1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/WinEHFuncInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 static cl::opt<int> ReciprocalEstimateRefinementSteps(
71 "x86-recip-refinement-steps", cl::init(1),
72 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
73 "result of the hardware reciprocal estimate instruction."),
76 // Forward declarations.
77 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
80 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
81 const X86Subtarget &STI)
82 : TargetLowering(TM), Subtarget(&STI) {
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
87 // Set up the TargetLowering object.
88 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
90 // X86 is weird. It always uses i8 for shift amounts and setcc results.
91 setBooleanContents(ZeroOrOneBooleanContent);
92 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
93 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
95 // For 64-bit, since we have so many registers, use the ILP scheduler.
96 // For 32-bit, use the register pressure specific scheduling.
97 // For Atom, always use ILP scheduling.
98 if (Subtarget->isAtom())
99 setSchedulingPreference(Sched::ILP);
100 else if (Subtarget->is64Bit())
101 setSchedulingPreference(Sched::ILP);
103 setSchedulingPreference(Sched::RegPressure);
104 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
105 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
107 // Bypass expensive divides on Atom when compiling with O2.
108 if (TM.getOptLevel() >= CodeGenOpt::Default) {
109 if (Subtarget->hasSlowDivide32())
110 addBypassSlowDiv(32, 8);
111 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
112 addBypassSlowDiv(64, 16);
115 if (Subtarget->isTargetKnownWindowsMSVC()) {
116 // Setup Windows compiler runtime calls.
117 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
118 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
119 setLibcallName(RTLIB::SREM_I64, "_allrem");
120 setLibcallName(RTLIB::UREM_I64, "_aullrem");
121 setLibcallName(RTLIB::MUL_I64, "_allmul");
122 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
123 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
124 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
125 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
126 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
128 // The _ftol2 runtime function has an unusual calling conv, which
129 // is modeled by a special pseudo-instruction.
130 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
131 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
132 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
133 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
136 if (Subtarget->isTargetDarwin()) {
137 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
138 setUseUnderscoreSetJmp(false);
139 setUseUnderscoreLongJmp(false);
140 } else if (Subtarget->isTargetWindowsGNU()) {
141 // MS runtime is weird: it exports _setjmp, but longjmp!
142 setUseUnderscoreSetJmp(true);
143 setUseUnderscoreLongJmp(false);
145 setUseUnderscoreSetJmp(true);
146 setUseUnderscoreLongJmp(true);
149 // Set up the register classes.
150 addRegisterClass(MVT::i8, &X86::GR8RegClass);
151 addRegisterClass(MVT::i16, &X86::GR16RegClass);
152 addRegisterClass(MVT::i32, &X86::GR32RegClass);
153 if (Subtarget->is64Bit())
154 addRegisterClass(MVT::i64, &X86::GR64RegClass);
156 for (MVT VT : MVT::integer_valuetypes())
157 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
159 // We don't accept any truncstore of integer registers.
160 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
161 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
162 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
163 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
164 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
165 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
167 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
169 // SETOEQ and SETUNE require checking two conditions.
170 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
171 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
172 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
173 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
174 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
175 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
177 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
179 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
180 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
181 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
183 if (Subtarget->is64Bit()) {
184 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
185 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
186 } else if (!Subtarget->useSoftFloat()) {
187 // We have an algorithm for SSE2->double, and we turn this into a
188 // 64-bit FILD followed by conditional FADD for other targets.
189 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
190 // We have an algorithm for SSE2, and we turn this into a 64-bit
191 // FILD for other targets.
192 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
195 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
197 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
200 if (!Subtarget->useSoftFloat()) {
201 // SSE has no i16 to fp conversion, only i32
202 if (X86ScalarSSEf32) {
203 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
204 // f32 and f64 cases are Legal, f80 case is not
205 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
207 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
208 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
211 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
212 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
215 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
216 // are Legal, f80 is custom lowered.
217 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
218 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
220 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
222 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
225 if (X86ScalarSSEf32) {
226 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
227 // f32 and f64 cases are Legal, f80 case is not
228 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
230 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
231 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
234 // Handle FP_TO_UINT by promoting the destination to a larger signed
236 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
237 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
238 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
240 if (Subtarget->is64Bit()) {
241 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
242 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
243 } else if (!Subtarget->useSoftFloat()) {
244 // Since AVX is a superset of SSE3, only check for SSE here.
245 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
246 // Expand FP_TO_UINT into a select.
247 // FIXME: We would like to use a Custom expander here eventually to do
248 // the optimal thing for SSE vs. the default expansion in the legalizer.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
251 // With SSE3 we can use fisttpll to convert to a signed i64; without
252 // SSE, we're stuck with a fistpll.
253 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
256 if (isTargetFTOL()) {
257 // Use the _ftol2 runtime function, which has a pseudo-instruction
258 // to handle its weird calling convention.
259 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
262 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
263 if (!X86ScalarSSEf64) {
264 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
265 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
266 if (Subtarget->is64Bit()) {
267 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
268 // Without SSE, i64->f64 goes through memory.
269 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
273 // Scalar integer divide and remainder are lowered to use operations that
274 // produce two results, to match the available instructions. This exposes
275 // the two-result form to trivial CSE, which is able to combine x/y and x%y
276 // into a single instruction.
278 // Scalar integer multiply-high is also lowered to use two-result
279 // operations, to match the available instructions. However, plain multiply
280 // (low) operations are left as Legal, as there are single-result
281 // instructions for this in x86. Using the two-result multiply instructions
282 // when both high and low results are needed must be arranged by dagcombine.
283 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
285 setOperationAction(ISD::MULHS, VT, Expand);
286 setOperationAction(ISD::MULHU, VT, Expand);
287 setOperationAction(ISD::SDIV, VT, Expand);
288 setOperationAction(ISD::UDIV, VT, Expand);
289 setOperationAction(ISD::SREM, VT, Expand);
290 setOperationAction(ISD::UREM, VT, Expand);
292 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
293 setOperationAction(ISD::ADDC, VT, Custom);
294 setOperationAction(ISD::ADDE, VT, Custom);
295 setOperationAction(ISD::SUBC, VT, Custom);
296 setOperationAction(ISD::SUBE, VT, Custom);
299 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
300 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
301 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
302 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
303 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
304 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
305 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
306 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
307 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
312 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
313 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
314 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
315 if (Subtarget->is64Bit())
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
318 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
319 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
320 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
321 setOperationAction(ISD::FREM , MVT::f32 , Expand);
322 setOperationAction(ISD::FREM , MVT::f64 , Expand);
323 setOperationAction(ISD::FREM , MVT::f80 , Expand);
324 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
326 // Promote the i8 variants and force them on up to i32 which has a shorter
328 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
329 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
331 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
332 if (Subtarget->hasBMI()) {
333 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
335 if (Subtarget->is64Bit())
336 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
338 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
339 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
340 if (Subtarget->is64Bit())
341 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
344 if (Subtarget->hasLZCNT()) {
345 // When promoting the i8 variants, force them to i32 for a shorter
347 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
348 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
350 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
352 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
353 if (Subtarget->is64Bit())
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
356 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
357 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
358 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
362 if (Subtarget->is64Bit()) {
363 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
364 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
368 // Special handling for half-precision floating point conversions.
369 // If we don't have F16C support, then lower half float conversions
370 // into library calls.
371 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
372 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
373 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
376 // There's never any support for operations beyond MVT::f32.
377 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
378 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
379 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
380 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
382 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
383 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
384 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
385 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
386 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
387 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
389 if (Subtarget->hasPOPCNT()) {
390 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
392 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
393 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
395 if (Subtarget->is64Bit())
396 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
399 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
401 if (!Subtarget->hasMOVBE())
402 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
404 // These should be promoted to a larger select which is supported.
405 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
406 // X86 wants to expand cmov itself.
407 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
408 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
409 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
411 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
412 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
414 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
415 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
417 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
418 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
419 if (Subtarget->is64Bit()) {
420 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
421 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
423 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
424 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
425 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
426 // support continuation, user-level threading, and etc.. As a result, no
427 // other SjLj exception interfaces are implemented and please don't build
428 // your own exception handling based on them.
429 // LLVM/Clang supports zero-cost DWARF exception handling.
430 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
431 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
434 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
435 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
436 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
438 if (Subtarget->is64Bit())
439 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
440 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
441 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
442 if (Subtarget->is64Bit()) {
443 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
444 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
445 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
446 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
447 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
449 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
450 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
451 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
452 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
453 if (Subtarget->is64Bit()) {
454 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
455 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
456 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
459 if (Subtarget->hasSSE1())
460 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
462 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
464 // Expand certain atomics
465 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
467 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
468 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
469 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
472 if (Subtarget->hasCmpxchg16b()) {
473 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
476 // FIXME - use subtarget debug flags
477 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
478 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
479 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
482 if (Subtarget->is64Bit()) {
483 setExceptionPointerRegister(X86::RAX);
484 setExceptionSelectorRegister(X86::RDX);
486 setExceptionPointerRegister(X86::EAX);
487 setExceptionSelectorRegister(X86::EDX);
489 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
490 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
492 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
493 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
495 setOperationAction(ISD::TRAP, MVT::Other, Legal);
496 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
498 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
499 setOperationAction(ISD::VASTART , MVT::Other, Custom);
500 setOperationAction(ISD::VAEND , MVT::Other, Expand);
501 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
502 // TargetInfo::X86_64ABIBuiltinVaList
503 setOperationAction(ISD::VAARG , MVT::Other, Custom);
504 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
506 // TargetInfo::CharPtrBuiltinVaList
507 setOperationAction(ISD::VAARG , MVT::Other, Expand);
508 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
511 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
512 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
514 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
516 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
517 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
518 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
520 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
521 // f32 and f64 use SSE.
522 // Set up the FP register classes.
523 addRegisterClass(MVT::f32, &X86::FR32RegClass);
524 addRegisterClass(MVT::f64, &X86::FR64RegClass);
526 // Use ANDPD to simulate FABS.
