))
9106 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9107 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9108 8, 24, 9, 25, 10, 26, 11, 27,
9110 12, 28, 13, 29, 14, 30, 15, 31}))
9111 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9113 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9114 // with PSHUFB. It is important to do this before we attempt to generate any
9115 // blends but after all of the single-input lowerings. If the single input
9116 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9117 // want to preserve that and we can DAG combine any longer sequences into
9118 // a PSHUFB in the end. But once we start blending from multiple inputs,
9119 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9120 // and there are *very* few patterns that would actually be faster than the
9121 // PSHUFB approach because of its ability to zero lanes.
9123 // FIXME: The only exceptions to the above are blends which are exact
9124 // interleavings with direct instructions supporting them. We currently don't
9125 // handle those well here.
9126 if (Subtarget->hasSSSE3()) {
9127 bool V1InUse = false;
9128 bool V2InUse = false;
9130 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9131 DAG, V1InUse, V2InUse);
9133 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9134 // do so. This avoids using them to handle blends-with-zero which is
9135 // important as a single pshufb is significantly faster for that.
9136 if (V1InUse && V2InUse) {
9137 if (Subtarget->hasSSE41())
9138 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9139 Mask, Subtarget, DAG))
9142 // We can use an unpack to do the blending rather than an or in some
9143 // cases. Even though the or may be (very minorly) more efficient, we
9144 // preference this lowering because there are common cases where part of
9145 // the complexity of the shuffles goes away when we do the final blend as
9147 // FIXME: It might be worth trying to detect if the unpack-feeding
9148 // shuffles will both be pshufb, in which case we shouldn't bother with
9150 if (SDValue Unpack =
9151 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
9158 // There are special ways we can lower some single-element blends.
9159 if (NumV2Elements == 1)
9160 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9161 Mask, Subtarget, DAG))
9164 if (SDValue BitBlend =
9165 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9168 // Check whether a compaction lowering can be done. This handles shuffles
9169 // which take every Nth element for some even N. See the helper function for
9172 // We special case these as they can be particularly efficiently handled with
9173 // the PACKUSB instruction on x86 and they show up in common patterns of
9174 // rearranging bytes to truncate wide elements.
9175 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9176 // NumEvenDrops is the power of two stride of the elements. Another way of
9177 // thinking about it is that we need to drop the even elements this many
9178 // times to get the original input.
9179 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9181 // First we need to zero all the dropped bytes.
9182 assert(NumEvenDrops <= 3 &&
9183 "No support for dropping even elements more than 3 times.");
9184 // We use the mask type to pick which bytes are preserved based on how many
9185 // elements are dropped.
9186 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9187 SDValue ByteClearMask = DAG.getBitcast(
9188 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9189 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9191 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9193 // Now pack things back together.
9194 V1 = DAG.getBitcast(MVT::v8i16, V1);
9195 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9196 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9197 for (int i = 1; i < NumEvenDrops; ++i) {
9198 Result = DAG.getBitcast(MVT::v8i16, Result);
9199 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9205 // Handle multi-input cases by blending single-input shuffles.
9206 if (NumV2Elements > 0)
9207 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9210 // The fallback path for single-input shuffles widens this into two v8i16
9211 // vectors with unpacks, shuffles those, and then pulls them back together
9215 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9216 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9217 for (int i = 0; i < 16; ++i)
9219 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9221 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9223 SDValue VLoHalf, VHiHalf;
9224 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9225 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9227 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9228 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9229 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9230 [](int M) { return M >= 0 && M % 2 == 1; })) {
9231 // Use a mask to drop the high bytes.
9232 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9233 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9234 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9236 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9237 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9239 // Squash the masks to point directly into VLoHalf.
9240 for (int &M : LoBlendMask)
9243 for (int &M : HiBlendMask)
9247 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9248 // VHiHalf so that we can blend them as i16s.
9249 VLoHalf = DAG.getBitcast(
9250 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9251 VHiHalf = DAG.getBitcast(
9252 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9255 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9256 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9258 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9261 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9263 /// This routine breaks down the specific type of 128-bit shuffle and
9264 /// dispatches to the lowering routines accordingly.
9265 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9266 MVT VT, const X86Subtarget *Subtarget,
9267 SelectionDAG &DAG) {
9268 switch (VT.SimpleTy) {
9270 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9272 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9274 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9276 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9278 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9280 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9283 llvm_unreachable("Unimplemented!");
9287 /// \brief Helper function to test whether a shuffle mask could be
9288 /// simplified by widening the elements being shuffled.
9290 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9291 /// leaves it in an unspecified state.
9293 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9294 /// shuffle masks. The latter have the special property of a '-2' representing
9295 /// a zero-ed lane of a vector.
9296 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9297 SmallVectorImpl<int> &WidenedMask) {
9298 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9299 // If both elements are undef, its trivial.
9300 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9301 WidenedMask.push_back(SM_SentinelUndef);
9305 // Check for an undef mask and a mask value properly aligned to fit with
9306 // a pair of values. If we find such a case, use the non-undef mask's value.
9307 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9308 WidenedMask.push_back(Mask[i + 1] / 2);
9311 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9312 WidenedMask.push_back(Mask[i] / 2);
9316 // When zeroing, we need to spread the zeroing across both lanes to widen.
9317 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9318 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9319 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9320 WidenedMask.push_back(SM_SentinelZero);
9326 // Finally check if the two mask values are adjacent and aligned with
9328 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9329 WidenedMask.push_back(Mask[i] / 2);
9333 // Otherwise we can't safely widen the elements used in this shuffle.
9336 assert(WidenedMask.size() == Mask.size() / 2 &&
9337 "Incorrect size of mask after widening the elements!");
9342 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9344 /// This routine just extracts two subvectors, shuffles them independently, and
9345 /// then concatenates them back together. This should work effectively with all
9346 /// AVX vector shuffle types.
9347 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9348 SDValue V2, ArrayRef<int> Mask,
9349 SelectionDAG &DAG) {
9350 assert(VT.getSizeInBits() >= 256 &&
9351 "Only for 256-bit or wider vector shuffles!");
9352 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9353 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9355 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9356 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9358 int NumElements = VT.getVectorNumElements();
9359 int SplitNumElements = NumElements / 2;
9360 MVT ScalarVT = VT.getScalarType();
9361 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9363 // Rather than splitting build-vectors, just build two narrower build
9364 // vectors. This helps shuffling with splats and zeros.
9365 auto SplitVector = [&](SDValue V) {
9366 while (V.getOpcode() == ISD::BITCAST)
9367 V = V->getOperand(0);
9369 MVT OrigVT = V.getSimpleValueType();
9370 int OrigNumElements = OrigVT.getVectorNumElements();
9371 int OrigSplitNumElements = OrigNumElements / 2;
9372 MVT OrigScalarVT = OrigVT.getScalarType();
9373 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9377 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9379 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9380 DAG.getIntPtrConstant(0, DL));
9381 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9382 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9385 SmallVector<SDValue, 16> LoOps, HiOps;
9386 for (int i = 0; i < OrigSplitNumElements; ++i) {
9387 LoOps.push_back(BV->getOperand(i));
9388 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9390 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9391 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9393 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9394 DAG.getBitcast(SplitVT, HiV));
9397 SDValue LoV1, HiV1, LoV2, HiV2;
9398 std::tie(LoV1, HiV1) = SplitVector(V1);
9399 std::tie(LoV2, HiV2) = SplitVector(V2);
9401 // Now create two 4-way blends of these half-width vectors.
9402 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9403 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9404 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9405 for (int i = 0; i < SplitNumElements; ++i) {
9406 int M = HalfMask[i];
9407 if (M >= NumElements) {
9408 if (M >= NumElements + SplitNumElements)
9412 V2BlendMask.push_back(M - NumElements);
9413 V1BlendMask.push_back(-1);
9414 BlendMask.push_back(SplitNumElements + i);
9415 } else if (M >= 0) {
9416 if (M >= SplitNumElements)
9420 V2BlendMask.push_back(-1);
9421 V1BlendMask.push_back(M);
9422 BlendMask.push_back(i);
9424 V2BlendMask.push_back(-1);
9425 V1BlendMask.push_back(-1);
9426 BlendMask.push_back(-1);
9430 // Because the lowering happens after all combining takes place, we need to
9431 // manually combine these blend masks as much as possible so that we create
9432 // a minimal number of high-level vector shuffle nodes.
9434 // First try just blending the halves of V1 or V2.
9435 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9436 return DAG.getUNDEF(SplitVT);
9437 if (!UseLoV2 && !UseHiV2)
9438 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9439 if (!UseLoV1 && !UseHiV1)
9440 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9442 SDValue V1Blend, V2Blend;
9443 if (UseLoV1 && UseHiV1) {
9445 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9447 // We only use half of V1 so map the usage down into the final blend mask.
9448 V1Blend = UseLoV1 ? LoV1 : HiV1;
9449 for (int i = 0; i < SplitNumElements; ++i)
9450 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9451 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9453 if (UseLoV2 && UseHiV2) {
9455 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9457 // We only use half of V2 so map the usage down into the final blend mask.
9458 V2Blend = UseLoV2 ? LoV2 : HiV2;
9459 for (int i = 0; i < SplitNumElements; ++i)
9460 if (BlendMask[i] >= SplitNumElements)
9461 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9463 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9465 SDValue Lo = HalfBlend(LoMask);
9466 SDValue Hi = HalfBlend(HiMask);
9467 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9470 /// \brief Either split a vector in halves or decompose the shuffles and the
9473 /// This is provided as a good fallback for many lowerings of non-single-input
9474 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9475 /// between splitting the shuffle into 128-bit components and stitching those
9476 /// back together vs. extracting the single-input shuffles and blending those
9478 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9479 SDValue V2, ArrayRef<int> Mask,
9480 SelectionDAG &DAG) {
9481 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9482 "lower single-input shuffles as it "
9483 "could then recurse on itself.");
9484 int Size = Mask.size();
9486 // If this can be modeled as a broadcast of two elements followed by a blend,
9487 // prefer that lowering. This is especially important because broadcasts can
9488 // often fold with memory operands.
9489 auto DoBothBroadcast = [&] {
9490 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9493 if (V2BroadcastIdx == -1)
9494 V2BroadcastIdx = M - Size;
9495 else if (M - Size != V2BroadcastIdx)
9497 } else if (M >= 0) {
9498 if (V1BroadcastIdx == -1)
9500 else if (M != V1BroadcastIdx)
9505 if (DoBothBroadcast())
9506 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9509 // If the inputs all stem from a single 128-bit lane of each input, then we
9510 // split them rather than blending because the split will decompose to
9511 // unusually few instructions.
9512 int LaneCount = VT.getSizeInBits() / 128;
9513 int LaneSize = Size / LaneCount;
9514 SmallBitVector LaneInputs[2];
9515 LaneInputs[0].resize(LaneCount, false);
9516 LaneInputs[1].resize(LaneCount, false);
9517 for (int i = 0; i < Size; ++i)
9519 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9520 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9521 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9523 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9524 // that the decomposed single-input shuffles don't end up here.
9525 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9528 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9529 /// a permutation and blend of those lanes.
9531 /// This essentially blends the out-of-lane inputs to each lane into the lane
9532 /// from a permuted copy of the vector. This lowering strategy results in four
9533 /// instructions in the worst case for a single-input cross lane shuffle which
9534 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9535 /// of. Special cases for each particular shuffle pattern should be handled
9536 /// prior to trying this lowering.
9537 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9538 SDValue V1, SDValue V2,
9540 SelectionDAG &DAG) {
9541 // FIXME: This should probably be generalized for 512-bit vectors as well.
9542 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9543 int LaneSize = Mask.size() / 2;
9545 // If there are only inputs from one 128-bit lane, splitting will in fact be
9546 // less expensive. The flags track whether the given lane contains an element
9547 // that crosses to another lane.
9548 bool LaneCrossing[2] = {false, false};
9549 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9550 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9551 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9552 if (!LaneCrossing[0] || !LaneCrossing[1])
9553 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9555 if (isSingleInputShuffleMask(Mask)) {
9556 SmallVector<int, 32> FlippedBlendMask;
9557 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9558 FlippedBlendMask.push_back(
9559 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9561 : Mask[i] % LaneSize +
9562 (i / LaneSize) * LaneSize + Size));
9564 // Flip the vector, and blend the results which should now be in-lane. The
9565 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9566 // 5 for the high source. The value 3 selects the high half of source 2 and
9567 // the value 2 selects the low half of source 2. We only use source 2 to
9568 // allow folding it into a memory operand.
9569 unsigned PERMMask = 3 | 2 << 4;
9570 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9571 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9572 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9575 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9576 // will be handled by the above logic and a blend of the results, much like
9577 // other patterns in AVX.
9578 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9581 /// \brief Handle lowering 2-lane 128-bit shuffles.
9582 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9583 SDValue V2, ArrayRef<int> Mask,
9584 const X86Subtarget *Subtarget,
9585 SelectionDAG &DAG) {
9586 // TODO: If minimizing size and one of the inputs is a zero vector and the
9587 // the zero vector has only one use, we could use a VPERM2X128 to save the
9588 // instruction bytes needed to explicitly generate the zero vector.
9590 // Blends are faster and handle all the non-lane-crossing cases.
9591 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9595 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9596 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9598 // If either input operand is a zero vector, use VPERM2X128 because its mask
9599 // allows us to replace the zero input with an implicit zero.
9600 if (!IsV1Zero && !IsV2Zero) {
9601 // Check for patterns which can be matched with a single insert of a 128-bit
9603 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9604 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9605 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9606 VT.getVectorNumElements() / 2);
9607 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9608 DAG.getIntPtrConstant(0, DL));
9609 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9610 OnlyUsesV1 ? V1 : V2,
9611 DAG.getIntPtrConstant(0, DL));
9612 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9616 // Otherwise form a 128-bit permutation. After accounting for undefs,
9617 // convert the 64-bit shuffle mask selection values into 128-bit
9618 // selection bits by dividing the indexes by 2 and shifting into positions
9619 // defined by a vperm2*128 instruction's immediate control byte.
9621 // The immediate permute control byte looks like this:
9622 // [1:0] - select 128 bits from sources for low half of destination
9624 // [3] - zero low half of destination
9625 // [5:4] - select 128 bits from sources for high half of destination
9627 // [7] - zero high half of destination
9629 int MaskLO = Mask[0];
9630 if (MaskLO == SM_SentinelUndef)
9631 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9633 int MaskHI = Mask[2];
9634 if (MaskHI == SM_SentinelUndef)
9635 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9637 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9639 // If either input is a zero vector, replace it with an undef input.
9640 // Shuffle mask values < 4 are selecting elements of V1.
9641 // Shuffle mask values >= 4 are selecting elements of V2.
9642 // Adjust each half of the permute mask by clearing the half that was
9643 // selecting the zero vector and setting the zero mask bit.
9645 V1 = DAG.getUNDEF(VT);
9647 PermMask = (PermMask & 0xf0) | 0x08;
9649 PermMask = (PermMask & 0x0f) | 0x80;
9652 V2 = DAG.getUNDEF(VT);
9654 PermMask = (PermMask & 0xf0) | 0x08;
9656 PermMask = (PermMask & 0x0f) | 0x80;
9659 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9660 DAG.getConstant(PermMask, DL, MVT::i8));
9663 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9664 /// shuffling each lane.
9666 /// This will only succeed when the result of fixing the 128-bit lanes results
9667 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9668 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9669 /// the lane crosses early and then use simpler shuffles within each lane.
9671 /// FIXME: It might be worthwhile at some point to support this without
9672 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9673 /// in x86 only floating point has interesting non-repeating shuffles, and even
9674 /// those are still *marginally* more expensive.
9675 static SDValue lowerVectorShuffleByMerging128BitLanes(
9676 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9677 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9678 assert(!isSingleInputShuffleMask(Mask) &&
9679 "This is only useful with multiple inputs.");
9681 int Size = Mask.size();
9682 int LaneSize = 128 / VT.getScalarSizeInBits();
9683 int NumLanes = Size / LaneSize;
9684 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9686 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9687 // check whether the in-128-bit lane shuffles share a repeating pattern.
9688 SmallVector<int, 4> Lanes;
9689 Lanes.resize(NumLanes, -1);
9690 SmallVector<int, 4> InLaneMask;
9691 InLaneMask.resize(LaneSize, -1);
9692 for (int i = 0; i < Size; ++i) {
9696 int j = i / LaneSize;
9699 // First entry we've seen for this lane.
9700 Lanes[j] = Mask[i] / LaneSize;
9701 } else if (Lanes[j] != Mask[i] / LaneSize) {
9702 // This doesn't match the lane selected previously!
9706 // Check that within each lane we have a consistent shuffle mask.
9707 int k = i % LaneSize;
9708 if (InLaneMask[k] < 0) {
9709 InLaneMask[k] = Mask[i] % LaneSize;
9710 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9711 // This doesn't fit a repeating in-lane mask.
9716 // First shuffle the lanes into place.
9717 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9718 VT.getSizeInBits() / 64);
9719 SmallVector<int, 8> LaneMask;
9720 LaneMask.resize(NumLanes * 2, -1);
9721 for (int i = 0; i < NumLanes; ++i)
9722 if (Lanes[i] >= 0) {
9723 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9724 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9727 V1 = DAG.getBitcast(LaneVT, V1);
9728 V2 = DAG.getBitcast(LaneVT, V2);
9729 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9731 // Cast it back to the type we actually want.
9732 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9734 // Now do a simple shuffle that isn't lane crossing.
9735 SmallVector<int, 8> NewMask;
9736 NewMask.resize(Size, -1);
9737 for (int i = 0; i < Size; ++i)
9739 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9740 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9741 "Must not introduce lane crosses at this point!");
9743 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9746 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9749 /// This returns true if the elements from a particular input are already in the
9750 /// slot required by the given mask and require no permutation.
9751 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9752 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9753 int Size = Mask.size();
9754 for (int i = 0; i < Size; ++i)
9755 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9761 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9762 ArrayRef<int> Mask, SDValue V1,
9763 SDValue V2, SelectionDAG &DAG) {
9765 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9766 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9767 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9768 int NumElts = VT.getVectorNumElements();
9769 bool ShufpdMask = true;
9770 bool CommutableMask = true;
9771 unsigned Immediate = 0;
9772 for (int i = 0; i < NumElts; ++i) {
9775 int Val = (i & 6) + NumElts * (i & 1);
9776 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9777 if (Mask[i] < Val || Mask[i] > Val + 1)
9779 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9780 CommutableMask = false;
9781 Immediate |= (Mask[i] % 2) << i;
9784 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9785 DAG.getConstant(Immediate, DL, MVT::i8));
9787 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9788 DAG.getConstant(Immediate, DL, MVT::i8));
9792 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9794 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9795 /// isn't available.
9796 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9797 const X86Subtarget *Subtarget,
9798 SelectionDAG &DAG) {
9800 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9801 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9802 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9803 ArrayRef<int> Mask = SVOp->getMask();
9804 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9806 SmallVector<int, 4> WidenedMask;
9807 if (canWidenShuffleElements(Mask, WidenedMask))
9808 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9811 if (isSingleInputShuffleMask(Mask)) {
9812 // Check for being able to broadcast a single element.
9813 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9814 Mask, Subtarget, DAG))
9817 // Use low duplicate instructions for masks that match their pattern.
9818 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9819 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9821 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9822 // Non-half-crossing single input shuffles can be lowerid with an
9823 // interleaved permutation.
9824 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9825 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9826 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9827 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9830 // With AVX2 we have direct support for this permutation.
9831 if (Subtarget->hasAVX2())
9832 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9833 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9835 // Otherwise, fall back.
9836 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9840 // X86 has dedicated unpack instructions that can handle specific blend
9841 // operations: UNPCKH and UNPCKL.
9842 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9843 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9844 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9845 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9846 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9847 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9848 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9849 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9851 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9855 // Check if the blend happens to exactly fit that of SHUFPD.
9857 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
9860 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9861 // shuffle. However, if we have AVX2 and either inputs are already in place,
9862 // we will be able to shuffle even across lanes the other input in a single
9863 // instruction so skip this pattern.
9864 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9865 isShuffleMaskInputInPlace(1, Mask))))
9866 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9867 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9870 // If we have AVX2 then we always want to lower with a blend because an v4 we
9871 // can fully permute the elements.
9872 if (Subtarget->hasAVX2())
9873 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9876 // Otherwise fall back on generic lowering.
9877 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9880 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9882 /// This routine is only called when we have AVX2 and thus a reasonable
9883 /// instruction set for v4i64 shuffling..
9884 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9885 const X86Subtarget *Subtarget,
9886 SelectionDAG &DAG) {
9888 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9889 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9890 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9891 ArrayRef<int> Mask = SVOp->getMask();
9892 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9893 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9895 SmallVector<int, 4> WidenedMask;
9896 if (canWidenShuffleElements(Mask, WidenedMask))
9897 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9900 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9904 // Check for being able to broadcast a single element.
9905 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9906 Mask, Subtarget, DAG))
9909 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9910 // use lower latency instructions that will operate on both 128-bit lanes.
9911 SmallVector<int, 2> RepeatedMask;
9912 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9913 if (isSingleInputShuffleMask(Mask)) {
9914 int PSHUFDMask[] = {-1, -1, -1, -1};
9915 for (int i = 0; i < 2; ++i)
9916 if (RepeatedMask[i] >= 0) {
9917 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9918 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9920 return DAG.getBitcast(
9922 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9923 DAG.getBitcast(MVT::v8i32, V1),
9924 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9928 // AVX2 provides a direct instruction for permuting a single input across
9930 if (isSingleInputShuffleMask(Mask))
9931 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9932 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9934 // Try to use shift instructions.
9936 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9939 // Use dedicated unpack instructions for masks that match their pattern.
9940 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9941 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9942 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9943 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9944 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9945 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9946 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9947 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9949 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9950 // shuffle. However, if we have AVX2 and either inputs are already in place,
9951 // we will be able to shuffle even across lanes the other input in a single
9952 // instruction so skip this pattern.
9953 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9954 isShuffleMaskInputInPlace(1, Mask))))
9955 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9956 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9959 // Otherwise fall back on generic blend lowering.
9960 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9964 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9966 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9967 /// isn't available.
9968 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9969 const X86Subtarget *Subtarget,
9970 SelectionDAG &DAG) {
9972 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9973 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9974 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9975 ArrayRef<int> Mask = SVOp->getMask();
9976 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9978 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9982 // Check for being able to broadcast a single element.
9983 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9984 Mask, Subtarget, DAG))
9987 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9988 // options to efficiently lower the shuffle.
9989 SmallVector<int, 4> RepeatedMask;
9990 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9991 assert(RepeatedMask.size() == 4 &&
9992 "Repeated masks must be half the mask width!");
9994 // Use even/odd duplicate instructions for masks that match their pattern.
9995 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9996 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9997 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9998 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10000 if (isSingleInputShuffleMask(Mask))
10001 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10002 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10004 // Use dedicated unpack instructions for masks that match their pattern.
10005 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10006 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10007 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10008 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10009 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10010 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10011 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10012 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10014 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10015 // have already handled any direct blends. We also need to squash the
10016 // repeated mask into a simulated v4f32 mask.
10017 for (int i = 0; i < 4; ++i)
10018 if (RepeatedMask[i] >= 8)
10019 RepeatedMask[i] -= 4;
10020 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10023 // If we have a single input shuffle with different shuffle patterns in the
10024 // two 128-bit lanes use the variable mask to VPERMILPS.
10025 if (isSingleInputShuffleMask(Mask)) {
10026 SDValue VPermMask[8];
10027 for (int i = 0; i < 8; ++i)
10028 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10029 : DAG.getConstant(Mask[i], DL, MVT::i32);
10030 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10031 return DAG.getNode(
10032 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10033 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10035 if (Subtarget->hasAVX2())
10036 return DAG.getNode(
10037 X86ISD::VPERMV, DL, MVT::v8f32,
10038 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10039 MVT::v8i32, VPermMask)),
10042 // Otherwise, fall back.
10043 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10047 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10049 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10050 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10053 // If we have AVX2 then we always want to lower with a blend because at v8 we
10054 // can fully permute the elements.
10055 if (Subtarget->hasAVX2())
10056 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10059 // Otherwise fall back on generic lowering.
10060 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10063 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10065 /// This routine is only called when we have AVX2 and thus a reasonable
10066 /// instruction set for v8i32 shuffling..
10067 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10068 const X86Subtarget *Subtarget,
10069 SelectionDAG &DAG) {
10071 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10072 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10073 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10074 ArrayRef<int> Mask = SVOp->getMask();
10075 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10076 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10078 // Whenever we can lower this as a zext, that instruction is strictly faster
10079 // than any alternative. It also allows us to fold memory operands into the
10080 // shuffle in many cases.
10081 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10082 Mask, Subtarget, DAG))
10085 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10089 // Check for being able to broadcast a single element.
10090 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10091 Mask, Subtarget, DAG))
10094 // If the shuffle mask is repeated in each 128-bit lane we can use more
10095 // efficient instructions that mirror the shuffles across the two 128-bit
10097 SmallVector<int, 4> RepeatedMask;
10098 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10099 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10100 if (isSingleInputShuffleMask(Mask))
10101 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10102 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10104 // Use dedicated unpack instructions for masks that match their pattern.
10105 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10106 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10107 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10108 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10109 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10110 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10111 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10112 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10115 // Try to use shift instructions.
10116 if (SDValue Shift =
10117 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10120 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10121 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10124 // If the shuffle patterns aren't repeated but it is a single input, directly
10125 // generate a cross-lane VPERMD instruction.
10126 if (isSingleInputShuffleMask(Mask)) {
10127 SDValue VPermMask[8];
10128 for (int i = 0; i < 8; ++i)
10129 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10130 : DAG.getConstant(Mask[i], DL, MVT::i32);
10131 return DAG.getNode(
10132 X86ISD::VPERMV, DL, MVT::v8i32,
10133 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10136 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10138 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10139 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10142 // Otherwise fall back on generic blend lowering.
10143 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10147 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10149 /// This routine is only called when we have AVX2 and thus a reasonable
10150 /// instruction set for v16i16 shuffling..
10151 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10152 const X86Subtarget *Subtarget,
10153 SelectionDAG &DAG) {
10155 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10156 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10157 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10158 ArrayRef<int> Mask = SVOp->getMask();
10159 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10160 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10162 // Whenever we can lower this as a zext, that instruction is strictly faster
10163 // than any alternative. It also allows us to fold memory operands into the
10164 // shuffle in many cases.
10165 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10166 Mask, Subtarget, DAG))
10169 // Check for being able to broadcast a single element.
10170 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10171 Mask, Subtarget, DAG))
10174 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10178 // Use dedicated unpack instructions for masks that match their pattern.
10179 if (isShuffleEquivalent(V1, V2, Mask,
10180 {// First 128-bit lane:
10181 0, 16, 1, 17, 2, 18, 3, 19,
10182 // Second 128-bit lane:
10183 8, 24, 9, 25, 10, 26, 11, 27}))
10184 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10185 if (isShuffleEquivalent(V1, V2, Mask,
10186 {// First 128-bit lane:
10187 4, 20, 5, 21, 6, 22, 7, 23,
10188 // Second 128-bit lane:
10189 12, 28, 13, 29, 14, 30, 15, 31}))
10190 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10192 // Try to use shift instructions.
10193 if (SDValue Shift =
10194 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10197 // Try to use byte rotation instructions.
10198 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10199 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10202 if (isSingleInputShuffleMask(Mask)) {
10203 // There are no generalized cross-lane shuffle operations available on i16
10205 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10206 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10209 SmallVector<int, 8> RepeatedMask;
10210 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10211 // As this is a single-input shuffle, the repeated mask should be
10212 // a strictly valid v8i16 mask that we can pass through to the v8i16
10213 // lowering to handle even the v16 case.
10214 return lowerV8I16GeneralSingleInputVectorShuffle(
10215 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10218 SDValue PSHUFBMask[32];
10219 for (int i = 0; i < 16; ++i) {
10220 if (Mask[i] == -1) {
10221 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10225 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10226 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10227 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10228 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10230 return DAG.getBitcast(MVT::v16i16,
10231 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10232 DAG.getBitcast(MVT::v32i8, V1),
10233 DAG.getNode(ISD::BUILD_VECTOR, DL,
10234 MVT::v32i8, PSHUFBMask)));
10237 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10239 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10240 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10243 // Otherwise fall back on generic lowering.
10244 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10247 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10249 /// This routine is only called when we have AVX2 and thus a reasonable
10250 /// instruction set for v32i8 shuffling..
10251 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10252 const X86Subtarget *Subtarget,
10253 SelectionDAG &DAG) {
10255 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10256 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10257 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10258 ArrayRef<int> Mask = SVOp->getMask();
10259 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10260 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10262 // Whenever we can lower this as a zext, that instruction is strictly faster
10263 // than any alternative. It also allows us to fold memory operands into the
10264 // shuffle in many cases.
10265 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10266 Mask, Subtarget, DAG))
10269 // Check for being able to broadcast a single element.
10270 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10271 Mask, Subtarget, DAG))
10274 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10278 // Use dedicated unpack instructions for masks that match their pattern.
10279 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10281 if (isShuffleEquivalent(
10283 {// First 128-bit lane:
10284 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10285 // Second 128-bit lane:
10286 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10287 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10288 if (isShuffleEquivalent(
10290 {// First 128-bit lane:
10291 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10292 // Second 128-bit lane:
10293 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10294 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10296 // Try to use shift instructions.
10297 if (SDValue Shift =
10298 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10301 // Try to use byte rotation instructions.
10302 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10303 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10306 if (isSingleInputShuffleMask(Mask)) {
10307 // There are no generalized cross-lane shuffle operations available on i8
10309 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10310 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10313 SDValue PSHUFBMask[32];
10314 for (int i = 0; i < 32; ++i)
10317 ? DAG.getUNDEF(MVT::i8)
10318 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10321 return DAG.getNode(
10322 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10323 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10326 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10328 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10329 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10332 // Otherwise fall back on generic lowering.
10333 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10336 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10338 /// This routine either breaks down the specific type of a 256-bit x86 vector
10339 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10340 /// together based on the available instructions.
10341 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10342 MVT VT, const X86Subtarget *Subtarget,
10343 SelectionDAG &DAG) {
10345 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10346 ArrayRef<int> Mask = SVOp->getMask();
10348 // If we have a single input to the zero element, insert that into V1 if we
10349 // can do so cheaply.
10350 int NumElts = VT.getVectorNumElements();
10351 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10352 return M >= NumElts;
10355 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10356 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10357 DL, VT, V1, V2, Mask, Subtarget, DAG))
10360 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10361 // check for those subtargets here and avoid much of the subtarget querying in
10362 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10363 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10364 // floating point types there eventually, just immediately cast everything to
10365 // a float and operate entirely in that domain.
10366 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10367 int ElementBits = VT.getScalarSizeInBits();
10368 if (ElementBits < 32)
10369 // No floating point type available, decompose into 128-bit vectors.
10370 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10372 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10373 VT.getVectorNumElements());
10374 V1 = DAG.getBitcast(FpVT, V1);
10375 V2 = DAG.getBitcast(FpVT, V2);
10376 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10379 switch (VT.SimpleTy) {
10381 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10383 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10385 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10387 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10389 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10391 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10394 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10398 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10399 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10400 const X86Subtarget *Subtarget,
10401 SelectionDAG &DAG) {
10403 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10404 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10405 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10406 ArrayRef<int> Mask = SVOp->getMask();
10407 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10409 // X86 has dedicated unpack instructions that can handle specific blend
10410 // operations: UNPCKH and UNPCKL.
10411 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10412 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10413 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10414 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10416 // FIXME: Implement direct support for this type!
10417 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10420 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10421 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10422 const X86Subtarget *Subtarget,
10423 SelectionDAG &DAG) {
10425 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10426 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10427 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10428 ArrayRef<int> Mask = SVOp->getMask();
10429 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10431 // Use dedicated unpack instructions for masks that match their pattern.
10432 if (isShuffleEquivalent(V1, V2, Mask,
10433 {// First 128-bit lane.
10434 0, 16, 1, 17, 4, 20, 5, 21,
10435 // Second 128-bit lane.
10436 8, 24, 9, 25, 12, 28, 13, 29}))
10437 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10438 if (isShuffleEquivalent(V1, V2, Mask,
10439 {// First 128-bit lane.
10440 2, 18, 3, 19, 6, 22, 7, 23,
10441 // Second 128-bit lane.
10442 10, 26, 11, 27, 14, 30, 15, 31}))
10443 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10445 // FIXME: Implement direct support for this type!
10446 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10449 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10450 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10451 const X86Subtarget *Subtarget,
10452 SelectionDAG &DAG) {
10454 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10455 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10456 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10457 ArrayRef<int> Mask = SVOp->getMask();
10458 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10460 // X86 has dedicated unpack instructions that can handle specific blend
10461 // operations: UNPCKH and UNPCKL.
10462 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10463 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10464 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10465 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10467 // FIXME: Implement direct support for this type!
10468 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10471 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10472 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10473 const X86Subtarget *Subtarget,
10474 SelectionDAG &DAG) {
10476 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10477 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10478 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10479 ArrayRef<int> Mask = SVOp->getMask();
10480 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10482 // Use dedicated unpack instructions for masks that match their pattern.
10483 if (isShuffleEquivalent(V1, V2, Mask,
10484 {// First 128-bit lane.
10485 0, 16, 1, 17, 4, 20, 5, 21,
10486 // Second 128-bit lane.
10487 8, 24, 9, 25, 12, 28, 13, 29}))
10488 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10489 if (isShuffleEquivalent(V1, V2, Mask,
10490 {// First 128-bit lane.
10491 2, 18, 3, 19, 6, 22, 7, 23,
10492 // Second 128-bit lane.
10493 10, 26, 11, 27, 14, 30, 15, 31}))
10494 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10496 // FIXME: Implement direct support for this type!
10497 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10500 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10501 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10502 const X86Subtarget *Subtarget,
10503 SelectionDAG &DAG) {
10505 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10506 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10507 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10508 ArrayRef<int> Mask = SVOp->getMask();
10509 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10510 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10512 // FIXME: Implement direct support for this type!
10513 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10516 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10517 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10518 const X86Subtarget *Subtarget,
10519 SelectionDAG &DAG) {
10521 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10522 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10523 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10524 ArrayRef<int> Mask = SVOp->getMask();
10525 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10526 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10528 // FIXME: Implement direct support for this type!
10529 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10532 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10534 /// This routine either breaks down the specific type of a 512-bit x86 vector
10535 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10536 /// together based on the available instructions.
10537 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10538 MVT VT, const X86Subtarget *Subtarget,
10539 SelectionDAG &DAG) {
10541 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10542 ArrayRef<int> Mask = SVOp->getMask();
10543 assert(Subtarget->hasAVX512() &&
10544 "Cannot lower 512-bit vectors w/ basic ISA!");
10546 // Check for being able to broadcast a single element.
10547 if (SDValue Broadcast =
10548 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10551 // Dispatch to each element type for lowering. If we don't have supprot for
10552 // specific element type shuffles at 512 bits, immediately split them and
10553 // lower them. Each lowering routine of a given type is allowed to assume that
10554 // the requisite ISA extensions for that element type are available.
10555 switch (VT.SimpleTy) {
10557 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10559 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10561 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10563 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10565 if (Subtarget->hasBWI())
10566 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10569 if (Subtarget->hasBWI())
10570 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10574 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10577 // Otherwise fall back on splitting.
10578 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10581 /// \brief Top-level lowering for x86 vector shuffles.
10583 /// This handles decomposition, canonicalization, and lowering of all x86
10584 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10585 /// above in helper routines. The canonicalization attempts to widen shuffles
10586 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10587 /// s.t. only one of the two inputs needs to be tested, etc.
10588 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10589 SelectionDAG &DAG) {
10590 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10591 ArrayRef<int> Mask = SVOp->getMask();
10592 SDValue V1 = Op.getOperand(0);
10593 SDValue V2 = Op.getOperand(1);
10594 MVT VT = Op.getSimpleValueType();
10595 int NumElements = VT.getVectorNumElements();
10598 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10600 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10601 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10602 if (V1IsUndef && V2IsUndef)
10603 return DAG.getUNDEF(VT);
10605 // When we create a shuffle node we put the UNDEF node to second operand,
10606 // but in some cases the first operand may be transformed to UNDEF.
10607 // In this case we should just commute the node.
10609 return DAG.getCommutedVectorShuffle(*SVOp);
10611 // Check for non-undef masks pointing at an undef vector and make the masks
10612 // undef as well. This makes it easier to match the shuffle based solely on
10616 if (M >= NumElements) {
10617 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10618 for (int &M : NewMask)
10619 if (M >= NumElements)
10621 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10624 // We actually see shuffles that are entirely re-arrangements of a set of
10625 // zero inputs. This mostly happens while decomposing complex shuffles into
10626 // simple ones. Directly lower these as a buildvector of zeros.
10627 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10628 if (Zeroable.all())
10629 return getZeroVector(VT, Subtarget, DAG, dl);
10631 // Try to collapse shuffles into using a vector type with fewer elements but
10632 // wider element types. We cap this to not form integers or floating point
10633 // elements wider than 64 bits, but it might be interesting to form i128
10634 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10635 SmallVector<int, 16> WidenedMask;
10636 if (VT.getScalarSizeInBits() < 64 &&
10637 canWidenShuffleElements(Mask, WidenedMask)) {
10638 MVT NewEltVT = VT.isFloatingPoint()
10639 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10640 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10641 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10642 // Make sure that the new vector type is legal. For example, v2f64 isn't
10644 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10645 V1 = DAG.getBitcast(NewVT, V1);
10646 V2 = DAG.getBitcast(NewVT, V2);
10647 return DAG.getBitcast(
10648 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10652 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10653 for (int M : SVOp->getMask())
10655 ++NumUndefElements;
10656 else if (M < NumElements)
10661 // Commute the shuffle as needed such that more elements come from V1 than
10662 // V2. This allows us to match the shuffle pattern strictly on how many
10663 // elements come from V1 without handling the symmetric cases.
10664 if (NumV2Elements > NumV1Elements)
10665 return DAG.getCommutedVectorShuffle(*SVOp);
10667 // When the number of V1 and V2 elements are the same, try to minimize the
10668 // number of uses of V2 in the low half of the vector. When that is tied,
10669 // ensure that the sum of indices for V1 is equal to or lower than the sum
10670 // indices for V2. When those are equal, try to ensure that the number of odd
10671 // indices for V1 is lower than the number of odd indices for V2.
10672 if (NumV1Elements == NumV2Elements) {
10673 int LowV1Elements = 0, LowV2Elements = 0;
10674 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10675 if (M >= NumElements)
10679 if (LowV2Elements > LowV1Elements) {
10680 return DAG.getCommutedVectorShuffle(*SVOp);
10681 } else if (LowV2Elements == LowV1Elements) {
10682 int SumV1Indices = 0, SumV2Indices = 0;
10683 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10684 if (SVOp->getMask()[i] >= NumElements)
10686 else if (SVOp->getMask()[i] >= 0)
10688 if (SumV2Indices < SumV1Indices) {
10689 return DAG.getCommutedVectorShuffle(*SVOp);
10690 } else if (SumV2Indices == SumV1Indices) {
10691 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10692 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10693 if (SVOp->getMask()[i] >= NumElements)
10694 NumV2OddIndices += i % 2;
10695 else if (SVOp->getMask()[i] >= 0)
10696 NumV1OddIndices += i % 2;
10697 if (NumV2OddIndices < NumV1OddIndices)
10698 return DAG.getCommutedVectorShuffle(*SVOp);
10703 // For each vector width, delegate to a specialized lowering routine.
10704 if (VT.getSizeInBits() == 128)
10705 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10707 if (VT.getSizeInBits() == 256)
10708 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10710 // Force AVX-512 vectors to be scalarized for now.
10711 // FIXME: Implement AVX-512 support!
10712 if (VT.getSizeInBits() == 512)
10713 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10715 llvm_unreachable("Unimplemented!");
10718 // This function assumes its argument is a BUILD_VECTOR of constants or
10719 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10721 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10722 unsigned &MaskValue) {
10724 unsigned NumElems = BuildVector->getNumOperands();
10725 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10726 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10727 unsigned NumElemsInLane = NumElems / NumLanes;
10729 // Blend for v16i16 should be symetric for the both lanes.
10730 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10731 SDValue EltCond = BuildVector->getOperand(i);
10732 SDValue SndLaneEltCond =
10733 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10735 int Lane1Cond = -1, Lane2Cond = -1;
10736 if (isa<ConstantSDNode>(EltCond))
10737 Lane1Cond = !isZero(EltCond);
10738 if (isa<ConstantSDNode>(SndLaneEltCond))
10739 Lane2Cond = !isZero(SndLaneEltCond);
10741 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10742 // Lane1Cond != 0, means we want the first argument.
10743 // Lane1Cond == 0, means we want the second argument.
10744 // The encoding of this argument is 0 for the first argument, 1
10745 // for the second. Therefore, invert the condition.
10746 MaskValue |= !Lane1Cond << i;
10747 else if (Lane1Cond < 0)
10748 MaskValue |= !Lane2Cond << i;
10755 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10756 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10757 const X86Subtarget *Subtarget,
10758 SelectionDAG &DAG) {
10759 SDValue Cond = Op.getOperand(0);
10760 SDValue LHS = Op.getOperand(1);
10761 SDValue RHS = Op.getOperand(2);
10763 MVT VT = Op.getSimpleValueType();
10765 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10767 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10769 // Only non-legal VSELECTs reach this lowering, convert those into generic
10770 // shuffles and re-use the shuffle lowering path for blends.
10771 SmallVector<int, 32> Mask;
10772 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10773 SDValue CondElt = CondBV->getOperand(i);
10775 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10777 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10780 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10781 // A vselect where all conditions and data are constants can be optimized into
10782 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10783 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10784 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10785 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10788 // Try to lower this to a blend-style vector shuffle. This can handle all
10789 // constant condition cases.
10790 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10793 // Variable blends are only legal from SSE4.1 onward.
10794 if (!Subtarget->hasSSE41())
10797 // Only some types will be legal on some subtargets. If we can emit a legal
10798 // VSELECT-matching blend, return Op, and but if we need to expand, return
10800 switch (Op.getSimpleValueType().SimpleTy) {
10802 // Most of the vector types have blends past SSE4.1.
10806 // The byte blends for AVX vectors were introduced only in AVX2.
10807 if (Subtarget->hasAVX2())
10814 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10815 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10818 // FIXME: We should custom lower this by fixing the condition and using i8
10824 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10825 MVT VT = Op.getSimpleValueType();
10828 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10831 if (VT.getSizeInBits() == 8) {
10832 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10833 Op.getOperand(0), Op.getOperand(1));
10834 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10835 DAG.getValueType(VT));
10836 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10839 if (VT.getSizeInBits() == 16) {
10840 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10841 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10843 return DAG.getNode(
10844 ISD::TRUNCATE, dl, MVT::i16,
10845 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10846 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10847 Op.getOperand(1)));
10848 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10849 Op.getOperand(0), Op.getOperand(1));
10850 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10851 DAG.getValueType(VT));
10852 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10855 if (VT == MVT::f32) {
10856 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10857 // the result back to FR32 register. It's only worth matching if the
10858 // result has a single use which is a store or a bitcast to i32. And in
10859 // the case of a store, it's not worth it if the index is a constant 0,
10860 // because a MOVSSmr can be used instead, which is smaller and faster.
10861 if (!Op.hasOneUse())
10863 SDNode *User = *Op.getNode()->use_begin();
10864 if ((User->getOpcode() != ISD::STORE ||
10865 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10866 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10867 (User->getOpcode() != ISD::BITCAST ||
10868 User->getValueType(0) != MVT::i32))
10870 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10871 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10873 return DAG.getBitcast(MVT::f32, Extract);
10876 if (VT == MVT::i32 || VT == MVT::i64) {
10877 // ExtractPS/pextrq works with constant index.
10878 if (isa<ConstantSDNode>(Op.getOperand(1)))
10884 /// Extract one bit from mask vector, like v16i1 or v8i1.
10885 /// AVX-512 feature.
10887 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10888 SDValue Vec = Op.getOperand(0);
10890 MVT VecVT = Vec.getSimpleValueType();
10891 SDValue Idx = Op.getOperand(1);
10892 MVT EltVT = Op.getSimpleValueType();
10894 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10895 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10896 "Unexpected vector type in ExtractBitFromMaskVector");
10898 // variable index can't be handled in mask registers,
10899 // extend vector to VR512
10900 if (!isa<ConstantSDNode>(Idx)) {
10901 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10902 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10903 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10904 ExtVT.getVectorElementType(), Ext, Idx);
10905 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10908 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10909 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10910 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10911 rc = getRegClassFor(MVT::v16i1);
10912 unsigned MaxSift = rc->getSize()*8 - 1;
10913 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10914 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10915 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10916 DAG.getConstant(MaxSift, dl, MVT::i8));
10917 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10918 DAG.getIntPtrConstant(0, dl));
10922 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10923 SelectionDAG &DAG) const {
10925 SDValue Vec = Op.getOperand(0);
10926 MVT VecVT = Vec.getSimpleValueType();
10927 SDValue Idx = Op.getOperand(1);
10929 if (Op.getSimpleValueType() == MVT::i1)
10930 return ExtractBitFromMaskVector(Op, DAG);
10932 if (!isa<ConstantSDNode>(Idx)) {
10933 if (VecVT.is512BitVector() ||
10934 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10935 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10938 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10939 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10940 MaskEltVT.getSizeInBits());
10942 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10943 auto PtrVT = getPointerTy(DAG.getDataLayout());
10944 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10945 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
10946 DAG.getConstant(0, dl, PtrVT));
10947 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10948 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
10949 DAG.getConstant(0, dl, PtrVT));
10954 // If this is a 256-bit vector result, first extract the 128-bit vector and
10955 // then extract the element from the 128-bit vector.
10956 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10958 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10959 // Get the 128-bit vector.
10960 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10961 MVT EltVT = VecVT.getVectorElementType();
10963 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10965 //if (IdxVal >= NumElems/2)
10966 // IdxVal -= NumElems/2;
10967 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10968 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10969 DAG.getConstant(IdxVal, dl, MVT::i32));
10972 assert(VecVT.is128BitVector() && "Unexpected vector length");
10974 if (Subtarget->hasSSE41())
10975 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
10978 MVT VT = Op.getSimpleValueType();
10979 // TODO: handle v16i8.
10980 if (VT.getSizeInBits() == 16) {
10981 SDValue Vec = Op.getOperand(0);
10982 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10984 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10985 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10986 DAG.getBitcast(MVT::v4i32, Vec),
10987 Op.getOperand(1)));
10988 // Transform it so it match pextrw which produces a 32-bit result.
10989 MVT EltVT = MVT::i32;
10990 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10991 Op.getOperand(0), Op.getOperand(1));
10992 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10993 DAG.getValueType(VT));
10994 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10997 if (VT.getSizeInBits() == 32) {
10998 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11002 // SHUFPS the element to the lowest double word, then movss.
11003 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11004 MVT VVT = Op.getOperand(0).getSimpleValueType();
11005 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11006 DAG.getUNDEF(VVT), Mask);
11007 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11008 DAG.getIntPtrConstant(0, dl));
11011 if (VT.getSizeInBits() == 64) {
11012 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11013 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11014 // to match extract_elt for f64.
11015 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11019 // UNPCKHPD the element to the lowest double word, then movsd.
11020 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11021 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11022 int Mask[2] = { 1, -1 };
11023 MVT VVT = Op.getOperand(0).getSimpleValueType();
11024 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11025 DAG.getUNDEF(VVT), Mask);
11026 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11027 DAG.getIntPtrConstant(0, dl));
11033 /// Insert one bit to mask vector, like v16i1 or v8i1.
11034 /// AVX-512 feature.
11036 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11038 SDValue Vec = Op.getOperand(0);
11039 SDValue Elt = Op.getOperand(1);
11040 SDValue Idx = Op.getOperand(2);
11041 MVT VecVT = Vec.getSimpleValueType();
11043 if (!isa<ConstantSDNode>(Idx)) {
11044 // Non constant index. Extend source and destination,
11045 // insert element and then truncate the result.
11046 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11047 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11048 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11049 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11050 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11051 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11054 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11055 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11057 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11058 DAG.getConstant(IdxVal, dl, MVT::i8));
11059 if (Vec.getOpcode() == ISD::UNDEF)
11061 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11064 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11065 SelectionDAG &DAG) const {
11066 MVT VT = Op.getSimpleValueType();
11067 MVT EltVT = VT.getVectorElementType();
11069 if (EltVT == MVT::i1)
11070 return InsertBitToMaskVector(Op, DAG);
11073 SDValue N0 = Op.getOperand(0);
11074 SDValue N1 = Op.getOperand(1);
11075 SDValue N2 = Op.getOperand(2);
11076 if (!isa<ConstantSDNode>(N2))
11078 auto *N2C = cast<ConstantSDNode>(N2);
11079 unsigned IdxVal = N2C->getZExtValue();
11081 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11082 // into that, and then insert the subvector back into the result.
11083 if (VT.is256BitVector() || VT.is512BitVector()) {
11084 // With a 256-bit vector, we can insert into the zero element efficiently
11085 // using a blend if we have AVX or AVX2 and the right data type.
11086 if (VT.is256BitVector() && IdxVal == 0) {
11087 // TODO: It is worthwhile to cast integer to floating point and back
11088 // and incur a domain crossing penalty if that's what we'll end up
11089 // doing anyway after extracting to a 128-bit vector.
11090 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11091 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11092 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11093 N2 = DAG.getIntPtrConstant(1, dl);
11094 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11098 // Get the desired 128-bit vector chunk.
11099 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11101 // Insert the element into the desired chunk.
11102 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11103 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11105 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11106 DAG.getConstant(IdxIn128, dl, MVT::i32));
11108 // Insert the changed part back into the bigger vector
11109 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11111 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11113 if (Subtarget->hasSSE41()) {
11114 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11116 if (VT == MVT::v8i16) {
11117 Opc = X86ISD::PINSRW;
11119 assert(VT == MVT::v16i8);
11120 Opc = X86ISD::PINSRB;
11123 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11125 if (N1.getValueType() != MVT::i32)
11126 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11127 if (N2.getValueType() != MVT::i32)
11128 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11129 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11132 if (EltVT == MVT::f32) {
11133 // Bits [7:6] of the constant are the source select. This will always be
11134 // zero here. The DAG Combiner may combine an extract_elt index into
11135 // these bits. For example (insert (extract, 3), 2) could be matched by
11136 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11137 // Bits [5:4] of the constant are the destination select. This is the
11138 // value of the incoming immediate.
11139 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11140 // combine either bitwise AND or insert of float 0.0 to set these bits.
11142 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11143 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11144 // If this is an insertion of 32-bits into the low 32-bits of
11145 // a vector, we prefer to generate a blend with immediate rather
11146 // than an insertps. Blends are simpler operations in hardware and so
11147 // will always have equal or better performance than insertps.
11148 // But if optimizing for size and there's a load folding opportunity,
11149 // generate insertps because blendps does not have a 32-bit memory
11151 N2 = DAG.getIntPtrConstant(1, dl);
11152 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11153 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11155 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11156 // Create this as a scalar to vector..
11157 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11158 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11161 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11162 // PINSR* works with constant index.
11167 if (EltVT == MVT::i8)
11170 if (EltVT.getSizeInBits() == 16) {
11171 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11172 // as its second argument.
11173 if (N1.getValueType() != MVT::i32)
11174 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11175 if (N2.getValueType() != MVT::i32)
11176 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11177 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11182 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11184 MVT OpVT = Op.getSimpleValueType();
11186 // If this is a 256-bit vector result, first insert into a 128-bit
11187 // vector and then insert into the 256-bit vector.
11188 if (!OpVT.is128BitVector()) {
11189 // Insert into a 128-bit vector.
11190 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11191 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11192 OpVT.getVectorNumElements() / SizeFactor);
11194 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11196 // Insert the 128-bit vector.
11197 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11200 if (OpVT == MVT::v1i64 &&
11201 Op.getOperand(0).getValueType() == MVT::i64)
11202 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11204 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11205 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11206 return DAG.getBitcast(
11207 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11210 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11211 // a simple subregister reference or explicit instructions to grab
11212 // upper bits of a vector.
11213 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11214 SelectionDAG &DAG) {
11216 SDValue In = Op.getOperand(0);
11217 SDValue Idx = Op.getOperand(1);
11218 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11219 MVT ResVT = Op.getSimpleValueType();
11220 MVT InVT = In.getSimpleValueType();
11222 if (Subtarget->hasFp256()) {
11223 if (ResVT.is128BitVector() &&
11224 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11225 isa<ConstantSDNode>(Idx)) {
11226 return Extract128BitVector(In, IdxVal, DAG, dl);
11228 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11229 isa<ConstantSDNode>(Idx)) {
11230 return Extract256BitVector(In, IdxVal, DAG, dl);
11236 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11237 // simple superregister reference or explicit instructions to insert
11238 // the upper bits of a vector.