527 setOperationAction(ISD::FABS , MVT::f64, Custom);
528 setOperationAction(ISD::FABS , MVT::f32, Custom);
530 // Use XORP to simulate FNEG.
531 setOperationAction(ISD::FNEG , MVT::f64, Custom);
532 setOperationAction(ISD::FNEG , MVT::f32, Custom);
534 // Use ANDPD and ORPD to simulate FCOPYSIGN.
535 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
536 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
538 // Lower this to FGETSIGNx86 plus an AND.
539 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
540 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
542 // We don't support sin/cos/fmod
543 setOperationAction(ISD::FSIN , MVT::f64, Expand);
544 setOperationAction(ISD::FCOS , MVT::f64, Expand);
545 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
546 setOperationAction(ISD::FSIN , MVT::f32, Expand);
547 setOperationAction(ISD::FCOS , MVT::f32, Expand);
548 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
550 // Expand FP immediates into loads from the stack, except for the special
552 addLegalFPImmediate(APFloat(+0.0)); // xorpd
553 addLegalFPImmediate(APFloat(+0.0f)); // xorps
554 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
555 // Use SSE for f32, x87 for f64.
556 // Set up the FP register classes.
557 addRegisterClass(MVT::f32, &X86::FR32RegClass);
558 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
560 // Use ANDPS to simulate FABS.
561 setOperationAction(ISD::FABS , MVT::f32, Custom);
563 // Use XORP to simulate FNEG.
564 setOperationAction(ISD::FNEG , MVT::f32, Custom);
566 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
568 // Use ANDPS and ORPS to simulate FCOPYSIGN.
569 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
570 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
572 // We don't support sin/cos/fmod
573 setOperationAction(ISD::FSIN , MVT::f32, Expand);
574 setOperationAction(ISD::FCOS , MVT::f32, Expand);
575 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
577 // Special cases we handle for FP constants.
578 addLegalFPImmediate(APFloat(+0.0f)); // xorps
579 addLegalFPImmediate(APFloat(+0.0)); // FLD0
580 addLegalFPImmediate(APFloat(+1.0)); // FLD1
581 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
582 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
584 if (!TM.Options.UnsafeFPMath) {
585 setOperationAction(ISD::FSIN , MVT::f64, Expand);
586 setOperationAction(ISD::FCOS , MVT::f64, Expand);
587 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
589 } else if (!Subtarget->useSoftFloat()) {
590 // f32 and f64 in x87.
591 // Set up the FP register classes.
592 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
593 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
595 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
596 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
598 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
600 if (!TM.Options.UnsafeFPMath) {
601 setOperationAction(ISD::FSIN , MVT::f64, Expand);
602 setOperationAction(ISD::FSIN , MVT::f32, Expand);
603 setOperationAction(ISD::FCOS , MVT::f64, Expand);
604 setOperationAction(ISD::FCOS , MVT::f32, Expand);
605 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
606 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
608 addLegalFPImmediate(APFloat(+0.0)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
612 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
613 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
614 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
615 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
618 // We don't support FMA.
619 setOperationAction(ISD::FMA, MVT::f64, Expand);
620 setOperationAction(ISD::FMA, MVT::f32, Expand);
622 // Long double always uses X87.
623 if (!Subtarget->useSoftFloat()) {
624 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
625 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
626 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
628 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
629 addLegalFPImmediate(TmpFlt); // FLD0
631 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
634 APFloat TmpFlt2(+1.0);
635 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
637 addLegalFPImmediate(TmpFlt2); // FLD1
638 TmpFlt2.changeSign();
639 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
642 if (!TM.Options.UnsafeFPMath) {
643 setOperationAction(ISD::FSIN , MVT::f80, Expand);
644 setOperationAction(ISD::FCOS , MVT::f80, Expand);
645 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
648 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
649 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
650 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
651 setOperationAction(ISD::FRINT, MVT::f80, Expand);
652 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
653 setOperationAction(ISD::FMA, MVT::f80, Expand);
656 // Always use a library call for pow.
657 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
658 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
659 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
661 setOperationAction(ISD::FLOG, MVT::f80, Expand);
662 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
663 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
664 setOperationAction(ISD::FEXP, MVT::f80, Expand);
665 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
666 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
667 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
669 // First set operation action for all vector types to either promote
670 // (for widening) or expand (for scalarization). Then we will selectively
671 // turn on ones that can be effectively codegen'd.
672 for (MVT VT : MVT::vector_valuetypes()) {
673 setOperationAction(ISD::ADD , VT, Expand);
674 setOperationAction(ISD::SUB , VT, Expand);
675 setOperationAction(ISD::FADD, VT, Expand);
676 setOperationAction(ISD::FNEG, VT, Expand);
677 setOperationAction(ISD::FSUB, VT, Expand);
678 setOperationAction(ISD::MUL , VT, Expand);
679 setOperationAction(ISD::FMUL, VT, Expand);
680 setOperationAction(ISD::SDIV, VT, Expand);
681 setOperationAction(ISD::UDIV, VT, Expand);
682 setOperationAction(ISD::FDIV, VT, Expand);
683 setOperationAction(ISD::SREM, VT, Expand);
684 setOperationAction(ISD::UREM, VT, Expand);
685 setOperationAction(ISD::LOAD, VT, Expand);
686 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
687 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
688 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
689 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
690 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
691 setOperationAction(ISD::FABS, VT, Expand);
692 setOperationAction(ISD::FSIN, VT, Expand);
693 setOperationAction(ISD::FSINCOS, VT, Expand);
694 setOperationAction(ISD::FCOS, VT, Expand);
695 setOperationAction(ISD::FSINCOS, VT, Expand);
696 setOperationAction(ISD::FREM, VT, Expand);
697 setOperationAction(ISD::FMA, VT, Expand);
698 setOperationAction(ISD::FPOWI, VT, Expand);
699 setOperationAction(ISD::FSQRT, VT, Expand);
700 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
701 setOperationAction(ISD::FFLOOR, VT, Expand);
702 setOperationAction(ISD::FCEIL, VT, Expand);
703 setOperationAction(ISD::FTRUNC, VT, Expand);
704 setOperationAction(ISD::FRINT, VT, Expand);
705 setOperationAction(ISD::FNEARBYINT, VT, Expand);
706 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
707 setOperationAction(ISD::MULHS, VT, Expand);
708 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
709 setOperationAction(ISD::MULHU, VT, Expand);
710 setOperationAction(ISD::SDIVREM, VT, Expand);
711 setOperationAction(ISD::UDIVREM, VT, Expand);
712 setOperationAction(ISD::FPOW, VT, Expand);
713 setOperationAction(ISD::CTPOP, VT, Expand);
714 setOperationAction(ISD::CTTZ, VT, Expand);
715 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
716 setOperationAction(ISD::CTLZ, VT, Expand);
717 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
718 setOperationAction(ISD::SHL, VT, Expand);
719 setOperationAction(ISD::SRA, VT, Expand);
720 setOperationAction(ISD::SRL, VT, Expand);
721 setOperationAction(ISD::ROTL, VT, Expand);
722 setOperationAction(ISD::ROTR, VT, Expand);
723 setOperationAction(ISD::BSWAP, VT, Expand);
724 setOperationAction(ISD::SETCC, VT, Expand);
725 setOperationAction(ISD::FLOG, VT, Expand);
726 setOperationAction(ISD::FLOG2, VT, Expand);
727 setOperationAction(ISD::FLOG10, VT, Expand);
728 setOperationAction(ISD::FEXP, VT, Expand);
729 setOperationAction(ISD::FEXP2, VT, Expand);
730 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
731 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
732 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
733 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
734 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
735 setOperationAction(ISD::TRUNCATE, VT, Expand);
736 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
737 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
738 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
739 setOperationAction(ISD::VSELECT, VT, Expand);
740 setOperationAction(ISD::SELECT_CC, VT, Expand);
741 for (MVT InnerVT : MVT::vector_valuetypes()) {
742 setTruncStoreAction(InnerVT, VT, Expand);
744 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
745 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
747 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
748 // types, we have to deal with them whether we ask for Expansion or not.
749 // Setting Expand causes its own optimisation problems though, so leave
751 if (VT.getVectorElementType() == MVT::i1)
752 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
754 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
755 // split/scalarized right now.
756 if (VT.getVectorElementType() == MVT::f16)
757 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
761 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
762 // with -msoft-float, disable use of MMX as well.
763 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
764 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
765 // No operations on x86mmx supported, everything uses intrinsics.
768 // MMX-sized vectors (other than x86mmx) are expected to be expanded
769 // into smaller operations.
770 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
771 setOperationAction(ISD::MULHS, MMXTy, Expand);
772 setOperationAction(ISD::AND, MMXTy, Expand);
773 setOperationAction(ISD::OR, MMXTy, Expand);
774 setOperationAction(ISD::XOR, MMXTy, Expand);
775 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
776 setOperationAction(ISD::SELECT, MMXTy, Expand);
777 setOperationAction(ISD::BITCAST, MMXTy, Expand);
779 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
781 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
782 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
784 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
785 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
786 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
787 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
788 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
789 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
790 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
791 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
792 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
793 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
794 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
796 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
797 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
800 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
801 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
803 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
804 // registers cannot be used even for integer operations.
805 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
806 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
807 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
808 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
810 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
811 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
812 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
813 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
814 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
815 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
816 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
817 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
818 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
819 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
820 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
821 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
822 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
823 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
824 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
825 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
826 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
827 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
828 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
829 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
830 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
831 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
832 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
834 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
835 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
836 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
837 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
839 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
840 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
845 // Only provide customized ctpop vector bit twiddling for vector types we
846 // know to perform better than using the popcnt instructions on each vector
847 // element. If popcnt isn't supported, always provide the custom version.
848 if (!Subtarget->hasPOPCNT()) {
849 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
850 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
853 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
854 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
855 MVT VT = (MVT::SimpleValueType)i;
856 // Do not attempt to custom lower non-power-of-2 vectors
857 if (!isPowerOf2_32(VT.getVectorNumElements()))
859 // Do not attempt to custom lower non-128-bit vectors
860 if (!VT.is128BitVector())
862 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
863 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
864 setOperationAction(ISD::VSELECT, VT, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
868 // We support custom legalizing of sext and anyext loads for specific
869 // memory vector types which we can load as a scalar (or sequence of
870 // scalars) and extend in-register to a legal 128-bit vector type. For sext
871 // loads these must work with a single scalar load.
872 for (MVT VT : MVT::integer_vector_valuetypes()) {
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
874 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
875 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
881 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
888 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
889 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
893 if (Subtarget->is64Bit()) {
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
895 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
898 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
899 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
900 MVT VT = (MVT::SimpleValueType)i;
902 // Do not attempt to promote non-128-bit vectors
903 if (!VT.is128BitVector())
906 setOperationAction(ISD::AND, VT, Promote);
907 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
908 setOperationAction(ISD::OR, VT, Promote);
909 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
910 setOperationAction(ISD::XOR, VT, Promote);
911 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
912 setOperationAction(ISD::LOAD, VT, Promote);
913 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
914 setOperationAction(ISD::SELECT, VT, Promote);
915 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
918 // Custom lower v2i64 and v2f64 selects.
919 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
920 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
921 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
922 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
924 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
925 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
927 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
928 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
929 // As there is no 64-bit GPR available, we need build a special custom
930 // sequence to convert from v2i32 to v2f32.
931 if (!Subtarget->is64Bit())
932 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
934 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
935 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
937 for (MVT VT : MVT::fp_vector_valuetypes())
938 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
940 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
941 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
942 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
945 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
946 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
947 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
948 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
949 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
950 setOperationAction(ISD::FRINT, RoundedTy, Legal);
951 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
954 // FIXME: Do we need to handle scalar-to-vector here?
955 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
957 // We directly match byte blends in the backend as they match the VSELECT
959 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
961 // SSE41 brings specific instructions for doing vector sign extend even in
962 // cases where we don't have SRA.
963 for (MVT VT : MVT::integer_vector_valuetypes()) {
964 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
965 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
966 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
969 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
970 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
971 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
972 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
973 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
974 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
977 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
978 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
979 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
980 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
981 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
984 // i8 and i16 vectors are custom because the source register and source
985 // source memory operand types are not the same width. f32 vectors are
986 // custom since the immediate controlling the insert encodes additional
988 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
989 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
990 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
991 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
993 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
994 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
995 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
998 // FIXME: these should be Legal, but that's only for the case where
999 // the index is constant. For now custom expand to deal with that.
1000 if (Subtarget->is64Bit()) {
1001 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1002 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1006 if (Subtarget->hasSSE2()) {
1007 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1008 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1009 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1011 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1012 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1014 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1015 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1017 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1018 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1020 // In the customized shift lowering, the legal cases in AVX2 will be
1022 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1023 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1025 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1026 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1028 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1031 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1032 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1033 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1034 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1035 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1036 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1037 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1039 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1040 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1041 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1043 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1044 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1045 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1046 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1047 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1048 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1049 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1050 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1051 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1052 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1053 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1054 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1056 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1057 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1058 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1059 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1060 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1063 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1064 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1065 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1066 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1067 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1069 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1070 // even though v8i16 is a legal type.
1071 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1072 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1073 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1075 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1076 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1077 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1079 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1080 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1082 for (MVT VT : MVT::fp_vector_valuetypes())
1083 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1085 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1086 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1088 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1089 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1091 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1092 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1094 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1095 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1096 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1097 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1099 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1100 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1101 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1103 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1104 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1105 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1106 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1107 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1108 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1109 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1110 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1111 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1112 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1113 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1114 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1116 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1117 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1118 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1119 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1120 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1121 setOperationAction(ISD::FMA, MVT::f32, Legal);
1122 setOperationAction(ISD::FMA, MVT::f64, Legal);
1125 if (Subtarget->hasInt256()) {
1126 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1127 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1128 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1129 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1131 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1132 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1133 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1134 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1136 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1137 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1138 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1139 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1141 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1142 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1143 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1144 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1146 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1147 // when we have a 256bit-wide blend with immediate.
1148 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1150 // Only provide customized ctpop vector bit twiddling for vector types we
1151 // know to perform better than using the popcnt instructions on each
1152 // vector element. If popcnt isn't supported, always provide the custom
1154 if (!Subtarget->hasPOPCNT())
1155 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1157 // Custom CTPOP always performs better on natively supported v8i32
1158 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1160 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1161 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1162 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1163 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1164 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1165 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1166 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1168 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1169 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1170 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1171 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1172 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1173 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1175 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1176 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1177 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1178 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1180 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1181 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1182 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1183 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1185 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1186 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1187 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1188 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1191 // In the customized shift lowering, the legal cases in AVX2 will be
1193 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1194 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1196 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1197 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1199 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1201 // Custom lower several nodes for 256-bit types.
1202 for (MVT VT : MVT::vector_valuetypes()) {
1203 if (VT.getScalarSizeInBits() >= 32) {
1204 setOperationAction(ISD::MLOAD, VT, Legal);
1205 setOperationAction(ISD::MSTORE, VT, Legal);
1207 // Extract subvector is special because the value type
1208 // (result) is 128-bit but the source is 256-bit wide.
1209 if (VT.is128BitVector()) {
1210 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1212 // Do not attempt to custom lower other non-256-bit vectors
1213 if (!VT.is256BitVector())
1216 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1217 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1218 setOperationAction(ISD::VSELECT, VT, Custom);
1219 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1220 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1221 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1222 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1223 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1226 if (Subtarget->hasInt256())
1227 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1230 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1231 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1232 MVT VT = (MVT::SimpleValueType)i;
1234 // Do not attempt to promote non-256-bit vectors
1235 if (!VT.is256BitVector())
1238 setOperationAction(ISD::AND, VT, Promote);
1239 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1240 setOperationAction(ISD::OR, VT, Promote);
1241 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1242 setOperationAction(ISD::XOR, VT, Promote);
1243 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1244 setOperationAction(ISD::LOAD, VT, Promote);
1245 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1246 setOperationAction(ISD::SELECT, VT, Promote);
1247 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1251 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1252 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1253 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1254 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1255 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1257 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1258 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1259 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1261 for (MVT VT : MVT::fp_vector_valuetypes())
1262 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1264 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1265 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1266 setOperationAction(ISD::XOR, MVT::i1, Legal);
1267 setOperationAction(ISD::OR, MVT::i1, Legal);
1268 setOperationAction(ISD::AND, MVT::i1, Legal);
1269 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1270 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1271 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1272 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1273 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1275 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1276 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1277 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1278 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1279 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1280 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1282 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1283 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1284 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1285 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1286 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1287 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1288 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1289 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1291 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1292 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1293 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1294 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1295 if (Subtarget->is64Bit()) {
1296 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1297 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1298 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1299 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1301 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1302 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1303 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1304 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1305 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1306 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1307 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1308 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1309 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1310 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1311 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1312 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1313 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1314 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1315 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1316 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1318 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1319 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1320 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1321 if (Subtarget->hasDQI()) {
1322 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1323 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1325 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1326 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1327 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1328 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1329 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1330 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1331 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1332 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1333 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1334 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1335 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1336 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1337 if (Subtarget->hasDQI()) {
1338 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1339 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1341 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1342 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1343 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1344 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1345 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1346 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1347 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1348 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1349 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1350 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1352 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1353 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1354 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1355 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1356 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1358 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1359 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1361 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1363 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1364 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1365 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1366 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1367 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1368 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1369 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1370 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1371 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1372 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1373 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1375 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1376 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1378 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1379 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1381 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1383 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1384 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1386 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1387 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1389 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1390 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1392 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1393 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1394 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1395 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1396 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1397 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1399 if (Subtarget->hasCDI()) {
1400 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1401 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1403 if (Subtarget->hasDQI()) {
1404 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1405 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1406 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1408 // Custom lower several nodes.
1409 for (MVT VT : MVT::vector_valuetypes()) {
1410 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1412 setOperationAction(ISD::AND, VT, Legal);
1413 setOperationAction(ISD::OR, VT, Legal);
1414 setOperationAction(ISD::XOR, VT, Legal);
1416 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1417 setOperationAction(ISD::MGATHER, VT, Custom);
1418 setOperationAction(ISD::MSCATTER, VT, Custom);
1420 // Extract subvector is special because the value type
1421 // (result) is 256/128-bit but the source is 512-bit wide.
1422 if (VT.is128BitVector() || VT.is256BitVector()) {
1423 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1425 if (VT.getVectorElementType() == MVT::i1)
1426 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1428 // Do not attempt to custom lower other non-512-bit vectors
1429 if (!VT.is512BitVector())
1432 if (EltSize >= 32) {
1433 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1434 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1435 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1436 setOperationAction(ISD::VSELECT, VT, Legal);
1437 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1438 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1439 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1440 setOperationAction(ISD::MLOAD, VT, Legal);
1441 setOperationAction(ISD::MSTORE, VT, Legal);
1444 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1445 MVT VT = (MVT::SimpleValueType)i;
1447 // Do not attempt to promote non-512-bit vectors.
1448 if (!VT.is512BitVector())
1451 setOperationAction(ISD::SELECT, VT, Promote);
1452 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1456 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1457 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1458 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1460 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1461 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1463 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1464 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1465 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1466 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1467 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1468 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1469 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1470 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1471 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1474 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1475 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1476 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1477 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1478 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1479 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1480 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1481 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1483 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1484 const MVT VT = (MVT::SimpleValueType)i;
1486 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1488 // Do not attempt to promote non-512-bit vectors.
1489 if (!VT.is512BitVector())
1493 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1494 setOperationAction(ISD::VSELECT, VT, Legal);
1499 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1500 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1501 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1503 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1504 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1505 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1506 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1507 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1508 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1509 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1510 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1511 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1512 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1514 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1515 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1516 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1517 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1518 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1519 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1520 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1521 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1524 // We want to custom lower some of our intrinsics.