11239 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11240 SelectionDAG &DAG) {
11241 if (!Subtarget->hasAVX())
11245 SDValue Vec = Op.getOperand(0);
11246 SDValue SubVec = Op.getOperand(1);
11247 SDValue Idx = Op.getOperand(2);
11249 if (!isa<ConstantSDNode>(Idx))
11252 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11253 MVT OpVT = Op.getSimpleValueType();
11254 MVT SubVecVT = SubVec.getSimpleValueType();
11256 // Fold two 16-byte subvector loads into one 32-byte load:
11257 // (insert_subvector (insert_subvector undef, (load addr), 0),
11258 // (load addr + 16), Elts/2)
11260 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11261 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11262 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
11263 !Subtarget->isUnalignedMem32Slow()) {
11264 SDValue SubVec2 = Vec.getOperand(1);
11265 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
11266 if (Idx2->getZExtValue() == 0) {
11267 SDValue Ops[] = { SubVec2, SubVec };
11268 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11274 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11275 SubVecVT.is128BitVector())
11276 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11278 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11279 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11281 if (OpVT.getVectorElementType() == MVT::i1) {
11282 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11284 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11285 SDValue Undef = DAG.getUNDEF(OpVT);
11286 unsigned NumElems = OpVT.getVectorNumElements();
11287 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11289 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11290 // Zero upper bits of the Vec
11291 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11292 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11294 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11296 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11297 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11300 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11302 // Zero upper bits of the Vec2
11303 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11304 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11305 // Zero lower bits of the Vec
11306 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11307 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11308 // Merge them together
11309 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11315 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11316 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11317 // one of the above mentioned nodes. It has to be wrapped because otherwise
11318 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11319 // be used to form addressing mode. These wrapped nodes will be selected
11322 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11323 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11325 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11326 // global base reg.
11327 unsigned char OpFlag = 0;
11328 unsigned WrapperKind = X86ISD::Wrapper;
11329 CodeModel::Model M = DAG.getTarget().getCodeModel();
11331 if (Subtarget->isPICStyleRIPRel() &&
11332 (M == CodeModel::Small || M == CodeModel::Kernel))
11333 WrapperKind = X86ISD::WrapperRIP;
11334 else if (Subtarget->isPICStyleGOT())
11335 OpFlag = X86II::MO_GOTOFF;
11336 else if (Subtarget->isPICStyleStubPIC())
11337 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11339 auto PtrVT = getPointerTy(DAG.getDataLayout());
11340 SDValue Result = DAG.getTargetConstantPool(
11341 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11343 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11344 // With PIC, the address is actually $g + Offset.
11347 DAG.getNode(ISD::ADD, DL, PtrVT,
11348 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11354 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11355 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11357 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11358 // global base reg.
11359 unsigned char OpFlag = 0;
11360 unsigned WrapperKind = X86ISD::Wrapper;
11361 CodeModel::Model M = DAG.getTarget().getCodeModel();
11363 if (Subtarget->isPICStyleRIPRel() &&
11364 (M == CodeModel::Small || M == CodeModel::Kernel))
11365 WrapperKind = X86ISD::WrapperRIP;
11366 else if (Subtarget->isPICStyleGOT())
11367 OpFlag = X86II::MO_GOTOFF;
11368 else if (Subtarget->isPICStyleStubPIC())
11369 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11371 auto PtrVT = getPointerTy(DAG.getDataLayout());
11372 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11374 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11376 // With PIC, the address is actually $g + Offset.
11379 DAG.getNode(ISD::ADD, DL, PtrVT,
11380 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11386 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11387 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11389 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11390 // global base reg.
11391 unsigned char OpFlag = 0;
11392 unsigned WrapperKind = X86ISD::Wrapper;
11393 CodeModel::Model M = DAG.getTarget().getCodeModel();
11395 if (Subtarget->isPICStyleRIPRel() &&
11396 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11397 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11398 OpFlag = X86II::MO_GOTPCREL;
11399 WrapperKind = X86ISD::WrapperRIP;
11400 } else if (Subtarget->isPICStyleGOT()) {
11401 OpFlag = X86II::MO_GOT;
11402 } else if (Subtarget->isPICStyleStubPIC()) {
11403 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11404 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11405 OpFlag = X86II::MO_DARWIN_NONLAZY;
11408 auto PtrVT = getPointerTy(DAG.getDataLayout());
11409 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11412 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11414 // With PIC, the address is actually $g + Offset.
11415 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11416 !Subtarget->is64Bit()) {
11418 DAG.getNode(ISD::ADD, DL, PtrVT,
11419 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11422 // For symbols that require a load from a stub to get the address, emit the
11424 if (isGlobalStubReference(OpFlag))
11425 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11426 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11427 false, false, false, 0);
11433 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11434 // Create the TargetBlockAddressAddress node.
11435 unsigned char OpFlags =
11436 Subtarget->ClassifyBlockAddressReference();
11437 CodeModel::Model M = DAG.getTarget().getCodeModel();
11438 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11439 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11441 auto PtrVT = getPointerTy(DAG.getDataLayout());
11442 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11444 if (Subtarget->isPICStyleRIPRel() &&
11445 (M == CodeModel::Small || M == CodeModel::Kernel))
11446 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11448 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11450 // With PIC, the address is actually $g + Offset.
11451 if (isGlobalRelativeToPICBase(OpFlags)) {
11452 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11453 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11460 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11461 int64_t Offset, SelectionDAG &DAG) const {
11462 // Create the TargetGlobalAddress node, folding in the constant
11463 // offset if it is legal.
11464 unsigned char OpFlags =
11465 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11466 CodeModel::Model M = DAG.getTarget().getCodeModel();
11467 auto PtrVT = getPointerTy(DAG.getDataLayout());
11469 if (OpFlags == X86II::MO_NO_FLAG &&
11470 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11471 // A direct static reference to a global.
11472 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11475 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11478 if (Subtarget->isPICStyleRIPRel() &&
11479 (M == CodeModel::Small || M == CodeModel::Kernel))
11480 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11482 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11484 // With PIC, the address is actually $g + Offset.
11485 if (isGlobalRelativeToPICBase(OpFlags)) {
11486 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11487 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11490 // For globals that require a load from a stub to get the address, emit the
11492 if (isGlobalStubReference(OpFlags))
11493 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11494 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11495 false, false, false, 0);
11497 // If there was a non-zero offset that we didn't fold, create an explicit
11498 // addition for it.
11500 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11501 DAG.getConstant(Offset, dl, PtrVT));
11507 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11508 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11509 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11510 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11514 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11515 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11516 unsigned char OperandFlags, bool LocalDynamic = false) {
11517 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11518 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11520 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11521 GA->getValueType(0),
11525 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11529 SDValue Ops[] = { Chain, TGA, *InFlag };
11530 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11532 SDValue Ops[] = { Chain, TGA };
11533 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11536 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11537 MFI->setAdjustsStack(true);
11538 MFI->setHasCalls(true);
11540 SDValue Flag = Chain.getValue(1);
11541 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11544 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11546 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11549 SDLoc dl(GA); // ? function entry point might be better
11550 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11551 DAG.getNode(X86ISD::GlobalBaseReg,
11552 SDLoc(), PtrVT), InFlag);
11553 InFlag = Chain.getValue(1);
11555 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11558 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11560 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11562 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11563 X86::RAX, X86II::MO_TLSGD);
11566 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11572 // Get the start address of the TLS block for this module.
11573 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11574 .getInfo<X86MachineFunctionInfo>();
11575 MFI->incNumLocalDynamicTLSAccesses();
11579 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11580 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11583 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11584 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11585 InFlag = Chain.getValue(1);
11586 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11587 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11590 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11594 unsigned char OperandFlags = X86II::MO_DTPOFF;
11595 unsigned WrapperKind = X86ISD::Wrapper;
11596 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11597 GA->getValueType(0),
11598 GA->getOffset(), OperandFlags);
11599 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11601 // Add x@dtpoff with the base.
11602 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11605 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11606 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11607 const EVT PtrVT, TLSModel::Model model,
11608 bool is64Bit, bool isPIC) {
11611 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11612 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11613 is64Bit ? 257 : 256));
11615 SDValue ThreadPointer =
11616 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11617 MachinePointerInfo(Ptr), false, false, false, 0);
11619 unsigned char OperandFlags = 0;
11620 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11622 unsigned WrapperKind = X86ISD::Wrapper;
11623 if (model == TLSModel::LocalExec) {
11624 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11625 } else if (model == TLSModel::InitialExec) {
11627 OperandFlags = X86II::MO_GOTTPOFF;
11628 WrapperKind = X86ISD::WrapperRIP;
11630 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11633 llvm_unreachable("Unexpected model");
11636 // emit "addl x@ntpoff,%eax" (local exec)
11637 // or "addl x@indntpoff,%eax" (initial exec)
11638 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11640 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11641 GA->getOffset(), OperandFlags);
11642 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11644 if (model == TLSModel::InitialExec) {
11645 if (isPIC && !is64Bit) {
11646 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11647 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11651 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11652 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11653 false, false, false, 0);
11656 // The address of the thread local variable is the add of the thread
11657 // pointer with the offset of the variable.
11658 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11662 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11664 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11665 const GlobalValue *GV = GA->getGlobal();
11666 auto PtrVT = getPointerTy(DAG.getDataLayout());
11668 if (Subtarget->isTargetELF()) {
11669 if (DAG.getTarget().Options.EmulatedTLS)
11670 return LowerToTLSEmulatedModel(GA, DAG);
11671 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11673 case TLSModel::GeneralDynamic:
11674 if (Subtarget->is64Bit())
11675 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11676 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11677 case TLSModel::LocalDynamic:
11678 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11679 Subtarget->is64Bit());
11680 case TLSModel::InitialExec:
11681 case TLSModel::LocalExec:
11682 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11683 DAG.getTarget().getRelocationModel() ==
11686 llvm_unreachable("Unknown TLS model.");
11689 if (Subtarget->isTargetDarwin()) {
11690 // Darwin only has one model of TLS. Lower to that.
11691 unsigned char OpFlag = 0;
11692 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11693 X86ISD::WrapperRIP : X86ISD::Wrapper;
11695 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11696 // global base reg.
11697 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11698 !Subtarget->is64Bit();
11700 OpFlag = X86II::MO_TLVP_PIC_BASE;
11702 OpFlag = X86II::MO_TLVP;
11704 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11705 GA->getValueType(0),
11706 GA->getOffset(), OpFlag);
11707 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11709 // With PIC32, the address is actually $g + Offset.
11711 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11712 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11715 // Lowering the machine isd will make sure everything is in the right
11717 SDValue Chain = DAG.getEntryNode();
11718 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11719 SDValue Args[] = { Chain, Offset };
11720 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11722 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11723 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11724 MFI->setAdjustsStack(true);
11726 // And our return value (tls address) is in the standard call return value
11728 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11729 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11732 if (Subtarget->isTargetKnownWindowsMSVC() ||
11733 Subtarget->isTargetWindowsGNU()) {
11734 // Just use the implicit TLS architecture
11735 // Need to generate someting similar to:
11736 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11738 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11739 // mov rcx, qword [rdx+rcx*8]
11740 // mov eax, .tls$:tlsvar
11741 // [rax+rcx] contains the address
11742 // Windows 64bit: gs:0x58
11743 // Windows 32bit: fs:__tls_array
11746 SDValue Chain = DAG.getEntryNode();
11748 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11749 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11750 // use its literal value of 0x2C.
11751 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11752 ? Type::getInt8PtrTy(*DAG.getContext(),
11754 : Type::getInt32PtrTy(*DAG.getContext(),
11757 SDValue TlsArray = Subtarget->is64Bit()
11758 ? DAG.getIntPtrConstant(0x58, dl)
11759 : (Subtarget->isTargetWindowsGNU()
11760 ? DAG.getIntPtrConstant(0x2C, dl)
11761 : DAG.getExternalSymbol("_tls_array", PtrVT));
11763 SDValue ThreadPointer =
11764 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
11768 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11769 res = ThreadPointer;
11771 // Load the _tls_index variable
11772 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
11773 if (Subtarget->is64Bit())
11774 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
11775 MachinePointerInfo(), MVT::i32, false, false,
11778 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
11781 auto &DL = DAG.getDataLayout();
11783 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
11784 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
11786 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
11789 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
11792 // Get the offset of start of .tls section
11793 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11794 GA->getValueType(0),
11795 GA->getOffset(), X86II::MO_SECREL);
11796 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
11798 // The address of the thread local variable is the add of the thread
11799 // pointer with the offset of the variable.
11800 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
11803 llvm_unreachable("TLS not implemented for this target.");
11806 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11807 /// and take a 2 x i32 value to shift plus a shift amount.
11808 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11809 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11810 MVT VT = Op.getSimpleValueType();
11811 unsigned VTBits = VT.getSizeInBits();
11813 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11814 SDValue ShOpLo = Op.getOperand(0);
11815 SDValue ShOpHi = Op.getOperand(1);
11816 SDValue ShAmt = Op.getOperand(2);
11817 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11818 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11820 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11821 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11822 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11823 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11824 : DAG.getConstant(0, dl, VT);
11826 SDValue Tmp2, Tmp3;
11827 if (Op.getOpcode() == ISD::SHL_PARTS) {
11828 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11829 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11831 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11832 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11835 // If the shift amount is larger or equal than the width of a part we can't
11836 // rely on the results of shld/shrd. Insert a test and select the appropriate
11837 // values for large shift amounts.
11838 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11839 DAG.getConstant(VTBits, dl, MVT::i8));
11840 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11841 AndNode, DAG.getConstant(0, dl, MVT::i8));
11844 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11845 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11846 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11848 if (Op.getOpcode() == ISD::SHL_PARTS) {
11849 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11850 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11852 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11853 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11856 SDValue Ops[2] = { Lo, Hi };
11857 return DAG.getMergeValues(Ops, dl);
11860 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11861 SelectionDAG &DAG) const {
11862 SDValue Src = Op.getOperand(0);
11863 MVT SrcVT = Src.getSimpleValueType();
11864 MVT VT = Op.getSimpleValueType();
11867 if (SrcVT.isVector()) {
11868 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
11869 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
11870 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
11871 DAG.getUNDEF(SrcVT)));
11873 if (SrcVT.getVectorElementType() == MVT::i1) {
11874 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11875 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11876 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
11881 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11882 "Unknown SINT_TO_FP to lower!");
11884 // These are really Legal; return the operand so the caller accepts it as
11886 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11888 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11889 Subtarget->is64Bit()) {
11893 unsigned Size = SrcVT.getSizeInBits()/8;
11894 MachineFunction &MF = DAG.getMachineFunction();
11895 auto PtrVT = getPointerTy(MF.getDataLayout());
11896 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11897 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11898 SDValue Chain = DAG.getStore(
11899 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
11900 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
11902 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11905 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11907 SelectionDAG &DAG) const {
11911 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11913 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11915 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11917 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11919 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11920 MachineMemOperand *MMO;
11922 int SSFI = FI->getIndex();
11923 MMO = DAG.getMachineFunction().getMachineMemOperand(
11924 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11925 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11927 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11928 StackSlot = StackSlot.getOperand(1);
11930 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11931 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11933 Tys, Ops, SrcVT, MMO);
11936 Chain = Result.getValue(1);
11937 SDValue InFlag = Result.getValue(2);
11939 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11940 // shouldn't be necessary except that RFP cannot be live across
11941 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11942 MachineFunction &MF = DAG.getMachineFunction();
11943 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11944 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11945 auto PtrVT = getPointerTy(MF.getDataLayout());
11946 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11947 Tys = DAG.getVTList(MVT::Other);
11949 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11951 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
11952 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11953 MachineMemOperand::MOStore, SSFISize, SSFISize);
11955 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11956 Ops, Op.getValueType(), MMO);
11957 Result = DAG.getLoad(
11958 Op.getValueType(), DL, Chain, StackSlot,
11959 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11960 false, false, false, 0);
11966 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11967 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11968 SelectionDAG &DAG) const {
11969 // This algorithm is not obvious. Here it is what we're trying to output:
11972 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11973 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11975 haddpd %xmm0, %xmm0
11977 pshufd $0x4e, %xmm0, %xmm1
11983 LLVMContext *Context = DAG.getContext();
11985 // Build some magic constants.
11986 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11987 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11988 auto PtrVT = getPointerTy(DAG.getDataLayout());
11989 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
11991 SmallVector<Constant*,2> CV1;
11993 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11994 APInt(64, 0x4330000000000000ULL))));
11996 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11997 APInt(64, 0x4530000000000000ULL))));
11998 Constant *C1 = ConstantVector::get(CV1);
11999 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12001 // Load the 64-bit value into an XMM register.
12002 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12005 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12006 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12007 false, false, false, 16);
12009 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12012 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12013 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12014 false, false, false, 16);
12015 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12016 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12019 if (Subtarget->hasSSE3()) {
12020 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12021 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12023 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12024 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12026 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12027 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12030 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12031 DAG.getIntPtrConstant(0, dl));
12034 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12035 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12036 SelectionDAG &DAG) const {
12038 // FP constant to bias correct the final result.
12039 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12042 // Load the 32-bit value into an XMM register.
12043 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12046 // Zero out the upper parts of the register.
12047 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12049 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12050 DAG.getBitcast(MVT::v2f64, Load),
12051 DAG.getIntPtrConstant(0, dl));
12053 // Or the load with the bias.
12054 SDValue Or = DAG.getNode(
12055 ISD::OR, dl, MVT::v2i64,
12056 DAG.getBitcast(MVT::v2i64,
12057 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12058 DAG.getBitcast(MVT::v2i64,
12059 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12061 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12062 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12064 // Subtract the bias.
12065 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12067 // Handle final rounding.
12068 EVT DestVT = Op.getValueType();
12070 if (DestVT.bitsLT(MVT::f64))
12071 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12072 DAG.getIntPtrConstant(0, dl));
12073 if (DestVT.bitsGT(MVT::f64))
12074 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12076 // Handle final rounding.
12080 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12081 const X86Subtarget &Subtarget) {
12082 // The algorithm is the following:
12083 // #ifdef __SSE4_1__
12084 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12085 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12086 // (uint4) 0x53000000, 0xaa);
12088 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12089 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12091 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12092 // return (float4) lo + fhi;
12095 SDValue V = Op->getOperand(0);
12096 EVT VecIntVT = V.getValueType();
12097 bool Is128 = VecIntVT == MVT::v4i32;
12098 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12099 // If we convert to something else than the supported type, e.g., to v4f64,
12101 if (VecFloatVT != Op->getValueType(0))
12104 unsigned NumElts = VecIntVT.getVectorNumElements();
12105 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12106 "Unsupported custom type");
12107 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12109 // In the #idef/#else code, we have in common:
12110 // - The vector of constants:
12116 // Create the splat vector for 0x4b000000.
12117 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12118 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12119 CstLow, CstLow, CstLow, CstLow};
12120 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12121 makeArrayRef(&CstLowArray[0], NumElts));
12122 // Create the splat vector for 0x53000000.
12123 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12124 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12125 CstHigh, CstHigh, CstHigh, CstHigh};
12126 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12127 makeArrayRef(&CstHighArray[0], NumElts));
12129 // Create the right shift.
12130 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12131 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12132 CstShift, CstShift, CstShift, CstShift};
12133 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12134 makeArrayRef(&CstShiftArray[0], NumElts));
12135 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12138 if (Subtarget.hasSSE41()) {
12139 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12140 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12141 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12142 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12143 // Low will be bitcasted right away, so do not bother bitcasting back to its
12145 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12146 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12147 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12148 // (uint4) 0x53000000, 0xaa);
12149 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12150 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12151 // High will be bitcasted right away, so do not bother bitcasting back to
12152 // its original type.
12153 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12154 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12156 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12157 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12158 CstMask, CstMask, CstMask);
12159 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12160 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12161 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12163 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12164 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12167 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12168 SDValue CstFAdd = DAG.getConstantFP(
12169 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12170 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12171 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12172 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12173 makeArrayRef(&CstFAddArray[0], NumElts));
12175 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12176 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12178 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12179 // return (float4) lo + fhi;
12180 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12181 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12184 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12185 SelectionDAG &DAG) const {
12186 SDValue N0 = Op.getOperand(0);
12187 MVT SVT = N0.getSimpleValueType();
12190 switch (SVT.SimpleTy) {
12192 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12197 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12198 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12199 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12203 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12206 if (Subtarget->hasAVX512())
12207 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12208 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12210 llvm_unreachable(nullptr);
12213 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12214 SelectionDAG &DAG) const {
12215 SDValue N0 = Op.getOperand(0);
12217 auto PtrVT = getPointerTy(DAG.getDataLayout());
12219 if (Op.getValueType().isVector())
12220 return lowerUINT_TO_FP_vec(Op, DAG);
12222 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12223 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12224 // the optimization here.
12225 if (DAG.SignBitIsZero(N0))
12226 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12228 MVT SrcVT = N0.getSimpleValueType();
12229 MVT DstVT = Op.getSimpleValueType();
12230 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12231 return LowerUINT_TO_FP_i64(Op, DAG);
12232 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12233 return LowerUINT_TO_FP_i32(Op, DAG);
12234 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12237 // Make a 64-bit buffer, and use it to build an FILD.
12238 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12239 if (SrcVT == MVT::i32) {
12240 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12241 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12242 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12243 StackSlot, MachinePointerInfo(),
12245 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12246 OffsetSlot, MachinePointerInfo(),
12248 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12252 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12253 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12254 StackSlot, MachinePointerInfo(),
12256 // For i64 source, we need to add the appropriate power of 2 if the input
12257 // was negative. This is the same as the optimization in
12258 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12259 // we must be careful to do the computation in x87 extended precision, not
12260 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12261 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12262 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12263 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12264 MachineMemOperand::MOLoad, 8, 8);
12266 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12267 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12268 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12271 APInt FF(32, 0x5F800000ULL);
12273 // Check whether the sign bit is set.
12274 SDValue SignSet = DAG.getSetCC(
12275 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12276 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12278 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12279 SDValue FudgePtr = DAG.getConstantPool(
12280 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12282 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12283 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12284 SDValue Four = DAG.getIntPtrConstant(4, dl);
12285 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12287 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12289 // Load the value out, extending it from f32 to f80.
12290 // FIXME: Avoid the extend by constructing the right constant pool?
12291 SDValue Fudge = DAG.getExtLoad(
12292 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12293 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12294 false, false, false, 4);
12295 // Extend everything to 80 bits to force it to be done on x87.
12296 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12297 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12298 DAG.getIntPtrConstant(0, dl));
12301 std::pair<SDValue,SDValue>
12302 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12303 bool IsSigned, bool IsReplace) const {
12306 EVT DstTy = Op.getValueType();
12307 auto PtrVT = getPointerTy(DAG.getDataLayout());
12309 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12310 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12314 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12315 DstTy.getSimpleVT() >= MVT::i16 &&
12316 "Unknown FP_TO_INT to lower!");
12318 // These are really Legal.
12319 if (DstTy == MVT::i32 &&
12320 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12321 return std::make_pair(SDValue(), SDValue());
12322 if (Subtarget->is64Bit() &&
12323 DstTy == MVT::i64 &&
12324 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12325 return std::make_pair(SDValue(), SDValue());
12327 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12328 // stack slot, or into the FTOL runtime function.
12329 MachineFunction &MF = DAG.getMachineFunction();
12330 unsigned MemSize = DstTy.getSizeInBits()/8;
12331 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12332 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12335 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12336 Opc = X86ISD::WIN_FTOL;
12338 switch (DstTy.getSimpleVT().SimpleTy) {
12339 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12340 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12341 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12342 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12345 SDValue Chain = DAG.getEntryNode();
12346 SDValue Value = Op.getOperand(0);
12347 EVT TheVT = Op.getOperand(0).getValueType();
12348 // FIXME This causes a redundant load/store if the SSE-class value is already
12349 // in memory, such as if it is on the callstack.
12350 if (isScalarFPTypeInSSEReg(TheVT)) {
12351 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12352 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12353 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12355 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12357 Chain, StackSlot, DAG.getValueType(TheVT)
12360 MachineMemOperand *MMO =
12361 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12362 MachineMemOperand::MOLoad, MemSize, MemSize);
12363 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12364 Chain = Value.getValue(1);
12365 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12366 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12369 MachineMemOperand *MMO =
12370 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12371 MachineMemOperand::MOStore, MemSize, MemSize);
12373 if (Opc != X86ISD::WIN_FTOL) {
12374 // Build the FP_TO_INT*_IN_MEM
12375 SDValue Ops[] = { Chain, Value, StackSlot };
12376 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12378 return std::make_pair(FIST, StackSlot);
12380 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12381 DAG.getVTList(MVT::Other, MVT::Glue),
12383 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12384 MVT::i32, ftol.getValue(1));
12385 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12386 MVT::i32, eax.getValue(2));
12387 SDValue Ops[] = { eax, edx };
12388 SDValue pair = IsReplace
12389 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12390 : DAG.getMergeValues(Ops, DL);
12391 return std::make_pair(pair, SDValue());
12395 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12396 const X86Subtarget *Subtarget) {
12397 MVT VT = Op->getSimpleValueType(0);
12398 SDValue In = Op->getOperand(0);
12399 MVT InVT = In.getSimpleValueType();
12402 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12403 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12405 // Optimize vectors in AVX mode:
12408 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12409 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12410 // Concat upper and lower parts.
12413 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12414 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12415 // Concat upper and lower parts.
12418 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12419 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12420 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12423 if (Subtarget->hasInt256())
12424 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12426 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12427 SDValue Undef = DAG.getUNDEF(InVT);
12428 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12429 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12430 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12432 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12433 VT.getVectorNumElements()/2);
12435 OpLo = DAG.getBitcast(HVT, OpLo);
12436 OpHi = DAG.getBitcast(HVT, OpHi);
12438 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12441 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12442 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12443 MVT VT = Op->getSimpleValueType(0);
12444 SDValue In = Op->getOperand(0);
12445 MVT InVT = In.getSimpleValueType();
12447 unsigned int NumElts = VT.getVectorNumElements();
12448 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12451 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12452 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12454 assert(InVT.getVectorElementType() == MVT::i1);
12455 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12457 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12459 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12461 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12462 if (VT.is512BitVector())
12464 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12467 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12468 SelectionDAG &DAG) {
12469 if (Subtarget->hasFp256())
12470 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12476 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12477 SelectionDAG &DAG) {
12479 MVT VT = Op.getSimpleValueType();
12480 SDValue In = Op.getOperand(0);
12481 MVT SVT = In.getSimpleValueType();
12483 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12484 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12486 if (Subtarget->hasFp256())
12487 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12490 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12491 VT.getVectorNumElements() != SVT.getVectorNumElements());
12495 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12497 MVT VT = Op.getSimpleValueType();
12498 SDValue In = Op.getOperand(0);
12499 MVT InVT = In.getSimpleValueType();
12501 if (VT == MVT::i1) {
12502 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12503 "Invalid scalar TRUNCATE operation");
12504 if (InVT.getSizeInBits() >= 32)
12506 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12507 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12509 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12510 "Invalid TRUNCATE operation");
12512 // move vector to mask - truncate solution for SKX
12513 if (VT.getVectorElementType() == MVT::i1) {
12514 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12515 Subtarget->hasBWI())
12516 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12517 if ((InVT.is256BitVector() || InVT.is128BitVector())
12518 && InVT.getScalarSizeInBits() <= 16 &&
12519 Subtarget->hasBWI() && Subtarget->hasVLX())
12520 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12521 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12522 Subtarget->hasDQI())
12523 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12524 if ((InVT.is256BitVector() || InVT.is128BitVector())
12525 && InVT.getScalarSizeInBits() >= 32 &&
12526 Subtarget->hasDQI() && Subtarget->hasVLX())
12527 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12530 if (VT.getVectorElementType() == MVT::i1) {
12531 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12532 unsigned NumElts = InVT.getVectorNumElements();
12533 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12534 if (InVT.getSizeInBits() < 512) {
12535 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12536 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12541 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12542 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12543 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12546 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12547 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12548 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12549 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12551 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12552 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12553 if (Subtarget->hasInt256()) {
12554 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12555 In = DAG.getBitcast(MVT::v8i32, In);
12556 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12558 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12559 DAG.getIntPtrConstant(0, DL));
12562 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12563 DAG.getIntPtrConstant(0, DL));
12564 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12565 DAG.getIntPtrConstant(2, DL));
12566 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12567 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12568 static const int ShufMask[] = {0, 2, 4, 6};
12569 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12572 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12573 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12574 if (Subtarget->hasInt256()) {
12575 In = DAG.getBitcast(MVT::v32i8, In);
12577 SmallVector<SDValue,32> pshufbMask;
12578 for (unsigned i = 0; i < 2; ++i) {
12579 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12580 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12581 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12582 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12583 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12584 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12585 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12586 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12587 for (unsigned j = 0; j < 8; ++j)
12588 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12590 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12591 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12592 In = DAG.getBitcast(MVT::v4i64, In);
12594 static const int ShufMask[] = {0, 2, -1, -1};
12595 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12597 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12598 DAG.getIntPtrConstant(0, DL));
12599 return DAG.getBitcast(VT, In);
12602 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12603 DAG.getIntPtrConstant(0, DL));
12605 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12606 DAG.getIntPtrConstant(4, DL));
12608 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12609 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12611 // The PSHUFB mask:
12612 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12613 -1, -1, -1, -1, -1, -1, -1, -1};
12615 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12616 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12617 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12619 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12620 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12622 // The MOVLHPS Mask:
12623 static const int ShufMask2[] = {0, 1, 4, 5};
12624 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12625 return DAG.getBitcast(MVT::v8i16, res);
12628 // Handle truncation of V256 to V128 using shuffles.
12629 if (!VT.is128BitVector() || !InVT.is256BitVector())
12632 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12634 unsigned NumElems = VT.getVectorNumElements();
12635 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12637 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12638 // Prepare truncation shuffle mask
12639 for (unsigned i = 0; i != NumElems; ++i)
12640 MaskVec[i] = i * 2;
12641 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12642 DAG.getUNDEF(NVT), &MaskVec[0]);
12643 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12644 DAG.getIntPtrConstant(0, DL));
12647 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12648 SelectionDAG &DAG) const {
12649 assert(!Op.getSimpleValueType().isVector());
12651 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12652 /*IsSigned=*/ true, /*IsReplace=*/ false);
12653 SDValue FIST = Vals.first, StackSlot = Vals.second;
12654 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12655 if (!FIST.getNode()) return Op;
12657 if (StackSlot.getNode())
12658 // Load the result.
12659 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12660 FIST, StackSlot, MachinePointerInfo(),
12661 false, false, false, 0);
12663 // The node is the result.
12667 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12668 SelectionDAG &DAG) const {
12669 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12670 /*IsSigned=*/ false, /*IsReplace=*/ false);
12671 SDValue FIST = Vals.first, StackSlot = Vals.second;
12672 assert(FIST.getNode() && "Unexpected failure");
12674 if (StackSlot.getNode())
12675 // Load the result.
12676 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12677 FIST, StackSlot, MachinePointerInfo(),
12678 false, false, false, 0);
12680 // The node is the result.
12684 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12686 MVT VT = Op.getSimpleValueType();
12687 SDValue In = Op.getOperand(0);
12688 MVT SVT = In.getSimpleValueType();
12690 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12692 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12693 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12694 In, DAG.getUNDEF(SVT)));
12697 /// The only differences between FABS and FNEG are the mask and the logic op.
12698 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12699 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12700 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12701 "Wrong opcode for lowering FABS or FNEG.");
12703 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12705 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12706 // into an FNABS. We'll lower the FABS after that if it is still in use.
12708 for (SDNode *User : Op->uses())
12709 if (User->getOpcode() == ISD::FNEG)
12713 MVT VT = Op.getSimpleValueType();
12715 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12716 // decide if we should generate a 16-byte constant mask when we only need 4 or
12717 // 8 bytes for the scalar case.
12723 if (VT.isVector()) {
12725 EltVT = VT.getVectorElementType();
12726 NumElts = VT.getVectorNumElements();
12728 // There are no scalar bitwise logical SSE/AVX instructions, so we
12729 // generate a 16-byte vector constant and logic op even for the scalar case.
12730 // Using a 16-byte mask allows folding the load of the mask with
12731 // the logic op, so it can save (~4 bytes) on code size.
12732 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12734 NumElts = (VT == MVT::f64) ? 2 : 4;
12737 unsigned EltBits = EltVT.getSizeInBits();
12738 LLVMContext *Context = DAG.getContext();
12739 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12741 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12742 Constant *C = ConstantInt::get(*Context, MaskElt);
12743 C = ConstantVector::getSplat(NumElts, C);
12744 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12745 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
12746 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12748 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12749 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12750 false, false, false, Alignment);
12752 SDValue Op0 = Op.getOperand(0);
12753 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12755 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12756 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12759 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12761 // For the scalar case extend to a 128-bit vector, perform the logic op,
12762 // and extract the scalar result back out.
12763 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
12764 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12765 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
12766 DAG.getIntPtrConstant(0, dl));
12769 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12770 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12771 LLVMContext *Context = DAG.getContext();
12772 SDValue Op0 = Op.getOperand(0);
12773 SDValue Op1 = Op.getOperand(1);
12775 MVT VT = Op.getSimpleValueType();
12776 MVT SrcVT = Op1.getSimpleValueType();
12778 // If second operand is smaller, extend it first.
12779 if (SrcVT.bitsLT(VT)) {
12780 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12783 // And if it is bigger, shrink it first.
12784 if (SrcVT.bitsGT(VT)) {
12785 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12789 // At this point the operands and the result should have the same
12790 // type, and that won't be f80 since that is not custom lowered.
12792 const fltSemantics &Sem =
12793 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12794 const unsigned SizeInBits = VT.getSizeInBits();
12796 SmallVector<Constant *, 4> CV(
12797 VT == MVT::f64 ? 2 : 4,
12798 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12800 // First, clear all bits but the sign bit from the second operand (sign).
12801 CV[0] = ConstantFP::get(*Context,
12802 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12803 Constant *C = ConstantVector::get(CV);
12804 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
12805 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12807 // Perform all logic operations as 16-byte vectors because there are no
12808 // scalar FP logic instructions in SSE. This allows load folding of the
12809 // constants into the logic instructions.
12810 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12812 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12813 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12814 false, false, false, 16);
12815 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
12816 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
12818 // Next, clear the sign bit from the first operand (magnitude).
12819 // If it's a constant, we can clear it here.
12820 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12821 APFloat APF = Op0CN->getValueAPF();
12822 // If the magnitude is a positive zero, the sign bit alone is enough.
12823 if (APF.isPosZero())
12824 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
12825 DAG.getIntPtrConstant(0, dl));
12827 CV[0] = ConstantFP::get(*Context, APF);
12829 CV[0] = ConstantFP::get(
12831 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12833 C = ConstantVector::get(CV);
12834 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12836 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12837 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12838 false, false, false, 16);
12839 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12840 if (!isa<ConstantFPSDNode>(Op0)) {
12841 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
12842 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
12844 // OR the magnitude value with the sign bit.
12845 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
12846 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
12847 DAG.getIntPtrConstant(0, dl));
12850 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12851 SDValue N0 = Op.getOperand(0);
12853 MVT VT = Op.getSimpleValueType();
12855 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12856 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12857 DAG.getConstant(1, dl, VT));
12858 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12861 // Check whether an OR'd tree is PTEST-able.
12862 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12863 SelectionDAG &DAG) {
12864 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12866 if (!Subtarget->hasSSE41())
12869 if (!Op->hasOneUse())
12872 SDNode *N = Op.getNode();
12875 SmallVector<SDValue, 8> Opnds;
12876 DenseMap<SDValue, unsigned> VecInMap;
12877 SmallVector<SDValue, 8> VecIns;
12878 EVT VT = MVT::Other;
12880 // Recognize a special case where a vector is casted into wide integer to
12882 Opnds.push_back(N->getOperand(0));
12883 Opnds.push_back(N->getOperand(1));
12885 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12886 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12887 // BFS traverse all OR'd operands.
12888 if (I->getOpcode() == ISD::OR) {
12889 Opnds.push_back(I->getOperand(0));
12890 Opnds.push_back(I->getOperand(1));
12891 // Re-evaluate the number of nodes to be traversed.
12892 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12896 // Quit if a non-EXTRACT_VECTOR_ELT
12897 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12900 // Quit if without a constant index.
12901 SDValue Idx = I->getOperand(1);
12902 if (!isa<ConstantSDNode>(Idx))
12905 SDValue ExtractedFromVec = I->getOperand(0);
12906 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12907 if (M == VecInMap.end()) {
12908 VT = ExtractedFromVec.getValueType();
12909 // Quit if not 128/256-bit vector.
12910 if (!VT.is128BitVector() && !VT.is256BitVector())
12912 // Quit if not the same type.
12913 if (VecInMap.begin() != VecInMap.end() &&
12914 VT != VecInMap.begin()->first.getValueType())
12916 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12917 VecIns.push_back(ExtractedFromVec);
12919 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12922 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12923 "Not extracted from 128-/256-bit vector.");
12925 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12927 for (DenseMap<SDValue, unsigned>::const_iterator
12928 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12929 // Quit if not all elements are used.
12930 if (I->second != FullMask)
12934 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12936 // Cast all vectors into TestVT for PTEST.
12937 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12938 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
12940 // If more than one full vectors are evaluated, OR them first before PTEST.
12941 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12942 // Each iteration will OR 2 nodes and append the result until there is only
12943 // 1 node left, i.e. the final OR'd value of all vectors.
12944 SDValue LHS = VecIns[Slot];
12945 SDValue RHS = VecIns[Slot + 1];
12946 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12949 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12950 VecIns.back(), VecIns.back());
12953 /// \brief return true if \c Op has a use that doesn't just read flags.
12954 static bool hasNonFlagsUse(SDValue Op) {
12955 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12957 SDNode *User = *UI;
12958 unsigned UOpNo = UI.getOperandNo();
12959 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12960 // Look pass truncate.
12961 UOpNo = User->use_begin().getOperandNo();
12962 User = *User->use_begin();
12965 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12966 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12972 /// Emit nodes that will be selected as "test Op0,Op0", or something
12974 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12975 SelectionDAG &DAG) const {
12976 if (Op.getValueType() == MVT::i1) {
12977 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12978 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12979 DAG.getConstant(0, dl, MVT::i8));
12981 // CF and OF aren't always set the way we want. Determine which
12982 // of these we need.
12983 bool NeedCF = false;
12984 bool NeedOF = false;
12987 case X86::COND_A: case X86::COND_AE:
12988 case X86::COND_B: case X86::COND_BE:
12991 case X86::COND_G: case X86::COND_GE:
12992 case X86::COND_L: case X86::COND_LE:
12993 case X86::COND_O: case X86::COND_NO: {
12994 // Check if we really need to set the
12995 // Overflow flag. If NoSignedWrap is present
12996 // that is not actually needed.
12997 switch (Op->getOpcode()) {
13002 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13003 if (BinNode->Flags.hasNoSignedWrap())
13013 // See if we can use the EFLAGS value from the operand instead of
13014 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13015 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13016 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13017 // Emit a CMP with 0, which is the TEST pattern.
13018 //if (Op.getValueType() == MVT::i1)
13019 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13020 // DAG.getConstant(0, MVT::i1));
13021 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13022 DAG.getConstant(0, dl, Op.getValueType()));
13024 unsigned Opcode = 0;
13025 unsigned NumOperands = 0;
13027 // Truncate operations may prevent the merge of the SETCC instruction
13028 // and the arithmetic instruction before it. Attempt to truncate the operands
13029 // of the arithmetic instruction and use a reduced bit-width instruction.
13030 bool NeedTruncation = false;
13031 SDValue ArithOp = Op;
13032 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13033 SDValue Arith = Op->getOperand(0);
13034 // Both the trunc and the arithmetic op need to have one user each.
13035 if (Arith->hasOneUse())
13036 switch (Arith.getOpcode()) {
13043 NeedTruncation = true;
13049 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13050 // which may be the result of a CAST. We use the variable 'Op', which is the
13051 // non-casted variable when we check for possible users.
13052 switch (ArithOp.getOpcode()) {
13054 // Due to an isel shortcoming, be conservative if this add is likely to be
13055 // selected as part of a load-modify-store instruction. When the root node
13056 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13057 // uses of other nodes in the match, such as the ADD in this case. This
13058 // leads to the ADD being left around and reselected, with the result being
13059 // two adds in the output. Alas, even if none our users are stores, that
13060 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13061 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13062 // climbing the DAG back to the root, and it doesn't seem to be worth the
13064 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13065 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13066 if (UI->getOpcode() != ISD::CopyToReg &&
13067 UI->getOpcode() != ISD::SETCC &&
13068 UI->getOpcode() != ISD::STORE)
13071 if (ConstantSDNode *C =
13072 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13073 // An add of one will be selected as an INC.
13074 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13075 Opcode = X86ISD::INC;
13080 // An add of negative one (subtract of one) will be selected as a DEC.
13081 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13082 Opcode = X86ISD::DEC;
13088 // Otherwise use a regular EFLAGS-setting add.
13089 Opcode = X86ISD::ADD;
13094 // If we have a constant logical shift that's only used in a comparison
13095 // against zero turn it into an equivalent AND. This allows turning it into
13096 // a TEST instruction later.
13097 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13098 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13099 EVT VT = Op.getValueType();
13100 unsigned BitWidth = VT.getSizeInBits();
13101 unsigned ShAmt = Op->getConstantOperandVal(1);
13102 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13104 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13105 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13106 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13107 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13109 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13110 DAG.getConstant(Mask, dl, VT));
13111 DAG.ReplaceAllUsesWith(Op, New);
13117 // If the primary and result isn't used, don't bother using X86ISD::AND,
13118 // because a TEST instruction will be better.
13119 if (!hasNonFlagsUse(Op))
13125 // Due to the ISEL shortcoming noted above, be conservative if this op is
13126 // likely to be selected as part of a load-modify-store instruction.
13127 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13128 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13129 if (UI->getOpcode() == ISD::STORE)
13132 // Otherwise use a regular EFLAGS-setting instruction.
13133 switch (ArithOp.getOpcode()) {
13134 default: llvm_unreachable("unexpected operator!");
13135 case ISD::SUB: Opcode = X86ISD::SUB; break;
13136 case ISD::XOR: Opcode = X86ISD::XOR; break;
13137 case ISD::AND: Opcode = X86ISD::AND; break;
13139 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13140 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13141 if (EFLAGS.getNode())
13144 Opcode = X86ISD::OR;
13158 return SDValue(Op.getNode(), 1);
13164 // If we found that truncation is beneficial, perform the truncation and
13166 if (NeedTruncation) {
13167 EVT VT = Op.getValueType();
13168 SDValue WideVal = Op->getOperand(0);
13169 EVT WideVT = WideVal.getValueType();
13170 unsigned ConvertedOp = 0;
13171 // Use a target machine opcode to prevent further DAGCombine
13172 // optimizations that may separate the arithmetic operations
13173 // from the setcc node.
13174 switch (WideVal.getOpcode()) {
13176 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13177 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13178 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13179 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13180 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13184 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13185 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13186 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13187 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13188 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13194 // Emit a CMP with 0, which is the TEST pattern.
13195 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13196 DAG.getConstant(0, dl, Op.getValueType()));
13198 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13199 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13201 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13202 DAG.ReplaceAllUsesWith(Op, New);
13203 return SDValue(New.getNode(), 1);
13206 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13208 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13209 SDLoc dl, SelectionDAG &DAG) const {
13210 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13211 if (C->getAPIntValue() == 0)
13212 return EmitTest(Op0, X86CC, dl, DAG);
13214 if (Op0.getValueType() == MVT::i1)
13215 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13218 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13219 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13220 // Do the comparison at i32 if it's smaller, besides the Atom case.
13221 // This avoids subregister aliasing issues. Keep the smaller reference
13222 // if we're optimizing for size, however, as that'll allow better folding
13223 // of memory operations.
13224 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13225 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13226 !Subtarget->isAtom()) {
13227 unsigned ExtendOp =
13228 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13229 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13230 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13232 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13233 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13234 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13236 return SDValue(Sub.getNode(), 1);
13238 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13241 /// Convert a comparison if required by the subtarget.
13242 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13243 SelectionDAG &DAG) const {
13244 // If the subtarget does not support the FUCOMI instruction, floating-point
13245 // comparisons have to be converted.
13246 if (Subtarget->hasCMov() ||
13247 Cmp.getOpcode() != X86ISD::CMP ||
13248 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13249 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13252 // The instruction selector will select an FUCOM instruction instead of
13253 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13254 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13255 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13257 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13258 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13259 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13260 DAG.getConstant(8, dl, MVT::i8));
13261 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13262 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13265 /// The minimum architected relative accuracy is 2^-12. We need one
13266 /// Newton-Raphson step to have a good float result (24 bits of precision).
13267 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13268 DAGCombinerInfo &DCI,
13269 unsigned &RefinementSteps,
13270 bool &UseOneConstNR) const {
13271 EVT VT = Op.getValueType();
13272 const char *RecipOp;
13274 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13275 // TODO: Add support for AVX512 (v16f32).
13276 // It is likely not profitable to do this for f64 because a double-precision
13277 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13278 // instructions: convert to single, rsqrtss, convert back to double, refine
13279 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13280 // along with FMA, this could be a throughput win.
13281 if (VT == MVT::f32 && Subtarget->hasSSE1())
13283 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13284 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13285 RecipOp = "vec-sqrtf";
13289 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13290 if (!Recips.isEnabled(RecipOp))
13293 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13294 UseOneConstNR = false;
13295 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13298 /// The minimum architected relative accuracy is 2^-12. We need one
13299 /// Newton-Raphson step to have a good float result (24 bits of precision).
13300 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13301 DAGCombinerInfo &DCI,
13302 unsigned &RefinementSteps) const {
13303 EVT VT = Op.getValueType();
13304 const char *RecipOp;
13306 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13307 // TODO: Add support for AVX512 (v16f32).
13308 // It is likely not profitable to do this for f64 because a double-precision
13309 // reciprocal estimate with refinement on x86 prior to FMA requires
13310 // 15 instructions: convert to single, rcpss, convert back to double, refine
13311 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13312 // along with FMA, this could be a throughput win.
13313 if (VT == MVT::f32 && Subtarget->hasSSE1())
13315 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13316 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13317 RecipOp = "vec-divf";
13321 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13322 if (!Recips.isEnabled(RecipOp))
13325 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13326 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13329 /// If we have at least two divisions that use the same divisor, convert to
13330 /// multplication by a reciprocal. This may need to be adjusted for a given
13331 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13332 /// This is because we still need one division to calculate the reciprocal and
13333 /// then we need two multiplies by that reciprocal as replacements for the
13334 /// original divisions.
13335 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13339 static bool isAllOnes(SDValue V) {
13340 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13341 return C && C->isAllOnesValue();
13344 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13345 /// if it's possible.
13346 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13347 SDLoc dl, SelectionDAG &DAG) const {
13348 SDValue Op0 = And.getOperand(0);
13349 SDValue Op1 = And.getOperand(1);
13350 if (Op0.getOpcode() == ISD::TRUNCATE)
13351 Op0 = Op0.getOperand(0);
13352 if (Op1.getOpcode() == ISD::TRUNCATE)
13353 Op1 = Op1.getOperand(0);
13356 if (Op1.getOpcode() == ISD::SHL)
13357 std::swap(Op0, Op1);
13358 if (Op0.getOpcode() == ISD::SHL) {
13359 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13360 if (And00C->getZExtValue() == 1) {
13361 // If we looked past a truncate, check that it's only truncating away
13363 unsigned BitWidth = Op0.getValueSizeInBits();
13364 unsigned AndBitWidth = And.getValueSizeInBits();
13365 if (BitWidth > AndBitWidth) {
13367 DAG.computeKnownBits(Op0, Zeros, Ones);
13368 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13372 RHS = Op0.getOperand(1);
13374 } else if (Op1.getOpcode() == ISD::Constant) {
13375 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13376 uint64_t AndRHSVal = AndRHS->getZExtValue();
13377 SDValue AndLHS = Op0;
13379 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13380 LHS = AndLHS.getOperand(0);
13381 RHS = AndLHS.getOperand(1);
13384 // Use BT if the immediate can't be encoded in a TEST instruction.
13385 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13387 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13391 if (LHS.getNode()) {
13392 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13393 // instruction. Since the shift amount is in-range-or-undefined, we know
13394 // that doing a bittest on the i32 value is ok. We extend to i32 because
13395 // the encoding for the i16 version is larger than the i32 version.
13396 // Also promote i16 to i32 for performance / code size reason.
13397 if (LHS.getValueType() == MVT::i8 ||
13398 LHS.getValueType() == MVT::i16)
13399 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13401 // If the operand types disagree, extend the shift amount to match. Since
13402 // BT ignores high bits (like shifts) we can use anyextend.
13403 if (LHS.getValueType() != RHS.getValueType())
13404 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13406 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13407 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13408 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13409 DAG.getConstant(Cond, dl, MVT::i8), BT);
13415 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13417 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13422 // SSE Condition code mapping:
13431 switch (SetCCOpcode) {
13432 default: llvm_unreachable("Unexpected SETCC condition");
13434 case ISD::SETEQ: SSECC = 0; break;
13436 case ISD::SETGT: Swap = true; // Fallthrough
13438 case ISD::SETOLT: SSECC = 1; break;
13440 case ISD::SETGE: Swap = true; // Fallthrough
13442 case ISD::SETOLE: SSECC = 2; break;
13443 case ISD::SETUO: SSECC = 3; break;
13445 case ISD::SETNE: SSECC = 4; break;
13446 case ISD::SETULE: Swap = true; // Fallthrough
13447 case ISD::SETUGE: SSECC = 5; break;
13448 case ISD::SETULT: Swap = true; // Fallthrough
13449 case ISD::SETUGT: SSECC = 6; break;
13450 case ISD::SETO: SSECC = 7; break;
13452 case ISD::SETONE: SSECC = 8; break;
13455 std::swap(Op0, Op1);
13460 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13461 // ones, and then concatenate the result back.
13462 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13463 MVT VT = Op.getSimpleValueType();
13465 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13466 "Unsupported value type for operation");
13468 unsigned NumElems = VT.getVectorNumElements();
13470 SDValue CC = Op.getOperand(2);
13472 // Extract the LHS vectors
13473 SDValue LHS = Op.getOperand(0);
13474 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13475 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13477 // Extract the RHS vectors
13478 SDValue RHS = Op.getOperand(1);
13479 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13480 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13482 // Issue the operation on the smaller types and concatenate the result back
13483 MVT EltVT = VT.getVectorElementType();
13484 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13485 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13486 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13487 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13490 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13491 SDValue Op0 = Op.getOperand(0);
13492 SDValue Op1 = Op.getOperand(1);
13493 SDValue CC = Op.getOperand(2);
13494 MVT VT = Op.getSimpleValueType();
13497 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13498 "Unexpected type for boolean compare operation");
13499 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13500 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13501 DAG.getConstant(-1, dl, VT));
13502 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13503 DAG.getConstant(-1, dl, VT));
13504 switch (SetCCOpcode) {
13505 default: llvm_unreachable("Unexpected SETCC condition");
13507 // (x == y) -> ~(x ^ y)
13508 return DAG.getNode(ISD::XOR, dl, VT,
13509 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13510 DAG.getConstant(-1, dl, VT));
13512 // (x != y) -> (x ^ y)
13513 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13516 // (x > y) -> (x & ~y)
13517 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13520 // (x < y) -> (~x & y)
13521 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13524 // (x <= y) -> (~x | y)
13525 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13528 // (x >=y) -> (x | ~y)
13529 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13533 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13534 const X86Subtarget *Subtarget) {
13535 SDValue Op0 = Op.getOperand(0);
13536 SDValue Op1 = Op.getOperand(1);
13537 SDValue CC = Op.getOperand(2);
13538 MVT VT = Op.getSimpleValueType();
13541 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13542 Op.getValueType().getScalarType() == MVT::i1 &&
13543 "Cannot set masked compare for this operation");
13545 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13547 bool Unsigned = false;
13550 switch (SetCCOpcode) {
13551 default: llvm_unreachable("Unexpected SETCC condition");
13552 case ISD::SETNE: SSECC = 4; break;
13553 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13554 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13555 case ISD::SETLT: Swap = true; //fall-through
13556 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13557 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13558 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13559 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13560 case ISD::SETULE: Unsigned = true; //fall-through
13561 case ISD::SETLE: SSECC = 2; break;
13565 std::swap(Op0, Op1);
13567 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13568 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13569 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13570 DAG.getConstant(SSECC, dl, MVT::i8));
13573 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13574 /// operand \p Op1. If non-trivial (for example because it's not constant)
13575 /// return an empty value.
13576 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13578 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13582 MVT VT = Op1.getSimpleValueType();
13583 MVT EVT = VT.getVectorElementType();
13584 unsigned n = VT.getVectorNumElements();
13585 SmallVector<SDValue, 8> ULTOp1;
13587 for (unsigned i = 0; i < n; ++i) {
13588 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13589 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13592 // Avoid underflow.
13593 APInt Val = Elt->getAPIntValue();
13597 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13600 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13603 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13604 SelectionDAG &DAG) {
13605 SDValue Op0 = Op.getOperand(0);
13606 SDValue Op1 = Op.getOperand(1);
13607 SDValue CC = Op.getOperand(2);
13608 MVT VT = Op.getSimpleValueType();
13609 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13610 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13615 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13616 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13619 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13620 unsigned Opc = X86ISD::CMPP;
13621 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13622 assert(VT.getVectorNumElements() <= 16);
13623 Opc = X86ISD::CMPM;
13625 // In the two special cases we can't handle, emit two comparisons.
13628 unsigned CombineOpc;
13629 if (SetCCOpcode == ISD::SETUEQ) {
13630 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13632 assert(SetCCOpcode == ISD::SETONE);
13633 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13636 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13637 DAG.getConstant(CC0, dl, MVT::i8));
13638 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13639 DAG.getConstant(CC1, dl, MVT::i8));
13640 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13642 // Handle all other FP comparisons here.
13643 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13644 DAG.getConstant(SSECC, dl, MVT::i8));
13647 // Break 256-bit integer vector compare into smaller ones.
13648 if (VT.is256BitVector() && !Subtarget->hasInt256())
13649 return Lower256IntVSETCC(Op, DAG);
13651 EVT OpVT = Op1.getValueType();
13652 if (OpVT.getVectorElementType() == MVT::i1)
13653 return LowerBoolVSETCC_AVX512(Op, DAG);
13655 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13656 if (Subtarget->hasAVX512()) {
13657 if (Op1.getValueType().is512BitVector() ||
13658 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13659 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13660 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13662 // In AVX-512 architecture setcc returns mask with i1 elements,
13663 // But there is no compare instruction for i8 and i16 elements in KNL.
13664 // We are not talking about 512-bit operands in this case, these
13665 // types are illegal.
13667 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13668 OpVT.getVectorElementType().getSizeInBits() >= 8))
13669 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13670 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13673 // We are handling one of the integer comparisons here. Since SSE only has
13674 // GT and EQ comparisons for integer, swapping operands and multiple
13675 // operations may be required for some comparisons.
13677 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13678 bool Subus = false;
13680 switch (SetCCOpcode) {
13681 default: llvm_unreachable("Unexpected SETCC condition");
13682 case ISD::SETNE: Invert = true;
13683 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13684 case ISD::SETLT: Swap = true;
13685 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13686 case ISD::SETGE: Swap = true;
13687 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13688 Invert = true; break;
13689 case ISD::SETULT: Swap = true;
13690 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13691 FlipSigns = true; break;
13692 case ISD::SETUGE: Swap = true;
13693 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13694 FlipSigns = true; Invert = true; break;
13697 // Special case: Use min/max operations for SETULE/SETUGE
13698 MVT VET = VT.getVectorElementType();
13700 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13701 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13704 switch (SetCCOpcode) {
13706 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
13707 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
13710 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13713 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13714 if (!MinMax && hasSubus) {
13715 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13717 // t = psubus Op0, Op1
13718 // pcmpeq t, <0..0>
13719 switch (SetCCOpcode) {
13721 case ISD::SETULT: {
13722 // If the comparison is against a constant we can turn this into a
13723 // setule. With psubus, setule does not require a swap. This is
13724 // beneficial because the constant in the register is no longer
13725 // destructed as the destination so it can be hoisted out of a loop.
13726 // Only do this pre-AVX since vpcmp* is no longer destructive.
13727 if (Subtarget->hasAVX())
13729 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13730 if (ULEOp1.getNode()) {
13732 Subus = true; Invert = false; Swap = false;
13736 // Psubus is better than flip-sign because it requires no inversion.
13737 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13738 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13742 Opc = X86ISD::SUBUS;
13748 std::swap(Op0, Op1);
13750 // Check that the operation in question is available (most are plain SSE2,
13751 // but PCMPGTQ and PCMPEQQ have different requirements).
13752 if (VT == MVT::v2i64) {
13753 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13754 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13756 // First cast everything to the right type.
13757 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13758 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13760 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13761 // bits of the inputs before performing those operations. The lower
13762 // compare is always unsigned.
13765 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13767 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13768 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13769 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13770 Sign, Zero, Sign, Zero);
13772 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13773 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13775 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13776 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13777 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13779 // Create masks for only the low parts/high parts of the 64 bit integers.
13780 static const int MaskHi[] = { 1, 1, 3, 3 };
13781 static const int MaskLo[] = { 0, 0, 2, 2 };
13782 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13783 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13784 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13786 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13787 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13790 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13792 return DAG.getBitcast(VT, Result);
13795 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13796 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13797 // pcmpeqd + pshufd + pand.
13798 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13800 // First cast everything to the right type.
13801 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13802 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13805 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13807 // Make sure the lower and upper halves are both all-ones.
13808 static const int Mask[] = { 1, 0, 3, 2 };
13809 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13810 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13813 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13815 return DAG.getBitcast(VT, Result);
13819 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13820 // bits of the inputs before performing those operations.
13822 EVT EltVT = VT.getVectorElementType();
13823 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13825 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13826 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13829 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13831 // If the logical-not of the result is required, perform that now.
13833 Result = DAG.getNOT(dl, Result, VT);
13836 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13839 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13840 getZeroVector(VT, Subtarget, DAG, dl));
13845 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13847 MVT VT = Op.getSimpleValueType();
13849 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13851 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13852 && "SetCC type must be 8-bit or 1-bit integer");
13853 SDValue Op0 = Op.getOperand(0);
13854 SDValue Op1 = Op.getOperand(1);
13856 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13858 // Optimize to BT if possible.
13859 // Lower (X & (1 << N)) == 0 to BT(X, N).
13860 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13861 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13862 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13863 Op1.getOpcode() == ISD::Constant &&
13864 cast<ConstantSDNode>(Op1)->isNullValue() &&
13865 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13866 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13867 if (NewSetCC.getNode()) {
13869 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13874 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13876 if (Op1.getOpcode() == ISD::Constant &&
13877 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13878 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13879 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13881 // If the input is a setcc, then reuse the input setcc or use a new one with
13882 // the inverted condition.
13883 if (Op0.getOpcode() == X86ISD::SETCC) {
13884 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13885 bool Invert = (CC == ISD::SETNE) ^
13886 cast<ConstantSDNode>(Op1)->isNullValue();
13890 CCode = X86::GetOppositeBranchCondition(CCode);
13891 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13892 DAG.getConstant(CCode, dl, MVT::i8),
13893 Op0.getOperand(1));
13895 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13899 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13900 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13901 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13903 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13904 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13907 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13908 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13909 if (X86CC == X86::COND_INVALID)
13912 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13913 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13914 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13915 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13917 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13921 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13922 static bool isX86LogicalCmp(SDValue Op) {
13923 unsigned Opc = Op.getNode()->getOpcode();
13924 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13925 Opc == X86ISD::SAHF)
13927 if (Op.getResNo() == 1 &&
13928 (Opc == X86ISD::ADD ||
13929 Opc == X86ISD::SUB ||
13930 Opc == X86ISD::ADC ||
13931 Opc == X86ISD::SBB ||
13932 Opc == X86ISD::SMUL ||
13933 Opc == X86ISD::UMUL ||
13934 Opc == X86ISD::INC ||
13935 Opc == X86ISD::DEC ||
13936 Opc == X86ISD::OR ||
13937 Opc == X86ISD::XOR ||
13938 Opc == X86ISD::AND))
13941 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13947 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13948 if (V.getOpcode() != ISD::TRUNCATE)
13951 SDValue VOp0 = V.getOperand(0);
13952 unsigned InBits = VOp0.getValueSizeInBits();
13953 unsigned Bits = V.getValueSizeInBits();
13954 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13957 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13958 bool addTest = true;
13959 SDValue Cond = Op.getOperand(0);
13960 SDValue Op1 = Op.getOperand(1);
13961 SDValue Op2 = Op.getOperand(2);
13963 EVT VT = Op1.getValueType();
13966 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13967 // are available or VBLENDV if AVX is available.
13968 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13969 if (Cond.getOpcode() == ISD::SETCC &&
13970 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13971 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13972 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13973 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13974 int SSECC = translateX86FSETCC(
13975 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13978 if (Subtarget->hasAVX512()) {
13979 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13980 DAG.getConstant(SSECC, DL, MVT::i8));
13981 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13984 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13985 DAG.getConstant(SSECC, DL, MVT::i8));
13987 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13988 // of 3 logic instructions for size savings and potentially speed.
13989 // Unfortunately, there is no scalar form of VBLENDV.
13991 // If either operand is a constant, don't try this. We can expect to
13992 // optimize away at least one of the logic instructions later in that
13993 // case, so that sequence would be faster than a variable blend.
13995 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13996 // uses XMM0 as the selection register. That may need just as many
13997 // instructions as the AND/ANDN/OR sequence due to register moves, so
14000 if (Subtarget->hasAVX() &&
14001 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14003 // Convert to vectors, do a VSELECT, and convert back to scalar.
14004 // All of the conversions should be optimized away.
14006 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14007 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14008 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14009 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14011 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14012 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14014 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14016 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14017 VSel, DAG.getIntPtrConstant(0, DL));
14019 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14020 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14021 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14025 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14027 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14028 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14029 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14030 Op1Scalar = Op1.getOperand(0);
14032 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14033 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14034 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14035 Op2Scalar = Op2.getOperand(0);
14036 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14037 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14038 Op1Scalar.getValueType(),
14039 Cond, Op1Scalar, Op2Scalar);
14040 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14041 return DAG.getBitcast(VT, newSelect);
14042 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14043 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14044 DAG.getIntPtrConstant(0, DL));
14048 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14049 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14050 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14051 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14052 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14053 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14054 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14056 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14059 if (Cond.getOpcode() == ISD::SETCC) {
14060 SDValue NewCond = LowerSETCC(Cond, DAG);
14061 if (NewCond.getNode())
14065 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14066 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14067 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14068 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14069 if (Cond.getOpcode() == X86ISD::SETCC &&
14070 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14071 isZero(Cond.getOperand(1).getOperand(1))) {
14072 SDValue Cmp = Cond.getOperand(1);
14074 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14076 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14077 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14078 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14080 SDValue CmpOp0 = Cmp.getOperand(0);
14081 // Apply further optimizations for special cases
14082 // (select (x != 0), -1, 0) -> neg & sbb
14083 // (select (x == 0), 0, -1) -> neg & sbb
14084 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14085 if (YC->isNullValue() &&
14086 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14087 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14088 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14089 DAG.getConstant(0, DL,
14090 CmpOp0.getValueType()),
14092 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14093 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14094 SDValue(Neg.getNode(), 1));
14098 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14099 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14100 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14102 SDValue Res = // Res = 0 or -1.
14103 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14104 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14106 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14107 Res = DAG.getNOT(DL, Res, Res.getValueType());
14109 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14110 if (!N2C || !N2C->isNullValue())
14111 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14116 // Look past (and (setcc_carry (cmp ...)), 1).
14117 if (Cond.getOpcode() == ISD::AND &&
14118 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14119 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14120 if (C && C->getAPIntValue() == 1)
14121 Cond = Cond.getOperand(0);
14124 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14125 // setting operand in place of the X86ISD::SETCC.
14126 unsigned CondOpcode = Cond.getOpcode();
14127 if (CondOpcode == X86ISD::SETCC ||
14128 CondOpcode == X86ISD::SETCC_CARRY) {
14129 CC = Cond.getOperand(0);
14131 SDValue Cmp = Cond.getOperand(1);
14132 unsigned Opc = Cmp.getOpcode();
14133 MVT VT = Op.getSimpleValueType();
14135 bool IllegalFPCMov = false;
14136 if (VT.isFloatingPoint() && !VT.isVector() &&
14137 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14138 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14140 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14141 Opc == X86ISD::BT) { // FIXME
14145 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14146 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14147 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14148 Cond.getOperand(0).getValueType() != MVT::i8)) {
14149 SDValue LHS = Cond.getOperand(0);
14150 SDValue RHS = Cond.getOperand(1);
14151 unsigned X86Opcode;
14154 switch (CondOpcode) {
14155 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14156 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14157 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14158 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14159 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14160 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14161 default: llvm_unreachable("unexpected overflowing operator");
14163 if (CondOpcode == ISD::UMULO)
14164 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14167 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14169 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14171 if (CondOpcode == ISD::UMULO)
14172 Cond = X86Op.getValue(2);
14174 Cond = X86Op.getValue(1);
14176 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14181 // Look past the truncate if the high bits are known zero.
14182 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14183 Cond = Cond.getOperand(0);
14185 // We know the result of AND is compared against zero. Try to match
14187 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14188 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14189 if (NewSetCC.getNode()) {
14190 CC = NewSetCC.getOperand(0);
14191 Cond = NewSetCC.getOperand(1);
14198 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14199 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14202 // a < b ? -1 : 0 -> RES = ~setcc_carry
14203 // a < b ? 0 : -1 -> RES = setcc_carry
14204 // a >= b ? -1 : 0 -> RES = setcc_carry
14205 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14206 if (Cond.getOpcode() == X86ISD::SUB) {
14207 Cond = ConvertCmpIfNecessary(Cond, DAG);
14208 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14210 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14211 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14212 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14213 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14215 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14216 return DAG.getNOT(DL, Res, Res.getValueType());
14221 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14222 // widen the cmov and push the truncate through. This avoids introducing a new
14223 // branch during isel and doesn't add any extensions.
14224 if (Op.getValueType() == MVT::i8 &&
14225 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14226 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14227 if (T1.getValueType() == T2.getValueType() &&
14228 // Blacklist CopyFromReg to avoid partial register stalls.
14229 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14230 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14231 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14232 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14236 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14237 // condition is true.
14238 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14239 SDValue Ops[] = { Op2, Op1, CC, Cond };
14240 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14243 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14244 const X86Subtarget *Subtarget,
14245 SelectionDAG &DAG) {
14246 MVT VT = Op->getSimpleValueType(0);
14247 SDValue In = Op->getOperand(0);
14248 MVT InVT = In.getSimpleValueType();
14249 MVT VTElt = VT.getVectorElementType();
14250 MVT InVTElt = InVT.getVectorElementType();
14254 if ((InVTElt == MVT::i1) &&
14255 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14256 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14258 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14259 VTElt.getSizeInBits() <= 16)) ||
14261 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14262 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14264 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14265 VTElt.getSizeInBits() >= 32))))
14266 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14268 unsigned int NumElts = VT.getVectorNumElements();
14270 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14273 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14274 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14275 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14276 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14279 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14280 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14282 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14285 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14287 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14288 if (VT.is512BitVector())
14290 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14293 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14294 const X86Subtarget *Subtarget,
14295 SelectionDAG &DAG) {
14296 SDValue In = Op->getOperand(0);
14297 MVT VT = Op->getSimpleValueType(0);
14298 MVT InVT = In.getSimpleValueType();
14299 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14301 MVT InSVT = InVT.getScalarType();
14302 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14304 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14306 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14311 // SSE41 targets can use the pmovsx* instructions directly.
14312 if (Subtarget->hasSSE41())
14313 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14315 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14319 // As SRAI is only available on i16/i32 types, we expand only up to i32
14320 // and handle i64 separately.
14321 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14322 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14323 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14324 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14325 Curr = DAG.getBitcast(CurrVT, Curr);
14328 SDValue SignExt = Curr;
14329 if (CurrVT != InVT) {
14330 unsigned SignExtShift =
14331 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14332 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14333 DAG.getConstant(SignExtShift, dl, MVT::i8));
14339 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14340 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14341 DAG.getConstant(31, dl, MVT::i8));
14342 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14343 return DAG.getBitcast(VT, Ext);
14349 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14350 SelectionDAG &DAG) {
14351 MVT VT = Op->getSimpleValueType(0);
14352 SDValue In = Op->getOperand(0);
14353 MVT InVT = In.getSimpleValueType();
14356 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14357 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14359 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14360 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14361 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14364 if (Subtarget->hasInt256())
14365 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14367 // Optimize vectors in AVX mode
14368 // Sign extend v8i16 to v8i32 and
14371 // Divide input vector into two parts
14372 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14373 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14374 // concat the vectors to original VT
14376 unsigned NumElems = InVT.getVectorNumElements();
14377 SDValue Undef = DAG.getUNDEF(InVT);
14379 SmallVector<int,8> ShufMask1(NumElems, -1);
14380 for (unsigned i = 0; i != NumElems/2; ++i)
14383 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14385 SmallVector<int,8> ShufMask2(NumElems, -1);
14386 for (unsigned i = 0; i != NumElems/2; ++i)
14387 ShufMask2[i] = i + NumElems/2;
14389 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14391 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14392 VT.getVectorNumElements()/2);
14394 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14395 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14397 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14400 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14401 // may emit an illegal shuffle but the expansion is still better than scalar
14402 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14403 // we'll emit a shuffle and a arithmetic shift.
14404 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14405 // TODO: It is possible to support ZExt by zeroing the undef values during
14406 // the shuffle phase or after the shuffle.
14407 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14408 SelectionDAG &DAG) {
14409 MVT RegVT = Op.getSimpleValueType();
14410 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14411 assert(RegVT.isInteger() &&
14412 "We only custom lower integer vector sext loads.");
14414 // Nothing useful we can do without SSE2 shuffles.
14415 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14417 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14419 EVT MemVT = Ld->getMemoryVT();
14420 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14421 unsigned RegSz = RegVT.getSizeInBits();
14423 ISD::LoadExtType Ext = Ld->getExtensionType();
14425 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14426 && "Only anyext and sext are currently implemented.");
14427 assert(MemVT != RegVT && "Cannot extend to the same type");
14428 assert(MemVT.isVector() && "Must load a vector from memory");
14430 unsigned NumElems = RegVT.getVectorNumElements();
14431 unsigned MemSz = MemVT.getSizeInBits();
14432 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14434 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14435 // The only way in which we have a legal 256-bit vector result but not the
14436 // integer 256-bit operations needed to directly lower a sextload is if we
14437 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14438 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14439 // correctly legalized. We do this late to allow the canonical form of
14440 // sextload to persist throughout the rest of the DAG combiner -- it wants
14441 // to fold together any extensions it can, and so will fuse a sign_extend
14442 // of an sextload into a sextload targeting a wider value.
14444 if (MemSz == 128) {
14445 // Just switch this to a normal load.
14446 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14447 "it must be a legal 128-bit vector "
14449 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14450 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14451 Ld->isInvariant(), Ld->getAlignment());
14453 assert(MemSz < 128 &&
14454 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14455 // Do an sext load to a 128-bit vector type. We want to use the same
14456 // number of elements, but elements half as wide. This will end up being
14457 // recursively lowered by this routine, but will succeed as we definitely
14458 // have all the necessary features if we're using AVX1.
14460 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14461 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14463 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14464 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14465 Ld->isNonTemporal(), Ld->isInvariant(),
14466 Ld->getAlignment());
14469 // Replace chain users with the new chain.
14470 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14471 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14473 // Finally, do a normal sign-extend to the desired register.
14474 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14477 // All sizes must be a power of two.
14478 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14479 "Non-power-of-two elements are not custom lowered!");
14481 // Attempt to load the original value using scalar loads.
14482 // Find the largest scalar type that divides the total loaded size.
14483 MVT SclrLoadTy = MVT::i8;
14484 for (MVT Tp : MVT::integer_valuetypes()) {
14485 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14490 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14491 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14493 SclrLoadTy = MVT::f64;
14495 // Calculate the number of scalar loads that we need to perform
14496 // in order to load our vector from memory.
14497 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14499 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14500 "Can only lower sext loads with a single scalar load!");
14502 unsigned loadRegZize = RegSz;
14503 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14506 // Represent our vector as a sequence of elements which are the
14507 // largest scalar that we can load.
14508 EVT LoadUnitVecVT = EVT::getVectorVT(
14509 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14511 // Represent the data using the same element type that is stored in
14512 // memory. In practice, we ''widen'' MemVT.
14514 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14515 loadRegZize / MemVT.getScalarType().getSizeInBits());
14517 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14518 "Invalid vector type");
14520 // We can't shuffle using an illegal type.
14521 assert(TLI.isTypeLegal(WideVecVT) &&
14522 "We only lower types that form legal widened vector types");
14524 SmallVector<SDValue, 8> Chains;
14525 SDValue Ptr = Ld->getBasePtr();
14526 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14527 TLI.getPointerTy(DAG.getDataLayout()));
14528 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14530 for (unsigned i = 0; i < NumLoads; ++i) {
14531 // Perform a single load.
14532 SDValue ScalarLoad =
14533 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14534 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14535 Ld->getAlignment());
14536 Chains.push_back(ScalarLoad.getValue(1));
14537 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14538 // another round of DAGCombining.
14540 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14542 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14543 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14545 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14548 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14550 // Bitcast the loaded value to a vector of the original element type, in
14551 // the size of the target vector type.
14552 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14553 unsigned SizeRatio = RegSz / MemSz;
14555 if (Ext == ISD::SEXTLOAD) {
14556 // If we have SSE4.1, we can directly emit a VSEXT node.
14557 if (Subtarget->hasSSE41()) {
14558 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14559 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14563 // Otherwise we'll shuffle the small elements in the high bits of the
14564 // larger type and perform an arithmetic shift. If the shift is not legal
14565 // it's better to scalarize.
14566 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14567 "We can't implement a sext load without an arithmetic right shift!");
14569 // Redistribute the loaded elements into the different locations.
14570 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14571 for (unsigned i = 0; i != NumElems; ++i)
14572 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14574 SDValue Shuff = DAG.getVectorShuffle(
14575 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14577 Shuff = DAG.getBitcast(RegVT, Shuff);
14579 // Build the arithmetic shift.
14580 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14581 MemVT.getVectorElementType().getSizeInBits();
14583 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14584 DAG.getConstant(Amt, dl, RegVT));
14586 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14590 // Redistribute the loaded elements into the different locations.
14591 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14592 for (unsigned i = 0; i != NumElems; ++i)
14593 ShuffleVec[i * SizeRatio] = i;
14595 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14596 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14598 // Bitcast to the requested type.
14599 Shuff = DAG.getBitcast(RegVT, Shuff);
14600 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14604 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14605 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14606 // from the AND / OR.
14607 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14608 Opc = Op.getOpcode();
14609 if (Opc != ISD::OR && Opc != ISD::AND)
14611 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14612 Op.getOperand(0).hasOneUse() &&
14613 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14614 Op.getOperand(1).hasOneUse());
14617 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14618 // 1 and that the SETCC node has a single use.
14619 static bool isXor1OfSetCC(SDValue Op) {
14620 if (Op.getOpcode() != ISD::XOR)
14622 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14623 if (N1C && N1C->getAPIntValue() == 1) {
14624 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14625 Op.getOperand(0).hasOneUse();
14630 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14631 bool addTest = true;
14632 SDValue Chain = Op.getOperand(0);
14633 SDValue Cond = Op.getOperand(1);
14634 SDValue Dest = Op.getOperand(2);
14637 bool Inverted = false;
14639 if (Cond.getOpcode() == ISD::SETCC) {
14640 // Check for setcc([su]{add,sub,mul}o == 0).
14641 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14642 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14643 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14644 Cond.getOperand(0).getResNo() == 1 &&
14645 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14646 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14647 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14648 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14649 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14650 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14652 Cond = Cond.getOperand(0);
14654 SDValue NewCond = LowerSETCC(Cond, DAG);
14655 if (NewCond.getNode())
14660 // FIXME: LowerXALUO doesn't handle these!!
14661 else if (Cond.getOpcode() == X86ISD::ADD ||
14662 Cond.getOpcode() == X86ISD::SUB ||
14663 Cond.getOpcode() == X86ISD::SMUL ||
14664 Cond.getOpcode() == X86ISD::UMUL)
14665 Cond = LowerXALUO(Cond, DAG);
14668 // Look pass (and (setcc_carry (cmp ...)), 1).
14669 if (Cond.getOpcode() == ISD::AND &&
14670 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14671 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14672 if (C && C->getAPIntValue() == 1)
14673 Cond = Cond.getOperand(0);
14676 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14677 // setting operand in place of the X86ISD::SETCC.
14678 unsigned CondOpcode = Cond.getOpcode();
14679 if (CondOpcode == X86ISD::SETCC ||
14680 CondOpcode == X86ISD::SETCC_CARRY) {
14681 CC = Cond.getOperand(0);
14683 SDValue Cmp = Cond.getOperand(1);
14684 unsigned Opc = Cmp.getOpcode();
14685 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14686 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14690 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14694 // These can only come from an arithmetic instruction with overflow,
14695 // e.g. SADDO, UADDO.
14696 Cond = Cond.getNode()->getOperand(1);
14702 CondOpcode = Cond.getOpcode();
14703 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14704 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14705 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14706 Cond.getOperand(0).getValueType() != MVT::i8)) {
14707 SDValue LHS = Cond.getOperand(0);
14708 SDValue RHS = Cond.getOperand(1);
14709 unsigned X86Opcode;
14712 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14713 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14715 switch (CondOpcode) {
14716 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14718 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14720 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14723 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14724 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14726 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14728 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14731 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14732 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14733 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14734 default: llvm_unreachable("unexpected overflowing operator");
14737 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14738 if (CondOpcode == ISD::UMULO)
14739 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14742 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14744 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14746 if (CondOpcode == ISD::UMULO)
14747 Cond = X86Op.getValue(2);
14749 Cond = X86Op.getValue(1);
14751 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14755 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14756 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14757 if (CondOpc == ISD::OR) {
14758 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14759 // two branches instead of an explicit OR instruction with a
14761 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14762 isX86LogicalCmp(Cmp)) {
14763 CC = Cond.getOperand(0).getOperand(0);
14764 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14765 Chain, Dest, CC, Cmp);
14766 CC = Cond.getOperand(1).getOperand(0);
14770 } else { // ISD::AND
14771 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14772 // two branches instead of an explicit AND instruction with a
14773 // separate test. However, we only do this if this block doesn't
14774 // have a fall-through edge, because this requires an explicit
14775 // jmp when the condition is false.
14776 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14777 isX86LogicalCmp(Cmp) &&
14778 Op.getNode()->hasOneUse()) {
14779 X86::CondCode CCode =
14780 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14781 CCode = X86::GetOppositeBranchCondition(CCode);
14782 CC = DAG.getConstant(CCode, dl, MVT::i8);
14783 SDNode *User = *Op.getNode()->use_begin();
14784 // Look for an unconditional branch following this conditional branch.
14785 // We need this because we need to reverse the successors in order
14786 // to implement FCMP_OEQ.
14787 if (User->getOpcode() == ISD::BR) {
14788 SDValue FalseBB = User->getOperand(1);
14790 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14791 assert(NewBR == User);
14795 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14796 Chain, Dest, CC, Cmp);
14797 X86::CondCode CCode =
14798 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14799 CCode = X86::GetOppositeBranchCondition(CCode);
14800 CC = DAG.getConstant(CCode, dl, MVT::i8);
14806 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14807 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14808 // It should be transformed during dag combiner except when the condition
14809 // is set by a arithmetics with overflow node.
14810 X86::CondCode CCode =
14811 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14812 CCode = X86::GetOppositeBranchCondition(CCode);
14813 CC = DAG.getConstant(CCode, dl, MVT::i8);
14814 Cond = Cond.getOperand(0).getOperand(1);
14816 } else if (Cond.getOpcode() == ISD::SETCC &&
14817 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14818 // For FCMP_OEQ, we can emit
14819 // two branches instead of an explicit AND instruction with a
14820 // separate test. However, we only do this if this block doesn't
14821 // have a fall-through edge, because this requires an explicit
14822 // jmp when the condition is false.
14823 if (Op.getNode()->hasOneUse()) {
14824 SDNode *User = *Op.getNode()->use_begin();
14825 // Look for an unconditional branch following this conditional branch.
14826 // We need this because we need to reverse the successors in order
14827 // to implement FCMP_OEQ.
14828 if (User->getOpcode() == ISD::BR) {
14829 SDValue FalseBB = User->getOperand(1);
14831 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14832 assert(NewBR == User);
14836 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14837 Cond.getOperand(0), Cond.getOperand(1));
14838 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14839 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14840 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14841 Chain, Dest, CC, Cmp);
14842 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14847 } else if (Cond.getOpcode() == ISD::SETCC &&
14848 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14849 // For FCMP_UNE, we can emit
14850 // two branches instead of an explicit AND instruction with a
14851 // separate test. However, we only do this if this block doesn't
14852 // have a fall-through edge, because this requires an explicit
14853 // jmp when the condition is false.
14854 if (Op.getNode()->hasOneUse()) {
14855 SDNode *User = *Op.getNode()->use_begin();
14856 // Look for an unconditional branch following this conditional branch.
14857 // We need this because we need to reverse the successors in order
14858 // to implement FCMP_UNE.
14859 if (User->getOpcode() == ISD::BR) {
14860 SDValue FalseBB = User->getOperand(1);
14862 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14863 assert(NewBR == User);
14866 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14867 Cond.getOperand(0), Cond.getOperand(1));
14868 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14869 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14870 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14871 Chain, Dest, CC, Cmp);
14872 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14882 // Look pass the truncate if the high bits are known zero.
14883 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14884 Cond = Cond.getOperand(0);
14886 // We know the result of AND is compared against zero. Try to match
14888 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14889 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14890 if (NewSetCC.getNode()) {
14891 CC = NewSetCC.getOperand(0);
14892 Cond = NewSetCC.getOperand(1);
14899 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14900 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14901 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14903 Cond = ConvertCmpIfNecessary(Cond, DAG);
14904 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14905 Chain, Dest, CC, Cond);
14908 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14909 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14910 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14911 // that the guard pages used by the OS virtual memory manager are allocated in
14912 // correct sequence.
14914 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14915 SelectionDAG &DAG) const {
14916 MachineFunction &MF = DAG.getMachineFunction();
14917 bool SplitStack = MF.shouldSplitStack();
14918 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14923 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14924 SDNode* Node = Op.getNode();
14926 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14927 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14928 " not tell us which reg is the stack pointer!");
14929 EVT VT = Node->getValueType(0);
14930 SDValue Tmp1 = SDValue(Node, 0);
14931 SDValue Tmp2 = SDValue(Node, 1);
14932 SDValue Tmp3 = Node->getOperand(2);
14933 SDValue Chain = Tmp1.getOperand(0);
14935 // Chain the dynamic stack allocation so that it doesn't modify the stack
14936 // pointer when other instructions are using the stack.
14937 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14940 SDValue Size = Tmp2.getOperand(1);
14941 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14942 Chain = SP.getValue(1);
14943 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14944 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14945 unsigned StackAlign = TFI.getStackAlignment();
14946 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14947 if (Align > StackAlign)
14948 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14949 DAG.getConstant(-(uint64_t)Align, dl, VT));
14950 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14952 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14953 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14956 SDValue Ops[2] = { Tmp1, Tmp2 };
14957 return DAG.getMergeValues(Ops, dl);
14961 SDValue Chain = Op.getOperand(0);
14962 SDValue Size = Op.getOperand(1);
14963 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14964 EVT VT = Op.getNode()->getValueType(0);
14966 bool Is64Bit = Subtarget->is64Bit();
14967 MVT SPTy = getPointerTy(DAG.getDataLayout());
14970 MachineRegisterInfo &MRI = MF.getRegInfo();
14973 // The 64 bit implementation of segmented stacks needs to clobber both r10
14974 // r11. This makes it impossible to use it along with nested parameters.
14975 const Function *F = MF.getFunction();
14977 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14979 if (I->hasNestAttr())
14980 report_fatal_error("Cannot use segmented stacks with functions that "
14981 "have nested arguments.");
14984 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
14985 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14986 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14987 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14988 DAG.getRegister(Vreg, SPTy));
14989 SDValue Ops1[2] = { Value, Chain };
14990 return DAG.getMergeValues(Ops1, dl);
14993 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14995 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14996 Flag = Chain.getValue(1);
14997 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14999 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15001 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15002 unsigned SPReg = RegInfo->getStackRegister();
15003 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15004 Chain = SP.getValue(1);
15007 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15008 DAG.getConstant(-(uint64_t)Align, dl, VT));
15009 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15012 SDValue Ops1[2] = { SP, Chain };
15013 return DAG.getMergeValues(Ops1, dl);
15017 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15018 MachineFunction &MF = DAG.getMachineFunction();
15019 auto PtrVT = getPointerTy(MF.getDataLayout());
15020 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15022 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15025 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
15026 // vastart just stores the address of the VarArgsFrameIndex slot into the
15027 // memory location argument.
15028 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15029 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15030 MachinePointerInfo(SV), false, false, 0);
15034 // gp_offset (0 - 6 * 8)
15035 // fp_offset (48 - 48 + 8 * 16)
15036 // overflow_arg_area (point to parameters coming in memory).
15038 SmallVector<SDValue, 8> MemOps;
15039 SDValue FIN = Op.getOperand(1);
15041 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15042 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15044 FIN, MachinePointerInfo(SV), false, false, 0);
15045 MemOps.push_back(Store);
15048 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15049 Store = DAG.getStore(Op.getOperand(0), DL,
15050 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15052 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15053 MemOps.push_back(Store);
15055 // Store ptr to overflow_arg_area
15056 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15057 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15058 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15059 MachinePointerInfo(SV, 8),
15061 MemOps.push_back(Store);
15063 // Store ptr to reg_save_area.
15064 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(8, DL));
15065 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15066 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15067 MachinePointerInfo(SV, 16), false, false, 0);
15068 MemOps.push_back(Store);
15069 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15072 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15073 assert(Subtarget->is64Bit() &&
15074 "LowerVAARG only handles 64-bit va_arg!");
15075 assert((Subtarget->isTargetLinux() ||
15076 Subtarget->isTargetDarwin()) &&
15077 "Unhandled target in LowerVAARG");
15078 assert(Op.getNode()->getNumOperands() == 4);
15079 SDValue Chain = Op.getOperand(0);
15080 SDValue SrcPtr = Op.getOperand(1);
15081 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15082 unsigned Align = Op.getConstantOperandVal(3);
15085 EVT ArgVT = Op.getNode()->getValueType(0);
15086 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15087 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15090 // Decide which area this value should be read from.
15091 // TODO: Implement the AMD64 ABI in its entirety. This simple
15092 // selection mechanism works only for the basic types.
15093 if (ArgVT == MVT::f80) {
15094 llvm_unreachable("va_arg for f80 not yet implemented");
15095 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15096 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15097 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15098 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15100 llvm_unreachable("Unhandled argument type in LowerVAARG");
15103 if (ArgMode == 2) {
15104 // Sanity Check: Make sure using fp_offset makes sense.
15105 assert(!Subtarget->useSoftFloat() &&
15106 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
15107 Attribute::NoImplicitFloat)) &&
15108 Subtarget->hasSSE1());
15111 // Insert VAARG_64 node into the DAG
15112 // VAARG_64 returns two values: Variable Argument Address, Chain
15113 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15114 DAG.getConstant(ArgMode, dl, MVT::i8),
15115 DAG.getConstant(Align, dl, MVT::i32)};
15116 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15117 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15118 VTs, InstOps, MVT::i64,
15119 MachinePointerInfo(SV),
15121 /*Volatile=*/false,
15123 /*WriteMem=*/true);
15124 Chain = VAARG.getValue(1);
15126 // Load the next argument and return it
15127 return DAG.getLoad(ArgVT, dl,
15130 MachinePointerInfo(),
15131 false, false, false, 0);
15134 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15135 SelectionDAG &DAG) {
15136 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
15137 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15138 SDValue Chain = Op.getOperand(0);
15139 SDValue DstPtr = Op.getOperand(1);
15140 SDValue SrcPtr = Op.getOperand(2);
15141 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15142 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15145 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15146 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15148 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15151 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15152 // amount is a constant. Takes immediate version of shift as input.
15153 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15154 SDValue SrcOp, uint64_t ShiftAmt,
15155 SelectionDAG &DAG) {
15156 MVT ElementType = VT.getVectorElementType();
15158 // Fold this packed shift into its first operand if ShiftAmt is 0.
15162 // Check for ShiftAmt >= element width
15163 if (ShiftAmt >= ElementType.getSizeInBits()) {
15164 if (Opc == X86ISD::VSRAI)
15165 ShiftAmt = ElementType.getSizeInBits() - 1;
15167 return DAG.getConstant(0, dl, VT);
15170 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15171 && "Unknown target vector shift-by-constant node");
15173 // Fold this packed vector shift into a build vector if SrcOp is a
15174 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15175 if (VT == SrcOp.getSimpleValueType() &&
15176 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15177 SmallVector<SDValue, 8> Elts;
15178 unsigned NumElts = SrcOp->getNumOperands();
15179 ConstantSDNode *ND;
15182 default: llvm_unreachable(nullptr);
15183 case X86ISD::VSHLI:
15184 for (unsigned i=0; i!=NumElts; ++i) {
15185 SDValue CurrentOp = SrcOp->getOperand(i);
15186 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15187 Elts.push_back(CurrentOp);
15190 ND = cast<ConstantSDNode>(CurrentOp);
15191 const APInt &C = ND->getAPIntValue();
15192 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15195 case X86ISD::VSRLI:
15196 for (unsigned i=0; i!=NumElts; ++i) {
15197 SDValue CurrentOp = SrcOp->getOperand(i);
15198 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15199 Elts.push_back(CurrentOp);
15202 ND = cast<ConstantSDNode>(CurrentOp);
15203 const APInt &C = ND->getAPIntValue();
15204 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15207 case X86ISD::VSRAI:
15208 for (unsigned i=0; i!=NumElts; ++i) {
15209 SDValue CurrentOp = SrcOp->getOperand(i);
15210 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15211 Elts.push_back(CurrentOp);
15214 ND = cast<ConstantSDNode>(CurrentOp);
15215 const APInt &C = ND->getAPIntValue();
15216 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15221 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15224 return DAG.getNode(Opc, dl, VT, SrcOp,
15225 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15228 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15229 // may or may not be a constant. Takes immediate version of shift as input.
15230 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15231 SDValue SrcOp, SDValue ShAmt,
15232 SelectionDAG &DAG) {
15233 MVT SVT = ShAmt.getSimpleValueType();
15234 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15236 // Catch shift-by-constant.
15237 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15238 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15239 CShAmt->getZExtValue(), DAG);
15241 // Change opcode to non-immediate version
15243 default: llvm_unreachable("Unknown target vector shift node");
15244 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15245 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15246 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15249 const X86Subtarget &Subtarget =
15250 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15251 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15252 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15253 // Let the shuffle legalizer expand this shift amount node.
15254 SDValue Op0 = ShAmt.getOperand(0);
15255 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15256 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15258 // Need to build a vector containing shift amount.
15259 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15260 SmallVector<SDValue, 4> ShOps;
15261 ShOps.push_back(ShAmt);
15262 if (SVT == MVT::i32) {
15263 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15264 ShOps.push_back(DAG.getUNDEF(SVT));
15266 ShOps.push_back(DAG.getUNDEF(SVT));
15268 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15269 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15272 // The return type has to be a 128-bit type with the same element
15273 // type as the input type.
15274 MVT EltVT = VT.getVectorElementType();
15275 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15277 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15278 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15281 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15282 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15283 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15284 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15285 SDValue PreservedSrc,
15286 const X86Subtarget *Subtarget,
15287 SelectionDAG &DAG) {
15288 EVT VT = Op.getValueType();
15289 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15290 MVT::i1, VT.getVectorNumElements());
15291 SDValue VMask = SDValue();
15292 unsigned OpcodeSelect = ISD::VSELECT;
15295 assert(MaskVT.isSimple() && "invalid mask type");
15297 if (isAllOnes(Mask))
15300 if (MaskVT.bitsGT(Mask.getValueType())) {
15301 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15302 MaskVT.getSizeInBits());
15303 VMask = DAG.getBitcast(MaskVT,
15304 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15306 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15307 Mask.getValueType().getSizeInBits());
15308 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15309 // are extracted by EXTRACT_SUBVECTOR.
15310 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15311 DAG.getBitcast(BitcastVT, Mask),
15312 DAG.getIntPtrConstant(0, dl));
15315 switch (Op.getOpcode()) {
15317 case X86ISD::PCMPEQM:
15318 case X86ISD::PCMPGTM:
15320 case X86ISD::CMPMU:
15321 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15322 case X86ISD::VTRUNC:
15323 case X86ISD::VTRUNCS:
15324 case X86ISD::VTRUNCUS:
15325 // We can't use ISD::VSELECT here because it is not always "Legal"
15326 // for the destination type. For example vpmovqb require only AVX512
15327 // and vselect that can operate on byte element type require BWI
15328 OpcodeSelect = X86ISD::SELECT;
15331 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15332 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15333 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15336 /// \brief Creates an SDNode for a predicated scalar operation.
15337 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15338 /// The mask is coming as MVT::i8 and it should be truncated
15339 /// to MVT::i1 while lowering masking intrinsics.
15340 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15341 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15342 /// for a scalar instruction.
15343 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15344 SDValue PreservedSrc,
15345 const X86Subtarget *Subtarget,
15346 SelectionDAG &DAG) {
15347 if (isAllOnes(Mask))
15350 EVT VT = Op.getValueType();
15352 // The mask should be of type MVT::i1
15353 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15355 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15356 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15357 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15360 static int getSEHRegistrationNodeSize(const Function *Fn) {
15361 if (!Fn->hasPersonalityFn())
15362 report_fatal_error(
15363 "querying registration node size for function without personality");
15364 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15365 // WinEHStatePass for the full struct definition.
15366 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15367 case EHPersonality::MSVC_X86SEH: return 24;
15368 case EHPersonality::MSVC_CXX: return 16;
15371 report_fatal_error("can only recover FP for MSVC EH personality functions");
15374 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15375 /// function or when returning to a parent frame after catching an exception, we
15376 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15377 /// Here's the math:
15378 /// RegNodeBase = EntryEBP - RegNodeSize
15379 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15380 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15381 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15382 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15383 SDValue EntryEBP) {
15384 MachineFunction &MF = DAG.getMachineFunction();
15387 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15388 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15390 // It's possible that the parent function no longer has a personality function
15391 // if the exceptional code was optimized away, in which case we just return
15392 // the incoming EBP.
15393 if (!Fn->hasPersonalityFn())
15396 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15398 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15400 MCSymbol *OffsetSym =
15401 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15402 GlobalValue::getRealLinkageName(Fn->getName()));
15403 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15404 SDValue RegNodeFrameOffset =
15405 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15407 // RegNodeBase = EntryEBP - RegNodeSize
15408 // ParentFP = RegNodeBase - RegNodeFrameOffset
15409 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15410 DAG.getConstant(RegNodeSize, dl, PtrVT));
15411 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15414 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15415 SelectionDAG &DAG) {
15417 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15418 EVT VT = Op.getValueType();
15419 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15421 switch(IntrData->Type) {
15422 case INTR_TYPE_1OP:
15423 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15424 case INTR_TYPE_2OP:
15425 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15427 case INTR_TYPE_3OP:
15428 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15429 Op.getOperand(2), Op.getOperand(3));
15430 case INTR_TYPE_4OP:
15431 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15432 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15433 case INTR_TYPE_1OP_MASK_RM: {
15434 SDValue Src = Op.getOperand(1);
15435 SDValue PassThru = Op.getOperand(2);
15436 SDValue Mask = Op.getOperand(3);
15437 SDValue RoundingMode;
15438 // We allways add rounding mode to the Node.
15439 // If the rounding mode is not specified, we add the
15440 // "current direction" mode.
15441 if (Op.getNumOperands() == 4)
15443 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15445 RoundingMode = Op.getOperand(4);
15446 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15447 if (IntrWithRoundingModeOpcode != 0)
15448 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15449 X86::STATIC_ROUNDING::CUR_DIRECTION)
15450 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15451 dl, Op.getValueType(), Src, RoundingMode),
15452 Mask, PassThru, Subtarget, DAG);
15453 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15455 Mask, PassThru, Subtarget, DAG);
15457 case INTR_TYPE_1OP_MASK: {
15458 SDValue Src = Op.getOperand(1);
15459 SDValue PassThru = Op.getOperand(2);
15460 SDValue Mask = Op.getOperand(3);
15461 // We add rounding mode to the Node when
15462 // - RM Opcode is specified and
15463 // - RM is not "current direction".
15464 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15465 if (IntrWithRoundingModeOpcode != 0) {
15466 SDValue Rnd = Op.getOperand(4);
15467 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15468 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15469 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15470 dl, Op.getValueType(),
15472 Mask, PassThru, Subtarget, DAG);
15475 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15476 Mask, PassThru, Subtarget, DAG);
15478 case INTR_TYPE_SCALAR_MASK_RM: {
15479 SDValue Src1 = Op.getOperand(1);
15480 SDValue Src2 = Op.getOperand(2);
15481 SDValue Src0 = Op.getOperand(3);
15482 SDValue Mask = Op.getOperand(4);
15483 // There are 2 kinds of intrinsics in this group:
15484 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
15485 // (2) With rounding mode and sae - 7 operands.
15486 if (Op.getNumOperands() == 6) {
15487 SDValue Sae = Op.getOperand(5);
15488 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15489 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15491 Mask, Src0, Subtarget, DAG);
15493 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15494 SDValue RoundingMode = Op.getOperand(5);
15495 SDValue Sae = Op.getOperand(6);
15496 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15497 RoundingMode, Sae),
15498 Mask, Src0, Subtarget, DAG);
15500 case INTR_TYPE_2OP_MASK: {
15501 SDValue Src1 = Op.getOperand(1);
15502 SDValue Src2 = Op.getOperand(2);
15503 SDValue PassThru = Op.getOperand(3);
15504 SDValue Mask = Op.getOperand(4);
15505 // We specify 2 possible opcodes for intrinsics with rounding modes.
15506 // First, we check if the intrinsic may have non-default rounding mode,
15507 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15508 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15509 if (IntrWithRoundingModeOpcode != 0) {
15510 SDValue Rnd = Op.getOperand(5);
15511 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15512 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15513 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15514 dl, Op.getValueType(),
15516 Mask, PassThru, Subtarget, DAG);
15519 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15521 Mask, PassThru, Subtarget, DAG);
15523 case INTR_TYPE_2OP_MASK_RM: {
15524 SDValue Src1 = Op.getOperand(1);
15525 SDValue Src2 = Op.getOperand(2);
15526 SDValue PassThru = Op.getOperand(3);
15527 SDValue Mask = Op.getOperand(4);
15528 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15529 // First, we check if the intrinsic have rounding mode (6 operands),
15530 // if not, we set rounding mode to "current".
15532 if (Op.getNumOperands() == 6)
15533 Rnd = Op.getOperand(5);
15535 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15536 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15538 Mask, PassThru, Subtarget, DAG);
15540 case INTR_TYPE_3OP_MASK_RM: {
15541 SDValue Src1 = Op.getOperand(1);
15542 SDValue Src2 = Op.getOperand(2);
15543 SDValue Imm = Op.getOperand(3);
15544 SDValue PassThru = Op.getOperand(4);
15545 SDValue Mask = Op.getOperand(5);
15546 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15547 // First, we check if the intrinsic have rounding mode (7 operands),
15548 // if not, we set rounding mode to "current".
15550 if (Op.getNumOperands() == 7)
15551 Rnd = Op.getOperand(6);
15553 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15554 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15555 Src1, Src2, Imm, Rnd),
15556 Mask, PassThru, Subtarget, DAG);
15558 case INTR_TYPE_3OP_MASK: {
15559 SDValue Src1 = Op.getOperand(1);
15560 SDValue Src2 = Op.getOperand(2);
15561 SDValue Src3 = Op.getOperand(3);
15562 SDValue PassThru = Op.getOperand(4);
15563 SDValue Mask = Op.getOperand(5);
15564 // We specify 2 possible opcodes for intrinsics with rounding modes.
15565 // First, we check if the intrinsic may have non-default rounding mode,
15566 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15567 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15568 if (IntrWithRoundingModeOpcode != 0) {
15569 SDValue Rnd = Op.getOperand(6);
15570 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15571 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15572 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15573 dl, Op.getValueType(),
15574 Src1, Src2, Src3, Rnd),
15575 Mask, PassThru, Subtarget, DAG);
15578 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15580 Mask, PassThru, Subtarget, DAG);
15582 case VPERM_3OP_MASKZ:
15583 case VPERM_3OP_MASK:
15586 case FMA_OP_MASK: {
15587 SDValue Src1 = Op.getOperand(1);
15588 SDValue Src2 = Op.getOperand(2);
15589 SDValue Src3 = Op.getOperand(3);
15590 SDValue Mask = Op.getOperand(4);
15591 EVT VT = Op.getValueType();
15592 SDValue PassThru = SDValue();
15594 // set PassThru element
15595 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15596 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15597 else if (IntrData->Type == FMA_OP_MASK3)
15602 // We specify 2 possible opcodes for intrinsics with rounding modes.
15603 // First, we check if the intrinsic may have non-default rounding mode,
15604 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15605 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15606 if (IntrWithRoundingModeOpcode != 0) {
15607 SDValue Rnd = Op.getOperand(5);
15608 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15609 X86::STATIC_ROUNDING::CUR_DIRECTION)
15610 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15611 dl, Op.getValueType(),
15612 Src1, Src2, Src3, Rnd),
15613 Mask, PassThru, Subtarget, DAG);
15615 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15616 dl, Op.getValueType(),
15618 Mask, PassThru, Subtarget, DAG);
15621 case CMP_MASK_CC: {
15622 // Comparison intrinsics with masks.
15623 // Example of transformation:
15624 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15625 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15627 // (v8i1 (insert_subvector undef,
15628 // (v2i1 (and (PCMPEQM %a, %b),
15629 // (extract_subvector
15630 // (v8i1 (bitcast %mask)), 0))), 0))))
15631 EVT VT = Op.getOperand(1).getValueType();
15632 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15633 VT.getVectorNumElements());
15634 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15635 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15636 Mask.getValueType().getSizeInBits());
15638 if (IntrData->Type == CMP_MASK_CC) {
15639 SDValue CC = Op.getOperand(3);
15640 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15641 // We specify 2 possible opcodes for intrinsics with rounding modes.
15642 // First, we check if the intrinsic may have non-default rounding mode,
15643 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15644 if (IntrData->Opc1 != 0) {
15645 SDValue Rnd = Op.getOperand(5);
15646 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15647 X86::STATIC_ROUNDING::CUR_DIRECTION)
15648 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15649 Op.getOperand(2), CC, Rnd);
15651 //default rounding mode
15653 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15654 Op.getOperand(2), CC);
15657 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15658 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15661 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15662 DAG.getTargetConstant(0, dl,
15665 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15666 DAG.getUNDEF(BitcastVT), CmpMask,
15667 DAG.getIntPtrConstant(0, dl));
15668 return DAG.getBitcast(Op.getValueType(), Res);
15670 case COMI: { // Comparison intrinsics
15671 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15672 SDValue LHS = Op.getOperand(1);
15673 SDValue RHS = Op.getOperand(2);
15674 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15675 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15676 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15677 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15678 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15679 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15682 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15683 Op.getOperand(1), Op.getOperand(2), DAG);
15685 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15686 Op.getSimpleValueType(),
15688 Op.getOperand(2), DAG),
15689 Op.getOperand(4), Op.getOperand(3), Subtarget,
15691 case COMPRESS_EXPAND_IN_REG: {
15692 SDValue Mask = Op.getOperand(3);
15693 SDValue DataToCompress = Op.getOperand(1);
15694 SDValue PassThru = Op.getOperand(2);
15695 if (isAllOnes(Mask)) // return data as is
15696 return Op.getOperand(1);
15698 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15700 Mask, PassThru, Subtarget, DAG);
15703 SDValue Mask = Op.getOperand(3);
15704 EVT VT = Op.getValueType();
15705 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15706 VT.getVectorNumElements());
15707 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15708 Mask.getValueType().getSizeInBits());
15710 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15711 DAG.getBitcast(BitcastVT, Mask),
15712 DAG.getIntPtrConstant(0, dl));
15713 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15722 default: return SDValue(); // Don't custom lower most intrinsics.
15724 case Intrinsic::x86_avx2_permd:
15725 case Intrinsic::x86_avx2_permps:
15726 // Operands intentionally swapped. Mask is last operand to intrinsic,
15727 // but second operand for node/instruction.
15728 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15729 Op.getOperand(2), Op.getOperand(1));
15731 // ptest and testp intrinsics. The intrinsic these come from are designed to
15732 // return an integer value, not just an instruction so lower it to the ptest
15733 // or testp pattern and a setcc for the result.