1525 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1526 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1527 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1528 if (!Subtarget->is64Bit())
1529 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1531 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1532 // handle type legalization for these operations here.
1534 // FIXME: We really should do custom legalization for addition and
1535 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1536 // than generic legalization for 64-bit multiplication-with-overflow, though.
1537 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1538 // Add/Sub/Mul with overflow operations are custom lowered.
1540 setOperationAction(ISD::SADDO, VT, Custom);
1541 setOperationAction(ISD::UADDO, VT, Custom);
1542 setOperationAction(ISD::SSUBO, VT, Custom);
1543 setOperationAction(ISD::USUBO, VT, Custom);
1544 setOperationAction(ISD::SMULO, VT, Custom);
1545 setOperationAction(ISD::UMULO, VT, Custom);
1549 if (!Subtarget->is64Bit()) {
1550 // These libcalls are not available in 32-bit.
1551 setLibcallName(RTLIB::SHL_I128, nullptr);
1552 setLibcallName(RTLIB::SRL_I128, nullptr);
1553 setLibcallName(RTLIB::SRA_I128, nullptr);
1556 // Combine sin / cos into one node or libcall if possible.
1557 if (Subtarget->hasSinCos()) {
1558 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1559 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1560 if (Subtarget->isTargetDarwin()) {
1561 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1562 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1563 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1564 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1568 if (Subtarget->isTargetWin64()) {
1569 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1570 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1571 setOperationAction(ISD::SREM, MVT::i128, Custom);
1572 setOperationAction(ISD::UREM, MVT::i128, Custom);
1573 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1574 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1577 // We have target-specific dag combine patterns for the following nodes:
1578 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1579 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1580 setTargetDAGCombine(ISD::BITCAST);
1581 setTargetDAGCombine(ISD::VSELECT);
1582 setTargetDAGCombine(ISD::SELECT);
1583 setTargetDAGCombine(ISD::SHL);
1584 setTargetDAGCombine(ISD::SRA);
1585 setTargetDAGCombine(ISD::SRL);
1586 setTargetDAGCombine(ISD::OR);
1587 setTargetDAGCombine(ISD::AND);
1588 setTargetDAGCombine(ISD::ADD);
1589 setTargetDAGCombine(ISD::FADD);
1590 setTargetDAGCombine(ISD::FSUB);
1591 setTargetDAGCombine(ISD::FMA);
1592 setTargetDAGCombine(ISD::SUB);
1593 setTargetDAGCombine(ISD::LOAD);
1594 setTargetDAGCombine(ISD::MLOAD);
1595 setTargetDAGCombine(ISD::STORE);
1596 setTargetDAGCombine(ISD::MSTORE);
1597 setTargetDAGCombine(ISD::ZERO_EXTEND);
1598 setTargetDAGCombine(ISD::ANY_EXTEND);
1599 setTargetDAGCombine(ISD::SIGN_EXTEND);
1600 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1601 setTargetDAGCombine(ISD::SINT_TO_FP);
1602 setTargetDAGCombine(ISD::SETCC);
1603 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1604 setTargetDAGCombine(ISD::BUILD_VECTOR);
1605 setTargetDAGCombine(ISD::MUL);
1606 setTargetDAGCombine(ISD::XOR);
1608 computeRegisterProperties(Subtarget->getRegisterInfo());
1610 // On Darwin, -Os means optimize for size without hurting performance,
1611 // do not reduce the limit.
1612 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1613 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1614 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1615 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1616 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1617 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1618 setPrefLoopAlignment(4); // 2^4 bytes.
1620 // Predictable cmov don't hurt on atom because it's in-order.
1621 PredictableSelectIsExpensive = !Subtarget->isAtom();
1622 EnableExtLdPromotion = true;
1623 setPrefFunctionAlignment(4); // 2^4 bytes.
1625 verifyIntrinsicTables();
1628 // This has so far only been implemented for 64-bit MachO.
1629 bool X86TargetLowering::useLoadStackGuardNode() const {
1630 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1633 TargetLoweringBase::LegalizeTypeAction
1634 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1635 if (ExperimentalVectorWideningLegalization &&
1636 VT.getVectorNumElements() != 1 &&
1637 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1638 return TypeWidenVector;
1640 return TargetLoweringBase::getPreferredVectorAction(VT);
1643 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1645 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1647 const unsigned NumElts = VT.getVectorNumElements();
1648 const EVT EltVT = VT.getVectorElementType();
1649 if (VT.is512BitVector()) {
1650 if (Subtarget->hasAVX512())
1651 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1652 EltVT == MVT::f32 || EltVT == MVT::f64)
1654 case 8: return MVT::v8i1;
1655 case 16: return MVT::v16i1;
1657 if (Subtarget->hasBWI())
1658 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1660 case 32: return MVT::v32i1;
1661 case 64: return MVT::v64i1;
1665 if (VT.is256BitVector() || VT.is128BitVector()) {
1666 if (Subtarget->hasVLX())
1667 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1668 EltVT == MVT::f32 || EltVT == MVT::f64)
1670 case 2: return MVT::v2i1;
1671 case 4: return MVT::v4i1;
1672 case 8: return MVT::v8i1;
1674 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1675 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1677 case 8: return MVT::v8i1;
1678 case 16: return MVT::v16i1;
1679 case 32: return MVT::v32i1;
1683 return VT.changeVectorElementTypeToInteger();
1686 /// Helper for getByValTypeAlignment to determine
1687 /// the desired ByVal argument alignment.
1688 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1691 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1692 if (VTy->getBitWidth() == 128)
1694 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1695 unsigned EltAlign = 0;
1696 getMaxByValAlign(ATy->getElementType(), EltAlign);
1697 if (EltAlign > MaxAlign)
1698 MaxAlign = EltAlign;
1699 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1700 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1701 unsigned EltAlign = 0;
1702 getMaxByValAlign(STy->getElementType(i), EltAlign);
1703 if (EltAlign > MaxAlign)
1704 MaxAlign = EltAlign;
1711 /// Return the desired alignment for ByVal aggregate
1712 /// function arguments in the caller parameter area. For X86, aggregates
1713 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1714 /// are at 4-byte boundaries.
1715 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1716 if (Subtarget->is64Bit()) {
1717 // Max of 8 and alignment of type.
1718 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1725 if (Subtarget->hasSSE1())
1726 getMaxByValAlign(Ty, Align);
1730 /// Returns the target specific optimal type for load
1731 /// and store operations as a result of memset, memcpy, and memmove
1732 /// lowering. If DstAlign is zero that means it's safe to destination
1733 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1734 /// means there isn't a need to check it against alignment requirement,
1735 /// probably because the source does not need to be loaded. If 'IsMemset' is
1736 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1737 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1738 /// source is constant so it does not need to be loaded.
1739 /// It returns EVT::Other if the type should be determined using generic
1740 /// target-independent logic.
1742 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1743 unsigned DstAlign, unsigned SrcAlign,
1744 bool IsMemset, bool ZeroMemset,
1746 MachineFunction &MF) const {
1747 const Function *F = MF.getFunction();
1748 if ((!IsMemset || ZeroMemset) &&
1749 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1751 (Subtarget->isUnalignedMemAccessFast() ||
1752 ((DstAlign == 0 || DstAlign >= 16) &&
1753 (SrcAlign == 0 || SrcAlign >= 16)))) {
1755 if (Subtarget->hasInt256())
1757 if (Subtarget->hasFp256())
1760 if (Subtarget->hasSSE2())
1762 if (Subtarget->hasSSE1())
1764 } else if (!MemcpyStrSrc && Size >= 8 &&
1765 !Subtarget->is64Bit() &&
1766 Subtarget->hasSSE2()) {
1767 // Do not use f64 to lower memcpy if source is string constant. It's
1768 // better to use i32 to avoid the loads.
1772 if (Subtarget->is64Bit() && Size >= 8)
1777 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1779 return X86ScalarSSEf32;
1780 else if (VT == MVT::f64)
1781 return X86ScalarSSEf64;
1786 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1791 *Fast = Subtarget->isUnalignedMemAccessFast();
1795 /// Return the entry encoding for a jump table in the
1796 /// current function. The returned value is a member of the
1797 /// MachineJumpTableInfo::JTEntryKind enum.
1798 unsigned X86TargetLowering::getJumpTableEncoding() const {
1799 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1801 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1802 Subtarget->isPICStyleGOT())
1803 return MachineJumpTableInfo::EK_Custom32;
1805 // Otherwise, use the normal jump table encoding heuristics.
1806 return TargetLowering::getJumpTableEncoding();
1809 bool X86TargetLowering::useSoftFloat() const {
1810 return Subtarget->useSoftFloat();
1814 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1815 const MachineBasicBlock *MBB,
1816 unsigned uid,MCContext &Ctx) const{
1817 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1818 Subtarget->isPICStyleGOT());
1819 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1821 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1822 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1825 /// Returns relocation base for the given PIC jumptable.
1826 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1827 SelectionDAG &DAG) const {
1828 if (!Subtarget->is64Bit())
1829 // This doesn't have SDLoc associated with it, but is not really the
1830 // same as a Register.
1831 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1835 /// This returns the relocation base for the given PIC jumptable,
1836 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1837 const MCExpr *X86TargetLowering::
1838 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1839 MCContext &Ctx) const {
1840 // X86-64 uses RIP relative addressing based on the jump table label.
1841 if (Subtarget->isPICStyleRIPRel())
1842 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1844 // Otherwise, the reference is relative to the PIC base.