15734 case Intrinsic::x86_sse41_ptestz:
15735 case Intrinsic::x86_sse41_ptestc:
15736 case Intrinsic::x86_sse41_ptestnzc:
15737 case Intrinsic::x86_avx_ptestz_256:
15738 case Intrinsic::x86_avx_ptestc_256:
15739 case Intrinsic::x86_avx_ptestnzc_256:
15740 case Intrinsic::x86_avx_vtestz_ps:
15741 case Intrinsic::x86_avx_vtestc_ps:
15742 case Intrinsic::x86_avx_vtestnzc_ps:
15743 case Intrinsic::x86_avx_vtestz_pd:
15744 case Intrinsic::x86_avx_vtestc_pd:
15745 case Intrinsic::x86_avx_vtestnzc_pd:
15746 case Intrinsic::x86_avx_vtestz_ps_256:
15747 case Intrinsic::x86_avx_vtestc_ps_256:
15748 case Intrinsic::x86_avx_vtestnzc_ps_256:
15749 case Intrinsic::x86_avx_vtestz_pd_256:
15750 case Intrinsic::x86_avx_vtestc_pd_256:
15751 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15752 bool IsTestPacked = false;
15755 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15756 case Intrinsic::x86_avx_vtestz_ps:
15757 case Intrinsic::x86_avx_vtestz_pd:
15758 case Intrinsic::x86_avx_vtestz_ps_256:
15759 case Intrinsic::x86_avx_vtestz_pd_256:
15760 IsTestPacked = true; // Fallthrough
15761 case Intrinsic::x86_sse41_ptestz:
15762 case Intrinsic::x86_avx_ptestz_256:
15764 X86CC = X86::COND_E;
15766 case Intrinsic::x86_avx_vtestc_ps:
15767 case Intrinsic::x86_avx_vtestc_pd:
15768 case Intrinsic::x86_avx_vtestc_ps_256:
15769 case Intrinsic::x86_avx_vtestc_pd_256:
15770 IsTestPacked = true; // Fallthrough
15771 case Intrinsic::x86_sse41_ptestc:
15772 case Intrinsic::x86_avx_ptestc_256:
15774 X86CC = X86::COND_B;
15776 case Intrinsic::x86_avx_vtestnzc_ps:
15777 case Intrinsic::x86_avx_vtestnzc_pd:
15778 case Intrinsic::x86_avx_vtestnzc_ps_256:
15779 case Intrinsic::x86_avx_vtestnzc_pd_256:
15780 IsTestPacked = true; // Fallthrough
15781 case Intrinsic::x86_sse41_ptestnzc:
15782 case Intrinsic::x86_avx_ptestnzc_256:
15784 X86CC = X86::COND_A;
15788 SDValue LHS = Op.getOperand(1);
15789 SDValue RHS = Op.getOperand(2);
15790 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15791 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15792 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15793 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15794 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15796 case Intrinsic::x86_avx512_kortestz_w:
15797 case Intrinsic::x86_avx512_kortestc_w: {
15798 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15799 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15800 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15801 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15802 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15803 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15804 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15807 case Intrinsic::x86_sse42_pcmpistria128:
15808 case Intrinsic::x86_sse42_pcmpestria128:
15809 case Intrinsic::x86_sse42_pcmpistric128:
15810 case Intrinsic::x86_sse42_pcmpestric128:
15811 case Intrinsic::x86_sse42_pcmpistrio128:
15812 case Intrinsic::x86_sse42_pcmpestrio128:
15813 case Intrinsic::x86_sse42_pcmpistris128:
15814 case Intrinsic::x86_sse42_pcmpestris128:
15815 case Intrinsic::x86_sse42_pcmpistriz128:
15816 case Intrinsic::x86_sse42_pcmpestriz128: {
15820 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15821 case Intrinsic::x86_sse42_pcmpistria128:
15822 Opcode = X86ISD::PCMPISTRI;
15823 X86CC = X86::COND_A;
15825 case Intrinsic::x86_sse42_pcmpestria128:
15826 Opcode = X86ISD::PCMPESTRI;
15827 X86CC = X86::COND_A;
15829 case Intrinsic::x86_sse42_pcmpistric128:
15830 Opcode = X86ISD::PCMPISTRI;
15831 X86CC = X86::COND_B;
15833 case Intrinsic::x86_sse42_pcmpestric128:
15834 Opcode = X86ISD::PCMPESTRI;
15835 X86CC = X86::COND_B;
15837 case Intrinsic::x86_sse42_pcmpistrio128:
15838 Opcode = X86ISD::PCMPISTRI;
15839 X86CC = X86::COND_O;
15841 case Intrinsic::x86_sse42_pcmpestrio128:
15842 Opcode = X86ISD::PCMPESTRI;
15843 X86CC = X86::COND_O;
15845 case Intrinsic::x86_sse42_pcmpistris128:
15846 Opcode = X86ISD::PCMPISTRI;
15847 X86CC = X86::COND_S;
15849 case Intrinsic::x86_sse42_pcmpestris128:
15850 Opcode = X86ISD::PCMPESTRI;
15851 X86CC = X86::COND_S;
15853 case Intrinsic::x86_sse42_pcmpistriz128:
15854 Opcode = X86ISD::PCMPISTRI;
15855 X86CC = X86::COND_E;
15857 case Intrinsic::x86_sse42_pcmpestriz128:
15858 Opcode = X86ISD::PCMPESTRI;
15859 X86CC = X86::COND_E;
15862 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15863 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15864 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15865 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15866 DAG.getConstant(X86CC, dl, MVT::i8),
15867 SDValue(PCMP.getNode(), 1));
15868 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15871 case Intrinsic::x86_sse42_pcmpistri128:
15872 case Intrinsic::x86_sse42_pcmpestri128: {
15874 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15875 Opcode = X86ISD::PCMPISTRI;
15877 Opcode = X86ISD::PCMPESTRI;
15879 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15880 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15881 return DAG.getNode(Opcode, dl, VTs, NewOps);
15884 case Intrinsic::x86_seh_lsda: {
15885 // Compute the symbol for the LSDA. We know it'll get emitted later.
15886 MachineFunction &MF = DAG.getMachineFunction();
15887 SDValue Op1 = Op.getOperand(1);
15888 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15889 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15890 GlobalValue::getRealLinkageName(Fn->getName()));
15892 // Generate a simple absolute symbol reference. This intrinsic is only
15893 // supported on 32-bit Windows, which isn't PIC.
15894 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
15895 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15898 case Intrinsic::x86_seh_recoverfp: {
15899 SDValue FnOp = Op.getOperand(1);
15900 SDValue IncomingFPOp = Op.getOperand(2);
15901 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
15902 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
15904 report_fatal_error(
15905 "llvm.x86.seh.recoverfp must take a function as the first argument");
15906 return recoverFramePointer(DAG, Fn, IncomingFPOp);
15909 case Intrinsic::localaddress: {
15910 // Returns one of the stack, base, or frame pointer registers, depending on
15911 // which is used to reference local variables.
15912 MachineFunction &MF = DAG.getMachineFunction();
15913 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15915 if (RegInfo->hasBasePointer(MF))
15916 Reg = RegInfo->getBaseRegister();
15917 else // This function handles the SP or FP case.
15918 Reg = RegInfo->getPtrSizedFrameRegister(MF);
15919 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
15924 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15925 SDValue Src, SDValue Mask, SDValue Base,
15926 SDValue Index, SDValue ScaleOp, SDValue Chain,
15927 const X86Subtarget * Subtarget) {
15929 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15931 llvm_unreachable("Invalid scale type");
15932 unsigned ScaleVal = C->getZExtValue();
15933 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15934 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15936 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15937 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15938 Index.getSimpleValueType().getVectorNumElements());
15940 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15942 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15944 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15945 Mask.getValueType().getSizeInBits());
15947 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15948 // are extracted by EXTRACT_SUBVECTOR.
15949 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15950 DAG.getBitcast(BitcastVT, Mask),
15951 DAG.getIntPtrConstant(0, dl));
15953 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15954 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15955 SDValue Segment = DAG.getRegister(0, MVT::i32);
15956 if (Src.getOpcode() == ISD::UNDEF)
15957 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15958 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15959 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15960 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15961 return DAG.getMergeValues(RetOps, dl);
15964 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15965 SDValue Src, SDValue Mask, SDValue Base,
15966 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15968 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15970 llvm_unreachable("Invalid scale type");
15971 unsigned ScaleVal = C->getZExtValue();
15972 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15973 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15975 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15976 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15977 SDValue Segment = DAG.getRegister(0, MVT::i32);
15978 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15979 Index.getSimpleValueType().getVectorNumElements());
15981 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15983 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15985 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15986 Mask.getValueType().getSizeInBits());
15988 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15989 // are extracted by EXTRACT_SUBVECTOR.
15990 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15991 DAG.getBitcast(BitcastVT, Mask),
15992 DAG.getIntPtrConstant(0, dl));
15994 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15995 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15996 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15997 return SDValue(Res, 1);
16000 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16001 SDValue Mask, SDValue Base, SDValue Index,
16002 SDValue ScaleOp, SDValue Chain) {
16004 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16005 assert(C && "Invalid scale type");
16006 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16007 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16008 SDValue Segment = DAG.getRegister(0, MVT::i32);
16010 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16012 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16014 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16016 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16017 //SDVTList VTs = DAG.getVTList(MVT::Other);
16018 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16019 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16020 return SDValue(Res, 0);
16023 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16024 // read performance monitor counters (x86_rdpmc).
16025 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16026 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16027 SmallVectorImpl<SDValue> &Results) {
16028 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16029 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16032 // The ECX register is used to select the index of the performance counter
16034 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16036 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16038 // Reads the content of a 64-bit performance counter and returns it in the
16039 // registers EDX:EAX.
16040 if (Subtarget->is64Bit()) {
16041 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16042 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16045 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16046 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16049 Chain = HI.getValue(1);
16051 if (Subtarget->is64Bit()) {
16052 // The EAX register is loaded with the low-order 32 bits. The EDX register
16053 // is loaded with the supported high-order bits of the counter.
16054 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16055 DAG.getConstant(32, DL, MVT::i8));
16056 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16057 Results.push_back(Chain);
16061 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16062 SDValue Ops[] = { LO, HI };
16063 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16064 Results.push_back(Pair);
16065 Results.push_back(Chain);
16068 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16069 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16070 // also used to custom lower READCYCLECOUNTER nodes.
16071 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16072 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16073 SmallVectorImpl<SDValue> &Results) {
16074 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16075 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16078 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16079 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16080 // and the EAX register is loaded with the low-order 32 bits.
16081 if (Subtarget->is64Bit()) {
16082 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16083 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16086 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16087 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16090 SDValue Chain = HI.getValue(1);
16092 if (Opcode == X86ISD::RDTSCP_DAG) {
16093 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16095 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16096 // the ECX register. Add 'ecx' explicitly to the chain.
16097 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16099 // Explicitly store the content of ECX at the location passed in input
16100 // to the 'rdtscp' intrinsic.
16101 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16102 MachinePointerInfo(), false, false, 0);
16105 if (Subtarget->is64Bit()) {
16106 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16107 // the EAX register is loaded with the low-order 32 bits.
16108 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16109 DAG.getConstant(32, DL, MVT::i8));
16110 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16111 Results.push_back(Chain);
16115 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16116 SDValue Ops[] = { LO, HI };
16117 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16118 Results.push_back(Pair);
16119 Results.push_back(Chain);
16122 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16123 SelectionDAG &DAG) {
16124 SmallVector<SDValue, 2> Results;
16126 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16128 return DAG.getMergeValues(Results, DL);
16131 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16132 SelectionDAG &DAG) {
16133 MachineFunction &MF = DAG.getMachineFunction();
16134 const Function *Fn = MF.getFunction();
16136 SDValue Chain = Op.getOperand(0);
16138 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16139 "using llvm.x86.seh.restoreframe requires a frame pointer");
16141 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16142 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16144 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16145 unsigned FrameReg =
16146 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16147 unsigned SPReg = RegInfo->getStackRegister();
16148 unsigned SlotSize = RegInfo->getSlotSize();
16150 // Get incoming EBP.
16151 SDValue IncomingEBP =
16152 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16154 // SP is saved in the first field of every registration node, so load
16155 // [EBP-RegNodeSize] into SP.
16156 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16157 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16158 DAG.getConstant(-RegNodeSize, dl, VT));
16160 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16161 false, VT.getScalarSizeInBits() / 8);
16162 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16164 if (!RegInfo->needsStackRealignment(MF)) {
16165 // Adjust EBP to point back to the original frame position.
16166 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16167 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16169 assert(RegInfo->hasBasePointer(MF) &&
16170 "functions with Win32 EH must use frame or base pointer register");
16172 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16173 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16174 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16176 // Reload the spilled EBP value, now that the stack and base pointers are
16178 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16179 X86FI->setHasSEHFramePtrSave(true);
16180 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16181 X86FI->setSEHFramePtrSaveIndex(FI);
16182 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16183 MachinePointerInfo(), false, false, false,
16184 VT.getScalarSizeInBits() / 8);
16185 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16191 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16192 /// return truncate Store/MaskedStore Node
16193 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16197 SDValue Mask = Op.getOperand(4);
16198 SDValue DataToTruncate = Op.getOperand(3);
16199 SDValue Addr = Op.getOperand(2);
16200 SDValue Chain = Op.getOperand(0);
16202 EVT VT = DataToTruncate.getValueType();
16203 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16204 ElementType, VT.getVectorNumElements());
16206 if (isAllOnes(Mask)) // return just a truncate store
16207 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16208 MachinePointerInfo(), SVT, false, false,
16209 SVT.getScalarSizeInBits()/8);
16211 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16212 MVT::i1, VT.getVectorNumElements());
16213 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16214 Mask.getValueType().getSizeInBits());
16215 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16216 // are extracted by EXTRACT_SUBVECTOR.
16217 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16218 DAG.getBitcast(BitcastVT, Mask),
16219 DAG.getIntPtrConstant(0, dl));
16221 MachineMemOperand *MMO = DAG.getMachineFunction().
16222 getMachineMemOperand(MachinePointerInfo(),
16223 MachineMemOperand::MOStore, SVT.getStoreSize(),
16224 SVT.getScalarSizeInBits()/8);
16226 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16227 VMask, SVT, MMO, true);
16230 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16231 SelectionDAG &DAG) {
16232 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16234 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16236 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16237 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16242 switch(IntrData->Type) {
16244 llvm_unreachable("Unknown Intrinsic Type");
16248 // Emit the node with the right value type.
16249 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16250 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16252 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16253 // Otherwise return the value from Rand, which is always 0, casted to i32.
16254 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16255 DAG.getConstant(1, dl, Op->getValueType(1)),
16256 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16257 SDValue(Result.getNode(), 1) };
16258 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16259 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16262 // Return { result, isValid, chain }.
16263 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16264 SDValue(Result.getNode(), 2));
16267 //gather(v1, mask, index, base, scale);
16268 SDValue Chain = Op.getOperand(0);
16269 SDValue Src = Op.getOperand(2);
16270 SDValue Base = Op.getOperand(3);
16271 SDValue Index = Op.getOperand(4);
16272 SDValue Mask = Op.getOperand(5);
16273 SDValue Scale = Op.getOperand(6);
16274 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16278 //scatter(base, mask, index, v1, scale);
16279 SDValue Chain = Op.getOperand(0);
16280 SDValue Base = Op.getOperand(2);
16281 SDValue Mask = Op.getOperand(3);
16282 SDValue Index = Op.getOperand(4);
16283 SDValue Src = Op.getOperand(5);
16284 SDValue Scale = Op.getOperand(6);
16285 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16289 SDValue Hint = Op.getOperand(6);
16290 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16291 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16292 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16293 SDValue Chain = Op.getOperand(0);
16294 SDValue Mask = Op.getOperand(2);
16295 SDValue Index = Op.getOperand(3);
16296 SDValue Base = Op.getOperand(4);
16297 SDValue Scale = Op.getOperand(5);
16298 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16300 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16302 SmallVector<SDValue, 2> Results;
16303 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16305 return DAG.getMergeValues(Results, dl);
16307 // Read Performance Monitoring Counters.
16309 SmallVector<SDValue, 2> Results;
16310 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16311 return DAG.getMergeValues(Results, dl);
16313 // XTEST intrinsics.
16315 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16316 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16317 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16318 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16320 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16321 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16322 Ret, SDValue(InTrans.getNode(), 1));
16326 SmallVector<SDValue, 2> Results;
16327 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16328 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16329 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16330 DAG.getConstant(-1, dl, MVT::i8));
16331 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16332 Op.getOperand(4), GenCF.getValue(1));
16333 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16334 Op.getOperand(5), MachinePointerInfo(),
16336 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16337 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16339 Results.push_back(SetCC);
16340 Results.push_back(Store);
16341 return DAG.getMergeValues(Results, dl);
16343 case COMPRESS_TO_MEM: {
16345 SDValue Mask = Op.getOperand(4);
16346 SDValue DataToCompress = Op.getOperand(3);
16347 SDValue Addr = Op.getOperand(2);
16348 SDValue Chain = Op.getOperand(0);
16350 EVT VT = DataToCompress.getValueType();
16351 if (isAllOnes(Mask)) // return just a store
16352 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16353 MachinePointerInfo(), false, false,
16354 VT.getScalarSizeInBits()/8);
16356 SDValue Compressed =
16357 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16358 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16359 return DAG.getStore(Chain, dl, Compressed, Addr,
16360 MachinePointerInfo(), false, false,
16361 VT.getScalarSizeInBits()/8);
16363 case TRUNCATE_TO_MEM_VI8:
16364 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16365 case TRUNCATE_TO_MEM_VI16:
16366 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16367 case TRUNCATE_TO_MEM_VI32:
16368 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16369 case EXPAND_FROM_MEM: {
16371 SDValue Mask = Op.getOperand(4);
16372 SDValue PassThru = Op.getOperand(3);
16373 SDValue Addr = Op.getOperand(2);
16374 SDValue Chain = Op.getOperand(0);
16375 EVT VT = Op.getValueType();
16377 if (isAllOnes(Mask)) // return just a load
16378 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16379 false, VT.getScalarSizeInBits()/8);
16381 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16382 false, false, false,
16383 VT.getScalarSizeInBits()/8);
16385 SDValue Results[] = {
16386 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16387 Mask, PassThru, Subtarget, DAG), Chain};
16388 return DAG.getMergeValues(Results, dl);
16393 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16394 SelectionDAG &DAG) const {
16395 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16396 MFI->setReturnAddressIsTaken(true);
16398 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16401 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16403 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16406 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16407 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16408 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16409 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16410 DAG.getNode(ISD::ADD, dl, PtrVT,
16411 FrameAddr, Offset),
16412 MachinePointerInfo(), false, false, false, 0);
16415 // Just load the return address.
16416 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16417 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16418 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16421 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16422 MachineFunction &MF = DAG.getMachineFunction();
16423 MachineFrameInfo *MFI = MF.getFrameInfo();
16424 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16425 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16426 EVT VT = Op.getValueType();
16428 MFI->setFrameAddressIsTaken(true);
16430 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16431 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16432 // is not possible to crawl up the stack without looking at the unwind codes
16434 int FrameAddrIndex = FuncInfo->getFAIndex();
16435 if (!FrameAddrIndex) {
16436 // Set up a frame object for the return address.
16437 unsigned SlotSize = RegInfo->getSlotSize();
16438 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16439 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16440 FuncInfo->setFAIndex(FrameAddrIndex);
16442 return DAG.getFrameIndex(FrameAddrIndex, VT);
16445 unsigned FrameReg =
16446 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16447 SDLoc dl(Op); // FIXME probably not meaningful
16448 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16449 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16450 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16451 "Invalid Frame Register!");
16452 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16454 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16455 MachinePointerInfo(),
16456 false, false, false, 0);
16460 // FIXME? Maybe this could be a TableGen attribute on some registers and
16461 // this table could be generated automatically from RegInfo.
16462 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16463 SelectionDAG &DAG) const {
16464 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16465 const MachineFunction &MF = DAG.getMachineFunction();
16467 unsigned Reg = StringSwitch<unsigned>(RegName)
16468 .Case("esp", X86::ESP)
16469 .Case("rsp", X86::RSP)
16470 .Case("ebp", X86::EBP)
16471 .Case("rbp", X86::RBP)
16474 if (Reg == X86::EBP || Reg == X86::RBP) {
16475 if (!TFI.hasFP(MF))
16476 report_fatal_error("register " + StringRef(RegName) +
16477 " is allocatable: function has no frame pointer");
16480 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16481 unsigned FrameReg =
16482 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16483 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16484 "Invalid Frame Register!");
16492 report_fatal_error("Invalid register name global variable");
16495 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16496 SelectionDAG &DAG) const {
16497 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16498 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16501 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16502 SDValue Chain = Op.getOperand(0);
16503 SDValue Offset = Op.getOperand(1);
16504 SDValue Handler = Op.getOperand(2);
16507 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16508 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16509 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16510 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16511 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16512 "Invalid Frame Register!");
16513 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16514 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16516 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16517 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16519 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16520 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16522 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16524 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16525 DAG.getRegister(StoreAddrReg, PtrVT));
16528 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16529 SelectionDAG &DAG) const {
16531 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16532 DAG.getVTList(MVT::i32, MVT::Other),
16533 Op.getOperand(0), Op.getOperand(1));
16536 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16537 SelectionDAG &DAG) const {
16539 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16540 Op.getOperand(0), Op.getOperand(1));
16543 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16544 return Op.getOperand(0);
16547 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16548 SelectionDAG &DAG) const {
16549 SDValue Root = Op.getOperand(0);
16550 SDValue Trmp = Op.getOperand(1); // trampoline
16551 SDValue FPtr = Op.getOperand(2); // nested function
16552 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16555 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16556 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16558 if (Subtarget->is64Bit()) {
16559 SDValue OutChains[6];
16561 // Large code-model.
16562 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16563 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16565 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16566 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16568 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16570 // Load the pointer to the nested function into R11.
16571 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16572 SDValue Addr = Trmp;
16573 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16574 Addr, MachinePointerInfo(TrmpAddr),
16577 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16578 DAG.getConstant(2, dl, MVT::i64));
16579 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16580 MachinePointerInfo(TrmpAddr, 2),
16583 // Load the 'nest' parameter value into R10.
16584 // R10 is specified in X86CallingConv.td
16585 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16586 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16587 DAG.getConstant(10, dl, MVT::i64));
16588 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16589 Addr, MachinePointerInfo(TrmpAddr, 10),
16592 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16593 DAG.getConstant(12, dl, MVT::i64));
16594 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16595 MachinePointerInfo(TrmpAddr, 12),
16598 // Jump to the nested function.
16599 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16600 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16601 DAG.getConstant(20, dl, MVT::i64));
16602 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16603 Addr, MachinePointerInfo(TrmpAddr, 20),
16606 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16607 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16608 DAG.getConstant(22, dl, MVT::i64));
16609 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16610 Addr, MachinePointerInfo(TrmpAddr, 22),
16613 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16615 const Function *Func =
16616 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16617 CallingConv::ID CC = Func->getCallingConv();
16622 llvm_unreachable("Unsupported calling convention");
16623 case CallingConv::C:
16624 case CallingConv::X86_StdCall: {
16625 // Pass 'nest' parameter in ECX.
16626 // Must be kept in sync with X86CallingConv.td
16627 NestReg = X86::ECX;
16629 // Check that ECX wasn't needed by an 'inreg' parameter.
16630 FunctionType *FTy = Func->getFunctionType();
16631 const AttributeSet &Attrs = Func->getAttributes();
16633 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16634 unsigned InRegCount = 0;
16637 for (FunctionType::param_iterator I = FTy->param_begin(),
16638 E = FTy->param_end(); I != E; ++I, ++Idx)
16639 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
16640 auto &DL = DAG.getDataLayout();
16641 // FIXME: should only count parameters that are lowered to integers.
16642 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
16645 if (InRegCount > 2) {
16646 report_fatal_error("Nest register in use - reduce number of inreg"
16652 case CallingConv::X86_FastCall:
16653 case CallingConv::X86_ThisCall:
16654 case CallingConv::Fast:
16655 // Pass 'nest' parameter in EAX.
16656 // Must be kept in sync with X86CallingConv.td
16657 NestReg = X86::EAX;
16661 SDValue OutChains[4];
16662 SDValue Addr, Disp;
16664 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16665 DAG.getConstant(10, dl, MVT::i32));
16666 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16668 // This is storing the opcode for MOV32ri.
16669 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16670 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16671 OutChains[0] = DAG.getStore(Root, dl,
16672 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16673 Trmp, MachinePointerInfo(TrmpAddr),
16676 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16677 DAG.getConstant(1, dl, MVT::i32));
16678 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16679 MachinePointerInfo(TrmpAddr, 1),
16682 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16683 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16684 DAG.getConstant(5, dl, MVT::i32));
16685 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16686 Addr, MachinePointerInfo(TrmpAddr, 5),
16689 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16690 DAG.getConstant(6, dl, MVT::i32));
16691 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16692 MachinePointerInfo(TrmpAddr, 6),
16695 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16699 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16700 SelectionDAG &DAG) const {
16702 The rounding mode is in bits 11:10 of FPSR, and has the following
16704 00 Round to nearest
16709 FLT_ROUNDS, on the other hand, expects the following:
16716 To perform the conversion, we do:
16717 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16720 MachineFunction &MF = DAG.getMachineFunction();
16721 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16722 unsigned StackAlignment = TFI.getStackAlignment();
16723 MVT VT = Op.getSimpleValueType();
16726 // Save FP Control Word to stack slot
16727 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16728 SDValue StackSlot =
16729 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
16731 MachineMemOperand *MMO =
16732 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
16733 MachineMemOperand::MOStore, 2, 2);
16735 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16736 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16737 DAG.getVTList(MVT::Other),
16738 Ops, MVT::i16, MMO);
16740 // Load FP Control Word from stack slot
16741 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16742 MachinePointerInfo(), false, false, false, 0);
16744 // Transform as necessary
16746 DAG.getNode(ISD::SRL, DL, MVT::i16,
16747 DAG.getNode(ISD::AND, DL, MVT::i16,
16748 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16749 DAG.getConstant(11, DL, MVT::i8));
16751 DAG.getNode(ISD::SRL, DL, MVT::i16,
16752 DAG.getNode(ISD::AND, DL, MVT::i16,
16753 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16754 DAG.getConstant(9, DL, MVT::i8));
16757 DAG.getNode(ISD::AND, DL, MVT::i16,
16758 DAG.getNode(ISD::ADD, DL, MVT::i16,
16759 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16760 DAG.getConstant(1, DL, MVT::i16)),
16761 DAG.getConstant(3, DL, MVT::i16));
16763 return DAG.getNode((VT.getSizeInBits() < 16 ?
16764 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16767 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16768 MVT VT = Op.getSimpleValueType();
16770 unsigned NumBits = VT.getSizeInBits();
16773 Op = Op.getOperand(0);
16774 if (VT == MVT::i8) {
16775 // Zero extend to i32 since there is not an i8 bsr.
16777 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16780 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16781 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16782 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16784 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16787 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16788 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16791 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16793 // Finally xor with NumBits-1.
16794 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16795 DAG.getConstant(NumBits - 1, dl, OpVT));
16798 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16802 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16803 MVT VT = Op.getSimpleValueType();
16805 unsigned NumBits = VT.getSizeInBits();
16808 Op = Op.getOperand(0);
16809 if (VT == MVT::i8) {
16810 // Zero extend to i32 since there is not an i8 bsr.
16812 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16815 // Issue a bsr (scan bits in reverse).
16816 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16817 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16819 // And xor with NumBits-1.
16820 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16821 DAG.getConstant(NumBits - 1, dl, OpVT));
16824 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16828 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16829 MVT VT = Op.getSimpleValueType();
16830 unsigned NumBits = VT.getSizeInBits();
16832 Op = Op.getOperand(0);
16834 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16835 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16836 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16838 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16841 DAG.getConstant(NumBits, dl, VT),
16842 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16845 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16848 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16849 // ones, and then concatenate the result back.
16850 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16851 MVT VT = Op.getSimpleValueType();
16853 assert(VT.is256BitVector() && VT.isInteger() &&
16854 "Unsupported value type for operation");
16856 unsigned NumElems = VT.getVectorNumElements();
16859 // Extract the LHS vectors
16860 SDValue LHS = Op.getOperand(0);
16861 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16862 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16864 // Extract the RHS vectors
16865 SDValue RHS = Op.getOperand(1);
16866 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16867 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16869 MVT EltVT = VT.getVectorElementType();
16870 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16872 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16873 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16874 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16877 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16878 if (Op.getValueType() == MVT::i1)
16879 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16880 Op.getOperand(0), Op.getOperand(1));
16881 assert(Op.getSimpleValueType().is256BitVector() &&
16882 Op.getSimpleValueType().isInteger() &&
16883 "Only handle AVX 256-bit vector integer operation");
16884 return Lower256IntArith(Op, DAG);
16887 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16888 if (Op.getValueType() == MVT::i1)
16889 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16890 Op.getOperand(0), Op.getOperand(1));
16891 assert(Op.getSimpleValueType().is256BitVector() &&
16892 Op.getSimpleValueType().isInteger() &&
16893 "Only handle AVX 256-bit vector integer operation");
16894 return Lower256IntArith(Op, DAG);
16897 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
16898 assert(Op.getSimpleValueType().is256BitVector() &&
16899 Op.getSimpleValueType().isInteger() &&
16900 "Only handle AVX 256-bit vector integer operation");
16901 return Lower256IntArith(Op, DAG);
16904 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16905 SelectionDAG &DAG) {
16907 MVT VT = Op.getSimpleValueType();
16910 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
16912 // Decompose 256-bit ops into smaller 128-bit ops.
16913 if (VT.is256BitVector() && !Subtarget->hasInt256())
16914 return Lower256IntArith(Op, DAG);
16916 SDValue A = Op.getOperand(0);
16917 SDValue B = Op.getOperand(1);
16919 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16920 // pairs, multiply and truncate.
16921 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16922 if (Subtarget->hasInt256()) {
16923 if (VT == MVT::v32i8) {
16924 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16925 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16926 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16927 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16928 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16929 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16930 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16931 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16932 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16933 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16936 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16937 return DAG.getNode(
16938 ISD::TRUNCATE, dl, VT,
16939 DAG.getNode(ISD::MUL, dl, ExVT,
16940 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16941 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16944 assert(VT == MVT::v16i8 &&
16945 "Pre-AVX2 support only supports v16i8 multiplication");
16946 MVT ExVT = MVT::v8i16;
16948 // Extract the lo parts and sign extend to i16
16950 if (Subtarget->hasSSE41()) {
16951 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16952 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16954 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16955 -1, 4, -1, 5, -1, 6, -1, 7};
16956 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16957 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16958 ALo = DAG.getBitcast(ExVT, ALo);
16959 BLo = DAG.getBitcast(ExVT, BLo);
16960 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16961 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16964 // Extract the hi parts and sign extend to i16
16966 if (Subtarget->hasSSE41()) {
16967 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16968 -1, -1, -1, -1, -1, -1, -1, -1};
16969 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16970 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16971 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16972 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16974 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16975 -1, 12, -1, 13, -1, 14, -1, 15};
16976 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16977 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16978 AHi = DAG.getBitcast(ExVT, AHi);
16979 BHi = DAG.getBitcast(ExVT, BHi);
16980 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16981 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16984 // Multiply, mask the lower 8bits of the lo/hi results and pack
16985 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16986 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16987 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16988 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16989 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16992 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16993 if (VT == MVT::v4i32) {
16994 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16995 "Should not custom lower when pmuldq is available!");
16997 // Extract the odd parts.
16998 static const int UnpackMask[] = { 1, -1, 3, -1 };
16999 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17000 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17002 // Multiply the even parts.
17003 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17004 // Now multiply odd parts.
17005 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17007 Evens = DAG.getBitcast(VT, Evens);
17008 Odds = DAG.getBitcast(VT, Odds);
17010 // Merge the two vectors back together with a shuffle. This expands into 2
17012 static const int ShufMask[] = { 0, 4, 2, 6 };
17013 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17016 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17017 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17019 // Ahi = psrlqi(a, 32);
17020 // Bhi = psrlqi(b, 32);
17022 // AloBlo = pmuludq(a, b);
17023 // AloBhi = pmuludq(a, Bhi);
17024 // AhiBlo = pmuludq(Ahi, b);
17026 // AloBhi = psllqi(AloBhi, 32);
17027 // AhiBlo = psllqi(AhiBlo, 32);
17028 // return AloBlo + AloBhi + AhiBlo;
17030 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17031 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17033 SDValue AhiBlo = Ahi;
17034 SDValue AloBhi = Bhi;
17035 // Bit cast to 32-bit vectors for MULUDQ
17036 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17037 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17038 A = DAG.getBitcast(MulVT, A);
17039 B = DAG.getBitcast(MulVT, B);
17040 Ahi = DAG.getBitcast(MulVT, Ahi);
17041 Bhi = DAG.getBitcast(MulVT, Bhi);
17043 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17044 // After shifting right const values the result may be all-zero.
17045 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17046 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17047 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17049 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17050 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17051 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17054 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17055 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17058 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17059 assert(Subtarget->isTargetWin64() && "Unexpected target");
17060 EVT VT = Op.getValueType();
17061 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17062 "Unexpected return type for lowering");
17066 switch (Op->getOpcode()) {
17067 default: llvm_unreachable("Unexpected request for libcall!");
17068 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17069 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17070 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17071 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17072 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17073 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17077 SDValue InChain = DAG.getEntryNode();
17079 TargetLowering::ArgListTy Args;
17080 TargetLowering::ArgListEntry Entry;
17081 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17082 EVT ArgVT = Op->getOperand(i).getValueType();
17083 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17084 "Unexpected argument type for lowering");
17085 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17086 Entry.Node = StackPtr;
17087 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17089 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17090 Entry.Ty = PointerType::get(ArgTy,0);
17091 Entry.isSExt = false;
17092 Entry.isZExt = false;
17093 Args.push_back(Entry);
17096 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17097 getPointerTy(DAG.getDataLayout()));
17099 TargetLowering::CallLoweringInfo CLI(DAG);
17100 CLI.setDebugLoc(dl).setChain(InChain)
17101 .setCallee(getLibcallCallingConv(LC),
17102 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17103 Callee, std::move(Args), 0)
17104 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17106 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17107 return DAG.getBitcast(VT, CallInfo.first);
17110 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17111 SelectionDAG &DAG) {
17112 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17113 EVT VT = Op0.getValueType();
17116 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17117 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17119 // PMULxD operations multiply each even value (starting at 0) of LHS with
17120 // the related value of RHS and produce a widen result.
17121 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17122 // => <2 x i64> <ae|cg>
17124 // In other word, to have all the results, we need to perform two PMULxD:
17125 // 1. one with the even values.
17126 // 2. one with the odd values.
17127 // To achieve #2, with need to place the odd values at an even position.
17129 // Place the odd value at an even position (basically, shift all values 1
17130 // step to the left):
17131 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17132 // <a|b|c|d> => <b|undef|d|undef>
17133 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17134 // <e|f|g|h> => <f|undef|h|undef>
17135 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17137 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17139 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17140 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17142 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17143 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17144 // => <2 x i64> <ae|cg>
17145 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17146 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17147 // => <2 x i64> <bf|dh>
17148 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17150 // Shuffle it back into the right order.
17151 SDValue Highs, Lows;
17152 if (VT == MVT::v8i32) {
17153 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17154 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17155 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17156 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17158 const int HighMask[] = {1, 5, 3, 7};
17159 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17160 const int LowMask[] = {0, 4, 2, 6};
17161 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17164 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17165 // unsigned multiply.
17166 if (IsSigned && !Subtarget->hasSSE41()) {
17167 SDValue ShAmt = DAG.getConstant(
17169 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17170 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17171 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17172 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17173 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17175 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17176 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17179 // The first result of MUL_LOHI is actually the low value, followed by the
17181 SDValue Ops[] = {Lows, Highs};
17182 return DAG.getMergeValues(Ops, dl);
17185 // Return true if the required (according to Opcode) shift-imm form is natively
17186 // supported by the Subtarget
17187 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17189 if (VT.getScalarSizeInBits() < 16)
17192 if (VT.is512BitVector() &&
17193 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17196 bool LShift = VT.is128BitVector() ||
17197 (VT.is256BitVector() && Subtarget->hasInt256());
17199 bool AShift = LShift && (Subtarget->hasVLX() ||
17200 (VT != MVT::v2i64 && VT != MVT::v4i64));
17201 return (Opcode == ISD::SRA) ? AShift : LShift;
17204 // The shift amount is a variable, but it is the same for all vector lanes.
17205 // These instructions are defined together with shift-immediate.
17207 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17209 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17212 // Return true if the required (according to Opcode) variable-shift form is
17213 // natively supported by the Subtarget
17214 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17217 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17220 // vXi16 supported only on AVX-512, BWI
17221 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17224 if (VT.is512BitVector() || Subtarget->hasVLX())
17227 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17228 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17229 return (Opcode == ISD::SRA) ? AShift : LShift;
17232 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17233 const X86Subtarget *Subtarget) {
17234 MVT VT = Op.getSimpleValueType();
17236 SDValue R = Op.getOperand(0);
17237 SDValue Amt = Op.getOperand(1);
17239 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17240 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17242 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17243 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17244 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17245 SDValue Ex = DAG.getBitcast(ExVT, R);
17247 if (ShiftAmt >= 32) {
17248 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17250 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17251 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17252 ShiftAmt - 32, DAG);
17253 if (VT == MVT::v2i64)
17254 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17255 if (VT == MVT::v4i64)
17256 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17257 {9, 1, 11, 3, 13, 5, 15, 7});
17259 // SRA upper i32, SHL whole i64 and select lower i32.
17260 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17263 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17264 Lower = DAG.getBitcast(ExVT, Lower);
17265 if (VT == MVT::v2i64)
17266 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17267 if (VT == MVT::v4i64)
17268 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17269 {8, 1, 10, 3, 12, 5, 14, 7});
17271 return DAG.getBitcast(VT, Ex);
17274 // Optimize shl/srl/sra with constant shift amount.
17275 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17276 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17277 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17279 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17280 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17282 // i64 SRA needs to be performed as partial shifts.
17283 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17284 Op.getOpcode() == ISD::SRA)
17285 return ArithmeticShiftRight64(ShiftAmt);
17287 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17288 unsigned NumElts = VT.getVectorNumElements();
17289 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17291 if (Op.getOpcode() == ISD::SHL) {
17292 // Simple i8 add case
17294 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17296 // Make a large shift.
17297 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17299 SHL = DAG.getBitcast(VT, SHL);
17300 // Zero out the rightmost bits.
17301 SmallVector<SDValue, 32> V(
17302 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17303 return DAG.getNode(ISD::AND, dl, VT, SHL,
17304 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17306 if (Op.getOpcode() == ISD::SRL) {
17307 // Make a large shift.
17308 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17310 SRL = DAG.getBitcast(VT, SRL);
17311 // Zero out the leftmost bits.
17312 SmallVector<SDValue, 32> V(
17313 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17314 return DAG.getNode(ISD::AND, dl, VT, SRL,
17315 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17317 if (Op.getOpcode() == ISD::SRA) {
17318 if (ShiftAmt == 7) {
17319 // ashr(R, 7) === cmp_slt(R, 0)
17320 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17321 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17324 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
17325 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17326 SmallVector<SDValue, 32> V(NumElts,
17327 DAG.getConstant(128 >> ShiftAmt, dl,
17329 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17330 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17331 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17334 llvm_unreachable("Unknown shift opcode.");
17339 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17340 if (!Subtarget->is64Bit() &&
17341 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
17343 // Peek through any splat that was introduced for i64 shift vectorization.
17344 int SplatIndex = -1;
17345 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
17346 if (SVN->isSplat()) {
17347 SplatIndex = SVN->getSplatIndex();
17348 Amt = Amt.getOperand(0);
17349 assert(SplatIndex < (int)VT.getVectorNumElements() &&
17350 "Splat shuffle referencing second operand");
17353 if (Amt.getOpcode() != ISD::BITCAST ||
17354 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
17357 Amt = Amt.getOperand(0);
17358 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17359 VT.getVectorNumElements();
17360 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17361 uint64_t ShiftAmt = 0;
17362 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
17363 for (unsigned i = 0; i != Ratio; ++i) {
17364 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
17368 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17371 // Check remaining shift amounts (if not a splat).
17372 if (SplatIndex < 0) {
17373 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17374 uint64_t ShAmt = 0;
17375 for (unsigned j = 0; j != Ratio; ++j) {
17376 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17380 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17382 if (ShAmt != ShiftAmt)
17387 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17388 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17390 if (Op.getOpcode() == ISD::SRA)
17391 return ArithmeticShiftRight64(ShiftAmt);
17397 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17398 const X86Subtarget* Subtarget) {
17399 MVT VT = Op.getSimpleValueType();
17401 SDValue R = Op.getOperand(0);
17402 SDValue Amt = Op.getOperand(1);
17404 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17405 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17407 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17408 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17410 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17412 EVT EltVT = VT.getVectorElementType();
17414 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17415 // Check if this build_vector node is doing a splat.
17416 // If so, then set BaseShAmt equal to the splat value.
17417 BaseShAmt = BV->getSplatValue();
17418 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17419 BaseShAmt = SDValue();
17421 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17422 Amt = Amt.getOperand(0);
17424 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17425 if (SVN && SVN->isSplat()) {
17426 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17427 SDValue InVec = Amt.getOperand(0);
17428 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17429 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17430 "Unexpected shuffle index found!");
17431 BaseShAmt = InVec.getOperand(SplatIdx);
17432 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17433 if (ConstantSDNode *C =
17434 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17435 if (C->getZExtValue() == SplatIdx)
17436 BaseShAmt = InVec.getOperand(1);
17441 // Avoid introducing an extract element from a shuffle.
17442 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17443 DAG.getIntPtrConstant(SplatIdx, dl));
17447 if (BaseShAmt.getNode()) {
17448 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17449 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17450 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17451 else if (EltVT.bitsLT(MVT::i32))
17452 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17454 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17458 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17459 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17460 Amt.getOpcode() == ISD::BITCAST &&
17461 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17462 Amt = Amt.getOperand(0);
17463 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17464 VT.getVectorNumElements();
17465 std::vector<SDValue> Vals(Ratio);
17466 for (unsigned i = 0; i != Ratio; ++i)
17467 Vals[i] = Amt.getOperand(i);
17468 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17469 for (unsigned j = 0; j != Ratio; ++j)
17470 if (Vals[j] != Amt.getOperand(i + j))
17474 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17475 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17480 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17481 SelectionDAG &DAG) {
17482 MVT VT = Op.getSimpleValueType();
17484 SDValue R = Op.getOperand(0);
17485 SDValue Amt = Op.getOperand(1);
17487 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17488 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17490 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17493 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17496 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17499 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17500 // shifts per-lane and then shuffle the partial results back together.
17501 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17502 // Splat the shift amounts so the scalar shifts above will catch it.
17503 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17504 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17505 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17506 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17507 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17510 // i64 vector arithmetic shift can be emulated with the transform:
17511 // M = lshr(SIGN_BIT, Amt)
17512 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
17513 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
17514 Op.getOpcode() == ISD::SRA) {
17515 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
17516 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
17517 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17518 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
17519 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
17523 // If possible, lower this packed shift into a vector multiply instead of
17524 // expanding it into a sequence of scalar shifts.
17525 // Do this only if the vector shift count is a constant build_vector.
17526 if (Op.getOpcode() == ISD::SHL &&
17527 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17528 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17529 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17530 SmallVector<SDValue, 8> Elts;
17531 EVT SVT = VT.getScalarType();
17532 unsigned SVTBits = SVT.getSizeInBits();
17533 const APInt &One = APInt(SVTBits, 1);
17534 unsigned NumElems = VT.getVectorNumElements();
17536 for (unsigned i=0; i !=NumElems; ++i) {
17537 SDValue Op = Amt->getOperand(i);
17538 if (Op->getOpcode() == ISD::UNDEF) {
17539 Elts.push_back(Op);
17543 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17544 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17545 uint64_t ShAmt = C.getZExtValue();
17546 if (ShAmt >= SVTBits) {
17547 Elts.push_back(DAG.getUNDEF(SVT));
17550 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17552 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17553 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17556 // Lower SHL with variable shift amount.
17557 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17558 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17560 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17561 DAG.getConstant(0x3f800000U, dl, VT));
17562 Op = DAG.getBitcast(MVT::v4f32, Op);
17563 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17564 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17567 // If possible, lower this shift as a sequence of two shifts by
17568 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17570 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17572 // Could be rewritten as:
17573 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17575 // The advantage is that the two shifts from the example would be
17576 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17577 // the vector shift into four scalar shifts plus four pairs of vector
17579 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17580 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17581 unsigned TargetOpcode = X86ISD::MOVSS;
17582 bool CanBeSimplified;
17583 // The splat value for the first packed shift (the 'X' from the example).
17584 SDValue Amt1 = Amt->getOperand(0);
17585 // The splat value for the second packed shift (the 'Y' from the example).
17586 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17587 Amt->getOperand(2);
17589 // See if it is possible to replace this node with a sequence of
17590 // two shifts followed by a MOVSS/MOVSD
17591 if (VT == MVT::v4i32) {
17592 // Check if it is legal to use a MOVSS.
17593 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17594 Amt2 == Amt->getOperand(3);
17595 if (!CanBeSimplified) {
17596 // Otherwise, check if we can still simplify this node using a MOVSD.
17597 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17598 Amt->getOperand(2) == Amt->getOperand(3);
17599 TargetOpcode = X86ISD::MOVSD;
17600 Amt2 = Amt->getOperand(2);
17603 // Do similar checks for the case where the machine value type
17605 CanBeSimplified = Amt1 == Amt->getOperand(1);
17606 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17607 CanBeSimplified = Amt2 == Amt->getOperand(i);
17609 if (!CanBeSimplified) {
17610 TargetOpcode = X86ISD::MOVSD;
17611 CanBeSimplified = true;
17612 Amt2 = Amt->getOperand(4);
17613 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17614 CanBeSimplified = Amt1 == Amt->getOperand(i);
17615 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17616 CanBeSimplified = Amt2 == Amt->getOperand(j);
17620 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17621 isa<ConstantSDNode>(Amt2)) {
17622 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17623 EVT CastVT = MVT::v4i32;
17625 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
17626 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17628 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
17629 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17630 if (TargetOpcode == X86ISD::MOVSD)
17631 CastVT = MVT::v2i64;
17632 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
17633 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
17634 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17636 return DAG.getBitcast(VT, Result);
17640 // v4i32 Non Uniform Shifts.
17641 // If the shift amount is constant we can shift each lane using the SSE2
17642 // immediate shifts, else we need to zero-extend each lane to the lower i64
17643 // and shift using the SSE2 variable shifts.
17644 // The separate results can then be blended together.
17645 if (VT == MVT::v4i32) {
17646 unsigned Opc = Op.getOpcode();
17647 SDValue Amt0, Amt1, Amt2, Amt3;
17648 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17649 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
17650 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
17651 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
17652 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
17654 // ISD::SHL is handled above but we include it here for completeness.
17657 llvm_unreachable("Unknown target vector shift node");
17659 Opc = X86ISD::VSHL;
17662 Opc = X86ISD::VSRL;
17665 Opc = X86ISD::VSRA;
17668 // The SSE2 shifts use the lower i64 as the same shift amount for
17669 // all lanes and the upper i64 is ignored. These shuffle masks
17670 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
17671 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17672 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
17673 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
17674 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
17675 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
17678 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
17679 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
17680 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
17681 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
17682 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
17683 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
17684 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
17687 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
17688 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
17689 unsigned ShiftOpcode = Op->getOpcode();
17691 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
17692 // On SSE41 targets we make use of the fact that VSELECT lowers
17693 // to PBLENDVB which selects bytes based just on the sign bit.
17694 if (Subtarget->hasSSE41()) {
17695 V0 = DAG.getBitcast(VT, V0);
17696 V1 = DAG.getBitcast(VT, V1);
17697 Sel = DAG.getBitcast(VT, Sel);
17698 return DAG.getBitcast(SelVT,
17699 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
17701 // On pre-SSE41 targets we test for the sign bit by comparing to
17702 // zero - a negative value will set all bits of the lanes to true
17703 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
17704 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
17705 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
17706 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
17709 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
17710 // We can safely do this using i16 shifts as we're only interested in
17711 // the 3 lower bits of each byte.
17712 Amt = DAG.getBitcast(ExtVT, Amt);
17713 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
17714 Amt = DAG.getBitcast(VT, Amt);
17716 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
17717 // r = VSELECT(r, shift(r, 4), a);
17719 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17720 R = SignBitSelect(VT, Amt, M, R);
17723 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17725 // r = VSELECT(r, shift(r, 2), a);
17726 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17727 R = SignBitSelect(VT, Amt, M, R);
17730 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17732 // return VSELECT(r, shift(r, 1), a);
17733 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17734 R = SignBitSelect(VT, Amt, M, R);
17738 if (Op->getOpcode() == ISD::SRA) {
17739 // For SRA we need to unpack each byte to the higher byte of a i16 vector
17740 // so we can correctly sign extend. We don't care what happens to the
17742 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
17743 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
17744 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
17745 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
17746 ALo = DAG.getBitcast(ExtVT, ALo);
17747 AHi = DAG.getBitcast(ExtVT, AHi);
17748 RLo = DAG.getBitcast(ExtVT, RLo);
17749 RHi = DAG.getBitcast(ExtVT, RHi);
17751 // r = VSELECT(r, shift(r, 4), a);
17752 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17753 DAG.getConstant(4, dl, ExtVT));
17754 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17755 DAG.getConstant(4, dl, ExtVT));
17756 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17757 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17760 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17761 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17763 // r = VSELECT(r, shift(r, 2), a);
17764 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17765 DAG.getConstant(2, dl, ExtVT));
17766 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17767 DAG.getConstant(2, dl, ExtVT));
17768 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17769 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17772 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17773 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17775 // r = VSELECT(r, shift(r, 1), a);
17776 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17777 DAG.getConstant(1, dl, ExtVT));
17778 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17779 DAG.getConstant(1, dl, ExtVT));
17780 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17781 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17783 // Logical shift the result back to the lower byte, leaving a zero upper
17785 // meaning that we can safely pack with PACKUSWB.
17787 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
17789 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
17790 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17794 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17795 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17796 // solution better.
17797 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17798 MVT ExtVT = MVT::v8i32;
17800 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17801 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
17802 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
17803 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17804 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
17807 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
17808 MVT ExtVT = MVT::v8i32;
17809 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17810 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
17811 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
17812 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
17813 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
17814 ALo = DAG.getBitcast(ExtVT, ALo);
17815 AHi = DAG.getBitcast(ExtVT, AHi);
17816 RLo = DAG.getBitcast(ExtVT, RLo);
17817 RHi = DAG.getBitcast(ExtVT, RHi);
17818 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
17819 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
17820 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
17821 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
17822 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
17825 if (VT == MVT::v8i16) {
17826 unsigned ShiftOpcode = Op->getOpcode();
17828 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
17829 // On SSE41 targets we make use of the fact that VSELECT lowers
17830 // to PBLENDVB which selects bytes based just on the sign bit.
17831 if (Subtarget->hasSSE41()) {
17832 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
17833 V0 = DAG.getBitcast(ExtVT, V0);
17834 V1 = DAG.getBitcast(ExtVT, V1);
17835 Sel = DAG.getBitcast(ExtVT, Sel);
17836 return DAG.getBitcast(
17837 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
17839 // On pre-SSE41 targets we splat the sign bit - a negative value will
17840 // set all bits of the lanes to true and VSELECT uses that in
17841 // its OR(AND(V0,C),AND(V1,~C)) lowering.
17843 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
17844 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
17847 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
17848 if (Subtarget->hasSSE41()) {
17849 // On SSE41 targets we need to replicate the shift mask in both
17850 // bytes for PBLENDVB.
17853 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
17854 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
17856 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
17859 // r = VSELECT(r, shift(r, 8), a);
17860 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
17861 R = SignBitSelect(Amt, M, R);
17864 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17866 // r = VSELECT(r, shift(r, 4), a);
17867 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17868 R = SignBitSelect(Amt, M, R);
17871 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17873 // r = VSELECT(r, shift(r, 2), a);
17874 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17875 R = SignBitSelect(Amt, M, R);
17878 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17880 // return VSELECT(r, shift(r, 1), a);
17881 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17882 R = SignBitSelect(Amt, M, R);
17886 // Decompose 256-bit shifts into smaller 128-bit shifts.
17887 if (VT.is256BitVector()) {
17888 unsigned NumElems = VT.getVectorNumElements();
17889 MVT EltVT = VT.getVectorElementType();
17890 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17892 // Extract the two vectors
17893 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17894 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17896 // Recreate the shift amount vectors
17897 SDValue Amt1, Amt2;
17898 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17899 // Constant shift amount
17900 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
17901 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
17902 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
17904 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17905 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17907 // Variable shift amount
17908 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17909 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17912 // Issue new vector shifts for the smaller types
17913 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17914 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17916 // Concatenate the result back
17917 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17923 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17924 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17925 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17926 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17927 // has only one use.
17928 SDNode *N = Op.getNode();
17929 SDValue LHS = N->getOperand(0);
17930 SDValue RHS = N->getOperand(1);
17931 unsigned BaseOp = 0;
17934 switch (Op.getOpcode()) {
17935 default: llvm_unreachable("Unknown ovf instruction!");
17937 // A subtract of one will be selected as a INC. Note that INC doesn't
17938 // set CF, so we can't do this for UADDO.
17939 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17941 BaseOp = X86ISD::INC;
17942 Cond = X86::COND_O;
17945 BaseOp = X86ISD::ADD;
17946 Cond = X86::COND_O;
17949 BaseOp = X86ISD::ADD;
17950 Cond = X86::COND_B;
17953 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17954 // set CF, so we can't do this for USUBO.
17955 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17957 BaseOp = X86ISD::DEC;
17958 Cond = X86::COND_O;
17961 BaseOp = X86ISD::SUB;
17962 Cond = X86::COND_O;
17965 BaseOp = X86ISD::SUB;
17966 Cond = X86::COND_B;
17969 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
17970 Cond = X86::COND_O;
17972 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
17973 if (N->getValueType(0) == MVT::i8) {
17974 BaseOp = X86ISD::UMUL8;
17975 Cond = X86::COND_O;
17978 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
17980 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
17983 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
17984 DAG.getConstant(X86::COND_O, DL, MVT::i32),
17985 SDValue(Sum.getNode(), 2));
17987 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17991 // Also sets EFLAGS.
17992 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
17993 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
17996 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17997 DAG.getConstant(Cond, DL, MVT::i32),
17998 SDValue(Sum.getNode(), 1));
18000 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18003 /// Returns true if the operand type is exactly twice the native width, and
18004 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18005 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18006 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18007 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18008 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18011 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18012 else if (OpWidth == 128)
18013 return Subtarget->hasCmpxchg16b();
18018 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18019 return needsCmpXchgNb(SI->getValueOperand()->getType());
18022 // Note: this turns large loads into lock cmpxchg8b/16b.
18023 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18024 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18025 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18026 return needsCmpXchgNb(PTy->getElementType());
18029 TargetLoweringBase::AtomicRMWExpansionKind
18030 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18031 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18032 Type *MemType = AI->getType();
18034 // If the operand is too big, we must see if cmpxchg8/16b is available
18035 // and default to library calls otherwise.
18036 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18037 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
18038 : AtomicRMWExpansionKind::None;
18041 AtomicRMWInst::BinOp Op = AI->getOperation();
18044 llvm_unreachable("Unknown atomic operation");
18045 case AtomicRMWInst::Xchg:
18046 case AtomicRMWInst::Add:
18047 case AtomicRMWInst::Sub:
18048 // It's better to use xadd, xsub or xchg for these in all cases.
18049 return AtomicRMWExpansionKind::None;
18050 case AtomicRMWInst::Or:
18051 case AtomicRMWInst::And:
18052 case AtomicRMWInst::Xor:
18053 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18054 // prefix to a normal instruction for these operations.
18055 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
18056 : AtomicRMWExpansionKind::None;
18057 case AtomicRMWInst::Nand:
18058 case AtomicRMWInst::Max:
18059 case AtomicRMWInst::Min:
18060 case AtomicRMWInst::UMax:
18061 case AtomicRMWInst::UMin:
18062 // These always require a non-trivial set of data operations on x86. We must
18063 // use a cmpxchg loop.
18064 return AtomicRMWExpansionKind::CmpXChg;
18068 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18069 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18070 // no-sse2). There isn't any reason to disable it if the target processor
18072 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18076 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18077 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18078 Type *MemType = AI->getType();
18079 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18080 // there is no benefit in turning such RMWs into loads, and it is actually
18081 // harmful as it introduces a mfence.
18082 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18085 auto Builder = IRBuilder<>(AI);
18086 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18087 auto SynchScope = AI->getSynchScope();
18088 // We must restrict the ordering to avoid generating loads with Release or
18089 // ReleaseAcquire orderings.
18090 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18091 auto Ptr = AI->getPointerOperand();
18093 // Before the load we need a fence. Here is an example lifted from
18094 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18097 // x.store(1, relaxed);
18098 // r1 = y.fetch_add(0, release);
18100 // y.fetch_add(42, acquire);
18101 // r2 = x.load(relaxed);
18102 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18103 // lowered to just a load without a fence. A mfence flushes the store buffer,
18104 // making the optimization clearly correct.
18105 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18106 // otherwise, we might be able to be more aggressive on relaxed idempotent
18107 // rmw. In practice, they do not look useful, so we don't try to be
18108 // especially clever.
18109 if (SynchScope == SingleThread)
18110 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18111 // the IR level, so we must wrap it in an intrinsic.
18114 if (!hasMFENCE(*Subtarget))
18115 // FIXME: it might make sense to use a locked operation here but on a
18116 // different cache-line to prevent cache-line bouncing. In practice it
18117 // is probably a small win, and x86 processors without mfence are rare
18118 // enough that we do not bother.
18122 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18123 Builder.CreateCall(MFence, {});
18125 // Finally we can emit the atomic load.
18126 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18127 AI->getType()->getPrimitiveSizeInBits());
18128 Loaded->setAtomic(Order, SynchScope);
18129 AI->replaceAllUsesWith(Loaded);
18130 AI->eraseFromParent();
18134 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18135 SelectionDAG &DAG) {
18137 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18138 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18139 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18140 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18142 // The only fence that needs an instruction is a sequentially-consistent
18143 // cross-thread fence.
18144 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18145 if (hasMFENCE(*Subtarget))
18146 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18148 SDValue Chain = Op.getOperand(0);
18149 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18151 DAG.getRegister(X86::ESP, MVT::i32), // Base
18152 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18153 DAG.getRegister(0, MVT::i32), // Index
18154 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18155 DAG.getRegister(0, MVT::i32), // Segment.
18159 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18160 return SDValue(Res, 0);
18163 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18164 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18167 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18168 SelectionDAG &DAG) {
18169 MVT T = Op.getSimpleValueType();
18173 switch(T.SimpleTy) {
18174 default: llvm_unreachable("Invalid value type!");
18175 case MVT::i8: Reg = X86::AL; size = 1; break;
18176 case MVT::i16: Reg = X86::AX; size = 2; break;
18177 case MVT::i32: Reg = X86::EAX; size = 4; break;
18179 assert(Subtarget->is64Bit() && "Node not type legal!");
18180 Reg = X86::RAX; size = 8;
18183 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18184 Op.getOperand(2), SDValue());
18185 SDValue Ops[] = { cpIn.getValue(0),
18188 DAG.getTargetConstant(size, DL, MVT::i8),
18189 cpIn.getValue(1) };
18190 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18191 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18192 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18196 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18197 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18198 MVT::i32, cpOut.getValue(2));
18199 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18200 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18203 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18204 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18205 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18209 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18210 SelectionDAG &DAG) {
18211 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18212 MVT DstVT = Op.getSimpleValueType();
18214 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18215 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18216 if (DstVT != MVT::f64)
18217 // This conversion needs to be expanded.
18220 SDValue InVec = Op->getOperand(0);
18222 unsigned NumElts = SrcVT.getVectorNumElements();
18223 EVT SVT = SrcVT.getVectorElementType();
18225 // Widen the vector in input in the case of MVT::v2i32.
18226 // Example: from MVT::v2i32 to MVT::v4i32.
18227 SmallVector<SDValue, 16> Elts;
18228 for (unsigned i = 0, e = NumElts; i != e; ++i)
18229 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18230 DAG.getIntPtrConstant(i, dl)));
18232 // Explicitly mark the extra elements as Undef.
18233 Elts.append(NumElts, DAG.getUNDEF(SVT));
18235 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18236 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18237 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18238 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18239 DAG.getIntPtrConstant(0, dl));
18242 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18243 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18244 assert((DstVT == MVT::i64 ||
18245 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18246 "Unexpected custom BITCAST");
18247 // i64 <=> MMX conversions are Legal.
18248 if (SrcVT==MVT::i64 && DstVT.isVector())
18250 if (DstVT==MVT::i64 && SrcVT.isVector())
18252 // MMX <=> MMX conversions are Legal.
18253 if (SrcVT.isVector() && DstVT.isVector())
18255 // All other conversions need to be expanded.
18259 /// Compute the horizontal sum of bytes in V for the elements of VT.
18261 /// Requires V to be a byte vector and VT to be an integer vector type with
18262 /// wider elements than V's type. The width of the elements of VT determines
18263 /// how many bytes of V are summed horizontally to produce each element of the
18265 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18266 const X86Subtarget *Subtarget,
18267 SelectionDAG &DAG) {
18269 MVT ByteVecVT = V.getSimpleValueType();
18270 MVT EltVT = VT.getVectorElementType();
18271 int NumElts = VT.getVectorNumElements();
18272 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18273 "Expected value to have byte element type.");
18274 assert(EltVT != MVT::i8 &&
18275 "Horizontal byte sum only makes sense for wider elements!");
18276 unsigned VecSize = VT.getSizeInBits();
18277 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18279 // PSADBW instruction horizontally add all bytes and leave the result in i64
18280 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18281 if (EltVT == MVT::i64) {
18282 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18283 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18284 return DAG.getBitcast(VT, V);
18287 if (EltVT == MVT::i32) {
18288 // We unpack the low half and high half into i32s interleaved with zeros so
18289 // that we can use PSADBW to horizontally sum them. The most useful part of
18290 // this is that it lines up the results of two PSADBW instructions to be
18291 // two v2i64 vectors which concatenated are the 4 population counts. We can
18292 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18293 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18294 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18295 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18297 // Do the horizontal sums into two v2i64s.
18298 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18299 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18300 DAG.getBitcast(ByteVecVT, Low), Zeros);
18301 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18302 DAG.getBitcast(ByteVecVT, High), Zeros);
18304 // Merge them together.
18305 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18306 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18307 DAG.getBitcast(ShortVecVT, Low),
18308 DAG.getBitcast(ShortVecVT, High));
18310 return DAG.getBitcast(VT, V);
18313 // The only element type left is i16.
18314 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18316 // To obtain pop count for each i16 element starting from the pop count for
18317 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18318 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18319 // directly supported.
18320 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18321 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18322 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18323 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18324 DAG.getBitcast(ByteVecVT, V));
18325 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18328 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18329 const X86Subtarget *Subtarget,
18330 SelectionDAG &DAG) {
18331 MVT VT = Op.getSimpleValueType();
18332 MVT EltVT = VT.getVectorElementType();
18333 unsigned VecSize = VT.getSizeInBits();
18335 // Implement a lookup table in register by using an algorithm based on:
18336 // http://wm.ite.pl/articles/sse-popcount.html
18338 // The general idea is that every lower byte nibble in the input vector is an
18339 // index into a in-register pre-computed pop count table. We then split up the
18340 // input vector in two new ones: (1) a vector with only the shifted-right
18341 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18342 // masked out higher ones) for each byte. PSHUB is used separately with both
18343 // to index the in-register table. Next, both are added and the result is a
18344 // i8 vector where each element contains the pop count for input byte.
18346 // To obtain the pop count for elements != i8, we follow up with the same
18347 // approach and use additional tricks as described below.
18349 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18350 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18351 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18352 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18354 int NumByteElts = VecSize / 8;
18355 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18356 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18357 SmallVector<SDValue, 16> LUTVec;
18358 for (int i = 0; i < NumByteElts; ++i)
18359 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18360 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18361 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18362 DAG.getConstant(0x0F, DL, MVT::i8));
18363 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18366 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18367 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18368 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18371 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18373 // The input vector is used as the shuffle mask that index elements into the
18374 // LUT. After counting low and high nibbles, add the vector to obtain the
18375 // final pop count per i8 element.
18376 SDValue HighPopCnt =
18377 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18378 SDValue LowPopCnt =
18379 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18380 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18382 if (EltVT == MVT::i8)
18385 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18388 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18389 const X86Subtarget *Subtarget,
18390 SelectionDAG &DAG) {
18391 MVT VT = Op.getSimpleValueType();
18392 assert(VT.is128BitVector() &&
18393 "Only 128-bit vector bitmath lowering supported.");
18395 int VecSize = VT.getSizeInBits();
18396 MVT EltVT = VT.getVectorElementType();
18397 int Len = EltVT.getSizeInBits();
18399 // This is the vectorized version of the "best" algorithm from
18400 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18401 // with a minor tweak to use a series of adds + shifts instead of vector
18402 // multiplications. Implemented for all integer vector types. We only use
18403 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18404 // much faster, even faster than using native popcnt instructions.
18406 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18407 MVT VT = V.getSimpleValueType();
18408 SmallVector<SDValue, 32> Shifters(
18409 VT.getVectorNumElements(),
18410 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18411 return DAG.getNode(OpCode, DL, VT, V,
18412 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18414 auto GetMask = [&](SDValue V, APInt Mask) {
18415 MVT VT = V.getSimpleValueType();
18416 SmallVector<SDValue, 32> Masks(
18417 VT.getVectorNumElements(),
18418 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18419 return DAG.getNode(ISD::AND, DL, VT, V,
18420 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18423 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18424 // x86, so set the SRL type to have elements at least i16 wide. This is
18425 // correct because all of our SRLs are followed immediately by a mask anyways
18426 // that handles any bits that sneak into the high bits of the byte elements.
18427 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18431 // v = v - ((v >> 1) & 0x55555555...)
18433 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18434 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18435 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18437 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18438 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18439 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18440 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18441 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18443 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18444 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18445 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18446 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18448 // At this point, V contains the byte-wise population count, and we are
18449 // merely doing a horizontal sum if necessary to get the wider element
18451 if (EltVT == MVT::i8)
18454 return LowerHorizontalByteSum(
18455 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18459 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18460 SelectionDAG &DAG) {
18461 MVT VT = Op.getSimpleValueType();
18462 // FIXME: Need to add AVX-512 support here!
18463 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18464 "Unknown CTPOP type to handle");
18465 SDLoc DL(Op.getNode());
18466 SDValue Op0 = Op.getOperand(0);
18468 if (!Subtarget->hasSSSE3()) {
18469 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18470 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18471 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18474 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18475 unsigned NumElems = VT.getVectorNumElements();
18477 // Extract each 128-bit vector, compute pop count and concat the result.
18478 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18479 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18481 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18482 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18483 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18486 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18489 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18490 SelectionDAG &DAG) {
18491 assert(Op.getValueType().isVector() &&
18492 "We only do custom lowering for vector population count.");
18493 return LowerVectorCTPOP(Op, Subtarget, DAG);
18496 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18497 SDNode *Node = Op.getNode();
18499 EVT T = Node->getValueType(0);
18500 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18501 DAG.getConstant(0, dl, T), Node->getOperand(2));
18502 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18503 cast<AtomicSDNode>(Node)->getMemoryVT(),
18504 Node->getOperand(0),
18505 Node->getOperand(1), negOp,
18506 cast<AtomicSDNode>(Node)->getMemOperand(),
18507 cast<AtomicSDNode>(Node)->getOrdering(),
18508 cast<AtomicSDNode>(Node)->getSynchScope());
18511 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18512 SDNode *Node = Op.getNode();
18514 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18516 // Convert seq_cst store -> xchg
18517 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18518 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18519 // (The only way to get a 16-byte store is cmpxchg16b)
18520 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18521 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18522 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18523 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18524 cast<AtomicSDNode>(Node)->getMemoryVT(),
18525 Node->getOperand(0),
18526 Node->getOperand(1), Node->getOperand(2),
18527 cast<AtomicSDNode>(Node)->getMemOperand(),
18528 cast<AtomicSDNode>(Node)->getOrdering(),
18529 cast<AtomicSDNode>(Node)->getSynchScope());
18530 return Swap.getValue(1);
18532 // Other atomic stores have a simple pattern.
18536 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18537 EVT VT = Op.getNode()->getSimpleValueType(0);
18539 // Let legalize expand this if it isn't a legal type yet.
18540 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18543 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18546 bool ExtraOp = false;
18547 switch (Op.getOpcode()) {
18548 default: llvm_unreachable("Invalid code");
18549 case ISD::ADDC: Opc = X86ISD::ADD; break;
18550 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18551 case ISD::SUBC: Opc = X86ISD::SUB; break;
18552 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18556 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18558 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18559 Op.getOperand(1), Op.getOperand(2));
18562 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18563 SelectionDAG &DAG) {
18564 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18566 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18567 // which returns the values as { float, float } (in XMM0) or
18568 // { double, double } (which is returned in XMM0, XMM1).
18570 SDValue Arg = Op.getOperand(0);
18571 EVT ArgVT = Arg.getValueType();
18572 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18574 TargetLowering::ArgListTy Args;
18575 TargetLowering::ArgListEntry Entry;
18579 Entry.isSExt = false;
18580 Entry.isZExt = false;
18581 Args.push_back(Entry);
18583 bool isF64 = ArgVT == MVT::f64;
18584 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18585 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18586 // the results are returned via SRet in memory.
18587 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18588 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18590 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
18592 Type *RetTy = isF64
18593 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
18594 : (Type*)VectorType::get(ArgTy, 4);
18596 TargetLowering::CallLoweringInfo CLI(DAG);
18597 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18598 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18600 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18603 // Returned in xmm0 and xmm1.
18604 return CallResult.first;
18606 // Returned in bits 0:31 and 32:64 xmm0.
18607 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18608 CallResult.first, DAG.getIntPtrConstant(0, dl));
18609 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18610 CallResult.first, DAG.getIntPtrConstant(1, dl));
18611 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18612 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18615 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
18616 SelectionDAG &DAG) {
18617 assert(Subtarget->hasAVX512() &&
18618 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18620 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
18621 EVT VT = N->getValue().getValueType();
18622 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
18625 // X86 scatter kills mask register, so its type should be added to
18626 // the list of return values
18627 if (N->getNumValues() == 1) {
18628 SDValue Index = N->getIndex();
18629 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18630 !Index.getValueType().is512BitVector())
18631 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18633 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
18634 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18635 N->getOperand(3), Index };
18637 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
18638 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
18639 return SDValue(NewScatter.getNode(), 0);
18644 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
18645 SelectionDAG &DAG) {
18646 assert(Subtarget->hasAVX512() &&
18647 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18649 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
18650 EVT VT = Op.getValueType();
18651 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
18654 SDValue Index = N->getIndex();
18655 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18656 !Index.getValueType().is512BitVector()) {
18657 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18658 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18659 N->getOperand(3), Index };
18660 DAG.UpdateNodeOperands(N, Ops);
18665 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
18666 SelectionDAG &DAG) const {
18667 // TODO: Eventually, the lowering of these nodes should be informed by or
18668 // deferred to the GC strategy for the function in which they appear. For
18669 // now, however, they must be lowered to something. Since they are logically
18670 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18671 // require special handling for these nodes), lower them as literal NOOPs for
18673 SmallVector<SDValue, 2> Ops;
18675 Ops.push_back(Op.getOperand(0));
18676 if (Op->getGluedNode())
18677 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18680 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18681 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18686 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
18687 SelectionDAG &DAG) const {
18688 // TODO: Eventually, the lowering of these nodes should be informed by or
18689 // deferred to the GC strategy for the function in which they appear. For
18690 // now, however, they must be lowered to something. Since they are logically
18691 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18692 // require special handling for these nodes), lower them as literal NOOPs for
18694 SmallVector<SDValue, 2> Ops;
18696 Ops.push_back(Op.getOperand(0));
18697 if (Op->getGluedNode())
18698 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18701 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18702 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18707 /// LowerOperation - Provide custom lowering hooks for some operations.
18709 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18710 switch (Op.getOpcode()) {
18711 default: llvm_unreachable("Should not custom lower this!");
18712 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18713 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18714 return LowerCMP_SWAP(Op, Subtarget, DAG);
18715 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
18716 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18717 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18718 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18719 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
18720 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
18721 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18722 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18723 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18724 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18725 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18726 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18727 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18728 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18729 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18730 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18731 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18732 case ISD::SHL_PARTS:
18733 case ISD::SRA_PARTS:
18734 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18735 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18736 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18737 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18738 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18739 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18740 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18741 case ISD::SIGN_EXTEND_VECTOR_INREG:
18742 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
18743 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18744 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18745 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18746 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18748 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18749 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18750 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18751 case ISD::SETCC: return LowerSETCC(Op, DAG);
18752 case ISD::SELECT: return LowerSELECT(Op, DAG);
18753 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18754 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18755 case ISD::VASTART: return LowerVASTART(Op, DAG);
18756 case ISD::VAARG: return LowerVAARG(Op, DAG);
18757 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18758 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
18759 case ISD::INTRINSIC_VOID:
18760 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18761 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18762 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18763 case ISD::FRAME_TO_ARGS_OFFSET:
18764 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18765 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18766 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18767 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18768 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18769 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18770 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18771 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18772 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18773 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18774 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18775 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18776 case ISD::UMUL_LOHI:
18777 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18780 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18786 case ISD::UMULO: return LowerXALUO(Op, DAG);
18787 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18788 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18792 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18793 case ISD::ADD: return LowerADD(Op, DAG);
18794 case ISD::SUB: return LowerSUB(Op, DAG);
18798 case ISD::UMIN: return LowerMINMAX(Op, DAG);
18799 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18800 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
18801 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
18802 case ISD::GC_TRANSITION_START:
18803 return LowerGC_TRANSITION_START(Op, DAG);
18804 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
18808 /// ReplaceNodeResults - Replace a node with an illegal result type
18809 /// with a new node built out of custom code.
18810 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
18811 SmallVectorImpl<SDValue>&Results,
18812 SelectionDAG &DAG) const {
18814 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18815 switch (N->getOpcode()) {
18817 llvm_unreachable("Do not know how to custom type legalize this operation!");
18818 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
18819 case X86ISD::FMINC:
18821 case X86ISD::FMAXC:
18822 case X86ISD::FMAX: {
18823 EVT VT = N->getValueType(0);
18824 if (VT != MVT::v2f32)
18825 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
18826 SDValue UNDEF = DAG.getUNDEF(VT);
18827 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18828 N->getOperand(0), UNDEF);
18829 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18830 N->getOperand(1), UNDEF);
18831 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
18834 case ISD::SIGN_EXTEND_INREG:
18839 // We don't want to expand or promote these.
18846 case ISD::UDIVREM: {
18847 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
18848 Results.push_back(V);
18851 case ISD::FP_TO_SINT:
18852 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
18853 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
18854 if (N->getOperand(0).getValueType() == MVT::f16)
18857 case ISD::FP_TO_UINT: {
18858 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
18860 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
18863 std::pair<SDValue,SDValue> Vals =
18864 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18865 SDValue FIST = Vals.first, StackSlot = Vals.second;
18866 if (FIST.getNode()) {
18867 EVT VT = N->getValueType(0);
18868 // Return a load from the stack slot.
18869 if (StackSlot.getNode())
18870 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18871 MachinePointerInfo(),
18872 false, false, false, 0));
18874 Results.push_back(FIST);
18878 case ISD::UINT_TO_FP: {
18879 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18880 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18881 N->getValueType(0) != MVT::v2f32)
18883 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18885 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
18887 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18888 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18889 DAG.getBitcast(MVT::v2i64, VBias));
18890 Or = DAG.getBitcast(MVT::v2f64, Or);
18891 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18892 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18895 case ISD::FP_ROUND: {
18896 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18898 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18899 Results.push_back(V);
18902 case ISD::FP_EXTEND: {
18903 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
18904 // No other ValueType for FP_EXTEND should reach this point.
18905 assert(N->getValueType(0) == MVT::v2f32 &&
18906 "Do not know how to legalize this Node");
18909 case ISD::INTRINSIC_W_CHAIN: {
18910 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18912 default : llvm_unreachable("Do not know how to custom type "
18913 "legalize this intrinsic operation!");
18914 case Intrinsic::x86_rdtsc:
18915 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18917 case Intrinsic::x86_rdtscp:
18918 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18920 case Intrinsic::x86_rdpmc:
18921 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18924 case ISD::READCYCLECOUNTER: {
18925 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18928 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18929 EVT T = N->getValueType(0);
18930 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18931 bool Regs64bit = T == MVT::i128;
18932 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18933 SDValue cpInL, cpInH;
18934 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18935 DAG.getConstant(0, dl, HalfT));
18936 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18937 DAG.getConstant(1, dl, HalfT));
18938 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18939 Regs64bit ? X86::RAX : X86::EAX,
18941 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18942 Regs64bit ? X86::RDX : X86::EDX,
18943 cpInH, cpInL.getValue(1));
18944 SDValue swapInL, swapInH;
18945 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18946 DAG.getConstant(0, dl, HalfT));
18947 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18948 DAG.getConstant(1, dl, HalfT));
18949 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18950 Regs64bit ? X86::RBX : X86::EBX,
18951 swapInL, cpInH.getValue(1));
18952 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18953 Regs64bit ? X86::RCX : X86::ECX,
18954 swapInH, swapInL.getValue(1));
18955 SDValue Ops[] = { swapInH.getValue(0),
18957 swapInH.getValue(1) };
18958 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18959 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18960 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18961 X86ISD::LCMPXCHG8_DAG;
18962 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18963 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
18964 Regs64bit ? X86::RAX : X86::EAX,
18965 HalfT, Result.getValue(1));
18966 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
18967 Regs64bit ? X86::RDX : X86::EDX,
18968 HalfT, cpOutL.getValue(2));
18969 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
18971 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
18972 MVT::i32, cpOutH.getValue(2));
18974 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18975 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
18976 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
18978 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
18979 Results.push_back(Success);
18980 Results.push_back(EFLAGS.getValue(1));
18983 case ISD::ATOMIC_SWAP:
18984 case ISD::ATOMIC_LOAD_ADD:
18985 case ISD::ATOMIC_LOAD_SUB:
18986 case ISD::ATOMIC_LOAD_AND:
18987 case ISD::ATOMIC_LOAD_OR:
18988 case ISD::ATOMIC_LOAD_XOR:
18989 case ISD::ATOMIC_LOAD_NAND:
18990 case ISD::ATOMIC_LOAD_MIN:
18991 case ISD::ATOMIC_LOAD_MAX:
18992 case ISD::ATOMIC_LOAD_UMIN:
18993 case ISD::ATOMIC_LOAD_UMAX:
18994 case ISD::ATOMIC_LOAD: {
18995 // Delegate to generic TypeLegalization. Situations we can really handle
18996 // should have already been dealt with by AtomicExpandPass.cpp.
18999 case ISD::BITCAST: {
19000 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19001 EVT DstVT = N->getValueType(0);
19002 EVT SrcVT = N->getOperand(0)->getValueType(0);
19004 if (SrcVT != MVT::f64 ||
19005 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19008 unsigned NumElts = DstVT.getVectorNumElements();
19009 EVT SVT = DstVT.getVectorElementType();
19010 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19011 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19012 MVT::v2f64, N->getOperand(0));
19013 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19015 if (ExperimentalVectorWideningLegalization) {
19016 // If we are legalizing vectors by widening, we already have the desired
19017 // legal vector type, just return it.
19018 Results.push_back(ToVecInt);
19022 SmallVector<SDValue, 8> Elts;
19023 for (unsigned i = 0, e = NumElts; i != e; ++i)
19024 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19025 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19027 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19032 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19033 switch ((X86ISD::NodeType)Opcode) {
19034 case X86ISD::FIRST_NUMBER: break;
19035 case X86ISD::BSF: return "X86ISD::BSF";
19036 case X86ISD::BSR: return "X86ISD::BSR";
19037 case X86ISD::SHLD: return "X86ISD::SHLD";
19038 case X86ISD::SHRD: return "X86ISD::SHRD";
19039 case X86ISD::FAND: return "X86ISD::FAND";
19040 case X86ISD::FANDN: return "X86ISD::FANDN";
19041 case X86ISD::FOR: return "X86ISD::FOR";
19042 case X86ISD::FXOR: return "X86ISD::FXOR";
19043 case X86ISD::FILD: return "X86ISD::FILD";
19044 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19045 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19046 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19047 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19048 case X86ISD::FLD: return "X86ISD::FLD";
19049 case X86ISD::FST: return "X86ISD::FST";
19050 case X86ISD::CALL: return "X86ISD::CALL";
19051 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19052 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19053 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19054 case X86ISD::BT: return "X86ISD::BT";
19055 case X86ISD::CMP: return "X86ISD::CMP";
19056 case X86ISD::COMI: return "X86ISD::COMI";
19057 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19058 case X86ISD::CMPM: return "X86ISD::CMPM";
19059 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19060 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19061 case X86ISD::SETCC: return "X86ISD::SETCC";
19062 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19063 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19064 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19065 case X86ISD::CMOV: return "X86ISD::CMOV";
19066 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19067 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19068 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19069 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19070 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19071 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19072 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19073 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19074 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19075 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19076 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19077 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19078 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19079 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19080 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19081 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19082 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19083 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19084 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19085 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19086 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19087 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19088 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19089 case X86ISD::HADD: return "X86ISD::HADD";
19090 case X86ISD::HSUB: return "X86ISD::HSUB";
19091 case X86ISD::FHADD: return "X86ISD::FHADD";
19092 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19093 case X86ISD::ABS: return "X86ISD::ABS";
19094 case X86ISD::FMAX: return "X86ISD::FMAX";
19095 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19096 case X86ISD::FMIN: return "X86ISD::FMIN";
19097 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19098 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19099 case X86ISD::FMINC: return "X86ISD::FMINC";
19100 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19101 case X86ISD::FRCP: return "X86ISD::FRCP";
19102 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19103 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19104 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19105 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19106 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19107 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19108 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19109 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19110 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19111 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19112 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19113 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19114 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19115 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19116 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19117 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19118 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19119 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19120 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19121 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19122 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19123 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19124 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19125 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19126 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19127 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19128 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19129 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19130 case X86ISD::VSHL: return "X86ISD::VSHL";
19131 case X86ISD::VSRL: return "X86ISD::VSRL";
19132 case X86ISD::VSRA: return "X86ISD::VSRA";
19133 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19134 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19135 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19136 case X86ISD::CMPP: return "X86ISD::CMPP";
19137 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19138 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19139 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19140 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19141 case X86ISD::ADD: return "X86ISD::ADD";
19142 case X86ISD::SUB: return "X86ISD::SUB";
19143 case X86ISD::ADC: return "X86ISD::ADC";
19144 case X86ISD::SBB: return "X86ISD::SBB";
19145 case X86ISD::SMUL: return "X86ISD::SMUL";
19146 case X86ISD::UMUL: return "X86ISD::UMUL";
19147 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19148 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19149 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19150 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19151 case X86ISD::INC: return "X86ISD::INC";
19152 case X86ISD::DEC: return "X86ISD::DEC";
19153 case X86ISD::OR: return "X86ISD::OR";
19154 case X86ISD::XOR: return "X86ISD::XOR";
19155 case X86ISD::AND: return "X86ISD::AND";
19156 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19157 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19158 case X86ISD::PTEST: return "X86ISD::PTEST";
19159 case X86ISD::TESTP: return "X86ISD::TESTP";
19160 case X86ISD::TESTM: return "X86ISD::TESTM";
19161 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19162 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19163 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19164 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19165 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19166 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19167 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19168 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19169 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19170 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19171 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19172 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19173 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19174 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19175 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19176 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19177 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19178 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19179 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19180 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19181 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19182 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19183 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19184 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19185 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19186 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19187 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19188 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19189 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19190 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19191 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19192 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19193 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19194 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19195 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19196 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19197 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19198 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19199 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19200 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19201 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19202 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19203 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19204 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19205 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19206 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19207 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
19208 case X86ISD::SAHF: return "X86ISD::SAHF";
19209 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19210 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19211 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19212 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19213 case X86ISD::FMADD: return "X86ISD::FMADD";
19214 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19215 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19216 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19217 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19218 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19219 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19220 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19221 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19222 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19223 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19224 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19225 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19226 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19227 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19228 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19229 case X86ISD::XTEST: return "X86ISD::XTEST";
19230 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19231 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19232 case X86ISD::SELECT: return "X86ISD::SELECT";
19233 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19234 case X86ISD::RCP28: return "X86ISD::RCP28";
19235 case X86ISD::EXP2: return "X86ISD::EXP2";
19236 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19237 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19238 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19239 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19240 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19241 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19242 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19243 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19244 case X86ISD::ADDS: return "X86ISD::ADDS";
19245 case X86ISD::SUBS: return "X86ISD::SUBS";
19246 case X86ISD::AVG: return "X86ISD::AVG";
19247 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19248 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19249 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19250 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19251 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19256 // isLegalAddressingMode - Return true if the addressing mode represented
19257 // by AM is legal for this target, for a load/store of the specified type.
19258 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19259 const AddrMode &AM, Type *Ty,
19260 unsigned AS) const {
19261 // X86 supports extremely general addressing modes.
19262 CodeModel::Model M = getTargetMachine().getCodeModel();
19263 Reloc::Model R = getTargetMachine().getRelocationModel();
19265 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19266 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19271 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19273 // If a reference to this global requires an extra load, we can't fold it.
19274 if (isGlobalStubReference(GVFlags))
19277 // If BaseGV requires a register for the PIC base, we cannot also have a
19278 // BaseReg specified.
19279 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19282 // If lower 4G is not available, then we must use rip-relative addressing.
19283 if ((M != CodeModel::Small || R != Reloc::Static) &&
19284 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19288 switch (AM.Scale) {
19294 // These scales always work.
19299 // These scales are formed with basereg+scalereg. Only accept if there is
19304 default: // Other stuff never works.
19311 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19312 unsigned Bits = Ty->getScalarSizeInBits();
19314 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19315 // particularly cheaper than those without.
19319 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19320 // variable shifts just as cheap as scalar ones.
19321 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19324 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19325 // fully general vector.
19329 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19330 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19332 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19333 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19334 return NumBits1 > NumBits2;
19337 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19338 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19341 if (!isTypeLegal(EVT::getEVT(Ty1)))
19344 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19346 // Assuming the caller doesn't have a zeroext or signext return parameter,
19347 // truncation all the way down to i1 is valid.
19351 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19352 return isInt<32>(Imm);
19355 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19356 // Can also use sub to handle negated immediates.
19357 return isInt<32>(Imm);
19360 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19361 if (!VT1.isInteger() || !VT2.isInteger())
19363 unsigned NumBits1 = VT1.getSizeInBits();
19364 unsigned NumBits2 = VT2.getSizeInBits();
19365 return NumBits1 > NumBits2;
19368 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19369 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19370 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19373 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19374 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19375 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19378 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19379 EVT VT1 = Val.getValueType();
19380 if (isZExtFree(VT1, VT2))
19383 if (Val.getOpcode() != ISD::LOAD)
19386 if (!VT1.isSimple() || !VT1.isInteger() ||
19387 !VT2.isSimple() || !VT2.isInteger())
19390 switch (VT1.getSimpleVT().SimpleTy) {
19395 // X86 has 8, 16, and 32-bit zero-extending loads.
19402 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19405 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19406 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19409 VT = VT.getScalarType();
19411 if (!VT.isSimple())
19414 switch (VT.getSimpleVT().SimpleTy) {
19425 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19426 // i16 instructions are longer (0x66 prefix) and potentially slower.
19427 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19430 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19431 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19432 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19433 /// are assumed to be legal.
19435 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19437 if (!VT.isSimple())
19440 // Not for i1 vectors
19441 if (VT.getScalarType() == MVT::i1)
19444 // Very little shuffling can be done for 64-bit vectors right now.
19445 if (VT.getSizeInBits() == 64)
19448 // We only care that the types being shuffled are legal. The lowering can
19449 // handle any possible shuffle mask that results.
19450 return isTypeLegal(VT.getSimpleVT());
19454 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19456 // Just delegate to the generic legality, clear masks aren't special.
19457 return isShuffleMaskLegal(Mask, VT);
19460 //===----------------------------------------------------------------------===//
19461 // X86 Scheduler Hooks
19462 //===----------------------------------------------------------------------===//
19464 /// Utility function to emit xbegin specifying the start of an RTM region.
19465 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19466 const TargetInstrInfo *TII) {
19467 DebugLoc DL = MI->getDebugLoc();
19469 const BasicBlock *BB = MBB->getBasicBlock();
19470 MachineFunction::iterator I = MBB;
19473 // For the v = xbegin(), we generate
19484 MachineBasicBlock *thisMBB = MBB;
19485 MachineFunction *MF = MBB->getParent();
19486 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19487 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19488 MF->insert(I, mainMBB);
19489 MF->insert(I, sinkMBB);
19491 // Transfer the remainder of BB and its successor edges to sinkMBB.
19492 sinkMBB->splice(sinkMBB->begin(), MBB,
19493 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19494 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19498 // # fallthrough to mainMBB
19499 // # abortion to sinkMBB
19500 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19501 thisMBB->addSuccessor(mainMBB);
19502 thisMBB->addSuccessor(sinkMBB);
19506 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19507 mainMBB->addSuccessor(sinkMBB);
19510 // EAX is live into the sinkMBB
19511 sinkMBB->addLiveIn(X86::EAX);
19512 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19513 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19516 MI->eraseFromParent();
19520 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19521 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19522 // in the .td file.
19523 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19524 const TargetInstrInfo *TII) {
19526 switch (MI->getOpcode()) {
19527 default: llvm_unreachable("illegal opcode!");
19528 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19529 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19530 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19531 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19532 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19533 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19534 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19535 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19538 DebugLoc dl = MI->getDebugLoc();
19539 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19541 unsigned NumArgs = MI->getNumOperands();
19542 for (unsigned i = 1; i < NumArgs; ++i) {
19543 MachineOperand &Op = MI->getOperand(i);
19544 if (!(Op.isReg() && Op.isImplicit()))
19545 MIB.addOperand(Op);
19547 if (MI->hasOneMemOperand())
19548 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19550 BuildMI(*BB, MI, dl,
19551 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19552 .addReg(X86::XMM0);
19554 MI->eraseFromParent();
19558 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19559 // defs in an instruction pattern
19560 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19561 const TargetInstrInfo *TII) {
19563 switch (MI->getOpcode()) {
19564 default: llvm_unreachable("illegal opcode!");
19565 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
19566 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
19567 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
19568 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
19569 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
19570 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
19571 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
19572 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
19575 DebugLoc dl = MI->getDebugLoc();
19576 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19578 unsigned NumArgs = MI->getNumOperands(); // remove the results
19579 for (unsigned i = 1; i < NumArgs; ++i) {
19580 MachineOperand &Op = MI->getOperand(i);
19581 if (!(Op.isReg() && Op.isImplicit()))
19582 MIB.addOperand(Op);
19584 if (MI->hasOneMemOperand())
19585 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19587 BuildMI(*BB, MI, dl,
19588 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19591 MI->eraseFromParent();
19595 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
19596 const X86Subtarget *Subtarget) {
19597 DebugLoc dl = MI->getDebugLoc();
19598 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19599 // Address into RAX/EAX, other two args into ECX, EDX.
19600 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
19601 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
19602 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
19603 for (int i = 0; i < X86::AddrNumOperands; ++i)
19604 MIB.addOperand(MI->getOperand(i));
19606 unsigned ValOps = X86::AddrNumOperands;
19607 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
19608 .addReg(MI->getOperand(ValOps).getReg());
19609 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
19610 .addReg(MI->getOperand(ValOps+1).getReg());
19612 // The instruction doesn't actually take any operands though.
19613 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
19615 MI->eraseFromParent(); // The pseudo is gone now.
19619 MachineBasicBlock *
19620 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
19621 MachineBasicBlock *MBB) const {
19622 // Emit va_arg instruction on X86-64.
19624 // Operands to this pseudo-instruction:
19625 // 0 ) Output : destination address (reg)
19626 // 1-5) Input : va_list address (addr, i64mem)
19627 // 6 ) ArgSize : Size (in bytes) of vararg type
19628 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
19629 // 8 ) Align : Alignment of type
19630 // 9 ) EFLAGS (implicit-def)
19632 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
19633 static_assert(X86::AddrNumOperands == 5,
19634 "VAARG_64 assumes 5 address operands");
19636 unsigned DestReg = MI->getOperand(0).getReg();
19637 MachineOperand &Base = MI->getOperand(1);
19638 MachineOperand &Scale = MI->getOperand(2);
19639 MachineOperand &Index = MI->getOperand(3);
19640 MachineOperand &Disp = MI->getOperand(4);
19641 MachineOperand &Segment = MI->getOperand(5);
19642 unsigned ArgSize = MI->getOperand(6).getImm();
19643 unsigned ArgMode = MI->getOperand(7).getImm();
19644 unsigned Align = MI->getOperand(8).getImm();
19646 // Memory Reference
19647 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19648 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19649 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19651 // Machine Information
19652 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19653 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19654 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19655 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19656 DebugLoc DL = MI->getDebugLoc();
19658 // struct va_list {
19661 // i64 overflow_area (address)
19662 // i64 reg_save_area (address)
19664 // sizeof(va_list) = 24
19665 // alignment(va_list) = 8
19667 unsigned TotalNumIntRegs = 6;
19668 unsigned TotalNumXMMRegs = 8;
19669 bool UseGPOffset = (ArgMode == 1);
19670 bool UseFPOffset = (ArgMode == 2);
19671 unsigned MaxOffset = TotalNumIntRegs * 8 +
19672 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19674 /* Align ArgSize to a multiple of 8 */
19675 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19676 bool NeedsAlign = (Align > 8);
19678 MachineBasicBlock *thisMBB = MBB;
19679 MachineBasicBlock *overflowMBB;
19680 MachineBasicBlock *offsetMBB;
19681 MachineBasicBlock *endMBB;
19683 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19684 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19685 unsigned OffsetReg = 0;
19687 if (!UseGPOffset && !UseFPOffset) {
19688 // If we only pull from the overflow region, we don't create a branch.
19689 // We don't need to alter control flow.
19690 OffsetDestReg = 0; // unused
19691 OverflowDestReg = DestReg;
19693 offsetMBB = nullptr;
19694 overflowMBB = thisMBB;
19697 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19698 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19699 // If not, pull from overflow_area. (branch to overflowMBB)
19704 // offsetMBB overflowMBB
19709 // Registers for the PHI in endMBB
19710 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19711 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19713 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19714 MachineFunction *MF = MBB->getParent();
19715 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19716 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19717 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19719 MachineFunction::iterator MBBIter = MBB;
19722 // Insert the new basic blocks
19723 MF->insert(MBBIter, offsetMBB);
19724 MF->insert(MBBIter, overflowMBB);
19725 MF->insert(MBBIter, endMBB);
19727 // Transfer the remainder of MBB and its successor edges to endMBB.
19728 endMBB->splice(endMBB->begin(), thisMBB,
19729 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19730 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19732 // Make offsetMBB and overflowMBB successors of thisMBB
19733 thisMBB->addSuccessor(offsetMBB);
19734 thisMBB->addSuccessor(overflowMBB);
19736 // endMBB is a successor of both offsetMBB and overflowMBB
19737 offsetMBB->addSuccessor(endMBB);
19738 overflowMBB->addSuccessor(endMBB);
19740 // Load the offset value into a register
19741 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19742 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19746 .addDisp(Disp, UseFPOffset ? 4 : 0)
19747 .addOperand(Segment)
19748 .setMemRefs(MMOBegin, MMOEnd);
19750 // Check if there is enough room left to pull this argument.
19751 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19753 .addImm(MaxOffset + 8 - ArgSizeA8);
19755 // Branch to "overflowMBB" if offset >= max
19756 // Fall through to "offsetMBB" otherwise
19757 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19758 .addMBB(overflowMBB);
19761 // In offsetMBB, emit code to use the reg_save_area.
19763 assert(OffsetReg != 0);
19765 // Read the reg_save_area address.
19766 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19767 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19772 .addOperand(Segment)
19773 .setMemRefs(MMOBegin, MMOEnd);
19775 // Zero-extend the offset
19776 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19777 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19780 .addImm(X86::sub_32bit);
19782 // Add the offset to the reg_save_area to get the final address.
19783 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19784 .addReg(OffsetReg64)
19785 .addReg(RegSaveReg);
19787 // Compute the offset for the next argument
19788 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19789 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19791 .addImm(UseFPOffset ? 16 : 8);
19793 // Store it back into the va_list.
19794 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19798 .addDisp(Disp, UseFPOffset ? 4 : 0)
19799 .addOperand(Segment)
19800 .addReg(NextOffsetReg)
19801 .setMemRefs(MMOBegin, MMOEnd);
19804 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
19809 // Emit code to use overflow area
19812 // Load the overflow_area address into a register.
19813 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
19814 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
19819 .addOperand(Segment)
19820 .setMemRefs(MMOBegin, MMOEnd);
19822 // If we need to align it, do so. Otherwise, just copy the address
19823 // to OverflowDestReg.
19825 // Align the overflow address
19826 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
19827 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
19829 // aligned_addr = (addr + (align-1)) & ~(align-1)
19830 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
19831 .addReg(OverflowAddrReg)
19834 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
19836 .addImm(~(uint64_t)(Align-1));
19838 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
19839 .addReg(OverflowAddrReg);
19842 // Compute the next overflow address after this argument.
19843 // (the overflow address should be kept 8-byte aligned)
19844 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
19845 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
19846 .addReg(OverflowDestReg)
19847 .addImm(ArgSizeA8);
19849 // Store the new overflow address.
19850 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
19855 .addOperand(Segment)
19856 .addReg(NextAddrReg)
19857 .setMemRefs(MMOBegin, MMOEnd);
19859 // If we branched, emit the PHI to the front of endMBB.
19861 BuildMI(*endMBB, endMBB->begin(), DL,
19862 TII->get(X86::PHI), DestReg)
19863 .addReg(OffsetDestReg).addMBB(offsetMBB)
19864 .addReg(OverflowDestReg).addMBB(overflowMBB);
19867 // Erase the pseudo instruction
19868 MI->eraseFromParent();
19873 MachineBasicBlock *
19874 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
19876 MachineBasicBlock *MBB) const {
19877 // Emit code to save XMM registers to the stack. The ABI says that the
19878 // number of registers to save is given in %al, so it's theoretically
19879 // possible to do an indirect jump trick to avoid saving all of them,
19880 // however this code takes a simpler approach and just executes all
19881 // of the stores if %al is non-zero. It's less code, and it's probably
19882 // easier on the hardware branch predictor, and stores aren't all that
19883 // expensive anyway.
19885 // Create the new basic blocks. One block contains all the XMM stores,
19886 // and one block is the final destination regardless of whether any
19887 // stores were performed.
19888 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19889 MachineFunction *F = MBB->getParent();
19890 MachineFunction::iterator MBBIter = MBB;
19892 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19893 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19894 F->insert(MBBIter, XMMSaveMBB);
19895 F->insert(MBBIter, EndMBB);
19897 // Transfer the remainder of MBB and its successor edges to EndMBB.
19898 EndMBB->splice(EndMBB->begin(), MBB,
19899 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19900 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19902 // The original block will now fall through to the XMM save block.
19903 MBB->addSuccessor(XMMSaveMBB);
19904 // The XMMSaveMBB will fall through to the end block.
19905 XMMSaveMBB->addSuccessor(EndMBB);
19907 // Now add the instructions.
19908 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19909 DebugLoc DL = MI->getDebugLoc();
19911 unsigned CountReg = MI->getOperand(0).getReg();
19912 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19913 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19915 if (!Subtarget->isTargetWin64()) {
19916 // If %al is 0, branch around the XMM save block.
19917 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19918 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
19919 MBB->addSuccessor(EndMBB);
19922 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19923 // that was just emitted, but clearly shouldn't be "saved".
19924 assert((MI->getNumOperands() <= 3 ||
19925 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19926 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19927 && "Expected last argument to be EFLAGS");
19928 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19929 // In the XMM save block, save all the XMM argument registers.
19930 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19931 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19932 MachineMemOperand *MMO = F->getMachineMemOperand(
19933 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
19934 MachineMemOperand::MOStore,
19935 /*Size=*/16, /*Align=*/16);
19936 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19937 .addFrameIndex(RegSaveFrameIndex)
19938 .addImm(/*Scale=*/1)
19939 .addReg(/*IndexReg=*/0)
19940 .addImm(/*Disp=*/Offset)
19941 .addReg(/*Segment=*/0)
19942 .addReg(MI->getOperand(i).getReg())
19943 .addMemOperand(MMO);
19946 MI->eraseFromParent(); // The pseudo instruction is gone now.
19951 // The EFLAGS operand of SelectItr might be missing a kill marker
19952 // because there were multiple uses of EFLAGS, and ISel didn't know
19953 // which to mark. Figure out whether SelectItr should have had a
19954 // kill marker, and set it if it should. Returns the correct kill
19956 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19957 MachineBasicBlock* BB,
19958 const TargetRegisterInfo* TRI) {
19959 // Scan forward through BB for a use/def of EFLAGS.
19960 MachineBasicBlock::iterator miI(std::next(SelectItr));
19961 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19962 const MachineInstr& mi = *miI;
19963 if (mi.readsRegister(X86::EFLAGS))
19965 if (mi.definesRegister(X86::EFLAGS))
19966 break; // Should have kill-flag - update below.
19969 // If we hit the end of the block, check whether EFLAGS is live into a
19971 if (miI == BB->end()) {
19972 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
19973 sEnd = BB->succ_end();
19974 sItr != sEnd; ++sItr) {
19975 MachineBasicBlock* succ = *sItr;
19976 if (succ->isLiveIn(X86::EFLAGS))
19981 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
19982 // out. SelectMI should have a kill flag on EFLAGS.
19983 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
19987 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
19988 // together with other CMOV pseudo-opcodes into a single basic-block with
19989 // conditional jump around it.
19990 static bool isCMOVPseudo(MachineInstr *MI) {
19991 switch (MI->getOpcode()) {
19992 case X86::CMOV_FR32:
19993 case X86::CMOV_FR64:
19994 case X86::CMOV_GR8:
19995 case X86::CMOV_GR16:
19996 case X86::CMOV_GR32:
19997 case X86::CMOV_RFP32:
19998 case X86::CMOV_RFP64:
19999 case X86::CMOV_RFP80:
20000 case X86::CMOV_V2F64:
20001 case X86::CMOV_V2I64:
20002 case X86::CMOV_V4F32:
20003 case X86::CMOV_V4F64:
20004 case X86::CMOV_V4I64:
20005 case X86::CMOV_V16F32:
20006 case X86::CMOV_V8F32:
20007 case X86::CMOV_V8F64:
20008 case X86::CMOV_V8I64:
20009 case X86::CMOV_V8I1:
20010 case X86::CMOV_V16I1:
20011 case X86::CMOV_V32I1:
20012 case X86::CMOV_V64I1:
20020 MachineBasicBlock *
20021 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20022 MachineBasicBlock *BB) const {
20023 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20024 DebugLoc DL = MI->getDebugLoc();
20026 // To "insert" a SELECT_CC instruction, we actually have to insert the
20027 // diamond control-flow pattern. The incoming instruction knows the
20028 // destination vreg to set, the condition code register to branch on, the
20029 // true/false values to select between, and a branch opcode to use.
20030 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20031 MachineFunction::iterator It = BB;
20037 // cmpTY ccX, r1, r2
20039 // fallthrough --> copy0MBB
20040 MachineBasicBlock *thisMBB = BB;
20041 MachineFunction *F = BB->getParent();
20043 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20044 // as described above, by inserting a BB, and then making a PHI at the join
20045 // point to select the true and false operands of the CMOV in the PHI.
20047 // The code also handles two different cases of multiple CMOV opcodes
20051 // In this case, there are multiple CMOVs in a row, all which are based on
20052 // the same condition setting (or the exact opposite condition setting).
20053 // In this case we can lower all the CMOVs using a single inserted BB, and
20054 // then make a number of PHIs at the join point to model the CMOVs. The only
20055 // trickiness here, is that in a case like:
20057 // t2 = CMOV cond1 t1, f1
20058 // t3 = CMOV cond1 t2, f2
20060 // when rewriting this into PHIs, we have to perform some renaming on the
20061 // temps since you cannot have a PHI operand refer to a PHI result earlier
20062 // in the same block. The "simple" but wrong lowering would be:
20064 // t2 = PHI t1(BB1), f1(BB2)
20065 // t3 = PHI t2(BB1), f2(BB2)
20067 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20068 // renaming is to note that on the path through BB1, t2 is really just a
20069 // copy of t1, and do that renaming, properly generating:
20071 // t2 = PHI t1(BB1), f1(BB2)
20072 // t3 = PHI t1(BB1), f2(BB2)
20074 // Case 2, we lower cascaded CMOVs such as
20076 // (CMOV (CMOV F, T, cc1), T, cc2)
20078 // to two successives branches. For that, we look for another CMOV as the
20079 // following instruction.
20081 // Without this, we would add a PHI between the two jumps, which ends up
20082 // creating a few copies all around. For instance, for
20084 // (sitofp (zext (fcmp une)))
20086 // we would generate:
20088 // ucomiss %xmm1, %xmm0
20089 // movss <1.0f>, %xmm0
20090 // movaps %xmm0, %xmm1
20092 // xorps %xmm1, %xmm1
20095 // movaps %xmm1, %xmm0
20099 // because this custom-inserter would have generated:
20111 // A: X = ...; Y = ...
20113 // C: Z = PHI [X, A], [Y, B]
20115 // E: PHI [X, C], [Z, D]
20117 // If we lower both CMOVs in a single step, we can instead generate:
20129 // A: X = ...; Y = ...
20131 // E: PHI [X, A], [X, C], [Y, D]
20133 // Which, in our sitofp/fcmp example, gives us something like:
20135 // ucomiss %xmm1, %xmm0
20136 // movss <1.0f>, %xmm0
20139 // xorps %xmm0, %xmm0
20143 MachineInstr *CascadedCMOV = nullptr;
20144 MachineInstr *LastCMOV = MI;
20145 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20146 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20147 MachineBasicBlock::iterator NextMIIt =
20148 std::next(MachineBasicBlock::iterator(MI));
20150 // Check for case 1, where there are multiple CMOVs with the same condition
20151 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
20152 // number of jumps the most.
20154 if (isCMOVPseudo(MI)) {
20155 // See if we have a string of CMOVS with the same condition.