1845 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1848 std::pair<const TargetRegisterClass *, uint8_t>
1849 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1851 const TargetRegisterClass *RRC = nullptr;
1853 switch (VT.SimpleTy) {
1855 return TargetLowering::findRepresentativeClass(TRI, VT);
1856 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1857 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1860 RRC = &X86::VR64RegClass;
1862 case MVT::f32: case MVT::f64:
1863 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1864 case MVT::v4f32: case MVT::v2f64:
1865 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1867 RRC = &X86::VR128RegClass;
1870 return std::make_pair(RRC, Cost);
1873 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1874 unsigned &Offset) const {
1875 if (!Subtarget->isTargetLinux())
1878 if (Subtarget->is64Bit()) {
1879 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1881 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1893 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1894 unsigned DestAS) const {
1895 assert(SrcAS != DestAS && "Expected different address spaces!");
1897 return SrcAS < 256 && DestAS < 256;
1900 //===----------------------------------------------------------------------===//
1901 // Return Value Calling Convention Implementation
1902 //===----------------------------------------------------------------------===//
1904 #include "X86GenCallingConv.inc"
1907 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1908 MachineFunction &MF, bool isVarArg,
1909 const SmallVectorImpl<ISD::OutputArg> &Outs,
1910 LLVMContext &Context) const {
1911 SmallVector<CCValAssign, 16> RVLocs;
1912 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1913 return CCInfo.CheckReturn(Outs, RetCC_X86);
1916 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
1917 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
1922 X86TargetLowering::LowerReturn(SDValue Chain,
1923 CallingConv::ID CallConv, bool isVarArg,
1924 const SmallVectorImpl<ISD::OutputArg> &Outs,
1925 const SmallVectorImpl<SDValue> &OutVals,
1926 SDLoc dl, SelectionDAG &DAG) const {
1927 MachineFunction &MF = DAG.getMachineFunction();
1928 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1930 SmallVector<CCValAssign, 16> RVLocs;
1931 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
1932 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1935 SmallVector<SDValue, 6> RetOps;
1936 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1937 // Operand #1 = Bytes To Pop
1938 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
1941 // Copy the result values into the output registers.
1942 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1943 CCValAssign &VA = RVLocs[i];
1944 assert(VA.isRegLoc() && "Can only return in registers!");
1945 SDValue ValToCopy = OutVals[i];
1946 EVT ValVT = ValToCopy.getValueType();
1948 // Promote values to the appropriate types.
1949 if (VA.getLocInfo() == CCValAssign::SExt)
1950 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1951 else if (VA.getLocInfo() == CCValAssign::ZExt)
1952 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1953 else if (VA.getLocInfo() == CCValAssign::AExt) {
1954 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
1955 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1957 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1959 else if (VA.getLocInfo() == CCValAssign::BCvt)
1960 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1962 assert(VA.getLocInfo() != CCValAssign::FPExt &&
1963 "Unexpected FP-extend for return value.");
1965 // If this is x86-64, and we disabled SSE, we can't return FP values,
1966 // or SSE or MMX vectors.
1967 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1968 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1969 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1970 report_fatal_error("SSE register return with SSE disabled");
1972 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1973 // llvm-gcc has never done it right and no one has noticed, so this
1974 // should be OK for now.
1975 if (ValVT == MVT::f64 &&
1976 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1977 report_fatal_error("SSE2 register return with SSE2 disabled");
1979 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1980 // the RET instruction and handled by the FP Stackifier.
1981 if (VA.getLocReg() == X86::FP0 ||
1982 VA.getLocReg() == X86::FP1) {
1983 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1984 // change the value to the FP stack register class.
1985 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1986 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1987 RetOps.push_back(ValToCopy);
1988 // Don't emit a copytoreg.
1992 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1993 // which is returned in RAX / RDX.
1994 if (Subtarget->is64Bit()) {
1995 if (ValVT == MVT::x86mmx) {
1996 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1997 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1998 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2000 // If we don't have SSE2 available, convert to v4f32 so the generated
2001 // register is legal.
2002 if (!Subtarget->hasSSE2())
2003 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2008 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2009 Flag = Chain.getValue(1);
2010 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2013 // All x86 ABIs require that for returning structs by value we copy
2014 // the sret argument into %rax/%eax (depending on ABI) for the return.
2015 // We saved the argument into a virtual register in the entry block,
2016 // so now we copy the value out and into %rax/%eax.
2018 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2019 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2020 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2021 // either case FuncInfo->setSRetReturnReg() will have been called.
2022 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2023 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2026 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2027 X86::RAX : X86::EAX;
2028 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2029 Flag = Chain.getValue(1);
2031 // RAX/EAX now acts like a return value.
2032 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2035 RetOps[0] = Chain; // Update chain.
2037 // Add the flag if we have it.
2039 RetOps.push_back(Flag);
2041 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2044 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2045 if (N->getNumValues() != 1)
2047 if (!N->hasNUsesOfValue(1, 0))
2050 SDValue TCChain = Chain;
2051 SDNode *Copy = *N->use_begin();
2052 if (Copy->getOpcode() == ISD::CopyToReg) {
2053 // If the copy has a glue operand, we conservatively assume it isn't safe to
2054 // perform a tail call.
2055 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2057 TCChain = Copy->getOperand(0);
2058 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2061 bool HasRet = false;
2062 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2064 if (UI->getOpcode() != X86ISD::RET_FLAG)
2066 // If we are returning more than one value, we can definitely
2067 // not make a tail call see PR19530
2068 if (UI->getNumOperands() > 4)
2070 if (UI->getNumOperands() == 4 &&
2071 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2084 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2085 ISD::NodeType ExtendKind) const {
2087 // TODO: Is this also valid on 32-bit?
2088 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2089 ReturnMVT = MVT::i8;
2091 ReturnMVT = MVT::i32;
2093 EVT MinVT = getRegisterType(Context, ReturnMVT);
2094 return VT.bitsLT(MinVT) ? MinVT : VT;
2097 /// Lower the result values of a call into the
2098 /// appropriate copies out of appropriate physical registers.
2101 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2102 CallingConv::ID CallConv, bool isVarArg,
2103 const SmallVectorImpl<ISD::InputArg> &Ins,
2104 SDLoc dl, SelectionDAG &DAG,
2105 SmallVectorImpl<SDValue> &InVals) const {
2107 // Assign locations to each value returned by this call.
2108 SmallVector<CCValAssign, 16> RVLocs;
2109 bool Is64Bit = Subtarget->is64Bit();
2110 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2112 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2114 // Copy all of the result registers out of their specified physreg.
2115 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2116 CCValAssign &VA = RVLocs[i];
2117 EVT CopyVT = VA.getLocVT();
2119 // If this is x86-64, and we disabled SSE, we can't return FP values
2120 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2121 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2122 report_fatal_error("SSE register return with SSE disabled");
2125 // If we prefer to use the value in xmm registers, copy it out as f80 and
2126 // use a truncate to move it from fp stack reg to xmm reg.
2127 bool RoundAfterCopy = false;
2128 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2129 isScalarFPTypeInSSEReg(VA.getValVT())) {
2131 RoundAfterCopy = (CopyVT != VA.getLocVT());
2134 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2135 CopyVT, InFlag).getValue(1);
2136 SDValue Val = Chain.getValue(0);
2139 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2140 // This truncation won't change the value.
2141 DAG.getIntPtrConstant(1, dl));
2143 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2144 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2146 InFlag = Chain.getValue(2);
2147 InVals.push_back(Val);
2153 //===----------------------------------------------------------------------===//
2154 // C & StdCall & Fast Calling Convention implementation
2155 //===----------------------------------------------------------------------===//
2156 // StdCall calling convention seems to be standard for many Windows' API
2157 // routines and around. It differs from C calling convention just a little:
2158 // callee should clean up the stack, not caller. Symbols should be also
2159 // decorated in some fancy way :) It doesn't support any vector arguments.
2160 // For info on fast calling convention see Fast Calling Convention (tail call)
2161 // implementation LowerX86_32FastCCCallTo.
2163 /// CallIsStructReturn - Determines whether a call uses struct return
2165 enum StructReturnType {
2170 static StructReturnType
2171 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2173 return NotStructReturn;
2175 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2176 if (!Flags.isSRet())
2177 return NotStructReturn;
2178 if (Flags.isInReg())
2179 return RegStructReturn;
2180 return StackStructReturn;
2183 /// Determines whether a function uses struct return semantics.
2184 static StructReturnType
2185 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2187 return NotStructReturn;
2189 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2190 if (!Flags.isSRet())
2191 return NotStructReturn;
2192 if (Flags.isInReg())
2193 return RegStructReturn;
2194 return StackStructReturn;
2197 /// Make a copy of an aggregate at address specified by "Src" to address
2198 /// "Dst" with size and alignment information specified by the specific
2199 /// parameter attribute. The copy will be passed as a byval function parameter.
2201 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2202 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2204 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2206 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2207 /*isVolatile*/false, /*AlwaysInline=*/true,
2208 /*isTailCall*/false,
2209 MachinePointerInfo(), MachinePointerInfo());
2212 /// Return true if the calling convention is one that
2213 /// supports tail call optimization.
2214 static bool IsTailCallConvention(CallingConv::ID CC) {
2215 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2216 CC == CallingConv::HiPE);
2219 /// \brief Return true if the calling convention is a C calling convention.
2220 static bool IsCCallConvention(CallingConv::ID CC) {
2221 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2222 CC == CallingConv::X86_64_SysV);
2225 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2226 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2230 CallingConv::ID CalleeCC = CS.getCallingConv();
2231 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2237 /// Return true if the function is being made into
2238 /// a tailcall target by changing its ABI.
2239 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2240 bool GuaranteedTailCallOpt) {
2241 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2245 X86TargetLowering::LowerMemArgument(SDValue Chain,
2246 CallingConv::ID CallConv,
2247 const SmallVectorImpl<ISD::InputArg> &Ins,
2248 SDLoc dl, SelectionDAG &DAG,
2249 const CCValAssign &VA,
2250 MachineFrameInfo *MFI,
2252 // Create the nodes corresponding to a load from this parameter slot.
2253 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2254 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2255 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2256 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2259 // If value is passed by pointer we have address passed instead of the value
2261 bool ExtendedInMem = VA.isExtInLoc() &&
2262 VA.getValVT().getScalarType() == MVT::i1;
2264 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2265 ValVT = VA.getLocVT();
2267 ValVT = VA.getValVT();
2269 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2270 // changed with more analysis.
2271 // In case of tail call optimization mark all arguments mutable. Since they
2272 // could be overwritten by lowering of arguments in case of a tail call.
2273 if (Flags.isByVal()) {
2274 unsigned Bytes = Flags.getByValSize();
2275 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2276 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2277 return DAG.getFrameIndex(FI, getPointerTy());
2279 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2280 VA.getLocMemOffset(), isImmutable);
2281 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2282 SDValue Val = DAG.getLoad(ValVT, dl, Chain, FIN,
2283 MachinePointerInfo::getFixedStack(FI),
2284 false, false, false, 0);
2285 return ExtendedInMem ?
2286 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2290 // FIXME: Get this from tablegen.
2291 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2292 const X86Subtarget *Subtarget) {
2293 assert(Subtarget->is64Bit());
2295 if (Subtarget->isCallingConvWin64(CallConv)) {
2296 static const MCPhysReg GPR64ArgRegsWin64[] = {
2297 X86::RCX, X86::RDX, X86::R8, X86::R9
2299 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2302 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2303 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2305 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2308 // FIXME: Get this from tablegen.
2309 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2310 CallingConv::ID CallConv,
2311 const X86Subtarget *Subtarget) {
2312 assert(Subtarget->is64Bit());
2313 if (Subtarget->isCallingConvWin64(CallConv)) {
2314 // The XMM registers which might contain var arg parameters are shadowed
2315 // in their paired GPR. So we only need to save the GPR to their home
2317 // TODO: __vectorcall will change this.
2321 const Function *Fn = MF.getFunction();
2322 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2323 bool isSoftFloat = Subtarget->useSoftFloat();
2324 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2325 "SSE register cannot be used when SSE is disabled!");
2326 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2327 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2331 static const MCPhysReg XMMArgRegs64Bit[] = {
2332 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2333 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2335 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2339 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2340 CallingConv::ID CallConv,
2342 const SmallVectorImpl<ISD::InputArg> &Ins,
2345 SmallVectorImpl<SDValue> &InVals)
2347 MachineFunction &MF = DAG.getMachineFunction();
2348 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2349 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2351 const Function* Fn = MF.getFunction();
2352 if (Fn->hasExternalLinkage() &&
2353 Subtarget->isTargetCygMing() &&
2354 Fn->getName() == "main")
2355 FuncInfo->setForceFramePointer(true);
2357 MachineFrameInfo *MFI = MF.getFrameInfo();
2358 bool Is64Bit = Subtarget->is64Bit();
2359 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2361 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2362 "Var args not supported with calling convention fastcc, ghc or hipe");
2364 // Assign locations to all of the incoming arguments.
2365 SmallVector<CCValAssign, 16> ArgLocs;
2366 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2368 // Allocate shadow area for Win64
2370 CCInfo.AllocateStack(32, 8);
2372 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2374 unsigned LastVal = ~0U;
2376 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2377 CCValAssign &VA = ArgLocs[i];
2378 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2380 assert(VA.getValNo() != LastVal &&
2381 "Don't support value assigned to multiple locs yet");
2383 LastVal = VA.getValNo();
2385 if (VA.isRegLoc()) {
2386 EVT RegVT = VA.getLocVT();
2387 const TargetRegisterClass *RC;
2388 if (RegVT == MVT::i32)
2389 RC = &X86::GR32RegClass;
2390 else if (Is64Bit && RegVT == MVT::i64)
2391 RC = &X86::GR64RegClass;
2392 else if (RegVT == MVT::f32)
2393 RC = &X86::FR32RegClass;
2394 else if (RegVT == MVT::f64)
2395 RC = &X86::FR64RegClass;
2396 else if (RegVT.is512BitVector())
2397 RC = &X86::VR512RegClass;
2398 else if (RegVT.is256BitVector())
2399 RC = &X86::VR256RegClass;
2400 else if (RegVT.is128BitVector())
2401 RC = &X86::VR128RegClass;
2402 else if (RegVT == MVT::x86mmx)
2403 RC = &X86::VR64RegClass;
2404 else if (RegVT == MVT::i1)
2405 RC = &X86::VK1RegClass;
2406 else if (RegVT == MVT::v8i1)
2407 RC = &X86::VK8RegClass;
2408 else if (RegVT == MVT::v16i1)
2409 RC = &X86::VK16RegClass;
2410 else if (RegVT == MVT::v32i1)
2411 RC = &X86::VK32RegClass;
2412 else if (RegVT == MVT::v64i1)
2413 RC = &X86::VK64RegClass;
2415 llvm_unreachable("Unknown argument type!");
2417 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2418 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2420 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2421 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2423 if (VA.getLocInfo() == CCValAssign::SExt)
2424 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2425 DAG.getValueType(VA.getValVT()));
2426 else if (VA.getLocInfo() == CCValAssign::ZExt)
2427 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2428 DAG.getValueType(VA.getValVT()));
2429 else if (VA.getLocInfo() == CCValAssign::BCvt)
2430 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2432 if (VA.isExtInLoc()) {
2433 // Handle MMX values passed in XMM regs.
2434 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2435 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2437 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2440 assert(VA.isMemLoc());
2441 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2444 // If value is passed via pointer - do a load.
2445 if (VA.getLocInfo() == CCValAssign::Indirect)
2446 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2447 MachinePointerInfo(), false, false, false, 0);
2449 InVals.push_back(ArgValue);
2452 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2453 // All x86 ABIs require that for returning structs by value we copy the
2454 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2455 // the argument into a virtual register so that we can access it from the
2457 if (Ins[i].Flags.isSRet()) {
2458 unsigned Reg = FuncInfo->getSRetReturnReg();
2460 MVT PtrTy = getPointerTy();
2461 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2462 FuncInfo->setSRetReturnReg(Reg);
2464 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2465 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2470 unsigned StackSize = CCInfo.getNextStackOffset();
2471 // Align stack specially for tail calls.
2472 if (FuncIsMadeTailCallSafe(CallConv,
2473 MF.getTarget().Options.GuaranteedTailCallOpt))
2474 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2476 // If the function takes variable number of arguments, make a frame index for
2477 // the start of the first vararg value... for expansion of llvm.va_start. We
2478 // can skip this if there are no va_start calls.
2479 if (MFI->hasVAStart() &&
2480 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2481 CallConv != CallingConv::X86_ThisCall))) {
2482 FuncInfo->setVarArgsFrameIndex(
2483 MFI->CreateFixedObject(1, StackSize, true));
2486 MachineModuleInfo &MMI = MF.getMMI();
2487 const Function *WinEHParent = nullptr;
2488 if (IsWin64 && MMI.hasWinEHFuncInfo(Fn))
2489 WinEHParent = MMI.getWinEHParent(Fn);
2490 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2491 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2493 // Figure out if XMM registers are in use.
2494 assert(!(Subtarget->useSoftFloat() &&
2495 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2496 "SSE register cannot be used when SSE is disabled!");
2498 // 64-bit calling conventions support varargs and register parameters, so we
2499 // have to do extra work to spill them in the prologue.
2500 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2501 // Find the first unallocated argument registers.
2502 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2503 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2504 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2505 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2506 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2507 "SSE register cannot be used when SSE is disabled!");
2509 // Gather all the live in physical registers.
2510 SmallVector<SDValue, 6> LiveGPRs;
2511 SmallVector<SDValue, 8> LiveXMMRegs;
2513 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2514 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2516 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2518 if (!ArgXMMs.empty()) {
2519 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2520 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2521 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2522 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2523 LiveXMMRegs.push_back(
2524 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2529 // Get to the caller-allocated home save location. Add 8 to account
2530 // for the return address.
2531 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2532 FuncInfo->setRegSaveFrameIndex(
2533 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2534 // Fixup to set vararg frame on shadow area (4 x i64).
2536 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2538 // For X86-64, if there are vararg parameters that are passed via
2539 // registers, then we must store them to their spots on the stack so
2540 // they may be loaded by deferencing the result of va_next.
2541 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2542 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2543 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2544 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2547 // Store the integer parameter registers.
2548 SmallVector<SDValue, 8> MemOps;
2549 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2551 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2552 for (SDValue Val : LiveGPRs) {
2553 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2554 DAG.getIntPtrConstant(Offset, dl));
2556 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2557 MachinePointerInfo::getFixedStack(
2558 FuncInfo->getRegSaveFrameIndex(), Offset),
2560 MemOps.push_back(Store);
2564 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2565 // Now store the XMM (fp + vector) parameter registers.
2566 SmallVector<SDValue, 12> SaveXMMOps;
2567 SaveXMMOps.push_back(Chain);
2568 SaveXMMOps.push_back(ALVal);
2569 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2570 FuncInfo->getRegSaveFrameIndex(), dl));
2571 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2572 FuncInfo->getVarArgsFPOffset(), dl));
2573 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2575 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2576 MVT::Other, SaveXMMOps));
2579 if (!MemOps.empty())
2580 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2581 } else if (IsWinEHOutlined) {
2582 // Get to the caller-allocated home save location. Add 8 to account
2583 // for the return address.
2584 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2585 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2586 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2588 MMI.getWinEHFuncInfo(Fn)
2589 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2590 FuncInfo->getRegSaveFrameIndex();
2592 // Store the second integer parameter (rdx) into rsp+16 relative to the
2593 // stack pointer at the entry of the function.
2595 DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), getPointerTy());
2596 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2597 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2598 Chain = DAG.getStore(
2599 Val.getValue(1), dl, Val, RSFIN,
2600 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2601 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2604 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2605 // Find the largest legal vector type.
2606 MVT VecVT = MVT::Other;
2607 // FIXME: Only some x86_32 calling conventions support AVX512.
2608 if (Subtarget->hasAVX512() &&
2609 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2610 CallConv == CallingConv::Intel_OCL_BI)))
2611 VecVT = MVT::v16f32;
2612 else if (Subtarget->hasAVX())
2614 else if (Subtarget->hasSSE2())
2617 // We forward some GPRs and some vector types.