20156 while (NextMIIt != BB->end() &&
20157 isCMOVPseudo(NextMIIt) &&
20158 (NextMIIt->getOperand(3).getImm() == CC ||
20159 NextMIIt->getOperand(3).getImm() == OppCC)) {
20160 LastCMOV = &*NextMIIt;
20165 // This checks for case 2, but only do this if we didn't already find
20166 // case 1, as indicated by LastCMOV == MI.
20167 if (LastCMOV == MI &&
20168 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20169 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20170 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
20171 CascadedCMOV = &*NextMIIt;
20174 MachineBasicBlock *jcc1MBB = nullptr;
20176 // If we have a cascaded CMOV, we lower it to two successive branches to
20177 // the same block. EFLAGS is used by both, so mark it as live in the second.
20178 if (CascadedCMOV) {
20179 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20180 F->insert(It, jcc1MBB);
20181 jcc1MBB->addLiveIn(X86::EFLAGS);
20184 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20185 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20186 F->insert(It, copy0MBB);
20187 F->insert(It, sinkMBB);
20189 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20190 // live into the sink and copy blocks.
20191 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20193 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
20194 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20195 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20196 copy0MBB->addLiveIn(X86::EFLAGS);
20197 sinkMBB->addLiveIn(X86::EFLAGS);
20200 // Transfer the remainder of BB and its successor edges to sinkMBB.
20201 sinkMBB->splice(sinkMBB->begin(), BB,
20202 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
20203 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20205 // Add the true and fallthrough blocks as its successors.
20206 if (CascadedCMOV) {
20207 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
20208 BB->addSuccessor(jcc1MBB);
20210 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20211 // jump to the sinkMBB.
20212 jcc1MBB->addSuccessor(copy0MBB);
20213 jcc1MBB->addSuccessor(sinkMBB);
20215 BB->addSuccessor(copy0MBB);
20218 // The true block target of the first (or only) branch is always sinkMBB.
20219 BB->addSuccessor(sinkMBB);
20221 // Create the conditional branch instruction.
20222 unsigned Opc = X86::GetCondBranchFromCond(CC);
20223 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20225 if (CascadedCMOV) {
20226 unsigned Opc2 = X86::GetCondBranchFromCond(
20227 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
20228 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20232 // %FalseValue = ...
20233 // # fallthrough to sinkMBB
20234 copy0MBB->addSuccessor(sinkMBB);
20237 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20239 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
20240 MachineBasicBlock::iterator MIItEnd =
20241 std::next(MachineBasicBlock::iterator(LastCMOV));
20242 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
20243 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
20244 MachineInstrBuilder MIB;
20246 // As we are creating the PHIs, we have to be careful if there is more than
20247 // one. Later CMOVs may reference the results of earlier CMOVs, but later
20248 // PHIs have to reference the individual true/false inputs from earlier PHIs.
20249 // That also means that PHI construction must work forward from earlier to
20250 // later, and that the code must maintain a mapping from earlier PHI's
20251 // destination registers, and the registers that went into the PHI.
20253 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
20254 unsigned DestReg = MIIt->getOperand(0).getReg();
20255 unsigned Op1Reg = MIIt->getOperand(1).getReg();
20256 unsigned Op2Reg = MIIt->getOperand(2).getReg();
20258 // If this CMOV we are generating is the opposite condition from
20259 // the jump we generated, then we have to swap the operands for the
20260 // PHI that is going to be generated.
20261 if (MIIt->getOperand(3).getImm() == OppCC)
20262 std::swap(Op1Reg, Op2Reg);
20264 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
20265 Op1Reg = RegRewriteTable[Op1Reg].first;
20267 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
20268 Op2Reg = RegRewriteTable[Op2Reg].second;
20270 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
20271 TII->get(X86::PHI), DestReg)
20272 .addReg(Op1Reg).addMBB(copy0MBB)
20273 .addReg(Op2Reg).addMBB(thisMBB);
20275 // Add this PHI to the rewrite table.
20276 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
20279 // If we have a cascaded CMOV, the second Jcc provides the same incoming
20280 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20281 if (CascadedCMOV) {
20282 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20283 // Copy the PHI result to the register defined by the second CMOV.
20284 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20285 DL, TII->get(TargetOpcode::COPY),
20286 CascadedCMOV->getOperand(0).getReg())
20287 .addReg(MI->getOperand(0).getReg());
20288 CascadedCMOV->eraseFromParent();
20291 // Now remove the CMOV(s).
20292 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
20293 (MIIt++)->eraseFromParent();
20298 MachineBasicBlock *
20299 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
20300 MachineBasicBlock *BB) const {
20301 // Combine the following atomic floating-point modification pattern:
20302 // a.store(reg OP a.load(acquire), release)
20303 // Transform them into:
20304 // OPss (%gpr), %xmm
20305 // movss %xmm, (%gpr)
20306 // Or sd equivalent for 64-bit operations.
20308 switch (MI->getOpcode()) {
20309 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
20310 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
20311 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
20313 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20314 DebugLoc DL = MI->getDebugLoc();
20315 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
20316 unsigned MSrc = MI->getOperand(0).getReg();
20317 unsigned VSrc = MI->getOperand(5).getReg();
20318 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
20319 .addReg(/*Base=*/MSrc)
20320 .addImm(/*Scale=*/1)
20321 .addReg(/*Index=*/0)
20324 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
20325 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
20327 .addReg(/*Base=*/MSrc)
20328 .addImm(/*Scale=*/1)
20329 .addReg(/*Index=*/0)
20330 .addImm(/*Disp=*/0)
20331 .addReg(/*Segment=*/0);
20332 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
20333 MI->eraseFromParent(); // The pseudo instruction is gone now.
20337 MachineBasicBlock *
20338 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20339 MachineBasicBlock *BB) const {
20340 MachineFunction *MF = BB->getParent();
20341 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20342 DebugLoc DL = MI->getDebugLoc();
20343 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20345 assert(MF->shouldSplitStack());
20347 const bool Is64Bit = Subtarget->is64Bit();
20348 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20350 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20351 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20354 // ... [Till the alloca]
20355 // If stacklet is not large enough, jump to mallocMBB
20358 // Allocate by subtracting from RSP
20359 // Jump to continueMBB
20362 // Allocate by call to runtime
20366 // [rest of original BB]
20369 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20370 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20371 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20373 MachineRegisterInfo &MRI = MF->getRegInfo();
20374 const TargetRegisterClass *AddrRegClass =
20375 getRegClassFor(getPointerTy(MF->getDataLayout()));
20377 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20378 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20379 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20380 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20381 sizeVReg = MI->getOperand(1).getReg(),
20382 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20384 MachineFunction::iterator MBBIter = BB;
20387 MF->insert(MBBIter, bumpMBB);
20388 MF->insert(MBBIter, mallocMBB);
20389 MF->insert(MBBIter, continueMBB);
20391 continueMBB->splice(continueMBB->begin(), BB,
20392 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20393 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20395 // Add code to the main basic block to check if the stack limit has been hit,
20396 // and if so, jump to mallocMBB otherwise to bumpMBB.
20397 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20398 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20399 .addReg(tmpSPVReg).addReg(sizeVReg);
20400 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20401 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20402 .addReg(SPLimitVReg);
20403 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20405 // bumpMBB simply decreases the stack pointer, since we know the current
20406 // stacklet has enough space.
20407 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20408 .addReg(SPLimitVReg);
20409 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20410 .addReg(SPLimitVReg);
20411 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20413 // Calls into a routine in libgcc to allocate more space from the heap.
20414 const uint32_t *RegMask =
20415 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20417 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20419 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20420 .addExternalSymbol("__morestack_allocate_stack_space")
20421 .addRegMask(RegMask)
20422 .addReg(X86::RDI, RegState::Implicit)
20423 .addReg(X86::RAX, RegState::ImplicitDefine);
20424 } else if (Is64Bit) {
20425 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20427 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20428 .addExternalSymbol("__morestack_allocate_stack_space")
20429 .addRegMask(RegMask)
20430 .addReg(X86::EDI, RegState::Implicit)
20431 .addReg(X86::EAX, RegState::ImplicitDefine);
20433 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20435 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20436 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20437 .addExternalSymbol("__morestack_allocate_stack_space")
20438 .addRegMask(RegMask)
20439 .addReg(X86::EAX, RegState::ImplicitDefine);
20443 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20446 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20447 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20448 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20450 // Set up the CFG correctly.
20451 BB->addSuccessor(bumpMBB);
20452 BB->addSuccessor(mallocMBB);
20453 mallocMBB->addSuccessor(continueMBB);
20454 bumpMBB->addSuccessor(continueMBB);
20456 // Take care of the PHI nodes.
20457 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20458 MI->getOperand(0).getReg())
20459 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20460 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20462 // Delete the original pseudo instruction.
20463 MI->eraseFromParent();
20466 return continueMBB;
20469 MachineBasicBlock *
20470 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20471 MachineBasicBlock *BB) const {
20472 DebugLoc DL = MI->getDebugLoc();
20474 assert(!Subtarget->isTargetMachO());
20476 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20479 MI->eraseFromParent(); // The pseudo instruction is gone now.
20483 MachineBasicBlock *
20484 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20485 MachineBasicBlock *BB) const {
20486 // This is pretty easy. We're taking the value that we received from
20487 // our load from the relocation, sticking it in either RDI (x86-64)
20488 // or EAX and doing an indirect call. The return value will then
20489 // be in the normal return register.
20490 MachineFunction *F = BB->getParent();
20491 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20492 DebugLoc DL = MI->getDebugLoc();
20494 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20495 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20497 // Get a register mask for the lowered call.
20498 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20499 // proper register mask.
20500 const uint32_t *RegMask =
20501 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20502 if (Subtarget->is64Bit()) {
20503 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20504 TII->get(X86::MOV64rm), X86::RDI)
20506 .addImm(0).addReg(0)
20507 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20508 MI->getOperand(3).getTargetFlags())
20510 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20511 addDirectMem(MIB, X86::RDI);
20512 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20513 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20514 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20515 TII->get(X86::MOV32rm), X86::EAX)
20517 .addImm(0).addReg(0)
20518 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20519 MI->getOperand(3).getTargetFlags())
20521 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20522 addDirectMem(MIB, X86::EAX);
20523 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20525 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20526 TII->get(X86::MOV32rm), X86::EAX)
20527 .addReg(TII->getGlobalBaseReg(F))
20528 .addImm(0).addReg(0)
20529 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20530 MI->getOperand(3).getTargetFlags())
20532 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20533 addDirectMem(MIB, X86::EAX);
20534 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20537 MI->eraseFromParent(); // The pseudo instruction is gone now.
20541 MachineBasicBlock *
20542 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20543 MachineBasicBlock *MBB) const {
20544 DebugLoc DL = MI->getDebugLoc();
20545 MachineFunction *MF = MBB->getParent();
20546 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20547 MachineRegisterInfo &MRI = MF->getRegInfo();
20549 const BasicBlock *BB = MBB->getBasicBlock();
20550 MachineFunction::iterator I = MBB;
20553 // Memory Reference
20554 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20555 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20558 unsigned MemOpndSlot = 0;
20560 unsigned CurOp = 0;
20562 DstReg = MI->getOperand(CurOp++).getReg();
20563 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20564 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20565 unsigned mainDstReg = MRI.createVirtualRegister(RC);
20566 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
20568 MemOpndSlot = CurOp;
20570 MVT PVT = getPointerTy(MF->getDataLayout());
20571 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20572 "Invalid Pointer Size!");
20574 // For v = setjmp(buf), we generate
20577 // buf[LabelOffset] = restoreMBB
20578 // SjLjSetup restoreMBB
20584 // v = phi(main, restore)
20587 // if base pointer being used, load it from frame
20590 MachineBasicBlock *thisMBB = MBB;
20591 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20592 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20593 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
20594 MF->insert(I, mainMBB);
20595 MF->insert(I, sinkMBB);
20596 MF->push_back(restoreMBB);
20598 MachineInstrBuilder MIB;
20600 // Transfer the remainder of BB and its successor edges to sinkMBB.
20601 sinkMBB->splice(sinkMBB->begin(), MBB,
20602 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20603 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20606 unsigned PtrStoreOpc = 0;
20607 unsigned LabelReg = 0;
20608 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20609 Reloc::Model RM = MF->getTarget().getRelocationModel();
20610 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
20611 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
20613 // Prepare IP either in reg or imm.
20614 if (!UseImmLabel) {
20615 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
20616 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
20617 LabelReg = MRI.createVirtualRegister(PtrRC);
20618 if (Subtarget->is64Bit()) {
20619 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
20623 .addMBB(restoreMBB)
20626 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
20627 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
20628 .addReg(XII->getGlobalBaseReg(MF))
20631 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
20635 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
20637 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
20638 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20639 if (i == X86::AddrDisp)
20640 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
20642 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
20645 MIB.addReg(LabelReg);
20647 MIB.addMBB(restoreMBB);
20648 MIB.setMemRefs(MMOBegin, MMOEnd);
20650 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
20651 .addMBB(restoreMBB);
20653 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20654 MIB.addRegMask(RegInfo->getNoPreservedMask());
20655 thisMBB->addSuccessor(mainMBB);
20656 thisMBB->addSuccessor(restoreMBB);
20660 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
20661 mainMBB->addSuccessor(sinkMBB);
20664 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20665 TII->get(X86::PHI), DstReg)
20666 .addReg(mainDstReg).addMBB(mainMBB)
20667 .addReg(restoreDstReg).addMBB(restoreMBB);
20670 if (RegInfo->hasBasePointer(*MF)) {
20671 const bool Uses64BitFramePtr =
20672 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
20673 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
20674 X86FI->setRestoreBasePointer(MF);
20675 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
20676 unsigned BasePtr = RegInfo->getBaseRegister();
20677 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
20678 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
20679 FramePtr, true, X86FI->getRestoreBasePointerOffset())
20680 .setMIFlag(MachineInstr::FrameSetup);
20682 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
20683 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
20684 restoreMBB->addSuccessor(sinkMBB);
20686 MI->eraseFromParent();
20690 MachineBasicBlock *
20691 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
20692 MachineBasicBlock *MBB) const {
20693 DebugLoc DL = MI->getDebugLoc();
20694 MachineFunction *MF = MBB->getParent();
20695 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20696 MachineRegisterInfo &MRI = MF->getRegInfo();
20698 // Memory Reference
20699 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20700 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20702 MVT PVT = getPointerTy(MF->getDataLayout());
20703 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20704 "Invalid Pointer Size!");
20706 const TargetRegisterClass *RC =
20707 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
20708 unsigned Tmp = MRI.createVirtualRegister(RC);
20709 // Since FP is only updated here but NOT referenced, it's treated as GPR.
20710 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20711 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
20712 unsigned SP = RegInfo->getStackRegister();
20714 MachineInstrBuilder MIB;
20716 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20717 const int64_t SPOffset = 2 * PVT.getStoreSize();
20719 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
20720 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
20723 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
20724 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
20725 MIB.addOperand(MI->getOperand(i));
20726 MIB.setMemRefs(MMOBegin, MMOEnd);
20728 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
20729 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20730 if (i == X86::AddrDisp)
20731 MIB.addDisp(MI->getOperand(i), LabelOffset);
20733 MIB.addOperand(MI->getOperand(i));
20735 MIB.setMemRefs(MMOBegin, MMOEnd);
20737 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
20738 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20739 if (i == X86::AddrDisp)
20740 MIB.addDisp(MI->getOperand(i), SPOffset);
20742 MIB.addOperand(MI->getOperand(i));
20744 MIB.setMemRefs(MMOBegin, MMOEnd);
20746 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
20748 MI->eraseFromParent();
20752 // Replace 213-type (isel default) FMA3 instructions with 231-type for
20753 // accumulator loops. Writing back to the accumulator allows the coalescer
20754 // to remove extra copies in the loop.
20755 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
20756 MachineBasicBlock *
20757 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
20758 MachineBasicBlock *MBB) const {
20759 MachineOperand &AddendOp = MI->getOperand(3);
20761 // Bail out early if the addend isn't a register - we can't switch these.
20762 if (!AddendOp.isReg())
20765 MachineFunction &MF = *MBB->getParent();
20766 MachineRegisterInfo &MRI = MF.getRegInfo();
20768 // Check whether the addend is defined by a PHI:
20769 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
20770 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
20771 if (!AddendDef.isPHI())
20774 // Look for the following pattern:
20776 // %addend = phi [%entry, 0], [%loop, %result]
20778 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
20782 // %addend = phi [%entry, 0], [%loop, %result]
20784 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
20786 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
20787 assert(AddendDef.getOperand(i).isReg());
20788 MachineOperand PHISrcOp = AddendDef.getOperand(i);
20789 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
20790 if (&PHISrcInst == MI) {
20791 // Found a matching instruction.
20792 unsigned NewFMAOpc = 0;
20793 switch (MI->getOpcode()) {
20794 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
20795 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
20796 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
20797 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
20798 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
20799 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
20800 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
20801 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
20802 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
20803 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
20804 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
20805 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
20806 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
20807 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
20808 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
20809 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
20810 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
20811 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
20812 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
20813 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
20815 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
20816 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
20817 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
20818 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
20819 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
20820 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
20821 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
20822 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
20823 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
20824 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
20825 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
20826 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
20827 default: llvm_unreachable("Unrecognized FMA variant.");
20830 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
20831 MachineInstrBuilder MIB =
20832 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
20833 .addOperand(MI->getOperand(0))
20834 .addOperand(MI->getOperand(3))
20835 .addOperand(MI->getOperand(2))
20836 .addOperand(MI->getOperand(1));
20837 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
20838 MI->eraseFromParent();
20845 MachineBasicBlock *
20846 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
20847 MachineBasicBlock *BB) const {
20848 switch (MI->getOpcode()) {
20849 default: llvm_unreachable("Unexpected instr type to insert");
20850 case X86::TAILJMPd64:
20851 case X86::TAILJMPr64:
20852 case X86::TAILJMPm64:
20853 case X86::TAILJMPd64_REX:
20854 case X86::TAILJMPr64_REX:
20855 case X86::TAILJMPm64_REX:
20856 llvm_unreachable("TAILJMP64 would not be touched here.");
20857 case X86::TCRETURNdi64:
20858 case X86::TCRETURNri64:
20859 case X86::TCRETURNmi64:
20861 case X86::WIN_ALLOCA:
20862 return EmitLoweredWinAlloca(MI, BB);
20863 case X86::SEG_ALLOCA_32:
20864 case X86::SEG_ALLOCA_64:
20865 return EmitLoweredSegAlloca(MI, BB);
20866 case X86::TLSCall_32:
20867 case X86::TLSCall_64:
20868 return EmitLoweredTLSCall(MI, BB);
20869 case X86::CMOV_FR32:
20870 case X86::CMOV_FR64:
20871 case X86::CMOV_GR8:
20872 case X86::CMOV_GR16:
20873 case X86::CMOV_GR32:
20874 case X86::CMOV_RFP32:
20875 case X86::CMOV_RFP64:
20876 case X86::CMOV_RFP80:
20877 case X86::CMOV_V2F64:
20878 case X86::CMOV_V2I64:
20879 case X86::CMOV_V4F32:
20880 case X86::CMOV_V4F64:
20881 case X86::CMOV_V4I64:
20882 case X86::CMOV_V16F32:
20883 case X86::CMOV_V8F32:
20884 case X86::CMOV_V8F64:
20885 case X86::CMOV_V8I64:
20886 case X86::CMOV_V8I1:
20887 case X86::CMOV_V16I1:
20888 case X86::CMOV_V32I1:
20889 case X86::CMOV_V64I1:
20890 return EmitLoweredSelect(MI, BB);
20892 case X86::RELEASE_FADD32mr:
20893 case X86::RELEASE_FADD64mr:
20894 return EmitLoweredAtomicFP(MI, BB);
20896 case X86::FP32_TO_INT16_IN_MEM:
20897 case X86::FP32_TO_INT32_IN_MEM:
20898 case X86::FP32_TO_INT64_IN_MEM:
20899 case X86::FP64_TO_INT16_IN_MEM:
20900 case X86::FP64_TO_INT32_IN_MEM:
20901 case X86::FP64_TO_INT64_IN_MEM:
20902 case X86::FP80_TO_INT16_IN_MEM:
20903 case X86::FP80_TO_INT32_IN_MEM:
20904 case X86::FP80_TO_INT64_IN_MEM: {
20905 MachineFunction *F = BB->getParent();
20906 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20907 DebugLoc DL = MI->getDebugLoc();
20909 // Change the floating point control register to use "round towards zero"
20910 // mode when truncating to an integer value.
20911 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
20912 addFrameReference(BuildMI(*BB, MI, DL,
20913 TII->get(X86::FNSTCW16m)), CWFrameIdx);
20915 // Load the old value of the high byte of the control word...
20917 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
20918 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
20921 // Set the high part to be round to zero...
20922 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
20925 // Reload the modified control word now...
20926 addFrameReference(BuildMI(*BB, MI, DL,
20927 TII->get(X86::FLDCW16m)), CWFrameIdx);
20929 // Restore the memory image of control word to original value
20930 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
20933 // Get the X86 opcode to use.
20935 switch (MI->getOpcode()) {
20936 default: llvm_unreachable("illegal opcode!");
20937 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
20938 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
20939 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
20940 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
20941 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
20942 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
20943 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
20944 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
20945 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
20949 MachineOperand &Op = MI->getOperand(0);
20951 AM.BaseType = X86AddressMode::RegBase;
20952 AM.Base.Reg = Op.getReg();
20954 AM.BaseType = X86AddressMode::FrameIndexBase;
20955 AM.Base.FrameIndex = Op.getIndex();
20957 Op = MI->getOperand(1);
20959 AM.Scale = Op.getImm();
20960 Op = MI->getOperand(2);
20962 AM.IndexReg = Op.getImm();
20963 Op = MI->getOperand(3);
20964 if (Op.isGlobal()) {
20965 AM.GV = Op.getGlobal();
20967 AM.Disp = Op.getImm();
20969 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
20970 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
20972 // Reload the original control word now.
20973 addFrameReference(BuildMI(*BB, MI, DL,
20974 TII->get(X86::FLDCW16m)), CWFrameIdx);
20976 MI->eraseFromParent(); // The pseudo instruction is gone now.
20979 // String/text processing lowering.
20980 case X86::PCMPISTRM128REG:
20981 case X86::VPCMPISTRM128REG:
20982 case X86::PCMPISTRM128MEM:
20983 case X86::VPCMPISTRM128MEM:
20984 case X86::PCMPESTRM128REG:
20985 case X86::VPCMPESTRM128REG:
20986 case X86::PCMPESTRM128MEM:
20987 case X86::VPCMPESTRM128MEM:
20988 assert(Subtarget->hasSSE42() &&
20989 "Target must have SSE4.2 or AVX features enabled");
20990 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
20992 // String/text processing lowering.
20993 case X86::PCMPISTRIREG:
20994 case X86::VPCMPISTRIREG:
20995 case X86::PCMPISTRIMEM:
20996 case X86::VPCMPISTRIMEM:
20997 case X86::PCMPESTRIREG:
20998 case X86::VPCMPESTRIREG:
20999 case X86::PCMPESTRIMEM:
21000 case X86::VPCMPESTRIMEM:
21001 assert(Subtarget->hasSSE42() &&
21002 "Target must have SSE4.2 or AVX features enabled");
21003 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21005 // Thread synchronization.
21007 return EmitMonitor(MI, BB, Subtarget);
21011 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21013 case X86::VASTART_SAVE_XMM_REGS:
21014 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21016 case X86::VAARG_64:
21017 return EmitVAARG64WithCustomInserter(MI, BB);
21019 case X86::EH_SjLj_SetJmp32:
21020 case X86::EH_SjLj_SetJmp64:
21021 return emitEHSjLjSetJmp(MI, BB);
21023 case X86::EH_SjLj_LongJmp32:
21024 case X86::EH_SjLj_LongJmp64:
21025 return emitEHSjLjLongJmp(MI, BB);
21027 case TargetOpcode::STATEPOINT:
21028 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21029 // this point in the process. We diverge later.
21030 return emitPatchPoint(MI, BB);
21032 case TargetOpcode::STACKMAP:
21033 case TargetOpcode::PATCHPOINT:
21034 return emitPatchPoint(MI, BB);
21036 case X86::VFMADDPDr213r:
21037 case X86::VFMADDPSr213r:
21038 case X86::VFMADDSDr213r:
21039 case X86::VFMADDSSr213r:
21040 case X86::VFMSUBPDr213r:
21041 case X86::VFMSUBPSr213r:
21042 case X86::VFMSUBSDr213r:
21043 case X86::VFMSUBSSr213r:
21044 case X86::VFNMADDPDr213r:
21045 case X86::VFNMADDPSr213r:
21046 case X86::VFNMADDSDr213r:
21047 case X86::VFNMADDSSr213r:
21048 case X86::VFNMSUBPDr213r:
21049 case X86::VFNMSUBPSr213r:
21050 case X86::VFNMSUBSDr213r:
21051 case X86::VFNMSUBSSr213r:
21052 case X86::VFMADDSUBPDr213r:
21053 case X86::VFMADDSUBPSr213r:
21054 case X86::VFMSUBADDPDr213r:
21055 case X86::VFMSUBADDPSr213r:
21056 case X86::VFMADDPDr213rY:
21057 case X86::VFMADDPSr213rY:
21058 case X86::VFMSUBPDr213rY:
21059 case X86::VFMSUBPSr213rY:
21060 case X86::VFNMADDPDr213rY:
21061 case X86::VFNMADDPSr213rY:
21062 case X86::VFNMSUBPDr213rY:
21063 case X86::VFNMSUBPSr213rY:
21064 case X86::VFMADDSUBPDr213rY:
21065 case X86::VFMADDSUBPSr213rY:
21066 case X86::VFMSUBADDPDr213rY:
21067 case X86::VFMSUBADDPSr213rY:
21068 return emitFMA3Instr(MI, BB);
21072 //===----------------------------------------------------------------------===//
21073 // X86 Optimization Hooks
21074 //===----------------------------------------------------------------------===//
21076 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21079 const SelectionDAG &DAG,
21080 unsigned Depth) const {
21081 unsigned BitWidth = KnownZero.getBitWidth();
21082 unsigned Opc = Op.getOpcode();
21083 assert((Opc >= ISD::BUILTIN_OP_END ||
21084 Opc == ISD::INTRINSIC_WO_CHAIN ||
21085 Opc == ISD::INTRINSIC_W_CHAIN ||
21086 Opc == ISD::INTRINSIC_VOID) &&
21087 "Should use MaskedValueIsZero if you don't know whether Op"
21088 " is a target node!");
21090 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21104 // These nodes' second result is a boolean.
21105 if (Op.getResNo() == 0)
21108 case X86ISD::SETCC:
21109 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21111 case ISD::INTRINSIC_WO_CHAIN: {
21112 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21113 unsigned NumLoBits = 0;
21116 case Intrinsic::x86_sse_movmsk_ps:
21117 case Intrinsic::x86_avx_movmsk_ps_256:
21118 case Intrinsic::x86_sse2_movmsk_pd:
21119 case Intrinsic::x86_avx_movmsk_pd_256:
21120 case Intrinsic::x86_mmx_pmovmskb:
21121 case Intrinsic::x86_sse2_pmovmskb_128:
21122 case Intrinsic::x86_avx2_pmovmskb: {
21123 // High bits of movmskp{s|d}, pmovmskb are known zero.
21125 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21126 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21127 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21128 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21129 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21130 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21131 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21132 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21134 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21143 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21145 const SelectionDAG &,
21146 unsigned Depth) const {
21147 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21148 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21149 return Op.getValueType().getScalarType().getSizeInBits();
21155 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21156 /// node is a GlobalAddress + offset.
21157 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21158 const GlobalValue* &GA,
21159 int64_t &Offset) const {
21160 if (N->getOpcode() == X86ISD::Wrapper) {
21161 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21162 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21163 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21167 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21170 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21171 /// same as extracting the high 128-bit part of 256-bit vector and then
21172 /// inserting the result into the low part of a new 256-bit vector
21173 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21174 EVT VT = SVOp->getValueType(0);
21175 unsigned NumElems = VT.getVectorNumElements();
21177 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21178 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21179 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21180 SVOp->getMaskElt(j) >= 0)
21186 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21187 /// same as extracting the low 128-bit part of 256-bit vector and then
21188 /// inserting the result into the high part of a new 256-bit vector
21189 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21190 EVT VT = SVOp->getValueType(0);
21191 unsigned NumElems = VT.getVectorNumElements();
21193 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21194 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21195 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21196 SVOp->getMaskElt(j) >= 0)
21202 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21203 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21204 TargetLowering::DAGCombinerInfo &DCI,
21205 const X86Subtarget* Subtarget) {
21207 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21208 SDValue V1 = SVOp->getOperand(0);
21209 SDValue V2 = SVOp->getOperand(1);
21210 EVT VT = SVOp->getValueType(0);
21211 unsigned NumElems = VT.getVectorNumElements();
21213 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21214 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21218 // V UNDEF BUILD_VECTOR UNDEF
21220 // CONCAT_VECTOR CONCAT_VECTOR
21223 // RESULT: V + zero extended
21225 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21226 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21227 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21230 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21233 // To match the shuffle mask, the first half of the mask should
21234 // be exactly the first vector, and all the rest a splat with the
21235 // first element of the second one.
21236 for (unsigned i = 0; i != NumElems/2; ++i)
21237 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21238 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21241 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21242 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21243 if (Ld->hasNUsesOfValue(1, 0)) {
21244 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21245 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21247 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21249 Ld->getPointerInfo(),
21250 Ld->getAlignment(),
21251 false/*isVolatile*/, true/*ReadMem*/,
21252 false/*WriteMem*/);
21254 // Make sure the newly-created LOAD is in the same position as Ld in
21255 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21256 // and update uses of Ld's output chain to use the TokenFactor.
21257 if (Ld->hasAnyUseOfValue(1)) {
21258 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21259 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21260 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21261 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21262 SDValue(ResNode.getNode(), 1));
21265 return DAG.getBitcast(VT, ResNode);
21269 // Emit a zeroed vector and insert the desired subvector on its
21271 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21272 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21273 return DCI.CombineTo(N, InsV);
21276 //===--------------------------------------------------------------------===//
21277 // Combine some shuffles into subvector extracts and inserts:
21280 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21281 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21282 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21283 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21284 return DCI.CombineTo(N, InsV);
21287 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21288 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21289 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21290 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21291 return DCI.CombineTo(N, InsV);
21297 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21300 /// This is the leaf of the recursive combinine below. When we have found some
21301 /// chain of single-use x86 shuffle instructions and accumulated the combined
21302 /// shuffle mask represented by them, this will try to pattern match that mask
21303 /// into either a single instruction if there is a special purpose instruction
21304 /// for this operation, or into a PSHUFB instruction which is a fully general
21305 /// instruction but should only be used to replace chains over a certain depth.
21306 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21307 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21308 TargetLowering::DAGCombinerInfo &DCI,
21309 const X86Subtarget *Subtarget) {
21310 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21312 // Find the operand that enters the chain. Note that multiple uses are OK
21313 // here, we're not going to remove the operand we find.
21314 SDValue Input = Op.getOperand(0);
21315 while (Input.getOpcode() == ISD::BITCAST)
21316 Input = Input.getOperand(0);
21318 MVT VT = Input.getSimpleValueType();
21319 MVT RootVT = Root.getSimpleValueType();
21322 // Just remove no-op shuffle masks.
21323 if (Mask.size() == 1) {
21324 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21329 // Use the float domain if the operand type is a floating point type.
21330 bool FloatDomain = VT.isFloatingPoint();
21332 // For floating point shuffles, we don't have free copies in the shuffle
21333 // instructions or the ability to load as part of the instruction, so
21334 // canonicalize their shuffles to UNPCK or MOV variants.
21336 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21337 // vectors because it can have a load folded into it that UNPCK cannot. This
21338 // doesn't preclude something switching to the shorter encoding post-RA.
21340 // FIXME: Should teach these routines about AVX vector widths.
21341 if (FloatDomain && VT.getSizeInBits() == 128) {
21342 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21343 bool Lo = Mask.equals({0, 0});
21346 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21347 // is no slower than UNPCKLPD but has the option to fold the input operand
21348 // into even an unaligned memory load.
21349 if (Lo && Subtarget->hasSSE3()) {
21350 Shuffle = X86ISD::MOVDDUP;
21351 ShuffleVT = MVT::v2f64;
21353 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21354 // than the UNPCK variants.
21355 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21356 ShuffleVT = MVT::v4f32;
21358 if (Depth == 1 && Root->getOpcode() == Shuffle)
21359 return false; // Nothing to do!
21360 Op = DAG.getBitcast(ShuffleVT, Input);
21361 DCI.AddToWorklist(Op.getNode());
21362 if (Shuffle == X86ISD::MOVDDUP)
21363 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21365 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21366 DCI.AddToWorklist(Op.getNode());
21367 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21371 if (Subtarget->hasSSE3() &&
21372 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21373 bool Lo = Mask.equals({0, 0, 2, 2});
21374 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21375 MVT ShuffleVT = MVT::v4f32;
21376 if (Depth == 1 && Root->getOpcode() == Shuffle)
21377 return false; // Nothing to do!
21378 Op = DAG.getBitcast(ShuffleVT, Input);
21379 DCI.AddToWorklist(Op.getNode());
21380 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21381 DCI.AddToWorklist(Op.getNode());
21382 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21386 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21387 bool Lo = Mask.equals({0, 0, 1, 1});
21388 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21389 MVT ShuffleVT = MVT::v4f32;
21390 if (Depth == 1 && Root->getOpcode() == Shuffle)
21391 return false; // Nothing to do!
21392 Op = DAG.getBitcast(ShuffleVT, Input);
21393 DCI.AddToWorklist(Op.getNode());
21394 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21395 DCI.AddToWorklist(Op.getNode());
21396 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21402 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21403 // variants as none of these have single-instruction variants that are
21404 // superior to the UNPCK formulation.
21405 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21406 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21407 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21408 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21410 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21411 bool Lo = Mask[0] == 0;
21412 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21413 if (Depth == 1 && Root->getOpcode() == Shuffle)
21414 return false; // Nothing to do!
21416 switch (Mask.size()) {
21418 ShuffleVT = MVT::v8i16;
21421 ShuffleVT = MVT::v16i8;
21424 llvm_unreachable("Impossible mask size!");
21426 Op = DAG.getBitcast(ShuffleVT, Input);
21427 DCI.AddToWorklist(Op.getNode());
21428 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21429 DCI.AddToWorklist(Op.getNode());
21430 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21435 // Don't try to re-form single instruction chains under any circumstances now
21436 // that we've done encoding canonicalization for them.
21440 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21441 // can replace them with a single PSHUFB instruction profitably. Intel's
21442 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21443 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21444 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21445 SmallVector<SDValue, 16> PSHUFBMask;
21446 int NumBytes = VT.getSizeInBits() / 8;
21447 int Ratio = NumBytes / Mask.size();
21448 for (int i = 0; i < NumBytes; ++i) {
21449 if (Mask[i / Ratio] == SM_SentinelUndef) {
21450 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21453 int M = Mask[i / Ratio] != SM_SentinelZero
21454 ? Ratio * Mask[i / Ratio] + i % Ratio
21456 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21458 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21459 Op = DAG.getBitcast(ByteVT, Input);
21460 DCI.AddToWorklist(Op.getNode());
21461 SDValue PSHUFBMaskOp =
21462 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21463 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21464 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21465 DCI.AddToWorklist(Op.getNode());
21466 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21471 // Failed to find any combines.
21475 /// \brief Fully generic combining of x86 shuffle instructions.
21477 /// This should be the last combine run over the x86 shuffle instructions. Once
21478 /// they have been fully optimized, this will recursively consider all chains
21479 /// of single-use shuffle instructions, build a generic model of the cumulative
21480 /// shuffle operation, and check for simpler instructions which implement this
21481 /// operation. We use this primarily for two purposes:
21483 /// 1) Collapse generic shuffles to specialized single instructions when
21484 /// equivalent. In most cases, this is just an encoding size win, but
21485 /// sometimes we will collapse multiple generic shuffles into a single
21486 /// special-purpose shuffle.
21487 /// 2) Look for sequences of shuffle instructions with 3 or more total
21488 /// instructions, and replace them with the slightly more expensive SSSE3
21489 /// PSHUFB instruction if available. We do this as the last combining step
21490 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21491 /// a suitable short sequence of other instructions. The PHUFB will either
21492 /// use a register or have to read from memory and so is slightly (but only
21493 /// slightly) more expensive than the other shuffle instructions.
21495 /// Because this is inherently a quadratic operation (for each shuffle in
21496 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21497 /// This should never be an issue in practice as the shuffle lowering doesn't
21498 /// produce sequences of more than 8 instructions.
21500 /// FIXME: We will currently miss some cases where the redundant shuffling
21501 /// would simplify under the threshold for PSHUFB formation because of
21502 /// combine-ordering. To fix this, we should do the redundant instruction
21503 /// combining in this recursive walk.
21504 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21505 ArrayRef<int> RootMask,
21506 int Depth, bool HasPSHUFB,
21508 TargetLowering::DAGCombinerInfo &DCI,
21509 const X86Subtarget *Subtarget) {
21510 // Bound the depth of our recursive combine because this is ultimately
21511 // quadratic in nature.
21515 // Directly rip through bitcasts to find the underlying operand.
21516 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21517 Op = Op.getOperand(0);
21519 MVT VT = Op.getSimpleValueType();
21520 if (!VT.isVector())
21521 return false; // Bail if we hit a non-vector.
21523 assert(Root.getSimpleValueType().isVector() &&
21524 "Shuffles operate on vector types!");
21525 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21526 "Can only combine shuffles of the same vector register size.");
21528 if (!isTargetShuffle(Op.getOpcode()))
21530 SmallVector<int, 16> OpMask;
21532 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21533 // We only can combine unary shuffles which we can decode the mask for.
21534 if (!HaveMask || !IsUnary)
21537 assert(VT.getVectorNumElements() == OpMask.size() &&
21538 "Different mask size from vector size!");
21539 assert(((RootMask.size() > OpMask.size() &&
21540 RootMask.size() % OpMask.size() == 0) ||
21541 (OpMask.size() > RootMask.size() &&
21542 OpMask.size() % RootMask.size() == 0) ||
21543 OpMask.size() == RootMask.size()) &&
21544 "The smaller number of elements must divide the larger.");
21545 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21546 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21547 assert(((RootRatio == 1 && OpRatio == 1) ||
21548 (RootRatio == 1) != (OpRatio == 1)) &&
21549 "Must not have a ratio for both incoming and op masks!");
21551 SmallVector<int, 16> Mask;
21552 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21554 // Merge this shuffle operation's mask into our accumulated mask. Note that
21555 // this shuffle's mask will be the first applied to the input, followed by the
21556 // root mask to get us all the way to the root value arrangement. The reason
21557 // for this order is that we are recursing up the operation chain.
21558 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21559 int RootIdx = i / RootRatio;
21560 if (RootMask[RootIdx] < 0) {
21561 // This is a zero or undef lane, we're done.
21562 Mask.push_back(RootMask[RootIdx]);
21566 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
21567 int OpIdx = RootMaskedIdx / OpRatio;
21568 if (OpMask[OpIdx] < 0) {
21569 // The incoming lanes are zero or undef, it doesn't matter which ones we
21571 Mask.push_back(OpMask[OpIdx]);
21575 // Ok, we have non-zero lanes, map them through.
21576 Mask.push_back(OpMask[OpIdx] * OpRatio +
21577 RootMaskedIdx % OpRatio);
21580 // See if we can recurse into the operand to combine more things.
21581 switch (Op.getOpcode()) {
21582 case X86ISD::PSHUFB:
21584 case X86ISD::PSHUFD:
21585 case X86ISD::PSHUFHW:
21586 case X86ISD::PSHUFLW:
21587 if (Op.getOperand(0).hasOneUse() &&
21588 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21589 HasPSHUFB, DAG, DCI, Subtarget))
21593 case X86ISD::UNPCKL:
21594 case X86ISD::UNPCKH:
21595 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
21596 // We can't check for single use, we have to check that this shuffle is the only user.
21597 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
21598 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21599 HasPSHUFB, DAG, DCI, Subtarget))
21604 // Minor canonicalization of the accumulated shuffle mask to make it easier
21605 // to match below. All this does is detect masks with squential pairs of
21606 // elements, and shrink them to the half-width mask. It does this in a loop
21607 // so it will reduce the size of the mask to the minimal width mask which
21608 // performs an equivalent shuffle.
21609 SmallVector<int, 16> WidenedMask;
21610 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
21611 Mask = std::move(WidenedMask);
21612 WidenedMask.clear();
21615 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
21619 /// \brief Get the PSHUF-style mask from PSHUF node.
21621 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
21622 /// PSHUF-style masks that can be reused with such instructions.
21623 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
21624 MVT VT = N.getSimpleValueType();
21625 SmallVector<int, 4> Mask;
21627 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
21631 // If we have more than 128-bits, only the low 128-bits of shuffle mask
21632 // matter. Check that the upper masks are repeats and remove them.
21633 if (VT.getSizeInBits() > 128) {
21634 int LaneElts = 128 / VT.getScalarSizeInBits();
21636 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
21637 for (int j = 0; j < LaneElts; ++j)
21638 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
21639 "Mask doesn't repeat in high 128-bit lanes!");
21641 Mask.resize(LaneElts);
21644 switch (N.getOpcode()) {
21645 case X86ISD::PSHUFD:
21647 case X86ISD::PSHUFLW:
21650 case X86ISD::PSHUFHW:
21651 Mask.erase(Mask.begin(), Mask.begin() + 4);
21652 for (int &M : Mask)
21656 llvm_unreachable("No valid shuffle instruction found!");
21660 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
21662 /// We walk up the chain and look for a combinable shuffle, skipping over
21663 /// shuffles that we could hoist this shuffle's transformation past without
21664 /// altering anything.
21666 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
21668 TargetLowering::DAGCombinerInfo &DCI) {
21669 assert(N.getOpcode() == X86ISD::PSHUFD &&
21670 "Called with something other than an x86 128-bit half shuffle!");
21673 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
21674 // of the shuffles in the chain so that we can form a fresh chain to replace
21676 SmallVector<SDValue, 8> Chain;
21677 SDValue V = N.getOperand(0);
21678 for (; V.hasOneUse(); V = V.getOperand(0)) {
21679 switch (V.getOpcode()) {
21681 return SDValue(); // Nothing combined!
21684 // Skip bitcasts as we always know the type for the target specific
21688 case X86ISD::PSHUFD:
21689 // Found another dword shuffle.
21692 case X86ISD::PSHUFLW:
21693 // Check that the low words (being shuffled) are the identity in the
21694 // dword shuffle, and the high words are self-contained.
21695 if (Mask[0] != 0 || Mask[1] != 1 ||
21696 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
21699 Chain.push_back(V);
21702 case X86ISD::PSHUFHW:
21703 // Check that the high words (being shuffled) are the identity in the
21704 // dword shuffle, and the low words are self-contained.
21705 if (Mask[2] != 2 || Mask[3] != 3 ||
21706 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
21709 Chain.push_back(V);
21712 case X86ISD::UNPCKL:
21713 case X86ISD::UNPCKH:
21714 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
21715 // shuffle into a preceding word shuffle.
21716 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
21717 V.getSimpleValueType().getScalarType() != MVT::i16)
21720 // Search for a half-shuffle which we can combine with.
21721 unsigned CombineOp =
21722 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
21723 if (V.getOperand(0) != V.getOperand(1) ||
21724 !V->isOnlyUserOf(V.getOperand(0).getNode()))
21726 Chain.push_back(V);
21727 V = V.getOperand(0);
21729 switch (V.getOpcode()) {
21731 return SDValue(); // Nothing to combine.
21733 case X86ISD::PSHUFLW:
21734 case X86ISD::PSHUFHW:
21735 if (V.getOpcode() == CombineOp)
21738 Chain.push_back(V);
21742 V = V.getOperand(0);
21746 } while (V.hasOneUse());
21749 // Break out of the loop if we break out of the switch.
21753 if (!V.hasOneUse())
21754 // We fell out of the loop without finding a viable combining instruction.
21757 // Merge this node's mask and our incoming mask.
21758 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21759 for (int &M : Mask)
21761 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
21762 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21764 // Rebuild the chain around this new shuffle.
21765 while (!Chain.empty()) {
21766 SDValue W = Chain.pop_back_val();
21768 if (V.getValueType() != W.getOperand(0).getValueType())
21769 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
21771 switch (W.getOpcode()) {
21773 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
21775 case X86ISD::UNPCKL:
21776 case X86ISD::UNPCKH:
21777 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
21780 case X86ISD::PSHUFD:
21781 case X86ISD::PSHUFLW:
21782 case X86ISD::PSHUFHW:
21783 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
21787 if (V.getValueType() != N.getValueType())
21788 V = DAG.getBitcast(N.getValueType(), V);
21790 // Return the new chain to replace N.
21794 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
21796 /// We walk up the chain, skipping shuffles of the other half and looking
21797 /// through shuffles which switch halves trying to find a shuffle of the same
21798 /// pair of dwords.
21799 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
21801 TargetLowering::DAGCombinerInfo &DCI) {
21803 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
21804 "Called with something other than an x86 128-bit half shuffle!");
21806 unsigned CombineOpcode = N.getOpcode();
21808 // Walk up a single-use chain looking for a combinable shuffle.
21809 SDValue V = N.getOperand(0);
21810 for (; V.hasOneUse(); V = V.getOperand(0)) {
21811 switch (V.getOpcode()) {
21813 return false; // Nothing combined!
21816 // Skip bitcasts as we always know the type for the target specific
21820 case X86ISD::PSHUFLW:
21821 case X86ISD::PSHUFHW:
21822 if (V.getOpcode() == CombineOpcode)
21825 // Other-half shuffles are no-ops.
21828 // Break out of the loop if we break out of the switch.
21832 if (!V.hasOneUse())
21833 // We fell out of the loop without finding a viable combining instruction.
21836 // Combine away the bottom node as its shuffle will be accumulated into
21837 // a preceding shuffle.
21838 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21840 // Record the old value.
21843 // Merge this node's mask and our incoming mask (adjusted to account for all
21844 // the pshufd instructions encountered).
21845 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21846 for (int &M : Mask)
21848 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
21849 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21851 // Check that the shuffles didn't cancel each other out. If not, we need to
21852 // combine to the new one.
21854 // Replace the combinable shuffle with the combined one, updating all users
21855 // so that we re-evaluate the chain here.
21856 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
21861 /// \brief Try to combine x86 target specific shuffles.
21862 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
21863 TargetLowering::DAGCombinerInfo &DCI,
21864 const X86Subtarget *Subtarget) {
21866 MVT VT = N.getSimpleValueType();
21867 SmallVector<int, 4> Mask;
21869 switch (N.getOpcode()) {
21870 case X86ISD::PSHUFD:
21871 case X86ISD::PSHUFLW:
21872 case X86ISD::PSHUFHW:
21873 Mask = getPSHUFShuffleMask(N);
21874 assert(Mask.size() == 4);
21880 // Nuke no-op shuffles that show up after combining.
21881 if (isNoopShuffleMask(Mask))
21882 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21884 // Look for simplifications involving one or two shuffle instructions.
21885 SDValue V = N.getOperand(0);
21886 switch (N.getOpcode()) {
21889 case X86ISD::PSHUFLW:
21890 case X86ISD::PSHUFHW:
21891 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
21893 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
21894 return SDValue(); // We combined away this shuffle, so we're done.
21896 // See if this reduces to a PSHUFD which is no more expensive and can
21897 // combine with more operations. Note that it has to at least flip the
21898 // dwords as otherwise it would have been removed as a no-op.
21899 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
21900 int DMask[] = {0, 1, 2, 3};
21901 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
21902 DMask[DOffset + 0] = DOffset + 1;
21903 DMask[DOffset + 1] = DOffset + 0;
21904 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
21905 V = DAG.getBitcast(DVT, V);
21906 DCI.AddToWorklist(V.getNode());
21907 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
21908 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
21909 DCI.AddToWorklist(V.getNode());
21910 return DAG.getBitcast(VT, V);
21913 // Look for shuffle patterns which can be implemented as a single unpack.
21914 // FIXME: This doesn't handle the location of the PSHUFD generically, and
21915 // only works when we have a PSHUFD followed by two half-shuffles.
21916 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
21917 (V.getOpcode() == X86ISD::PSHUFLW ||
21918 V.getOpcode() == X86ISD::PSHUFHW) &&
21919 V.getOpcode() != N.getOpcode() &&
21921 SDValue D = V.getOperand(0);
21922 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
21923 D = D.getOperand(0);
21924 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
21925 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21926 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
21927 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21928 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21930 for (int i = 0; i < 4; ++i) {
21931 WordMask[i + NOffset] = Mask[i] + NOffset;
21932 WordMask[i + VOffset] = VMask[i] + VOffset;
21934 // Map the word mask through the DWord mask.
21936 for (int i = 0; i < 8; ++i)
21937 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
21938 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21939 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
21940 // We can replace all three shuffles with an unpack.
21941 V = DAG.getBitcast(VT, D.getOperand(0));
21942 DCI.AddToWorklist(V.getNode());
21943 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
21952 case X86ISD::PSHUFD:
21953 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
21962 /// \brief Try to combine a shuffle into a target-specific add-sub node.
21964 /// We combine this directly on the abstract vector shuffle nodes so it is
21965 /// easier to generically match. We also insert dummy vector shuffle nodes for
21966 /// the operands which explicitly discard the lanes which are unused by this
21967 /// operation to try to flow through the rest of the combiner the fact that
21968 /// they're unused.
21969 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
21971 EVT VT = N->getValueType(0);
21973 // We only handle target-independent shuffles.
21974 // FIXME: It would be easy and harmless to use the target shuffle mask
21975 // extraction tool to support more.
21976 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
21979 auto *SVN = cast<ShuffleVectorSDNode>(N);
21980 ArrayRef<int> Mask = SVN->getMask();
21981 SDValue V1 = N->getOperand(0);
21982 SDValue V2 = N->getOperand(1);
21984 // We require the first shuffle operand to be the SUB node, and the second to
21985 // be the ADD node.
21986 // FIXME: We should support the commuted patterns.
21987 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
21990 // If there are other uses of these operations we can't fold them.
21991 if (!V1->hasOneUse() || !V2->hasOneUse())
21994 // Ensure that both operations have the same operands. Note that we can
21995 // commute the FADD operands.
21996 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
21997 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
21998 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22001 // We're looking for blends between FADD and FSUB nodes. We insist on these
22002 // nodes being lined up in a specific expected pattern.
22003 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22004 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22005 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22008 // Only specific types are legal at this point, assert so we notice if and
22009 // when these change.
22010 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22011 VT == MVT::v4f64) &&
22012 "Unknown vector type encountered!");
22014 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22017 /// PerformShuffleCombine - Performs several different shuffle combines.
22018 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22019 TargetLowering::DAGCombinerInfo &DCI,
22020 const X86Subtarget *Subtarget) {
22022 SDValue N0 = N->getOperand(0);
22023 SDValue N1 = N->getOperand(1);
22024 EVT VT = N->getValueType(0);
22026 // Don't create instructions with illegal types after legalize types has run.
22027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22028 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22031 // If we have legalized the vector types, look for blends of FADD and FSUB
22032 // nodes that we can fuse into an ADDSUB node.
22033 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22034 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22037 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22038 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22039 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22040 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22042 // During Type Legalization, when promoting illegal vector types,
22043 // the backend might introduce new shuffle dag nodes and bitcasts.
22045 // This code performs the following transformation:
22046 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22047 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22049 // We do this only if both the bitcast and the BINOP dag nodes have
22050 // one use. Also, perform this transformation only if the new binary
22051 // operation is legal. This is to avoid introducing dag nodes that
22052 // potentially need to be further expanded (or custom lowered) into a
22053 // less optimal sequence of dag nodes.
22054 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22055 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22056 N0.getOpcode() == ISD::BITCAST) {
22057 SDValue BC0 = N0.getOperand(0);
22058 EVT SVT = BC0.getValueType();
22059 unsigned Opcode = BC0.getOpcode();
22060 unsigned NumElts = VT.getVectorNumElements();
22062 if (BC0.hasOneUse() && SVT.isVector() &&
22063 SVT.getVectorNumElements() * 2 == NumElts &&
22064 TLI.isOperationLegal(Opcode, VT)) {
22065 bool CanFold = false;
22077 unsigned SVTNumElts = SVT.getVectorNumElements();
22078 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22079 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22080 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22081 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22082 CanFold = SVOp->getMaskElt(i) < 0;
22085 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22086 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22087 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22088 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22093 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22094 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22095 // consecutive, non-overlapping, and in the right order.
22096 SmallVector<SDValue, 16> Elts;
22097 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22098 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22100 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22103 if (isTargetShuffle(N->getOpcode())) {
22105 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22106 if (Shuffle.getNode())
22109 // Try recursively combining arbitrary sequences of x86 shuffle
22110 // instructions into higher-order shuffles. We do this after combining
22111 // specific PSHUF instruction sequences into their minimal form so that we
22112 // can evaluate how many specialized shuffle instructions are involved in
22113 // a particular chain.
22114 SmallVector<int, 1> NonceMask; // Just a placeholder.
22115 NonceMask.push_back(0);
22116 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22117 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22119 return SDValue(); // This routine will use CombineTo to replace N.
22125 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22126 /// specific shuffle of a load can be folded into a single element load.
22127 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22128 /// shuffles have been custom lowered so we need to handle those here.
22129 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22130 TargetLowering::DAGCombinerInfo &DCI) {
22131 if (DCI.isBeforeLegalizeOps())
22134 SDValue InVec = N->getOperand(0);
22135 SDValue EltNo = N->getOperand(1);
22137 if (!isa<ConstantSDNode>(EltNo))
22140 EVT OriginalVT = InVec.getValueType();
22142 if (InVec.getOpcode() == ISD::BITCAST) {
22143 // Don't duplicate a load with other uses.
22144 if (!InVec.hasOneUse())
22146 EVT BCVT = InVec.getOperand(0).getValueType();
22147 if (!BCVT.isVector() ||
22148 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22150 InVec = InVec.getOperand(0);
22153 EVT CurrentVT = InVec.getValueType();
22155 if (!isTargetShuffle(InVec.getOpcode()))
22158 // Don't duplicate a load with other uses.
22159 if (!InVec.hasOneUse())
22162 SmallVector<int, 16> ShuffleMask;
22164 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22165 ShuffleMask, UnaryShuffle))
22168 // Select the input vector, guarding against out of range extract vector.
22169 unsigned NumElems = CurrentVT.getVectorNumElements();
22170 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22171 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22172 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22173 : InVec.getOperand(1);
22175 // If inputs to shuffle are the same for both ops, then allow 2 uses
22176 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22177 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22179 if (LdNode.getOpcode() == ISD::BITCAST) {
22180 // Don't duplicate a load with other uses.
22181 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22184 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22185 LdNode = LdNode.getOperand(0);
22188 if (!ISD::isNormalLoad(LdNode.getNode()))
22191 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22193 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22196 EVT EltVT = N->getValueType(0);
22197 // If there's a bitcast before the shuffle, check if the load type and
22198 // alignment is valid.
22199 unsigned Align = LN0->getAlignment();
22200 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22201 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
22202 EltVT.getTypeForEVT(*DAG.getContext()));
22204 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22207 // All checks match so transform back to vector_shuffle so that DAG combiner
22208 // can finish the job
22211 // Create shuffle node taking into account the case that its a unary shuffle
22212 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22213 : InVec.getOperand(1);
22214 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22215 InVec.getOperand(0), Shuffle,
22217 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
22218 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22222 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
22223 /// special and don't usually play with other vector types, it's better to
22224 /// handle them early to be sure we emit efficient code by avoiding
22225 /// store-load conversions.
22226 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
22227 if (N->getValueType(0) != MVT::x86mmx ||
22228 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
22229 N->getOperand(0)->getValueType(0) != MVT::v2i32)
22232 SDValue V = N->getOperand(0);
22233 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22234 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22235 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22236 N->getValueType(0), V.getOperand(0));
22241 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22242 /// generation and convert it from being a bunch of shuffles and extracts
22243 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22244 /// storing the value and loading scalars back, while for x64 we should
22245 /// use 64-bit extracts and shifts.
22246 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22247 TargetLowering::DAGCombinerInfo &DCI) {
22248 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22251 SDValue InputVector = N->getOperand(0);
22252 SDLoc dl(InputVector);
22253 // Detect mmx to i32 conversion through a v2i32 elt extract.
22254 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22255 N->getValueType(0) == MVT::i32 &&
22256 InputVector.getValueType() == MVT::v2i32) {
22258 // The bitcast source is a direct mmx result.
22259 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22260 if (MMXSrc.getValueType() == MVT::x86mmx)
22261 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22262 N->getValueType(0),
22263 InputVector.getNode()->getOperand(0));
22265 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22266 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22267 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22268 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22269 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22270 MMXSrcOp.getValueType() == MVT::v1i64 &&
22271 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22272 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22273 N->getValueType(0),
22274 MMXSrcOp.getOperand(0));
22277 EVT VT = N->getValueType(0);
22279 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22280 InputVector.getOpcode() == ISD::BITCAST &&
22281 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22282 uint64_t ExtractedElt =
22283 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22284 uint64_t InputValue =
22285 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22286 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22287 return DAG.getConstant(Res, dl, MVT::i1);
22289 // Only operate on vectors of 4 elements, where the alternative shuffling
22290 // gets to be more expensive.
22291 if (InputVector.getValueType() != MVT::v4i32)
22294 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22295 // single use which is a sign-extend or zero-extend, and all elements are
22297 SmallVector<SDNode *, 4> Uses;
22298 unsigned ExtractedElements = 0;
22299 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22300 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22301 if (UI.getUse().getResNo() != InputVector.getResNo())
22304 SDNode *Extract = *UI;
22305 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22308 if (Extract->getValueType(0) != MVT::i32)
22310 if (!Extract->hasOneUse())
22312 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22313 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22315 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22318 // Record which element was extracted.
22319 ExtractedElements |=
22320 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22322 Uses.push_back(Extract);
22325 // If not all the elements were used, this may not be worthwhile.
22326 if (ExtractedElements != 15)
22329 // Ok, we've now decided to do the transformation.
22330 // If 64-bit shifts are legal, use the extract-shift sequence,
22331 // otherwise bounce the vector off the cache.
22332 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22335 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22336 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22337 auto &DL = DAG.getDataLayout();
22338 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22339 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22340 DAG.getConstant(0, dl, VecIdxTy));
22341 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22342 DAG.getConstant(1, dl, VecIdxTy));
22344 SDValue ShAmt = DAG.getConstant(
22345 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22346 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22347 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22348 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22349 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22350 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22351 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22353 // Store the value to a temporary stack slot.
22354 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22355 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22356 MachinePointerInfo(), false, false, 0);
22358 EVT ElementType = InputVector.getValueType().getVectorElementType();
22359 unsigned EltSize = ElementType.getSizeInBits() / 8;
22361 // Replace each use (extract) with a load of the appropriate element.
22362 for (unsigned i = 0; i < 4; ++i) {
22363 uint64_t Offset = EltSize * i;
22364 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22365 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22367 SDValue ScalarAddr =
22368 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22370 // Load the scalar.
22371 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22372 ScalarAddr, MachinePointerInfo(),
22373 false, false, false, 0);
22378 // Replace the extracts
22379 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22380 UE = Uses.end(); UI != UE; ++UI) {
22381 SDNode *Extract = *UI;
22383 SDValue Idx = Extract->getOperand(1);
22384 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22385 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22388 // The replacement was made in place; don't return anything.
22393 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22394 const X86Subtarget *Subtarget) {
22396 SDValue Cond = N->getOperand(0);
22397 SDValue LHS = N->getOperand(1);
22398 SDValue RHS = N->getOperand(2);
22400 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22401 SDValue CondSrc = Cond->getOperand(0);
22402 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22403 Cond = CondSrc->getOperand(0);
22406 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22409 // A vselect where all conditions and data are constants can be optimized into
22410 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22411 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22412 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22415 unsigned MaskValue = 0;
22416 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22419 MVT VT = N->getSimpleValueType(0);
22420 unsigned NumElems = VT.getVectorNumElements();
22421 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22422 for (unsigned i = 0; i < NumElems; ++i) {
22423 // Be sure we emit undef where we can.
22424 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22425 ShuffleMask[i] = -1;
22427 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22430 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22431 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22433 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22436 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22438 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22439 TargetLowering::DAGCombinerInfo &DCI,
22440 const X86Subtarget *Subtarget) {
22442 SDValue Cond = N->getOperand(0);
22443 // Get the LHS/RHS of the select.
22444 SDValue LHS = N->getOperand(1);
22445 SDValue RHS = N->getOperand(2);
22446 EVT VT = LHS.getValueType();
22447 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22449 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22450 // instructions match the semantics of the common C idiom x<y?x:y but not
22451 // x<=y?x:y, because of how they handle negative zero (which can be
22452 // ignored in unsafe-math mode).
22453 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22454 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22455 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22456 (Subtarget->hasSSE2() ||
22457 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22458 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22460 unsigned Opcode = 0;
22461 // Check for x CC y ? x : y.
22462 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22463 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22467 // Converting this to a min would handle NaNs incorrectly, and swapping
22468 // the operands would cause it to handle comparisons between positive
22469 // and negative zero incorrectly.
22470 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22471 if (!DAG.getTarget().Options.UnsafeFPMath &&
22472 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22474 std::swap(LHS, RHS);
22476 Opcode = X86ISD::FMIN;
22479 // Converting this to a min would handle comparisons between positive
22480 // and negative zero incorrectly.
22481 if (!DAG.getTarget().Options.UnsafeFPMath &&
22482 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22484 Opcode = X86ISD::FMIN;
22487 // Converting this to a min would handle both negative zeros and NaNs
22488 // incorrectly, but we can swap the operands to fix both.
22489 std::swap(LHS, RHS);
22493 Opcode = X86ISD::FMIN;
22497 // Converting this to a max would handle comparisons between positive
22498 // and negative zero incorrectly.
22499 if (!DAG.getTarget().Options.UnsafeFPMath &&
22500 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22502 Opcode = X86ISD::FMAX;
22505 // Converting this to a max would handle NaNs incorrectly, and swapping
22506 // the operands would cause it to handle comparisons between positive
22507 // and negative zero incorrectly.
22508 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22509 if (!DAG.getTarget().Options.UnsafeFPMath &&
22510 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22512 std::swap(LHS, RHS);
22514 Opcode = X86ISD::FMAX;
22517 // Converting this to a max would handle both negative zeros and NaNs
22518 // incorrectly, but we can swap the operands to fix both.
22519 std::swap(LHS, RHS);
22523 Opcode = X86ISD::FMAX;
22526 // Check for x CC y ? y : x -- a min/max with reversed arms.
22527 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22528 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22532 // Converting this to a min would handle comparisons between positive
22533 // and negative zero incorrectly, and swapping the operands would
22534 // cause it to handle NaNs incorrectly.
22535 if (!DAG.getTarget().Options.UnsafeFPMath &&
22536 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22537 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22539 std::swap(LHS, RHS);
22541 Opcode = X86ISD::FMIN;
22544 // Converting this to a min would handle NaNs incorrectly.
22545 if (!DAG.getTarget().Options.UnsafeFPMath &&
22546 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22548 Opcode = X86ISD::FMIN;
22551 // Converting this to a min would handle both negative zeros and NaNs
22552 // incorrectly, but we can swap the operands to fix both.
22553 std::swap(LHS, RHS);
22557 Opcode = X86ISD::FMIN;
22561 // Converting this to a max would handle NaNs incorrectly.
22562 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22564 Opcode = X86ISD::FMAX;
22567 // Converting this to a max would handle comparisons between positive
22568 // and negative zero incorrectly, and swapping the operands would
22569 // cause it to handle NaNs incorrectly.
22570 if (!DAG.getTarget().Options.UnsafeFPMath &&
22571 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
22572 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22574 std::swap(LHS, RHS);
22576 Opcode = X86ISD::FMAX;
22579 // Converting this to a max would handle both negative zeros and NaNs
22580 // incorrectly, but we can swap the operands to fix both.
22581 std::swap(LHS, RHS);
22585 Opcode = X86ISD::FMAX;
22591 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
22594 EVT CondVT = Cond.getValueType();
22595 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
22596 CondVT.getVectorElementType() == MVT::i1) {
22597 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
22598 // lowering on KNL. In this case we convert it to
22599 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
22600 // The same situation for all 128 and 256-bit vectors of i8 and i16.
22601 // Since SKX these selects have a proper lowering.
22602 EVT OpVT = LHS.getValueType();
22603 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
22604 (OpVT.getVectorElementType() == MVT::i8 ||
22605 OpVT.getVectorElementType() == MVT::i16) &&
22606 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
22607 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
22608 DCI.AddToWorklist(Cond.getNode());
22609 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
22612 // If this is a select between two integer constants, try to do some
22614 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
22615 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
22616 // Don't do this for crazy integer types.
22617 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
22618 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
22619 // so that TrueC (the true value) is larger than FalseC.
22620 bool NeedsCondInvert = false;
22622 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
22623 // Efficiently invertible.
22624 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
22625 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
22626 isa<ConstantSDNode>(Cond.getOperand(1))))) {
22627 NeedsCondInvert = true;
22628 std::swap(TrueC, FalseC);
22631 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
22632 if (FalseC->getAPIntValue() == 0 &&
22633 TrueC->getAPIntValue().isPowerOf2()) {
22634 if (NeedsCondInvert) // Invert the condition if needed.
22635 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22636 DAG.getConstant(1, DL, Cond.getValueType()));
22638 // Zero extend the condition if needed.
22639 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
22641 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22642 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
22643 DAG.getConstant(ShAmt, DL, MVT::i8));
22646 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
22647 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22648 if (NeedsCondInvert) // Invert the condition if needed.
22649 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22650 DAG.getConstant(1, DL, Cond.getValueType()));
22652 // Zero extend the condition if needed.
22653 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22654 FalseC->getValueType(0), Cond);
22655 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22656 SDValue(FalseC, 0));
22659 // Optimize cases that will turn into an LEA instruction. This requires
22660 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22661 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22662 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22663 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22665 bool isFastMultiplier = false;
22667 switch ((unsigned char)Diff) {
22669 case 1: // result = add base, cond
22670 case 2: // result = lea base( , cond*2)
22671 case 3: // result = lea base(cond, cond*2)
22672 case 4: // result = lea base( , cond*4)
22673 case 5: // result = lea base(cond, cond*4)
22674 case 8: // result = lea base( , cond*8)
22675 case 9: // result = lea base(cond, cond*8)
22676 isFastMultiplier = true;
22681 if (isFastMultiplier) {
22682 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22683 if (NeedsCondInvert) // Invert the condition if needed.
22684 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22685 DAG.getConstant(1, DL, Cond.getValueType()));
22687 // Zero extend the condition if needed.
22688 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22690 // Scale the condition by the difference.
22692 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22693 DAG.getConstant(Diff, DL,
22694 Cond.getValueType()));
22696 // Add the base if non-zero.
22697 if (FalseC->getAPIntValue() != 0)
22698 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22699 SDValue(FalseC, 0));
22706 // Canonicalize max and min:
22707 // (x > y) ? x : y -> (x >= y) ? x : y
22708 // (x < y) ? x : y -> (x <= y) ? x : y
22709 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
22710 // the need for an extra compare
22711 // against zero. e.g.
22712 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
22714 // testl %edi, %edi
22716 // cmovgl %edi, %eax
22720 // cmovsl %eax, %edi
22721 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
22722 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22723 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22724 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22729 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
22730 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
22731 Cond.getOperand(0), Cond.getOperand(1), NewCC);
22732 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
22737 // Early exit check
22738 if (!TLI.isTypeLegal(VT))
22741 // Match VSELECTs into subs with unsigned saturation.
22742 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
22743 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
22744 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
22745 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
22746 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22748 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
22749 // left side invert the predicate to simplify logic below.
22751 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
22753 CC = ISD::getSetCCInverse(CC, true);
22754 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
22758 if (Other.getNode() && Other->getNumOperands() == 2 &&
22759 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
22760 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
22761 SDValue CondRHS = Cond->getOperand(1);
22763 // Look for a general sub with unsigned saturation first.
22764 // x >= y ? x-y : 0 --> subus x, y
22765 // x > y ? x-y : 0 --> subus x, y
22766 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
22767 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
22768 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
22770 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
22771 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
22772 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
22773 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
22774 // If the RHS is a constant we have to reverse the const
22775 // canonicalization.
22776 // x > C-1 ? x+-C : 0 --> subus x, C
22777 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
22778 CondRHSConst->getAPIntValue() ==
22779 (-OpRHSConst->getAPIntValue() - 1))
22780 return DAG.getNode(
22781 X86ISD::SUBUS, DL, VT, OpLHS,
22782 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
22784 // Another special case: If C was a sign bit, the sub has been
22785 // canonicalized into a xor.
22786 // FIXME: Would it be better to use computeKnownBits to determine
22787 // whether it's safe to decanonicalize the xor?
22788 // x s< 0 ? x^C : 0 --> subus x, C
22789 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
22790 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
22791 OpRHSConst->getAPIntValue().isSignBit())
22792 // Note that we have to rebuild the RHS constant here to ensure we
22793 // don't rely on particular values of undef lanes.
22794 return DAG.getNode(
22795 X86ISD::SUBUS, DL, VT, OpLHS,
22796 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
22801 // Simplify vector selection if condition value type matches vselect
22803 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
22804 assert(Cond.getValueType().isVector() &&
22805 "vector select expects a vector selector!");
22807 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
22808 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
22810 // Try invert the condition if true value is not all 1s and false value
22812 if (!TValIsAllOnes && !FValIsAllZeros &&
22813 // Check if the selector will be produced by CMPP*/PCMP*
22814 Cond.getOpcode() == ISD::SETCC &&
22815 // Check if SETCC has already been promoted
22816 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
22818 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
22819 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
22821 if (TValIsAllZeros || FValIsAllOnes) {
22822 SDValue CC = Cond.getOperand(2);
22823 ISD::CondCode NewCC =
22824 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
22825 Cond.getOperand(0).getValueType().isInteger());
22826 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
22827 std::swap(LHS, RHS);
22828 TValIsAllOnes = FValIsAllOnes;
22829 FValIsAllZeros = TValIsAllZeros;
22833 if (TValIsAllOnes || FValIsAllZeros) {
22836 if (TValIsAllOnes && FValIsAllZeros)
22838 else if (TValIsAllOnes)
22840 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
22841 else if (FValIsAllZeros)
22842 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
22843 DAG.getBitcast(CondVT, LHS));
22845 return DAG.getBitcast(VT, Ret);
22849 // We should generate an X86ISD::BLENDI from a vselect if its argument
22850 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
22851 // constants. This specific pattern gets generated when we split a
22852 // selector for a 512 bit vector in a machine without AVX512 (but with
22853 // 256-bit vectors), during legalization:
22855 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
22857 // Iff we find this pattern and the build_vectors are built from
22858 // constants, we translate the vselect into a shuffle_vector that we
22859 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
22860 if ((N->getOpcode() == ISD::VSELECT ||
22861 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
22862 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
22863 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
22864 if (Shuffle.getNode())
22868 // If this is a *dynamic* select (non-constant condition) and we can match
22869 // this node with one of the variable blend instructions, restructure the
22870 // condition so that the blends can use the high bit of each element and use
22871 // SimplifyDemandedBits to simplify the condition operand.
22872 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
22873 !DCI.isBeforeLegalize() &&
22874 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
22875 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
22877 // Don't optimize vector selects that map to mask-registers.
22881 // We can only handle the cases where VSELECT is directly legal on the
22882 // subtarget. We custom lower VSELECT nodes with constant conditions and
22883 // this makes it hard to see whether a dynamic VSELECT will correctly
22884 // lower, so we both check the operation's status and explicitly handle the
22885 // cases where a *dynamic* blend will fail even though a constant-condition
22886 // blend could be custom lowered.
22887 // FIXME: We should find a better way to handle this class of problems.
22888 // Potentially, we should combine constant-condition vselect nodes
22889 // pre-legalization into shuffles and not mark as many types as custom
22891 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
22893 // FIXME: We don't support i16-element blends currently. We could and
22894 // should support them by making *all* the bits in the condition be set
22895 // rather than just the high bit and using an i8-element blend.
22896 if (VT.getScalarType() == MVT::i16)
22898 // Dynamic blending was only available from SSE4.1 onward.
22899 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
22901 // Byte blends are only available in AVX2
22902 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
22903 !Subtarget->hasAVX2())
22906 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
22907 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
22909 APInt KnownZero, KnownOne;
22910 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
22911 DCI.isBeforeLegalizeOps());
22912 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
22913 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
22915 // If we changed the computation somewhere in the DAG, this change
22916 // will affect all users of Cond.
22917 // Make sure it is fine and update all the nodes so that we do not
22918 // use the generic VSELECT anymore. Otherwise, we may perform
22919 // wrong optimizations as we messed up with the actual expectation
22920 // for the vector boolean values.
22921 if (Cond != TLO.Old) {
22922 // Check all uses of that condition operand to check whether it will be
22923 // consumed by non-BLEND instructions, which may depend on all bits are
22925 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22927 if (I->getOpcode() != ISD::VSELECT)
22928 // TODO: Add other opcodes eventually lowered into BLEND.
22931 // Update all the users of the condition, before committing the change,
22932 // so that the VSELECT optimizations that expect the correct vector
22933 // boolean value will not be triggered.
22934 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22936 DAG.ReplaceAllUsesOfValueWith(
22938 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
22939 Cond, I->getOperand(1), I->getOperand(2)));
22940 DCI.CommitTargetLoweringOpt(TLO);
22943 // At this point, only Cond is changed. Change the condition
22944 // just for N to keep the opportunity to optimize all other
22945 // users their own way.
22946 DAG.ReplaceAllUsesOfValueWith(
22948 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
22949 TLO.New, N->getOperand(1), N->getOperand(2)));
22957 // Check whether a boolean test is testing a boolean value generated by
22958 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22961 // Simplify the following patterns:
22962 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22963 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
22964 // to (Op EFLAGS Cond)
22966 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
22967 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
22968 // to (Op EFLAGS !Cond)
22970 // where Op could be BRCOND or CMOV.
22972 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
22973 // Quit if not CMP and SUB with its value result used.
22974 if (Cmp.getOpcode() != X86ISD::CMP &&
22975 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
22978 // Quit if not used as a boolean value.
22979 if (CC != X86::COND_E && CC != X86::COND_NE)
22982 // Check CMP operands. One of them should be 0 or 1 and the other should be
22983 // an SetCC or extended from it.
22984 SDValue Op1 = Cmp.getOperand(0);
22985 SDValue Op2 = Cmp.getOperand(1);
22988 const ConstantSDNode* C = nullptr;
22989 bool needOppositeCond = (CC == X86::COND_E);
22990 bool checkAgainstTrue = false; // Is it a comparison against 1?
22992 if ((C = dyn_cast<ConstantSDNode>(Op1)))
22994 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
22996 else // Quit if all operands are not constants.
22999 if (C->getZExtValue() == 1) {
23000 needOppositeCond = !needOppositeCond;
23001 checkAgainstTrue = true;
23002 } else if (C->getZExtValue() != 0)
23003 // Quit if the constant is neither 0 or 1.
23006 bool truncatedToBoolWithAnd = false;
23007 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23008 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23009 SetCC.getOpcode() == ISD::TRUNCATE ||
23010 SetCC.getOpcode() == ISD::AND) {
23011 if (SetCC.getOpcode() == ISD::AND) {
23013 ConstantSDNode *CS;
23014 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23015 CS->getZExtValue() == 1)
23017 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23018 CS->getZExtValue() == 1)
23022 SetCC = SetCC.getOperand(OpIdx);
23023 truncatedToBoolWithAnd = true;
23025 SetCC = SetCC.getOperand(0);
23028 switch (SetCC.getOpcode()) {
23029 case X86ISD::SETCC_CARRY:
23030 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23031 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23032 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23033 // truncated to i1 using 'and'.
23034 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23036 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23037 "Invalid use of SETCC_CARRY!");
23039 case X86ISD::SETCC:
23040 // Set the condition code or opposite one if necessary.
23041 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23042 if (needOppositeCond)
23043 CC = X86::GetOppositeBranchCondition(CC);
23044 return SetCC.getOperand(1);
23045 case X86ISD::CMOV: {
23046 // Check whether false/true value has canonical one, i.e. 0 or 1.
23047 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23048 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23049 // Quit if true value is not a constant.
23052 // Quit if false value is not a constant.
23054 SDValue Op = SetCC.getOperand(0);
23055 // Skip 'zext' or 'trunc' node.
23056 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23057 Op.getOpcode() == ISD::TRUNCATE)
23058 Op = Op.getOperand(0);
23059 // A special case for rdrand/rdseed, where 0 is set if false cond is
23061 if ((Op.getOpcode() != X86ISD::RDRAND &&
23062 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23065 // Quit if false value is not the constant 0 or 1.
23066 bool FValIsFalse = true;
23067 if (FVal && FVal->getZExtValue() != 0) {
23068 if (FVal->getZExtValue() != 1)
23070 // If FVal is 1, opposite cond is needed.
23071 needOppositeCond = !needOppositeCond;
23072 FValIsFalse = false;
23074 // Quit if TVal is not the constant opposite of FVal.
23075 if (FValIsFalse && TVal->getZExtValue() != 1)
23077 if (!FValIsFalse && TVal->getZExtValue() != 0)
23079 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23080 if (needOppositeCond)
23081 CC = X86::GetOppositeBranchCondition(CC);
23082 return SetCC.getOperand(3);
23089 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23091 /// (X86or (X86setcc) (X86setcc))
23092 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23093 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23094 X86::CondCode &CC1, SDValue &Flags,
23096 if (Cond->getOpcode() == X86ISD::CMP) {
23097 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23098 if (!CondOp1C || !CondOp1C->isNullValue())
23101 Cond = Cond->getOperand(0);
23106 SDValue SetCC0, SetCC1;
23107 switch (Cond->getOpcode()) {
23108 default: return false;
23115 SetCC0 = Cond->getOperand(0);
23116 SetCC1 = Cond->getOperand(1);
23120 // Make sure we have SETCC nodes, using the same flags value.
23121 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23122 SetCC1.getOpcode() != X86ISD::SETCC ||
23123 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23126 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23127 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23128 Flags = SetCC0->getOperand(1);
23132 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23133 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23134 TargetLowering::DAGCombinerInfo &DCI,
23135 const X86Subtarget *Subtarget) {
23138 // If the flag operand isn't dead, don't touch this CMOV.
23139 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23142 SDValue FalseOp = N->getOperand(0);
23143 SDValue TrueOp = N->getOperand(1);
23144 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23145 SDValue Cond = N->getOperand(3);
23147 if (CC == X86::COND_E || CC == X86::COND_NE) {
23148 switch (Cond.getOpcode()) {
23152 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23153 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23154 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23160 Flags = checkBoolTestSetCCCombine(Cond, CC);
23161 if (Flags.getNode() &&
23162 // Extra check as FCMOV only supports a subset of X86 cond.
23163 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23164 SDValue Ops[] = { FalseOp, TrueOp,
23165 DAG.getConstant(CC, DL, MVT::i8), Flags };
23166 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23169 // If this is a select between two integer constants, try to do some
23170 // optimizations. Note that the operands are ordered the opposite of SELECT
23172 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23173 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23174 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23175 // larger than FalseC (the false value).
23176 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23177 CC = X86::GetOppositeBranchCondition(CC);
23178 std::swap(TrueC, FalseC);
23179 std::swap(TrueOp, FalseOp);
23182 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23183 // This is efficient for any integer data type (including i8/i16) and
23185 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23186 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23187 DAG.getConstant(CC, DL, MVT::i8), Cond);
23189 // Zero extend the condition if needed.
23190 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23192 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23193 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23194 DAG.getConstant(ShAmt, DL, MVT::i8));
23195 if (N->getNumValues() == 2) // Dead flag value?
23196 return DCI.CombineTo(N, Cond, SDValue());
23200 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23201 // for any integer data type, including i8/i16.
23202 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23203 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23204 DAG.getConstant(CC, DL, MVT::i8), Cond);
23206 // Zero extend the condition if needed.
23207 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23208 FalseC->getValueType(0), Cond);
23209 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23210 SDValue(FalseC, 0));
23212 if (N->getNumValues() == 2) // Dead flag value?
23213 return DCI.CombineTo(N, Cond, SDValue());
23217 // Optimize cases that will turn into an LEA instruction. This requires
23218 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23219 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23220 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23221 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23223 bool isFastMultiplier = false;
23225 switch ((unsigned char)Diff) {
23227 case 1: // result = add base, cond
23228 case 2: // result = lea base( , cond*2)
23229 case 3: // result = lea base(cond, cond*2)
23230 case 4: // result = lea base( , cond*4)
23231 case 5: // result = lea base(cond, cond*4)
23232 case 8: // result = lea base( , cond*8)
23233 case 9: // result = lea base(cond, cond*8)
23234 isFastMultiplier = true;
23239 if (isFastMultiplier) {
23240 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23241 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23242 DAG.getConstant(CC, DL, MVT::i8), Cond);
23243 // Zero extend the condition if needed.
23244 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23246 // Scale the condition by the difference.
23248 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23249 DAG.getConstant(Diff, DL, Cond.getValueType()));
23251 // Add the base if non-zero.
23252 if (FalseC->getAPIntValue() != 0)
23253 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23254 SDValue(FalseC, 0));
23255 if (N->getNumValues() == 2) // Dead flag value?
23256 return DCI.CombineTo(N, Cond, SDValue());
23263 // Handle these cases:
23264 // (select (x != c), e, c) -> select (x != c), e, x),
23265 // (select (x == c), c, e) -> select (x == c), x, e)
23266 // where the c is an integer constant, and the "select" is the combination
23267 // of CMOV and CMP.
23269 // The rationale for this change is that the conditional-move from a constant
23270 // needs two instructions, however, conditional-move from a register needs
23271 // only one instruction.
23273 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23274 // some instruction-combining opportunities. This opt needs to be
23275 // postponed as late as possible.
23277 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23278 // the DCI.xxxx conditions are provided to postpone the optimization as
23279 // late as possible.
23281 ConstantSDNode *CmpAgainst = nullptr;
23282 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23283 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23284 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23286 if (CC == X86::COND_NE &&
23287 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23288 CC = X86::GetOppositeBranchCondition(CC);
23289 std::swap(TrueOp, FalseOp);
23292 if (CC == X86::COND_E &&
23293 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23294 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23295 DAG.getConstant(CC, DL, MVT::i8), Cond };
23296 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23301 // Fold and/or of setcc's to double CMOV:
23302 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23303 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23305 // This combine lets us generate:
23306 // cmovcc1 (jcc1 if we don't have CMOV)
23312 // cmovne (jne if we don't have CMOV)
23313 // When we can't use the CMOV instruction, it might increase branch
23315 // When we can use CMOV, or when there is no mispredict, this improves
23316 // throughput and reduces register pressure.
23318 if (CC == X86::COND_NE) {
23320 X86::CondCode CC0, CC1;
23322 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23324 std::swap(FalseOp, TrueOp);
23325 CC0 = X86::GetOppositeBranchCondition(CC0);
23326 CC1 = X86::GetOppositeBranchCondition(CC1);
23329 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23331 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23332 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23333 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23334 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23342 /// PerformMulCombine - Optimize a single multiply with constant into two
23343 /// in order to implement it with two cheaper instructions, e.g.
23344 /// LEA + SHL, LEA + LEA.
23345 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23346 TargetLowering::DAGCombinerInfo &DCI) {
23347 // An imul is usually smaller than the alternative sequence.
23348 if (DAG.getMachineFunction().getFunction()->optForMinSize())
23351 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23354 EVT VT = N->getValueType(0);
23355 if (VT != MVT::i64 && VT != MVT::i32)
23358 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23361 uint64_t MulAmt = C->getZExtValue();
23362 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23365 uint64_t MulAmt1 = 0;
23366 uint64_t MulAmt2 = 0;
23367 if ((MulAmt % 9) == 0) {
23369 MulAmt2 = MulAmt / 9;
23370 } else if ((MulAmt % 5) == 0) {
23372 MulAmt2 = MulAmt / 5;
23373 } else if ((MulAmt % 3) == 0) {
23375 MulAmt2 = MulAmt / 3;
23378 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23381 if (isPowerOf2_64(MulAmt2) &&
23382 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23383 // If second multiplifer is pow2, issue it first. We want the multiply by
23384 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23386 std::swap(MulAmt1, MulAmt2);
23389 if (isPowerOf2_64(MulAmt1))
23390 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23391 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23393 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23394 DAG.getConstant(MulAmt1, DL, VT));
23396 if (isPowerOf2_64(MulAmt2))
23397 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23398 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23400 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23401 DAG.getConstant(MulAmt2, DL, VT));
23403 // Do not add new nodes to DAG combiner worklist.
23404 DCI.CombineTo(N, NewMul, false);
23409 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23410 SDValue N0 = N->getOperand(0);
23411 SDValue N1 = N->getOperand(1);
23412 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23413 EVT VT = N0.getValueType();
23415 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23416 // since the result of setcc_c is all zero's or all ones.
23417 if (VT.isInteger() && !VT.isVector() &&
23418 N1C && N0.getOpcode() == ISD::AND &&
23419 N0.getOperand(1).getOpcode() == ISD::Constant) {
23420 SDValue N00 = N0.getOperand(0);
23421 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
23422 ((N00.getOpcode() == ISD::ANY_EXTEND ||
23423 N00.getOpcode() == ISD::ZERO_EXTEND) &&
23424 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
23425 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23426 APInt ShAmt = N1C->getAPIntValue();
23427 Mask = Mask.shl(ShAmt);
23430 return DAG.getNode(ISD::AND, DL, VT,
23431 N00, DAG.getConstant(Mask, DL, VT));
23436 // Hardware support for vector shifts is sparse which makes us scalarize the
23437 // vector operations in many cases. Also, on sandybridge ADD is faster than
23439 // (shl V, 1) -> add V,V
23440 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23441 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23442 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23443 // We shift all of the values by one. In many cases we do not have
23444 // hardware support for this operation. This is better expressed as an ADD
23446 if (N1SplatC->getAPIntValue() == 1)
23447 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23453 /// \brief Returns a vector of 0s if the node in input is a vector logical
23454 /// shift by a constant amount which is known to be bigger than or equal
23455 /// to the vector element size in bits.
23456 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23457 const X86Subtarget *Subtarget) {
23458 EVT VT = N->getValueType(0);
23460 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23461 (!Subtarget->hasInt256() ||
23462 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23465 SDValue Amt = N->getOperand(1);
23467 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23468 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23469 APInt ShiftAmt = AmtSplat->getAPIntValue();
23470 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23472 // SSE2/AVX2 logical shifts always return a vector of 0s
23473 // if the shift amount is bigger than or equal to
23474 // the element size. The constant shift amount will be
23475 // encoded as a 8-bit immediate.
23476 if (ShiftAmt.trunc(8).uge(MaxAmount))
23477 return getZeroVector(VT, Subtarget, DAG, DL);
23483 /// PerformShiftCombine - Combine shifts.
23484 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23485 TargetLowering::DAGCombinerInfo &DCI,
23486 const X86Subtarget *Subtarget) {
23487 if (N->getOpcode() == ISD::SHL)
23488 if (SDValue V = PerformSHLCombine(N, DAG))
23491 // Try to fold this logical shift into a zero vector.
23492 if (N->getOpcode() != ISD::SRA)
23493 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23499 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23500 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23501 // and friends. Likewise for OR -> CMPNEQSS.
23502 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23503 TargetLowering::DAGCombinerInfo &DCI,
23504 const X86Subtarget *Subtarget) {
23507 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23508 // we're requiring SSE2 for both.
23509 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23510 SDValue N0 = N->getOperand(0);
23511 SDValue N1 = N->getOperand(1);
23512 SDValue CMP0 = N0->getOperand(1);
23513 SDValue CMP1 = N1->getOperand(1);
23516 // The SETCCs should both refer to the same CMP.
23517 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23520 SDValue CMP00 = CMP0->getOperand(0);
23521 SDValue CMP01 = CMP0->getOperand(1);
23522 EVT VT = CMP00.getValueType();
23524 if (VT == MVT::f32 || VT == MVT::f64) {
23525 bool ExpectingFlags = false;
23526 // Check for any users that want flags:
23527 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23528 !ExpectingFlags && UI != UE; ++UI)
23529 switch (UI->getOpcode()) {
23534 ExpectingFlags = true;
23536 case ISD::CopyToReg:
23537 case ISD::SIGN_EXTEND:
23538 case ISD::ZERO_EXTEND:
23539 case ISD::ANY_EXTEND:
23543 if (!ExpectingFlags) {
23544 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23545 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23547 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23548 X86::CondCode tmp = cc0;
23553 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
23554 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
23555 // FIXME: need symbolic constants for these magic numbers.
23556 // See X86ATTInstPrinter.cpp:printSSECC().
23557 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
23558 if (Subtarget->hasAVX512()) {
23559 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
23561 DAG.getConstant(x86cc, DL, MVT::i8));
23562 if (N->getValueType(0) != MVT::i1)
23563 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
23567 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
23568 CMP00.getValueType(), CMP00, CMP01,
23569 DAG.getConstant(x86cc, DL,
23572 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
23573 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
23575 if (is64BitFP && !Subtarget->is64Bit()) {
23576 // On a 32-bit target, we cannot bitcast the 64-bit float to a
23577 // 64-bit integer, since that's not a legal type. Since
23578 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
23579 // bits, but can do this little dance to extract the lowest 32 bits
23580 // and work with those going forward.
23581 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
23583 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
23584 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
23585 Vector32, DAG.getIntPtrConstant(0, DL));
23589 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
23590 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
23591 DAG.getConstant(1, DL, IntVT));
23592 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
23594 return OneBitOfTruth;
23602 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
23603 /// so it can be folded inside ANDNP.
23604 static bool CanFoldXORWithAllOnes(const SDNode *N) {
23605 EVT VT = N->getValueType(0);
23607 // Match direct AllOnes for 128 and 256-bit vectors
23608 if (ISD::isBuildVectorAllOnes(N))
23611 // Look through a bit convert.
23612 if (N->getOpcode() == ISD::BITCAST)
23613 N = N->getOperand(0).getNode();
23615 // Sometimes the operand may come from a insert_subvector building a 256-bit
23617 if (VT.is256BitVector() &&
23618 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
23619 SDValue V1 = N->getOperand(0);
23620 SDValue V2 = N->getOperand(1);
23622 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
23623 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
23624 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
23625 ISD::isBuildVectorAllOnes(V2.getNode()))
23632 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
23633 // register. In most cases we actually compare or select YMM-sized registers
23634 // and mixing the two types creates horrible code. This method optimizes
23635 // some of the transition sequences.
23636 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
23637 TargetLowering::DAGCombinerInfo &DCI,
23638 const X86Subtarget *Subtarget) {
23639 EVT VT = N->getValueType(0);
23640 if (!VT.is256BitVector())
23643 assert((N->getOpcode() == ISD::ANY_EXTEND ||
23644 N->getOpcode() == ISD::ZERO_EXTEND ||
23645 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
23647 SDValue Narrow = N->getOperand(0);
23648 EVT NarrowVT = Narrow->getValueType(0);
23649 if (!NarrowVT.is128BitVector())
23652 if (Narrow->getOpcode() != ISD::XOR &&
23653 Narrow->getOpcode() != ISD::AND &&
23654 Narrow->getOpcode() != ISD::OR)
23657 SDValue N0 = Narrow->getOperand(0);
23658 SDValue N1 = Narrow->getOperand(1);
23661 // The Left side has to be a trunc.
23662 if (N0.getOpcode() != ISD::TRUNCATE)
23665 // The type of the truncated inputs.
23666 EVT WideVT = N0->getOperand(0)->getValueType(0);
23670 // The right side has to be a 'trunc' or a constant vector.
23671 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
23672 ConstantSDNode *RHSConstSplat = nullptr;
23673 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
23674 RHSConstSplat = RHSBV->getConstantSplatNode();
23675 if (!RHSTrunc && !RHSConstSplat)
23678 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23680 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
23683 // Set N0 and N1 to hold the inputs to the new wide operation.
23684 N0 = N0->getOperand(0);
23685 if (RHSConstSplat) {
23686 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
23687 SDValue(RHSConstSplat, 0));
23688 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
23689 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
23690 } else if (RHSTrunc) {
23691 N1 = N1->getOperand(0);
23694 // Generate the wide operation.
23695 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
23696 unsigned Opcode = N->getOpcode();
23698 case ISD::ANY_EXTEND:
23700 case ISD::ZERO_EXTEND: {
23701 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
23702 APInt Mask = APInt::getAllOnesValue(InBits);
23703 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
23704 return DAG.getNode(ISD::AND, DL, VT,
23705 Op, DAG.getConstant(Mask, DL, VT));
23707 case ISD::SIGN_EXTEND:
23708 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
23709 Op, DAG.getValueType(NarrowVT));
23711 llvm_unreachable("Unexpected opcode");
23715 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
23716 TargetLowering::DAGCombinerInfo &DCI,
23717 const X86Subtarget *Subtarget) {
23718 SDValue N0 = N->getOperand(0);
23719 SDValue N1 = N->getOperand(1);
23722 // A vector zext_in_reg may be represented as a shuffle,
23723 // feeding into a bitcast (this represents anyext) feeding into
23724 // an and with a mask.
23725 // We'd like to try to combine that into a shuffle with zero
23726 // plus a bitcast, removing the and.
23727 if (N0.getOpcode() != ISD::BITCAST ||
23728 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
23731 // The other side of the AND should be a splat of 2^C, where C
23732 // is the number of bits in the source type.
23733 if (N1.getOpcode() == ISD::BITCAST)
23734 N1 = N1.getOperand(0);
23735 if (N1.getOpcode() != ISD::BUILD_VECTOR)
23737 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
23739 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
23740 EVT SrcType = Shuffle->getValueType(0);
23742 // We expect a single-source shuffle
23743 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
23746 unsigned SrcSize = SrcType.getScalarSizeInBits();
23748 APInt SplatValue, SplatUndef;
23749 unsigned SplatBitSize;
23751 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
23752 SplatBitSize, HasAnyUndefs))
23755 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
23756 // Make sure the splat matches the mask we expect
23757 if (SplatBitSize > ResSize ||
23758 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
23761 // Make sure the input and output size make sense
23762 if (SrcSize >= ResSize || ResSize % SrcSize)
23765 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
23766 // The number of u's between each two values depends on the ratio between
23767 // the source and dest type.
23768 unsigned ZextRatio = ResSize / SrcSize;
23769 bool IsZext = true;
23770 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
23771 if (i % ZextRatio) {
23772 if (Shuffle->getMaskElt(i) > 0) {
23778 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
23779 // Expected element number
23789 // Ok, perform the transformation - replace the shuffle with
23790 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
23791 // (instead of undef) where the k elements come from the zero vector.
23792 SmallVector<int, 8> Mask;
23793 unsigned NumElems = SrcType.getVectorNumElements();
23794 for (unsigned i = 0; i < NumElems; ++i)
23796 Mask.push_back(NumElems);
23798 Mask.push_back(i / ZextRatio);
23800 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
23801 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
23802 return DAG.getBitcast(N0.getValueType(), NewShuffle);
23805 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
23806 TargetLowering::DAGCombinerInfo &DCI,
23807 const X86Subtarget *Subtarget) {
23808 if (DCI.isBeforeLegalizeOps())
23811 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
23814 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23817 EVT VT = N->getValueType(0);
23818 SDValue N0 = N->getOperand(0);
23819 SDValue N1 = N->getOperand(1);
23822 // Create BEXTR instructions
23823 // BEXTR is ((X >> imm) & (2**size-1))
23824 if (VT == MVT::i32 || VT == MVT::i64) {
23825 // Check for BEXTR.
23826 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
23827 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
23828 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
23829 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23830 if (MaskNode && ShiftNode) {
23831 uint64_t Mask = MaskNode->getZExtValue();
23832 uint64_t Shift = ShiftNode->getZExtValue();
23833 if (isMask_64(Mask)) {
23834 uint64_t MaskSize = countPopulation(Mask);
23835 if (Shift + MaskSize <= VT.getSizeInBits())
23836 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
23837 DAG.getConstant(Shift | (MaskSize << 8), DL,
23846 // Want to form ANDNP nodes:
23847 // 1) In the hopes of then easily combining them with OR and AND nodes
23848 // to form PBLEND/PSIGN.
23849 // 2) To match ANDN packed intrinsics
23850 if (VT != MVT::v2i64 && VT != MVT::v4i64)
23853 // Check LHS for vnot
23854 if (N0.getOpcode() == ISD::XOR &&
23855 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
23856 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
23857 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
23859 // Check RHS for vnot
23860 if (N1.getOpcode() == ISD::XOR &&
23861 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
23862 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
23863 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
23868 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
23869 TargetLowering::DAGCombinerInfo &DCI,
23870 const X86Subtarget *Subtarget) {
23871 if (DCI.isBeforeLegalizeOps())
23874 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23877 SDValue N0 = N->getOperand(0);
23878 SDValue N1 = N->getOperand(1);
23879 EVT VT = N->getValueType(0);
23881 // look for psign/blend
23882 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
23883 if (!Subtarget->hasSSSE3() ||
23884 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
23887 // Canonicalize pandn to RHS
23888 if (N0.getOpcode() == X86ISD::ANDNP)
23890 // or (and (m, y), (pandn m, x))
23891 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
23892 SDValue Mask = N1.getOperand(0);
23893 SDValue X = N1.getOperand(1);
23895 if (N0.getOperand(0) == Mask)
23896 Y = N0.getOperand(1);
23897 if (N0.getOperand(1) == Mask)
23898 Y = N0.getOperand(0);
23900 // Check to see if the mask appeared in both the AND and ANDNP and
23904 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
23905 // Look through mask bitcast.
23906 if (Mask.getOpcode() == ISD::BITCAST)
23907 Mask = Mask.getOperand(0);
23908 if (X.getOpcode() == ISD::BITCAST)
23909 X = X.getOperand(0);
23910 if (Y.getOpcode() == ISD::BITCAST)
23911 Y = Y.getOperand(0);
23913 EVT MaskVT = Mask.getValueType();
23915 // Validate that the Mask operand is a vector sra node.
23916 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
23917 // there is no psrai.b
23918 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
23919 unsigned SraAmt = ~0;
23920 if (Mask.getOpcode() == ISD::SRA) {
23921 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
23922 if (auto *AmtConst = AmtBV->getConstantSplatNode())
23923 SraAmt = AmtConst->getZExtValue();
23924 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
23925 SDValue SraC = Mask.getOperand(1);
23926 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
23928 if ((SraAmt + 1) != EltBits)
23933 // Now we know we at least have a plendvb with the mask val. See if
23934 // we can form a psignb/w/d.
23935 // psign = x.type == y.type == mask.type && y = sub(0, x);
23936 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
23937 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
23938 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
23939 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
23940 "Unsupported VT for PSIGN");
23941 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
23942 return DAG.getBitcast(VT, Mask);
23944 // PBLENDVB only available on SSE 4.1
23945 if (!Subtarget->hasSSE41())
23948 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
23950 X = DAG.getBitcast(BlendVT, X);
23951 Y = DAG.getBitcast(BlendVT, Y);
23952 Mask = DAG.getBitcast(BlendVT, Mask);
23953 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
23954 return DAG.getBitcast(VT, Mask);
23958 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
23961 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
23962 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
23964 // SHLD/SHRD instructions have lower register pressure, but on some
23965 // platforms they have higher latency than the equivalent
23966 // series of shifts/or that would otherwise be generated.
23967 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
23968 // have higher latencies and we are not optimizing for size.
23969 if (!OptForSize && Subtarget->isSHLDSlow())
23972 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
23974 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
23976 if (!N0.hasOneUse() || !N1.hasOneUse())
23979 SDValue ShAmt0 = N0.getOperand(1);
23980 if (ShAmt0.getValueType() != MVT::i8)
23982 SDValue ShAmt1 = N1.getOperand(1);
23983 if (ShAmt1.getValueType() != MVT::i8)
23985 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
23986 ShAmt0 = ShAmt0.getOperand(0);
23987 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
23988 ShAmt1 = ShAmt1.getOperand(0);
23991 unsigned Opc = X86ISD::SHLD;
23992 SDValue Op0 = N0.getOperand(0);
23993 SDValue Op1 = N1.getOperand(0);
23994 if (ShAmt0.getOpcode() == ISD::SUB) {
23995 Opc = X86ISD::SHRD;
23996 std::swap(Op0, Op1);
23997 std::swap(ShAmt0, ShAmt1);
24000 unsigned Bits = VT.getSizeInBits();
24001 if (ShAmt1.getOpcode() == ISD::SUB) {
24002 SDValue Sum = ShAmt1.getOperand(0);
24003 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24004 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24005 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24006 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24007 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24008 return DAG.getNode(Opc, DL, VT,
24010 DAG.getNode(ISD::TRUNCATE, DL,
24013 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24014 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24016 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24017 return DAG.getNode(Opc, DL, VT,
24018 N0.getOperand(0), N1.getOperand(0),
24019 DAG.getNode(ISD::TRUNCATE, DL,
24026 // Generate NEG and CMOV for integer abs.
24027 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24028 EVT VT = N->getValueType(0);
24030 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24031 // 8-bit integer abs to NEG and CMOV.
24032 if (VT.isInteger() && VT.getSizeInBits() == 8)
24035 SDValue N0 = N->getOperand(0);
24036 SDValue N1 = N->getOperand(1);
24039 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24040 // and change it to SUB and CMOV.
24041 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24042 N0.getOpcode() == ISD::ADD &&
24043 N0.getOperand(1) == N1 &&
24044 N1.getOpcode() == ISD::SRA &&
24045 N1.getOperand(0) == N0.getOperand(0))
24046 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24047 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24048 // Generate SUB & CMOV.
24049 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24050 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24052 SDValue Ops[] = { N0.getOperand(0), Neg,
24053 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24054 SDValue(Neg.getNode(), 1) };
24055 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24060 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
24061 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24062 TargetLowering::DAGCombinerInfo &DCI,
24063 const X86Subtarget *Subtarget) {
24064 if (DCI.isBeforeLegalizeOps())
24067 if (Subtarget->hasCMov())
24068 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24074 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24075 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24076 TargetLowering::DAGCombinerInfo &DCI,
24077 const X86Subtarget *Subtarget) {
24078 LoadSDNode *Ld = cast<LoadSDNode>(N);
24079 EVT RegVT = Ld->getValueType(0);
24080 EVT MemVT = Ld->getMemoryVT();
24082 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24084 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24085 // into two 16-byte operations.
24086 ISD::LoadExtType Ext = Ld->getExtensionType();
24087 unsigned Alignment = Ld->getAlignment();
24088 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
24089 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24090 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
24091 unsigned NumElems = RegVT.getVectorNumElements();
24095 SDValue Ptr = Ld->getBasePtr();
24096 SDValue Increment =
24097 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24099 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24101 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24102 Ld->getPointerInfo(), Ld->isVolatile(),
24103 Ld->isNonTemporal(), Ld->isInvariant(),
24105 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24106 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24107 Ld->getPointerInfo(), Ld->isVolatile(),
24108 Ld->isNonTemporal(), Ld->isInvariant(),
24109 std::min(16U, Alignment));
24110 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24112 Load2.getValue(1));
24114 SDValue NewVec = DAG.getUNDEF(RegVT);
24115 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24116 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24117 return DCI.CombineTo(N, NewVec, TF, true);
24123 /// PerformMLOADCombine - Resolve extending loads
24124 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24125 TargetLowering::DAGCombinerInfo &DCI,
24126 const X86Subtarget *Subtarget) {
24127 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24128 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24131 EVT VT = Mld->getValueType(0);
24132 unsigned NumElems = VT.getVectorNumElements();
24133 EVT LdVT = Mld->getMemoryVT();
24136 assert(LdVT != VT && "Cannot extend to the same type");
24137 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24138 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24139 // From, To sizes and ElemCount must be pow of two
24140 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24141 "Unexpected size for extending masked load");
24143 unsigned SizeRatio = ToSz / FromSz;
24144 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24146 // Create a type on which we perform the shuffle
24147 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24148 LdVT.getScalarType(), NumElems*SizeRatio);
24149 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24151 // Convert Src0 value
24152 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24153 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24154 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24155 for (unsigned i = 0; i != NumElems; ++i)
24156 ShuffleVec[i] = i * SizeRatio;
24158 // Can't shuffle using an illegal type.
24159 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24160 && "WideVecVT should be legal");
24161 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24162 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24164 // Prepare the new mask
24166 SDValue Mask = Mld->getMask();
24167 if (Mask.getValueType() == VT) {
24168 // Mask and original value have the same type
24169 NewMask = DAG.getBitcast(WideVecVT, Mask);
24170 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24171 for (unsigned i = 0; i != NumElems; ++i)
24172 ShuffleVec[i] = i * SizeRatio;
24173 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24174 ShuffleVec[i] = NumElems*SizeRatio;
24175 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24176 DAG.getConstant(0, dl, WideVecVT),
24180 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24181 unsigned WidenNumElts = NumElems*SizeRatio;
24182 unsigned MaskNumElts = VT.getVectorNumElements();
24183 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24186 unsigned NumConcat = WidenNumElts / MaskNumElts;
24187 SmallVector<SDValue, 16> Ops(NumConcat);
24188 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24190 for (unsigned i = 1; i != NumConcat; ++i)
24193 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24196 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24197 Mld->getBasePtr(), NewMask, WideSrc0,
24198 Mld->getMemoryVT(), Mld->getMemOperand(),
24200 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24201 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24204 /// PerformMSTORECombine - Resolve truncating stores
24205 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24206 const X86Subtarget *Subtarget) {
24207 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24208 if (!Mst->isTruncatingStore())
24211 EVT VT = Mst->getValue().getValueType();
24212 unsigned NumElems = VT.getVectorNumElements();
24213 EVT StVT = Mst->getMemoryVT();
24216 assert(StVT != VT && "Cannot truncate to the same type");
24217 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24218 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24222 // The truncating store is legal in some cases. For example
24223 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24224 // are designated for truncate store.