2618 SmallVector<MVT, 2> RegParmTypes;
2619 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2620 RegParmTypes.push_back(IntVT);
2621 if (VecVT != MVT::Other)
2622 RegParmTypes.push_back(VecVT);
2624 // Compute the set of forwarded registers. The rest are scratch.
2625 SmallVectorImpl<ForwardedRegister> &Forwards =
2626 FuncInfo->getForwardedMustTailRegParms();
2627 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2629 // Conservatively forward AL on x86_64, since it might be used for varargs.
2630 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2631 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2632 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2635 // Copy all forwards from physical to virtual registers.
2636 for (ForwardedRegister &F : Forwards) {
2637 // FIXME: Can we use a less constrained schedule?
2638 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2639 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2640 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2644 // Some CCs need callee pop.
2645 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2646 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2647 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2649 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2650 // If this is an sret function, the return should pop the hidden pointer.
2651 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2652 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2653 argsAreStructReturn(Ins) == StackStructReturn)
2654 FuncInfo->setBytesToPopOnReturn(4);
2658 // RegSaveFrameIndex is X86-64 only.
2659 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2660 if (CallConv == CallingConv::X86_FastCall ||
2661 CallConv == CallingConv::X86_ThisCall)
2662 // fastcc functions can't have varargs.
2663 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2666 FuncInfo->setArgumentStackSize(StackSize);
2668 if (IsWinEHParent) {
2669 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2670 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2671 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2672 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2673 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2674 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2675 /*isVolatile=*/true,
2676 /*isNonTemporal=*/false, /*Alignment=*/0);
2683 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2684 SDValue StackPtr, SDValue Arg,
2685 SDLoc dl, SelectionDAG &DAG,
2686 const CCValAssign &VA,
2687 ISD::ArgFlagsTy Flags) const {
2688 unsigned LocMemOffset = VA.getLocMemOffset();
2689 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2690 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2691 if (Flags.isByVal())
2692 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2694 return DAG.getStore(Chain, dl, Arg, PtrOff,
2695 MachinePointerInfo::getStack(LocMemOffset),
2699 /// Emit a load of return address if tail call
2700 /// optimization is performed and it is required.
2702 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2703 SDValue &OutRetAddr, SDValue Chain,
2704 bool IsTailCall, bool Is64Bit,
2705 int FPDiff, SDLoc dl) const {
2706 // Adjust the Return address stack slot.
2707 EVT VT = getPointerTy();
2708 OutRetAddr = getReturnAddressFrameIndex(DAG);
2710 // Load the "old" Return address.
2711 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2712 false, false, false, 0);
2713 return SDValue(OutRetAddr.getNode(), 1);
2716 /// Emit a store of the return address if tail call
2717 /// optimization is performed and it is required (FPDiff!=0).
2718 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2719 SDValue Chain, SDValue RetAddrFrIdx,
2720 EVT PtrVT, unsigned SlotSize,
2721 int FPDiff, SDLoc dl) {
2722 // Store the return address to the appropriate stack slot.
2723 if (!FPDiff) return Chain;
2724 // Calculate the new stack slot for the return address.
2725 int NewReturnAddrFI =
2726 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2728 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2729 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2730 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2736 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2737 SmallVectorImpl<SDValue> &InVals) const {
2738 SelectionDAG &DAG = CLI.DAG;
2740 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2741 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2742 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2743 SDValue Chain = CLI.Chain;
2744 SDValue Callee = CLI.Callee;
2745 CallingConv::ID CallConv = CLI.CallConv;
2746 bool &isTailCall = CLI.IsTailCall;
2747 bool isVarArg = CLI.IsVarArg;
2749 MachineFunction &MF = DAG.getMachineFunction();
2750 bool Is64Bit = Subtarget->is64Bit();
2751 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2752 StructReturnType SR = callIsStructReturn(Outs);
2753 bool IsSibcall = false;
2754 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2756 if (MF.getTarget().Options.DisableTailCalls)
2759 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2761 // Force this to be a tail call. The verifier rules are enough to ensure
2762 // that we can lower this successfully without moving the return address
2765 } else if (isTailCall) {
2766 // Check if it's really possible to do a tail call.
2767 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2768 isVarArg, SR != NotStructReturn,
2769 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2770 Outs, OutVals, Ins, DAG);
2772 // Sibcalls are automatically detected tailcalls which do not require
2774 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2781 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2782 "Var args not supported with calling convention fastcc, ghc or hipe");
2784 // Analyze operands of the call, assigning locations to each operand.
2785 SmallVector<CCValAssign, 16> ArgLocs;
2786 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2788 // Allocate shadow area for Win64
2790 CCInfo.AllocateStack(32, 8);
2792 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2794 // Get a count of how many bytes are to be pushed on the stack.
2795 unsigned NumBytes = CCInfo.getNextStackOffset();
2797 // This is a sibcall. The memory operands are available in caller's
2798 // own caller's stack.
2800 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2801 IsTailCallConvention(CallConv))
2802 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2805 if (isTailCall && !IsSibcall && !IsMustTail) {
2806 // Lower arguments at fp - stackoffset + fpdiff.
2807 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2809 FPDiff = NumBytesCallerPushed - NumBytes;
2811 // Set the delta of movement of the returnaddr stackslot.
2812 // But only set if delta is greater than previous delta.
2813 if (FPDiff < X86Info->getTCReturnAddrDelta())
2814 X86Info->setTCReturnAddrDelta(FPDiff);
2817 unsigned NumBytesToPush = NumBytes;
2818 unsigned NumBytesToPop = NumBytes;
2820 // If we have an inalloca argument, all stack space has already been allocated
2821 // for us and be right at the top of the stack. We don't support multiple
2822 // arguments passed in memory when using inalloca.
2823 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2825 if (!ArgLocs.back().isMemLoc())
2826 report_fatal_error("cannot use inalloca attribute on a register "
2828 if (ArgLocs.back().getLocMemOffset() != 0)
2829 report_fatal_error("any parameter with the inalloca attribute must be "
2830 "the only memory argument");
2834 Chain = DAG.getCALLSEQ_START(
2835 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
2837 SDValue RetAddrFrIdx;
2838 // Load return address for tail calls.
2839 if (isTailCall && FPDiff)
2840 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2841 Is64Bit, FPDiff, dl);
2843 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2844 SmallVector<SDValue, 8> MemOpChains;
2847 // Walk the register/memloc assignments, inserting copies/loads. In the case
2848 // of tail call optimization arguments are handle later.
2849 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2850 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2851 // Skip inalloca arguments, they have already been written.
2852 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2853 if (Flags.isInAlloca())
2856 CCValAssign &VA = ArgLocs[i];
2857 EVT RegVT = VA.getLocVT();
2858 SDValue Arg = OutVals[i];
2859 bool isByVal = Flags.isByVal();
2861 // Promote the value if needed.
2862 switch (VA.getLocInfo()) {
2863 default: llvm_unreachable("Unknown loc info!");
2864 case CCValAssign::Full: break;
2865 case CCValAssign::SExt:
2866 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2868 case CCValAssign::ZExt:
2869 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2871 case CCValAssign::AExt:
2872 if (Arg.getValueType().isVector() &&
2873 Arg.getValueType().getScalarType() == MVT::i1)
2874 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2875 else if (RegVT.is128BitVector()) {
2876 // Special case: passing MMX values in XMM registers.
2877 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2878 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2879 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2881 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2883 case CCValAssign::BCvt:
2884 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2886 case CCValAssign::Indirect: {
2887 // Store the argument.
2888 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2889 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2890 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2891 MachinePointerInfo::getFixedStack(FI),
2898 if (VA.isRegLoc()) {
2899 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2900 if (isVarArg && IsWin64) {
2901 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2902 // shadow reg if callee is a varargs function.
2903 unsigned ShadowReg = 0;
2904 switch (VA.getLocReg()) {
2905 case X86::XMM0: ShadowReg = X86::RCX; break;
2906 case X86::XMM1: ShadowReg = X86::RDX; break;
2907 case X86::XMM2: ShadowReg = X86::R8; break;
2908 case X86::XMM3: ShadowReg = X86::R9; break;
2911 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2913 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2914 assert(VA.isMemLoc());
2915 if (!StackPtr.getNode())
2916 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2918 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2919 dl, DAG, VA, Flags));
2923 if (!MemOpChains.empty())
2924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2926 if (Subtarget->isPICStyleGOT()) {
2927 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2930 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2931 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2933 // If we are tail calling and generating PIC/GOT style code load the
2934 // address of the callee into ECX. The value in ecx is used as target of
2935 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2936 // for tail calls on PIC/GOT architectures. Normally we would just put the
2937 // address of GOT into ebx and then call target@PLT. But for tail calls
2938 // ebx would be restored (since ebx is callee saved) before jumping to the
2941 // Note: The actual moving to ECX is done further down.
2942 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2943 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2944 !G->getGlobal()->hasProtectedVisibility())
2945 Callee = LowerGlobalAddress(Callee, DAG);
2946 else if (isa<ExternalSymbolSDNode>(Callee))
2947 Callee = LowerExternalSymbol(Callee, DAG);
2951 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2952 // From AMD64 ABI document:
2953 // For calls that may call functions that use varargs or stdargs
2954 // (prototype-less calls or calls to functions containing ellipsis (...) in
2955 // the declaration) %al is used as hidden argument to specify the number
2956 // of SSE registers used. The contents of %al do not need to match exactly
2957 // the number of registers, but must be an ubound on the number of SSE
2958 // registers used and is in the range 0 - 8 inclusive.
2960 // Count the number of XMM registers allocated.
2961 static const MCPhysReg XMMArgRegs[] = {
2962 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2963 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2965 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2966 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2967 && "SSE registers cannot be used when SSE is disabled");
2969 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2970 DAG.getConstant(NumXMMRegs, dl,
2974 if (isVarArg && IsMustTail) {
2975 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
2976 for (const auto &F : Forwards) {
2977 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2978 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
2982 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
2983 // don't need this because the eligibility check rejects calls that require
2984 // shuffling arguments passed in memory.