24225 // In this case we don't need any further transformations.
24226 if (TLI.isTruncStoreLegal(VT, StVT))
24229 // From, To sizes and ElemCount must be pow of two
24230 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24231 "Unexpected size for truncating masked store");
24232 // We are going to use the original vector elt for storing.
24233 // Accumulated smaller vector elements must be a multiple of the store size.
24234 assert (((NumElems * FromSz) % ToSz) == 0 &&
24235 "Unexpected ratio for truncating masked store");
24237 unsigned SizeRatio = FromSz / ToSz;
24238 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24240 // Create a type on which we perform the shuffle
24241 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24242 StVT.getScalarType(), NumElems*SizeRatio);
24244 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24246 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24247 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24248 for (unsigned i = 0; i != NumElems; ++i)
24249 ShuffleVec[i] = i * SizeRatio;
24251 // Can't shuffle using an illegal type.
24252 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24253 && "WideVecVT should be legal");
24255 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24256 DAG.getUNDEF(WideVecVT),
24260 SDValue Mask = Mst->getMask();
24261 if (Mask.getValueType() == VT) {
24262 // Mask and original value have the same type
24263 NewMask = DAG.getBitcast(WideVecVT, Mask);
24264 for (unsigned i = 0; i != NumElems; ++i)
24265 ShuffleVec[i] = i * SizeRatio;
24266 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24267 ShuffleVec[i] = NumElems*SizeRatio;
24268 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24269 DAG.getConstant(0, dl, WideVecVT),
24273 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24274 unsigned WidenNumElts = NumElems*SizeRatio;
24275 unsigned MaskNumElts = VT.getVectorNumElements();
24276 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24279 unsigned NumConcat = WidenNumElts / MaskNumElts;
24280 SmallVector<SDValue, 16> Ops(NumConcat);
24281 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24283 for (unsigned i = 1; i != NumConcat; ++i)
24286 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24289 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24290 NewMask, StVT, Mst->getMemOperand(), false);
24292 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24293 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24294 const X86Subtarget *Subtarget) {
24295 StoreSDNode *St = cast<StoreSDNode>(N);
24296 EVT VT = St->getValue().getValueType();
24297 EVT StVT = St->getMemoryVT();
24299 SDValue StoredVal = St->getOperand(1);
24300 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24302 // If we are saving a concatenation of two XMM registers and 32-byte stores
24303 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24304 unsigned Alignment = St->getAlignment();
24305 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
24306 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24307 StVT == VT && !IsAligned) {
24308 unsigned NumElems = VT.getVectorNumElements();
24312 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24313 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24316 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24317 SDValue Ptr0 = St->getBasePtr();
24318 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24320 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24321 St->getPointerInfo(), St->isVolatile(),
24322 St->isNonTemporal(), Alignment);
24323 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24324 St->getPointerInfo(), St->isVolatile(),
24325 St->isNonTemporal(),
24326 std::min(16U, Alignment));
24327 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24330 // Optimize trunc store (of multiple scalars) to shuffle and store.
24331 // First, pack all of the elements in one place. Next, store to memory
24332 // in fewer chunks.
24333 if (St->isTruncatingStore() && VT.isVector()) {
24334 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24335 unsigned NumElems = VT.getVectorNumElements();
24336 assert(StVT != VT && "Cannot truncate to the same type");
24337 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24338 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24340 // The truncating store is legal in some cases. For example
24341 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24342 // are designated for truncate store.
24343 // In this case we don't need any further transformations.
24344 if (TLI.isTruncStoreLegal(VT, StVT))
24347 // From, To sizes and ElemCount must be pow of two
24348 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24349 // We are going to use the original vector elt for storing.
24350 // Accumulated smaller vector elements must be a multiple of the store size.
24351 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24353 unsigned SizeRatio = FromSz / ToSz;
24355 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24357 // Create a type on which we perform the shuffle
24358 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24359 StVT.getScalarType(), NumElems*SizeRatio);
24361 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24363 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24364 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24365 for (unsigned i = 0; i != NumElems; ++i)
24366 ShuffleVec[i] = i * SizeRatio;
24368 // Can't shuffle using an illegal type.
24369 if (!TLI.isTypeLegal(WideVecVT))
24372 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24373 DAG.getUNDEF(WideVecVT),
24375 // At this point all of the data is stored at the bottom of the
24376 // register. We now need to save it to mem.
24378 // Find the largest store unit
24379 MVT StoreType = MVT::i8;
24380 for (MVT Tp : MVT::integer_valuetypes()) {
24381 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24385 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24386 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24387 (64 <= NumElems * ToSz))
24388 StoreType = MVT::f64;
24390 // Bitcast the original vector into a vector of store-size units
24391 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24392 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24393 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24394 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24395 SmallVector<SDValue, 8> Chains;
24396 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24397 TLI.getPointerTy(DAG.getDataLayout()));
24398 SDValue Ptr = St->getBasePtr();
24400 // Perform one or more big stores into memory.
24401 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24402 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24403 StoreType, ShuffWide,
24404 DAG.getIntPtrConstant(i, dl));
24405 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24406 St->getPointerInfo(), St->isVolatile(),
24407 St->isNonTemporal(), St->getAlignment());
24408 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24409 Chains.push_back(Ch);
24412 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24415 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24416 // the FP state in cases where an emms may be missing.
24417 // A preferable solution to the general problem is to figure out the right
24418 // places to insert EMMS. This qualifies as a quick hack.
24420 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24421 if (VT.getSizeInBits() != 64)
24424 const Function *F = DAG.getMachineFunction().getFunction();
24425 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24427 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24428 if ((VT.isVector() ||
24429 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24430 isa<LoadSDNode>(St->getValue()) &&
24431 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24432 St->getChain().hasOneUse() && !St->isVolatile()) {
24433 SDNode* LdVal = St->getValue().getNode();
24434 LoadSDNode *Ld = nullptr;
24435 int TokenFactorIndex = -1;
24436 SmallVector<SDValue, 8> Ops;
24437 SDNode* ChainVal = St->getChain().getNode();
24438 // Must be a store of a load. We currently handle two cases: the load
24439 // is a direct child, and it's under an intervening TokenFactor. It is
24440 // possible to dig deeper under nested TokenFactors.
24441 if (ChainVal == LdVal)
24442 Ld = cast<LoadSDNode>(St->getChain());
24443 else if (St->getValue().hasOneUse() &&
24444 ChainVal->getOpcode() == ISD::TokenFactor) {
24445 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24446 if (ChainVal->getOperand(i).getNode() == LdVal) {
24447 TokenFactorIndex = i;
24448 Ld = cast<LoadSDNode>(St->getValue());
24450 Ops.push_back(ChainVal->getOperand(i));
24454 if (!Ld || !ISD::isNormalLoad(Ld))
24457 // If this is not the MMX case, i.e. we are just turning i64 load/store
24458 // into f64 load/store, avoid the transformation if there are multiple
24459 // uses of the loaded value.
24460 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24465 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24466 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24468 if (Subtarget->is64Bit() || F64IsLegal) {
24469 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24470 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24471 Ld->getPointerInfo(), Ld->isVolatile(),
24472 Ld->isNonTemporal(), Ld->isInvariant(),
24473 Ld->getAlignment());
24474 SDValue NewChain = NewLd.getValue(1);
24475 if (TokenFactorIndex != -1) {
24476 Ops.push_back(NewChain);
24477 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24479 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24480 St->getPointerInfo(),
24481 St->isVolatile(), St->isNonTemporal(),
24482 St->getAlignment());
24485 // Otherwise, lower to two pairs of 32-bit loads / stores.
24486 SDValue LoAddr = Ld->getBasePtr();
24487 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24488 DAG.getConstant(4, LdDL, MVT::i32));
24490 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24491 Ld->getPointerInfo(),
24492 Ld->isVolatile(), Ld->isNonTemporal(),
24493 Ld->isInvariant(), Ld->getAlignment());
24494 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24495 Ld->getPointerInfo().getWithOffset(4),
24496 Ld->isVolatile(), Ld->isNonTemporal(),
24498 MinAlign(Ld->getAlignment(), 4));
24500 SDValue NewChain = LoLd.getValue(1);
24501 if (TokenFactorIndex != -1) {
24502 Ops.push_back(LoLd);
24503 Ops.push_back(HiLd);
24504 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24507 LoAddr = St->getBasePtr();
24508 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24509 DAG.getConstant(4, StDL, MVT::i32));
24511 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24512 St->getPointerInfo(),
24513 St->isVolatile(), St->isNonTemporal(),
24514 St->getAlignment());
24515 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24516 St->getPointerInfo().getWithOffset(4),
24518 St->isNonTemporal(),
24519 MinAlign(St->getAlignment(), 4));
24520 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24523 // This is similar to the above case, but here we handle a scalar 64-bit
24524 // integer store that is extracted from a vector on a 32-bit target.
24525 // If we have SSE2, then we can treat it like a floating-point double
24526 // to get past legalization. The execution dependencies fixup pass will
24527 // choose the optimal machine instruction for the store if this really is
24528 // an integer or v2f32 rather than an f64.
24529 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
24530 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
24531 SDValue OldExtract = St->getOperand(1);
24532 SDValue ExtOp0 = OldExtract.getOperand(0);
24533 unsigned VecSize = ExtOp0.getValueSizeInBits();
24534 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
24535 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
24536 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
24537 BitCast, OldExtract.getOperand(1));
24538 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
24539 St->getPointerInfo(), St->isVolatile(),
24540 St->isNonTemporal(), St->getAlignment());
24546 /// Return 'true' if this vector operation is "horizontal"
24547 /// and return the operands for the horizontal operation in LHS and RHS. A
24548 /// horizontal operation performs the binary operation on successive elements
24549 /// of its first operand, then on successive elements of its second operand,
24550 /// returning the resulting values in a vector. For example, if
24551 /// A = < float a0, float a1, float a2, float a3 >
24553 /// B = < float b0, float b1, float b2, float b3 >
24554 /// then the result of doing a horizontal operation on A and B is
24555 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24556 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24557 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24558 /// set to A, RHS to B, and the routine returns 'true'.
24559 /// Note that the binary operation should have the property that if one of the
24560 /// operands is UNDEF then the result is UNDEF.
24561 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24562 // Look for the following pattern: if
24563 // A = < float a0, float a1, float a2, float a3 >
24564 // B = < float b0, float b1, float b2, float b3 >
24566 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24567 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24568 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24569 // which is A horizontal-op B.
24571 // At least one of the operands should be a vector shuffle.
24572 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24573 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24576 MVT VT = LHS.getSimpleValueType();
24578 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24579 "Unsupported vector type for horizontal add/sub");
24581 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24582 // operate independently on 128-bit lanes.
24583 unsigned NumElts = VT.getVectorNumElements();
24584 unsigned NumLanes = VT.getSizeInBits()/128;
24585 unsigned NumLaneElts = NumElts / NumLanes;
24586 assert((NumLaneElts % 2 == 0) &&
24587 "Vector type should have an even number of elements in each lane");
24588 unsigned HalfLaneElts = NumLaneElts/2;
24590 // View LHS in the form
24591 // LHS = VECTOR_SHUFFLE A, B, LMask
24592 // If LHS is not a shuffle then pretend it is the shuffle
24593 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24594 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24597 SmallVector<int, 16> LMask(NumElts);
24598 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24599 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24600 A = LHS.getOperand(0);
24601 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24602 B = LHS.getOperand(1);
24603 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24604 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24606 if (LHS.getOpcode() != ISD::UNDEF)
24608 for (unsigned i = 0; i != NumElts; ++i)
24612 // Likewise, view RHS in the form
24613 // RHS = VECTOR_SHUFFLE C, D, RMask
24615 SmallVector<int, 16> RMask(NumElts);
24616 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24617 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24618 C = RHS.getOperand(0);
24619 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24620 D = RHS.getOperand(1);
24621 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24622 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24624 if (RHS.getOpcode() != ISD::UNDEF)
24626 for (unsigned i = 0; i != NumElts; ++i)
24630 // Check that the shuffles are both shuffling the same vectors.
24631 if (!(A == C && B == D) && !(A == D && B == C))
24634 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
24635 if (!A.getNode() && !B.getNode())
24638 // If A and B occur in reverse order in RHS, then "swap" them (which means
24639 // rewriting the mask).
24641 ShuffleVectorSDNode::commuteMask(RMask);
24643 // At this point LHS and RHS are equivalent to
24644 // LHS = VECTOR_SHUFFLE A, B, LMask
24645 // RHS = VECTOR_SHUFFLE A, B, RMask
24646 // Check that the masks correspond to performing a horizontal operation.
24647 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
24648 for (unsigned i = 0; i != NumLaneElts; ++i) {
24649 int LIdx = LMask[i+l], RIdx = RMask[i+l];
24651 // Ignore any UNDEF components.
24652 if (LIdx < 0 || RIdx < 0 ||
24653 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
24654 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
24657 // Check that successive elements are being operated on. If not, this is
24658 // not a horizontal operation.
24659 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
24660 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
24661 if (!(LIdx == Index && RIdx == Index + 1) &&
24662 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
24667 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
24668 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
24672 /// Do target-specific dag combines on floating point adds.
24673 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
24674 const X86Subtarget *Subtarget) {
24675 EVT VT = N->getValueType(0);
24676 SDValue LHS = N->getOperand(0);
24677 SDValue RHS = N->getOperand(1);
24679 // Try to synthesize horizontal adds from adds of shuffles.
24680 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24681 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24682 isHorizontalBinOp(LHS, RHS, true))
24683 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
24687 /// Do target-specific dag combines on floating point subs.
24688 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
24689 const X86Subtarget *Subtarget) {
24690 EVT VT = N->getValueType(0);
24691 SDValue LHS = N->getOperand(0);
24692 SDValue RHS = N->getOperand(1);
24694 // Try to synthesize horizontal subs from subs of shuffles.
24695 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24696 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24697 isHorizontalBinOp(LHS, RHS, false))
24698 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
24702 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
24703 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
24704 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
24706 // F[X]OR(0.0, x) -> x
24707 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24708 if (C->getValueAPF().isPosZero())
24709 return N->getOperand(1);
24711 // F[X]OR(x, 0.0) -> x
24712 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24713 if (C->getValueAPF().isPosZero())
24714 return N->getOperand(0);
24718 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
24719 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
24720 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
24722 // Only perform optimizations if UnsafeMath is used.
24723 if (!DAG.getTarget().Options.UnsafeFPMath)
24726 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
24727 // into FMINC and FMAXC, which are Commutative operations.
24728 unsigned NewOp = 0;
24729 switch (N->getOpcode()) {
24730 default: llvm_unreachable("unknown opcode");
24731 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
24732 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
24735 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
24736 N->getOperand(0), N->getOperand(1));
24739 /// Do target-specific dag combines on X86ISD::FAND nodes.
24740 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
24741 // FAND(0.0, x) -> 0.0
24742 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24743 if (C->getValueAPF().isPosZero())
24744 return N->getOperand(0);
24746 // FAND(x, 0.0) -> 0.0
24747 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24748 if (C->getValueAPF().isPosZero())
24749 return N->getOperand(1);
24754 /// Do target-specific dag combines on X86ISD::FANDN nodes
24755 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
24756 // FANDN(0.0, x) -> x
24757 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24758 if (C->getValueAPF().isPosZero())
24759 return N->getOperand(1);
24761 // FANDN(x, 0.0) -> 0.0
24762 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24763 if (C->getValueAPF().isPosZero())
24764 return N->getOperand(1);
24769 static SDValue PerformBTCombine(SDNode *N,
24771 TargetLowering::DAGCombinerInfo &DCI) {
24772 // BT ignores high bits in the bit index operand.
24773 SDValue Op1 = N->getOperand(1);
24774 if (Op1.hasOneUse()) {
24775 unsigned BitWidth = Op1.getValueSizeInBits();
24776 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
24777 APInt KnownZero, KnownOne;
24778 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
24779 !DCI.isBeforeLegalizeOps());
24780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24781 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
24782 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
24783 DCI.CommitTargetLoweringOpt(TLO);
24788 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
24789 SDValue Op = N->getOperand(0);
24790 if (Op.getOpcode() == ISD::BITCAST)
24791 Op = Op.getOperand(0);
24792 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
24793 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
24794 VT.getVectorElementType().getSizeInBits() ==
24795 OpVT.getVectorElementType().getSizeInBits()) {
24796 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
24801 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
24802 const X86Subtarget *Subtarget) {
24803 EVT VT = N->getValueType(0);
24804 if (!VT.isVector())
24807 SDValue N0 = N->getOperand(0);
24808 SDValue N1 = N->getOperand(1);
24809 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
24812 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
24813 // both SSE and AVX2 since there is no sign-extended shift right
24814 // operation on a vector with 64-bit elements.
24815 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
24816 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
24817 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
24818 N0.getOpcode() == ISD::SIGN_EXTEND)) {
24819 SDValue N00 = N0.getOperand(0);
24821 // EXTLOAD has a better solution on AVX2,
24822 // it may be replaced with X86ISD::VSEXT node.
24823 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
24824 if (!ISD::isNormalLoad(N00.getNode()))
24827 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
24828 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
24830 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
24836 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
24837 TargetLowering::DAGCombinerInfo &DCI,
24838 const X86Subtarget *Subtarget) {
24839 SDValue N0 = N->getOperand(0);
24840 EVT VT = N->getValueType(0);
24841 EVT SVT = VT.getScalarType();
24842 EVT InVT = N0.getValueType();
24843 EVT InSVT = InVT.getScalarType();
24846 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
24847 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
24848 // This exposes the sext to the sdivrem lowering, so that it directly extends
24849 // from AH (which we otherwise need to do contortions to access).
24850 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
24851 InVT == MVT::i8 && VT == MVT::i32) {
24852 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24853 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
24854 N0.getOperand(0), N0.getOperand(1));
24855 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24856 return R.getValue(1);
24859 if (!DCI.isBeforeLegalizeOps()) {
24860 if (InVT == MVT::i1) {
24861 SDValue Zero = DAG.getConstant(0, DL, VT);
24863 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
24864 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
24869 if (VT.isVector() && Subtarget->hasSSE2()) {
24870 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
24871 EVT InVT = N.getValueType();
24872 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
24873 Size / InVT.getScalarSizeInBits());
24874 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
24875 DAG.getUNDEF(InVT));
24877 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
24880 // If target-size is less than 128-bits, extend to a type that would extend
24881 // to 128 bits, extend that and extract the original target vector.
24882 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
24883 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24884 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24885 unsigned Scale = 128 / VT.getSizeInBits();
24887 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
24888 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
24889 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
24890 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
24891 DAG.getIntPtrConstant(0, DL));
24894 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
24895 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
24896 if (VT.getSizeInBits() == 128 &&
24897 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24898 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24899 SDValue ExOp = ExtendVecSize(DL, N0, 128);
24900 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
24903 // On pre-AVX2 targets, split into 128-bit nodes of
24904 // ISD::SIGN_EXTEND_VECTOR_INREG.
24905 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
24906 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24907 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24908 unsigned NumVecs = VT.getSizeInBits() / 128;
24909 unsigned NumSubElts = 128 / SVT.getSizeInBits();
24910 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
24911 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
24913 SmallVector<SDValue, 8> Opnds;
24914 for (unsigned i = 0, Offset = 0; i != NumVecs;
24915 ++i, Offset += NumSubElts) {
24916 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
24917 DAG.getIntPtrConstant(Offset, DL));
24918 SrcVec = ExtendVecSize(DL, SrcVec, 128);
24919 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
24920 Opnds.push_back(SrcVec);
24922 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
24926 if (!Subtarget->hasFp256())
24929 if (VT.isVector() && VT.getSizeInBits() == 256)
24930 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
24936 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
24937 const X86Subtarget* Subtarget) {
24939 EVT VT = N->getValueType(0);
24941 // Let legalize expand this if it isn't a legal type yet.
24942 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
24945 EVT ScalarVT = VT.getScalarType();
24946 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
24947 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
24948 !Subtarget->hasAVX512()))
24951 SDValue A = N->getOperand(0);
24952 SDValue B = N->getOperand(1);
24953 SDValue C = N->getOperand(2);
24955 bool NegA = (A.getOpcode() == ISD::FNEG);
24956 bool NegB = (B.getOpcode() == ISD::FNEG);
24957 bool NegC = (C.getOpcode() == ISD::FNEG);
24959 // Negative multiplication when NegA xor NegB
24960 bool NegMul = (NegA != NegB);
24962 A = A.getOperand(0);
24964 B = B.getOperand(0);
24966 C = C.getOperand(0);
24970 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
24972 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
24974 return DAG.getNode(Opcode, dl, VT, A, B, C);
24977 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
24978 TargetLowering::DAGCombinerInfo &DCI,
24979 const X86Subtarget *Subtarget) {
24980 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
24981 // (and (i32 x86isd::setcc_carry), 1)
24982 // This eliminates the zext. This transformation is necessary because
24983 // ISD::SETCC is always legalized to i8.
24985 SDValue N0 = N->getOperand(0);
24986 EVT VT = N->getValueType(0);
24988 if (N0.getOpcode() == ISD::AND &&
24990 N0.getOperand(0).hasOneUse()) {
24991 SDValue N00 = N0.getOperand(0);
24992 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24993 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24994 if (!C || C->getZExtValue() != 1)
24996 return DAG.getNode(ISD::AND, dl, VT,
24997 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24998 N00.getOperand(0), N00.getOperand(1)),
24999 DAG.getConstant(1, dl, VT));
25003 if (N0.getOpcode() == ISD::TRUNCATE &&
25005 N0.getOperand(0).hasOneUse()) {
25006 SDValue N00 = N0.getOperand(0);
25007 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25008 return DAG.getNode(ISD::AND, dl, VT,
25009 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25010 N00.getOperand(0), N00.getOperand(1)),
25011 DAG.getConstant(1, dl, VT));
25015 if (VT.is256BitVector())
25016 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25019 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25020 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25021 // This exposes the zext to the udivrem lowering, so that it directly extends
25022 // from AH (which we otherwise need to do contortions to access).
25023 if (N0.getOpcode() == ISD::UDIVREM &&
25024 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25025 (VT == MVT::i32 || VT == MVT::i64)) {
25026 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25027 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25028 N0.getOperand(0), N0.getOperand(1));
25029 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25030 return R.getValue(1);
25036 // Optimize x == -y --> x+y == 0
25037 // x != -y --> x+y != 0
25038 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25039 const X86Subtarget* Subtarget) {
25040 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25041 SDValue LHS = N->getOperand(0);
25042 SDValue RHS = N->getOperand(1);
25043 EVT VT = N->getValueType(0);
25046 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25047 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25048 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25049 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25050 LHS.getOperand(1));
25051 return DAG.getSetCC(DL, N->getValueType(0), addV,
25052 DAG.getConstant(0, DL, addV.getValueType()), CC);
25054 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25055 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25056 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25057 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25058 RHS.getOperand(1));
25059 return DAG.getSetCC(DL, N->getValueType(0), addV,
25060 DAG.getConstant(0, DL, addV.getValueType()), CC);
25063 if (VT.getScalarType() == MVT::i1 &&
25064 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25066 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25067 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25068 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25070 if (!IsSEXT0 || !IsVZero1) {
25071 // Swap the operands and update the condition code.
25072 std::swap(LHS, RHS);
25073 CC = ISD::getSetCCSwappedOperands(CC);
25075 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25076 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25077 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25080 if (IsSEXT0 && IsVZero1) {
25081 assert(VT == LHS.getOperand(0).getValueType() &&
25082 "Uexpected operand type");
25083 if (CC == ISD::SETGT)
25084 return DAG.getConstant(0, DL, VT);
25085 if (CC == ISD::SETLE)
25086 return DAG.getConstant(1, DL, VT);
25087 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25088 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25090 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25091 "Unexpected condition code!");
25092 return LHS.getOperand(0);
25099 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25100 SelectionDAG &DAG) {
25102 MVT VT = Load->getSimpleValueType(0);
25103 MVT EVT = VT.getVectorElementType();
25104 SDValue Addr = Load->getOperand(1);
25105 SDValue NewAddr = DAG.getNode(
25106 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25107 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25108 Addr.getSimpleValueType()));
25111 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25112 DAG.getMachineFunction().getMachineMemOperand(
25113 Load->getMemOperand(), 0, EVT.getStoreSize()));
25117 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25118 const X86Subtarget *Subtarget) {
25120 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25121 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25122 "X86insertps is only defined for v4x32");
25124 SDValue Ld = N->getOperand(1);
25125 if (MayFoldLoad(Ld)) {
25126 // Extract the countS bits from the immediate so we can get the proper
25127 // address when narrowing the vector load to a specific element.
25128 // When the second source op is a memory address, insertps doesn't use
25129 // countS and just gets an f32 from that address.
25130 unsigned DestIndex =
25131 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25133 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25135 // Create this as a scalar to vector to match the instruction pattern.
25136 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25137 // countS bits are ignored when loading from memory on insertps, which
25138 // means we don't need to explicitly set them to 0.
25139 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25140 LoadScalarToVector, N->getOperand(2));
25145 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25146 SDValue V0 = N->getOperand(0);
25147 SDValue V1 = N->getOperand(1);
25149 EVT VT = N->getValueType(0);
25151 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25152 // operands and changing the mask to 1. This saves us a bunch of
25153 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25154 // x86InstrInfo knows how to commute this back after instruction selection
25155 // if it would help register allocation.
25157 // TODO: If optimizing for size or a processor that doesn't suffer from
25158 // partial register update stalls, this should be transformed into a MOVSD
25159 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25161 if (VT == MVT::v2f64)
25162 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25163 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25164 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25165 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25171 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25172 // as "sbb reg,reg", since it can be extended without zext and produces
25173 // an all-ones bit which is more useful than 0/1 in some cases.
25174 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25177 return DAG.getNode(ISD::AND, DL, VT,
25178 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25179 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25181 DAG.getConstant(1, DL, VT));
25182 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25183 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25184 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25185 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25189 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25190 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25191 TargetLowering::DAGCombinerInfo &DCI,
25192 const X86Subtarget *Subtarget) {
25194 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25195 SDValue EFLAGS = N->getOperand(1);
25197 if (CC == X86::COND_A) {
25198 // Try to convert COND_A into COND_B in an attempt to facilitate
25199 // materializing "setb reg".
25201 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25202 // cannot take an immediate as its first operand.
25204 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25205 EFLAGS.getValueType().isInteger() &&
25206 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25207 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25208 EFLAGS.getNode()->getVTList(),
25209 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25210 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25211 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25215 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25216 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25218 if (CC == X86::COND_B)
25219 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25221 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25222 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25223 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25229 // Optimize branch condition evaluation.
25231 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25232 TargetLowering::DAGCombinerInfo &DCI,
25233 const X86Subtarget *Subtarget) {
25235 SDValue Chain = N->getOperand(0);
25236 SDValue Dest = N->getOperand(1);
25237 SDValue EFLAGS = N->getOperand(3);
25238 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25240 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25241 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25242 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25249 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25250 SelectionDAG &DAG) {
25251 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25252 // optimize away operation when it's from a constant.
25254 // The general transformation is:
25255 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25256 // AND(VECTOR_CMP(x,y), constant2)
25257 // constant2 = UNARYOP(constant)
25259 // Early exit if this isn't a vector operation, the operand of the
25260 // unary operation isn't a bitwise AND, or if the sizes of the operations
25261 // aren't the same.
25262 EVT VT = N->getValueType(0);
25263 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25264 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25265 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25268 // Now check that the other operand of the AND is a constant. We could
25269 // make the transformation for non-constant splats as well, but it's unclear
25270 // that would be a benefit as it would not eliminate any operations, just
25271 // perform one more step in scalar code before moving to the vector unit.
25272 if (BuildVectorSDNode *BV =
25273 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25274 // Bail out if the vector isn't a constant.
25275 if (!BV->isConstant())
25278 // Everything checks out. Build up the new and improved node.
25280 EVT IntVT = BV->getValueType(0);
25281 // Create a new constant of the appropriate type for the transformed
25283 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25284 // The AND node needs bitcasts to/from an integer vector type around it.
25285 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25286 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25287 N->getOperand(0)->getOperand(0), MaskConst);
25288 SDValue Res = DAG.getBitcast(VT, NewAnd);
25295 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25296 const X86Subtarget *Subtarget) {
25297 SDValue Op0 = N->getOperand(0);
25298 EVT VT = N->getValueType(0);
25299 EVT InVT = Op0.getValueType();
25300 EVT InSVT = InVT.getScalarType();
25301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25303 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25304 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25305 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25307 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25308 InVT.getVectorNumElements());
25309 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25311 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25312 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25314 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25320 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25321 const X86Subtarget *Subtarget) {
25322 // First try to optimize away the conversion entirely when it's
25323 // conditionally from a constant. Vectors only.
25324 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25327 // Now move on to more general possibilities.
25328 SDValue Op0 = N->getOperand(0);
25329 EVT VT = N->getValueType(0);
25330 EVT InVT = Op0.getValueType();
25331 EVT InSVT = InVT.getScalarType();
25333 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25334 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25335 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25337 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25338 InVT.getVectorNumElements());
25339 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25340 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25343 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25344 // a 32-bit target where SSE doesn't support i64->FP operations.
25345 if (Op0.getOpcode() == ISD::LOAD) {
25346 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25347 EVT LdVT = Ld->getValueType(0);
25349 // This transformation is not supported if the result type is f16
25350 if (VT == MVT::f16)
25353 if (!Ld->isVolatile() && !VT.isVector() &&
25354 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25355 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25356 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25357 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25358 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25365 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25366 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25367 X86TargetLowering::DAGCombinerInfo &DCI) {
25368 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25369 // the result is either zero or one (depending on the input carry bit).
25370 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25371 if (X86::isZeroNode(N->getOperand(0)) &&
25372 X86::isZeroNode(N->getOperand(1)) &&
25373 // We don't have a good way to replace an EFLAGS use, so only do this when
25375 SDValue(N, 1).use_empty()) {
25377 EVT VT = N->getValueType(0);
25378 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25379 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25380 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25381 DAG.getConstant(X86::COND_B, DL,
25384 DAG.getConstant(1, DL, VT));
25385 return DCI.CombineTo(N, Res1, CarryOut);
25391 // fold (add Y, (sete X, 0)) -> adc 0, Y
25392 // (add Y, (setne X, 0)) -> sbb -1, Y
25393 // (sub (sete X, 0), Y) -> sbb 0, Y
25394 // (sub (setne X, 0), Y) -> adc -1, Y
25395 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25398 // Look through ZExts.
25399 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25400 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25403 SDValue SetCC = Ext.getOperand(0);
25404 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25407 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25408 if (CC != X86::COND_E && CC != X86::COND_NE)
25411 SDValue Cmp = SetCC.getOperand(1);
25412 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25413 !X86::isZeroNode(Cmp.getOperand(1)) ||
25414 !Cmp.getOperand(0).getValueType().isInteger())
25417 SDValue CmpOp0 = Cmp.getOperand(0);
25418 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25419 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25421 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25422 if (CC == X86::COND_NE)
25423 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25424 DL, OtherVal.getValueType(), OtherVal,
25425 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25427 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25428 DL, OtherVal.getValueType(), OtherVal,
25429 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25432 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25433 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25434 const X86Subtarget *Subtarget) {
25435 EVT VT = N->getValueType(0);
25436 SDValue Op0 = N->getOperand(0);
25437 SDValue Op1 = N->getOperand(1);
25439 // Try to synthesize horizontal adds from adds of shuffles.
25440 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25441 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25442 isHorizontalBinOp(Op0, Op1, true))
25443 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25445 return OptimizeConditionalInDecrement(N, DAG);
25448 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25449 const X86Subtarget *Subtarget) {
25450 SDValue Op0 = N->getOperand(0);
25451 SDValue Op1 = N->getOperand(1);
25453 // X86 can't encode an immediate LHS of a sub. See if we can push the
25454 // negation into a preceding instruction.
25455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25456 // If the RHS of the sub is a XOR with one use and a constant, invert the
25457 // immediate. Then add one to the LHS of the sub so we can turn
25458 // X-Y -> X+~Y+1, saving one register.
25459 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25460 isa<ConstantSDNode>(Op1.getOperand(1))) {
25461 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25462 EVT VT = Op0.getValueType();
25463 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25465 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25466 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25467 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25471 // Try to synthesize horizontal adds from adds of shuffles.
25472 EVT VT = N->getValueType(0);
25473 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25474 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25475 isHorizontalBinOp(Op0, Op1, true))
25476 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25478 return OptimizeConditionalInDecrement(N, DAG);
25481 /// performVZEXTCombine - Performs build vector combines
25482 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25483 TargetLowering::DAGCombinerInfo &DCI,
25484 const X86Subtarget *Subtarget) {
25486 MVT VT = N->getSimpleValueType(0);
25487 SDValue Op = N->getOperand(0);
25488 MVT OpVT = Op.getSimpleValueType();
25489 MVT OpEltVT = OpVT.getVectorElementType();
25490 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25492 // (vzext (bitcast (vzext (x)) -> (vzext x)
25494 while (V.getOpcode() == ISD::BITCAST)
25495 V = V.getOperand(0);
25497 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25498 MVT InnerVT = V.getSimpleValueType();
25499 MVT InnerEltVT = InnerVT.getVectorElementType();
25501 // If the element sizes match exactly, we can just do one larger vzext. This
25502 // is always an exact type match as vzext operates on integer types.
25503 if (OpEltVT == InnerEltVT) {
25504 assert(OpVT == InnerVT && "Types must match for vzext!");
25505 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25508 // The only other way we can combine them is if only a single element of the
25509 // inner vzext is used in the input to the outer vzext.
25510 if (InnerEltVT.getSizeInBits() < InputBits)
25513 // In this case, the inner vzext is completely dead because we're going to
25514 // only look at bits inside of the low element. Just do the outer vzext on
25515 // a bitcast of the input to the inner.
25516 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
25519 // Check if we can bypass extracting and re-inserting an element of an input
25520 // vector. Essentially:
25521 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25522 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25523 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25524 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25525 SDValue ExtractedV = V.getOperand(0);
25526 SDValue OrigV = ExtractedV.getOperand(0);
25527 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25528 if (ExtractIdx->getZExtValue() == 0) {
25529 MVT OrigVT = OrigV.getSimpleValueType();
25530 // Extract a subvector if necessary...
25531 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25532 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25533 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25534 OrigVT.getVectorNumElements() / Ratio);
25535 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25536 DAG.getIntPtrConstant(0, DL));
25538 Op = DAG.getBitcast(OpVT, OrigV);
25539 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25546 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25547 DAGCombinerInfo &DCI) const {
25548 SelectionDAG &DAG = DCI.DAG;
25549 switch (N->getOpcode()) {
25551 case ISD::EXTRACT_VECTOR_ELT:
25552 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25555 case X86ISD::SHRUNKBLEND:
25556 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25557 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
25558 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25559 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25560 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25561 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25562 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25565 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25566 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25567 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25568 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25569 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25570 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
25571 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25572 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
25573 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
25574 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
25575 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25576 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25578 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25580 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25581 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25582 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25583 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25584 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25585 case ISD::ANY_EXTEND:
25586 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25587 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25588 case ISD::SIGN_EXTEND_INREG:
25589 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25590 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25591 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25592 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25593 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25594 case X86ISD::SHUFP: // Handle all target specific shuffles
25595 case X86ISD::PALIGNR:
25596 case X86ISD::UNPCKH:
25597 case X86ISD::UNPCKL:
25598 case X86ISD::MOVHLPS:
25599 case X86ISD::MOVLHPS:
25600 case X86ISD::PSHUFB:
25601 case X86ISD::PSHUFD:
25602 case X86ISD::PSHUFHW:
25603 case X86ISD::PSHUFLW:
25604 case X86ISD::MOVSS:
25605 case X86ISD::MOVSD:
25606 case X86ISD::VPERMILPI:
25607 case X86ISD::VPERM2X128:
25608 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25609 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25610 case X86ISD::INSERTPS: {
25611 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
25612 return PerformINSERTPSCombine(N, DAG, Subtarget);
25615 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
25621 /// isTypeDesirableForOp - Return true if the target has native support for
25622 /// the specified value type and it is 'desirable' to use the type for the
25623 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25624 /// instruction encodings are longer and some i16 instructions are slow.
25625 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25626 if (!isTypeLegal(VT))
25628 if (VT != MVT::i16)
25635 case ISD::SIGN_EXTEND:
25636 case ISD::ZERO_EXTEND:
25637 case ISD::ANY_EXTEND:
25650 /// IsDesirableToPromoteOp - This method query the target whether it is
25651 /// beneficial for dag combiner to promote the specified node. If true, it
25652 /// should return the desired promotion type by reference.
25653 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25654 EVT VT = Op.getValueType();
25655 if (VT != MVT::i16)
25658 bool Promote = false;
25659 bool Commute = false;
25660 switch (Op.getOpcode()) {
25663 LoadSDNode *LD = cast<LoadSDNode>(Op);
25664 // If the non-extending load has a single use and it's not live out, then it
25665 // might be folded.
25666 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25667 Op.hasOneUse()*/) {
25668 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25669 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25670 // The only case where we'd want to promote LOAD (rather then it being
25671 // promoted as an operand is when it's only use is liveout.
25672 if (UI->getOpcode() != ISD::CopyToReg)
25679 case ISD::SIGN_EXTEND:
25680 case ISD::ZERO_EXTEND:
25681 case ISD::ANY_EXTEND:
25686 SDValue N0 = Op.getOperand(0);
25687 // Look out for (store (shl (load), x)).
25688 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25701 SDValue N0 = Op.getOperand(0);
25702 SDValue N1 = Op.getOperand(1);
25703 if (!Commute && MayFoldLoad(N1))
25705 // Avoid disabling potential load folding opportunities.
25706 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25708 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25718 //===----------------------------------------------------------------------===//
25719 // X86 Inline Assembly Support
25720 //===----------------------------------------------------------------------===//
25722 // Helper to match a string separated by whitespace.
25723 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
25724 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
25726 for (StringRef Piece : Pieces) {
25727 if (!S.startswith(Piece)) // Check if the piece matches.
25730 S = S.substr(Piece.size());
25731 StringRef::size_type Pos = S.find_first_not_of(" \t");
25732 if (Pos == 0) // We matched a prefix.
25741 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
25743 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
25744 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
25745 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
25746 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
25748 if (AsmPieces.size() == 3)
25750 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
25757 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
25758 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
25760 std::string AsmStr = IA->getAsmString();
25762 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
25763 if (!Ty || Ty->getBitWidth() % 16 != 0)
25766 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
25767 SmallVector<StringRef, 4> AsmPieces;
25768 SplitString(AsmStr, AsmPieces, ";\n");
25770 switch (AsmPieces.size()) {
25771 default: return false;
25773 // FIXME: this should verify that we are targeting a 486 or better. If not,
25774 // we will turn this bswap into something that will be lowered to logical
25775 // ops instead of emitting the bswap asm. For now, we don't support 486 or
25776 // lower so don't worry about this.
25778 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
25779 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
25780 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
25781 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
25782 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
25783 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
25784 // No need to check constraints, nothing other than the equivalent of
25785 // "=r,0" would be valid here.
25786 return IntrinsicLowering::LowerToByteSwap(CI);
25789 // rorw $$8, ${0:w} --> llvm.bswap.i16
25790 if (CI->getType()->isIntegerTy(16) &&
25791 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25792 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
25793 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
25795 StringRef ConstraintsStr = IA->getConstraintString();
25796 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25797 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25798 if (clobbersFlagRegisters(AsmPieces))
25799 return IntrinsicLowering::LowerToByteSwap(CI);
25803 if (CI->getType()->isIntegerTy(32) &&
25804 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25805 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
25806 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
25807 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
25809 StringRef ConstraintsStr = IA->getConstraintString();
25810 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25811 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25812 if (clobbersFlagRegisters(AsmPieces))
25813 return IntrinsicLowering::LowerToByteSwap(CI);
25816 if (CI->getType()->isIntegerTy(64)) {
25817 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
25818 if (Constraints.size() >= 2 &&
25819 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
25820 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
25821 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
25822 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
25823 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
25824 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
25825 return IntrinsicLowering::LowerToByteSwap(CI);
25833 /// getConstraintType - Given a constraint letter, return the type of
25834 /// constraint it is for this target.
25835 X86TargetLowering::ConstraintType
25836 X86TargetLowering::getConstraintType(StringRef Constraint) const {
25837 if (Constraint.size() == 1) {
25838 switch (Constraint[0]) {
25849 return C_RegisterClass;
25873 return TargetLowering::getConstraintType(Constraint);
25876 /// Examine constraint type and operand type and determine a weight value.
25877 /// This object must already have been set up with the operand type
25878 /// and the current alternative constraint selected.
25879 TargetLowering::ConstraintWeight
25880 X86TargetLowering::getSingleConstraintMatchWeight(
25881 AsmOperandInfo &info, const char *constraint) const {
25882 ConstraintWeight weight = CW_Invalid;
25883 Value *CallOperandVal = info.CallOperandVal;
25884 // If we don't have a value, we can't do a match,
25885 // but allow it at the lowest weight.
25886 if (!CallOperandVal)
25888 Type *type = CallOperandVal->getType();
25889 // Look at the constraint type.
25890 switch (*constraint) {
25892 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
25903 if (CallOperandVal->getType()->isIntegerTy())
25904 weight = CW_SpecificReg;
25909 if (type->isFloatingPointTy())
25910 weight = CW_SpecificReg;
25913 if (type->isX86_MMXTy() && Subtarget->hasMMX())
25914 weight = CW_SpecificReg;
25918 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
25919 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
25920 weight = CW_Register;
25923 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
25924 if (C->getZExtValue() <= 31)
25925 weight = CW_Constant;
25929 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25930 if (C->getZExtValue() <= 63)
25931 weight = CW_Constant;
25935 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25936 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
25937 weight = CW_Constant;
25941 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25942 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
25943 weight = CW_Constant;
25947 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25948 if (C->getZExtValue() <= 3)
25949 weight = CW_Constant;
25953 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25954 if (C->getZExtValue() <= 0xff)
25955 weight = CW_Constant;
25960 if (isa<ConstantFP>(CallOperandVal)) {
25961 weight = CW_Constant;
25965 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25966 if ((C->getSExtValue() >= -0x80000000LL) &&
25967 (C->getSExtValue() <= 0x7fffffffLL))
25968 weight = CW_Constant;
25972 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25973 if (C->getZExtValue() <= 0xffffffff)
25974 weight = CW_Constant;
25981 /// LowerXConstraint - try to replace an X constraint, which matches anything,
25982 /// with another that has more specific requirements based on the type of the
25983 /// corresponding operand.
25984 const char *X86TargetLowering::
25985 LowerXConstraint(EVT ConstraintVT) const {
25986 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
25987 // 'f' like normal targets.
25988 if (ConstraintVT.isFloatingPoint()) {
25989 if (Subtarget->hasSSE2())
25991 if (Subtarget->hasSSE1())
25995 return TargetLowering::LowerXConstraint(ConstraintVT);
25998 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
25999 /// vector. If it is invalid, don't add anything to Ops.
26000 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26001 std::string &Constraint,
26002 std::vector<SDValue>&Ops,
26003 SelectionDAG &DAG) const {
26006 // Only support length 1 constraints for now.
26007 if (Constraint.length() > 1) return;
26009 char ConstraintLetter = Constraint[0];
26010 switch (ConstraintLetter) {
26013 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26014 if (C->getZExtValue() <= 31) {
26015 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26016 Op.getValueType());
26022 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26023 if (C->getZExtValue() <= 63) {
26024 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26025 Op.getValueType());
26031 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26032 if (isInt<8>(C->getSExtValue())) {
26033 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26034 Op.getValueType());
26040 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26041 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26042 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26043 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26044 Op.getValueType());
26050 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26051 if (C->getZExtValue() <= 3) {
26052 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26053 Op.getValueType());
26059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26060 if (C->getZExtValue() <= 255) {
26061 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26062 Op.getValueType());
26068 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26069 if (C->getZExtValue() <= 127) {
26070 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26071 Op.getValueType());
26077 // 32-bit signed value
26078 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26079 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26080 C->getSExtValue())) {
26081 // Widen to 64 bits here to get it sign extended.
26082 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26085 // FIXME gcc accepts some relocatable values here too, but only in certain
26086 // memory models; it's complicated.
26091 // 32-bit unsigned value
26092 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26093 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26094 C->getZExtValue())) {
26095 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26096 Op.getValueType());
26100 // FIXME gcc accepts some relocatable values here too, but only in certain
26101 // memory models; it's complicated.
26105 // Literal immediates are always ok.
26106 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26107 // Widen to 64 bits here to get it sign extended.
26108 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26112 // In any sort of PIC mode addresses need to be computed at runtime by
26113 // adding in a register or some sort of table lookup. These can't
26114 // be used as immediates.
26115 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26118 // If we are in non-pic codegen mode, we allow the address of a global (with
26119 // an optional displacement) to be used with 'i'.
26120 GlobalAddressSDNode *GA = nullptr;
26121 int64_t Offset = 0;
26123 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26125 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26126 Offset += GA->getOffset();
26128 } else if (Op.getOpcode() == ISD::ADD) {
26129 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26130 Offset += C->getZExtValue();
26131 Op = Op.getOperand(0);
26134 } else if (Op.getOpcode() == ISD::SUB) {
26135 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26136 Offset += -C->getZExtValue();
26137 Op = Op.getOperand(0);
26142 // Otherwise, this isn't something we can handle, reject it.
26146 const GlobalValue *GV = GA->getGlobal();
26147 // If we require an extra load to get this address, as in PIC mode, we
26148 // can't accept it.
26149 if (isGlobalStubReference(
26150 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26153 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26154 GA->getValueType(0), Offset);
26159 if (Result.getNode()) {
26160 Ops.push_back(Result);
26163 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26166 std::pair<unsigned, const TargetRegisterClass *>
26167 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26168 StringRef Constraint,
26170 // First, see if this is a constraint that directly corresponds to an LLVM
26172 if (Constraint.size() == 1) {
26173 // GCC Constraint Letters
26174 switch (Constraint[0]) {
26176 // TODO: Slight differences here in allocation order and leaving
26177 // RIP in the class. Do they matter any more here than they do
26178 // in the normal allocation?
26179 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26180 if (Subtarget->is64Bit()) {
26181 if (VT == MVT::i32 || VT == MVT::f32)
26182 return std::make_pair(0U, &X86::GR32RegClass);
26183 if (VT == MVT::i16)
26184 return std::make_pair(0U, &X86::GR16RegClass);
26185 if (VT == MVT::i8 || VT == MVT::i1)
26186 return std::make_pair(0U, &X86::GR8RegClass);
26187 if (VT == MVT::i64 || VT == MVT::f64)
26188 return std::make_pair(0U, &X86::GR64RegClass);
26191 // 32-bit fallthrough
26192 case 'Q': // Q_REGS
26193 if (VT == MVT::i32 || VT == MVT::f32)
26194 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26195 if (VT == MVT::i16)
26196 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26197 if (VT == MVT::i8 || VT == MVT::i1)
26198 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26199 if (VT == MVT::i64)
26200 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26202 case 'r': // GENERAL_REGS
26203 case 'l': // INDEX_REGS
26204 if (VT == MVT::i8 || VT == MVT::i1)
26205 return std::make_pair(0U, &X86::GR8RegClass);
26206 if (VT == MVT::i16)
26207 return std::make_pair(0U, &X86::GR16RegClass);
26208 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26209 return std::make_pair(0U, &X86::GR32RegClass);
26210 return std::make_pair(0U, &X86::GR64RegClass);
26211 case 'R': // LEGACY_REGS
26212 if (VT == MVT::i8 || VT == MVT::i1)
26213 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26214 if (VT == MVT::i16)
26215 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26216 if (VT == MVT::i32 || !Subtarget->is64Bit())
26217 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26218 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26219 case 'f': // FP Stack registers.
26220 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26221 // value to the correct fpstack register class.
26222 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26223 return std::make_pair(0U, &X86::RFP32RegClass);
26224 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26225 return std::make_pair(0U, &X86::RFP64RegClass);
26226 return std::make_pair(0U, &X86::RFP80RegClass);
26227 case 'y': // MMX_REGS if MMX allowed.
26228 if (!Subtarget->hasMMX()) break;
26229 return std::make_pair(0U, &X86::VR64RegClass);
26230 case 'Y': // SSE_REGS if SSE2 allowed
26231 if (!Subtarget->hasSSE2()) break;
26233 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26234 if (!Subtarget->hasSSE1()) break;
26236 switch (VT.SimpleTy) {
26238 // Scalar SSE types.
26241 return std::make_pair(0U, &X86::FR32RegClass);
26244 return std::make_pair(0U, &X86::FR64RegClass);
26252 return std::make_pair(0U, &X86::VR128RegClass);
26260 return std::make_pair(0U, &X86::VR256RegClass);
26265 return std::make_pair(0U, &X86::VR512RegClass);
26271 // Use the default implementation in TargetLowering to convert the register
26272 // constraint into a member of a register class.
26273 std::pair<unsigned, const TargetRegisterClass*> Res;
26274 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26276 // Not found as a standard register?
26278 // Map st(0) -> st(7) -> ST0
26279 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26280 tolower(Constraint[1]) == 's' &&
26281 tolower(Constraint[2]) == 't' &&
26282 Constraint[3] == '(' &&
26283 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26284 Constraint[5] == ')' &&
26285 Constraint[6] == '}') {
26287 Res.first = X86::FP0+Constraint[4]-'0';
26288 Res.second = &X86::RFP80RegClass;
26292 // GCC allows "st(0)" to be called just plain "st".
26293 if (StringRef("{st}").equals_lower(Constraint)) {
26294 Res.first = X86::FP0;
26295 Res.second = &X86::RFP80RegClass;
26300 if (StringRef("{flags}").equals_lower(Constraint)) {
26301 Res.first = X86::EFLAGS;
26302 Res.second = &X86::CCRRegClass;
26306 // 'A' means EAX + EDX.
26307 if (Constraint == "A") {
26308 Res.first = X86::EAX;
26309 Res.second = &X86::GR32_ADRegClass;
26315 // Otherwise, check to see if this is a register class of the wrong value
26316 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26317 // turn into {ax},{dx}.
26318 // MVT::Other is used to specify clobber names.
26319 if (Res.second->hasType(VT) || VT == MVT::Other)
26320 return Res; // Correct type already, nothing to do.
26322 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26323 // return "eax". This should even work for things like getting 64bit integer
26324 // registers when given an f64 type.
26325 const TargetRegisterClass *Class = Res.second;
26326 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26327 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26328 unsigned Size = VT.getSizeInBits();
26329 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26330 : Size == 16 ? MVT::i16
26331 : Size == 32 ? MVT::i32
26332 : Size == 64 ? MVT::i64
26334 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26336 Res.first = DestReg;
26337 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26338 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26339 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26340 : &X86::GR64RegClass;
26341 assert(Res.second->contains(Res.first) && "Register in register class");
26343 // No register found/type mismatch.
26345 Res.second = nullptr;
26347 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26348 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26349 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26350 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26351 Class == &X86::VR512RegClass) {
26352 // Handle references to XMM physical registers that got mapped into the
26353 // wrong class. This can happen with constraints like {xmm0} where the
26354 // target independent register mapper will just pick the first match it can
26355 // find, ignoring the required type.
26357 if (VT == MVT::f32 || VT == MVT::i32)
26358 Res.second = &X86::FR32RegClass;
26359 else if (VT == MVT::f64 || VT == MVT::i64)
26360 Res.second = &X86::FR64RegClass;
26361 else if (X86::VR128RegClass.hasType(VT))
26362 Res.second = &X86::VR128RegClass;
26363 else if (X86::VR256RegClass.hasType(VT))
26364 Res.second = &X86::VR256RegClass;
26365 else if (X86::VR512RegClass.hasType(VT))
26366 Res.second = &X86::VR512RegClass;
26368 // Type mismatch and not a clobber: Return an error;
26370 Res.second = nullptr;
26377 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26378 const AddrMode &AM, Type *Ty,
26379 unsigned AS) const {
26380 // Scaling factors are not free at all.
26381 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26382 // will take 2 allocations in the out of order engine instead of 1
26383 // for plain addressing mode, i.e. inst (reg1).
26385 // vaddps (%rsi,%drx), %ymm0, %ymm1
26386 // Requires two allocations (one for the load, one for the computation)
26388 // vaddps (%rsi), %ymm0, %ymm1
26389 // Requires just 1 allocation, i.e., freeing allocations for other operations
26390 // and having less micro operations to execute.
26392 // For some X86 architectures, this is even worse because for instance for
26393 // stores, the complex addressing mode forces the instruction to use the
26394 // "load" ports instead of the dedicated "store" port.
26395 // E.g., on Haswell:
26396 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26397 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26398 if (isLegalAddressingMode(DL, AM, Ty, AS))
26399 // Scale represents reg2 * scale, thus account for 1
26400 // as soon as we use a second register.
26401 return AM.Scale != 0;
26405 bool X86TargetLowering::isTargetFTOL() const {
26406 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();