2985 if (!IsSibcall && isTailCall) {
2986 // Force all the incoming stack arguments to be loaded from the stack
2987 // before any new outgoing arguments are stored to the stack, because the
2988 // outgoing stack slots may alias the incoming argument stack slots, and
2989 // the alias isn't otherwise explicit. This is slightly more conservative
2990 // than necessary, because it means that each store effectively depends
2991 // on every argument instead of just those arguments it would clobber.
2992 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2994 SmallVector<SDValue, 8> MemOpChains2;
2997 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2998 CCValAssign &VA = ArgLocs[i];
3001 assert(VA.isMemLoc());
3002 SDValue Arg = OutVals[i];
3003 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3004 // Skip inalloca arguments. They don't require any work.
3005 if (Flags.isInAlloca())
3007 // Create frame index.
3008 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3009 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3010 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3011 FIN = DAG.getFrameIndex(FI, getPointerTy());
3013 if (Flags.isByVal()) {
3014 // Copy relative to framepointer.
3015 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3016 if (!StackPtr.getNode())
3017 StackPtr = DAG.getCopyFromReg(Chain, dl,
3018 RegInfo->getStackRegister(),
3020 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3022 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3026 // Store relative to framepointer.
3027 MemOpChains2.push_back(
3028 DAG.getStore(ArgChain, dl, Arg, FIN,
3029 MachinePointerInfo::getFixedStack(FI),
3034 if (!MemOpChains2.empty())
3035 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3037 // Store the return address to the appropriate stack slot.
3038 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3039 getPointerTy(), RegInfo->getSlotSize(),
3043 // Build a sequence of copy-to-reg nodes chained together with token chain
3044 // and flag operands which copy the outgoing args into registers.
3046 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3047 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3048 RegsToPass[i].second, InFlag);
3049 InFlag = Chain.getValue(1);
3052 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3053 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3054 // In the 64-bit large code model, we have to make all calls
3055 // through a register, since the call instruction's 32-bit
3056 // pc-relative offset may not be large enough to hold the whole
3058 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3059 // If the callee is a GlobalAddress node (quite common, every direct call
3060 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3062 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3064 // We should use extra load for direct calls to dllimported functions in
3066 const GlobalValue *GV = G->getGlobal();
3067 if (!GV->hasDLLImportStorageClass()) {
3068 unsigned char OpFlags = 0;
3069 bool ExtraLoad = false;
3070 unsigned WrapperKind = ISD::DELETED_NODE;
3072 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3073 // external symbols most go through the PLT in PIC mode. If the symbol
3074 // has hidden or protected visibility, or if it is static or local, then
3075 // we don't need to use the PLT - we can directly call it.
3076 if (Subtarget->isTargetELF() &&
3077 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3078 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3079 OpFlags = X86II::MO_PLT;
3080 } else if (Subtarget->isPICStyleStubAny() &&
3081 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3082 (!Subtarget->getTargetTriple().isMacOSX() ||
3083 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3084 // PC-relative references to external symbols should go through $stub,
3085 // unless we're building with the leopard linker or later, which
3086 // automatically synthesizes these stubs.
3087 OpFlags = X86II::MO_DARWIN_STUB;
3088 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3089 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3090 // If the function is marked as non-lazy, generate an indirect call
3091 // which loads from the GOT directly. This avoids runtime overhead
3092 // at the cost of eager binding (and one extra byte of encoding).
3093 OpFlags = X86II::MO_GOTPCREL;
3094 WrapperKind = X86ISD::WrapperRIP;
3098 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3099 G->getOffset(), OpFlags);
3101 // Add a wrapper if needed.
3102 if (WrapperKind != ISD::DELETED_NODE)
3103 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3104 // Add extra indirection if needed.
3106 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3107 MachinePointerInfo::getGOT(),
3108 false, false, false, 0);
3110 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3111 unsigned char OpFlags = 0;
3113 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3114 // external symbols should go through the PLT.
3115 if (Subtarget->isTargetELF() &&
3116 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3117 OpFlags = X86II::MO_PLT;
3118 } else if (Subtarget->isPICStyleStubAny() &&
3119 (!Subtarget->getTargetTriple().isMacOSX() ||
3120 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3121 // PC-relative references to external symbols should go through $stub,
3122 // unless we're building with the leopard linker or later, which
3123 // automatically synthesizes these stubs.
3124 OpFlags = X86II::MO_DARWIN_STUB;
3127 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3129 } else if (Subtarget->isTarget64BitILP32() &&
3130 Callee->getValueType(0) == MVT::i32) {
3131 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3132 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3135 // Returns a chain & a flag for retval copy to use.
3136 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3137 SmallVector<SDValue, 8> Ops;
3139 if (!IsSibcall && isTailCall) {
3140 Chain = DAG.getCALLSEQ_END(Chain,
3141 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3142 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3143 InFlag = Chain.getValue(1);
3146 Ops.push_back(Chain);
3147 Ops.push_back(Callee);
3150 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3152 // Add argument registers to the end of the list so that they are known live
3154 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3155 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3156 RegsToPass[i].second.getValueType()));
3158 // Add a register mask operand representing the call-preserved registers.
3159 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3160 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
3161 assert(Mask && "Missing call preserved mask for calling convention");
3162 Ops.push_back(DAG.getRegisterMask(Mask));
3164 if (InFlag.getNode())
3165 Ops.push_back(InFlag);
3169 //// If this is the first return lowered for this function, add the regs
3170 //// to the liveout set for the function.
3171 // This isn't right, although it's probably harmless on x86; liveouts
3172 // should be computed from returns not tail calls. Consider a void
3173 // function making a tail call to a function returning int.
3174 MF.getFrameInfo()->setHasTailCall();
3175 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3178 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3179 InFlag = Chain.getValue(1);
3181 // Create the CALLSEQ_END node.
3182 unsigned NumBytesForCalleeToPop;
3183 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3184 DAG.getTarget().Options.GuaranteedTailCallOpt))
3185 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3186 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3187 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3188 SR == StackStructReturn)
3189 // If this is a call to a struct-return function, the callee
3190 // pops the hidden struct pointer, so we have to push it back.
3191 // This is common for Darwin/X86, Linux & Mingw32 targets.
3192 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3193 NumBytesForCalleeToPop = 4;
3195 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3197 // Returns a flag for retval copy to use.
3199 Chain = DAG.getCALLSEQ_END(Chain,
3200 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3201 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3204 InFlag = Chain.getValue(1);
3207 // Handle result values, copying them out of physregs into vregs that we
3209 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3210 Ins, dl, DAG, InVals);
3213 //===----------------------------------------------------------------------===//
3214 // Fast Calling Convention (tail call) implementation
3215 //===----------------------------------------------------------------------===//
3217 // Like std call, callee cleans arguments, convention except that ECX is
3218 // reserved for storing the tail called function address. Only 2 registers are
3219 // free for argument passing (inreg). Tail call optimization is performed
3221 // * tailcallopt is enabled
3222 // * caller/callee are fastcc
3223 // On X86_64 architecture with GOT-style position independent code only local
3224 // (within module) calls are supported at the moment.
3225 // To keep the stack aligned according to platform abi the function
3226 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3227 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3228 // If a tail called function callee has more arguments than the caller the
3229 // caller needs to make sure that there is room to move the RETADDR to. This is
3230 // achieved by reserving an area the size of the argument delta right after the
3231 // original RETADDR, but before the saved framepointer or the spilled registers
3232 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3244 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3245 /// for a 16 byte align requirement.
3247 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3248 SelectionDAG& DAG) const {
3249 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3250 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3251 unsigned StackAlignment = TFI.getStackAlignment();
3252 uint64_t AlignMask = StackAlignment - 1;
3253 int64_t Offset = StackSize;
3254 unsigned SlotSize = RegInfo->getSlotSize();
3255 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3256 // Number smaller than 12 so just add the difference.
3257 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3259 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3260 Offset = ((~AlignMask) & Offset) + StackAlignment +
3261 (StackAlignment-SlotSize);
3266 /// MatchingStackOffset - Return true if the given stack call argument is
3267 /// already available in the same position (relatively) of the caller's
3268 /// incoming argument stack.
3270 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3271 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3272 const X86InstrInfo *TII) {
3273 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3275 if (Arg.getOpcode() == ISD::CopyFromReg) {
3276 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3277 if (!TargetRegisterInfo::isVirtualRegister(VR))
3279 MachineInstr *Def = MRI->getVRegDef(VR);
3282 if (!Flags.isByVal()) {
3283 if (!TII->isLoadFromStackSlot(Def, FI))
3286 unsigned Opcode = Def->getOpcode();
3287 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3288 Opcode == X86::LEA64_32r) &&
3289 Def->getOperand(1).isFI()) {
3290 FI = Def->getOperand(1).getIndex();
3291 Bytes = Flags.getByValSize();
3295 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3296 if (Flags.isByVal())
3297 // ByVal argument is passed in as a pointer but it's now being
3298 // dereferenced. e.g.
3299 // define @foo(%struct.X* %A) {
3300 // tail call @bar(%struct.X* byval %A)
3303 SDValue Ptr = Ld->getBasePtr();
3304 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3307 FI = FINode->getIndex();
3308 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3309 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3310 FI = FINode->getIndex();
3311 Bytes = Flags.getByValSize();
3315 assert(FI != INT_MAX);
3316 if (!MFI->isFixedObjectIndex(FI))
3318 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3321 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3322 /// for tail call optimization. Targets which want to do tail call
3323 /// optimization should implement this function.
3325 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3326 CallingConv::ID CalleeCC,
3328 bool isCalleeStructRet,
3329 bool isCallerStructRet,
3331 const SmallVectorImpl<ISD::OutputArg> &Outs,
3332 const SmallVectorImpl<SDValue> &OutVals,
3333 const SmallVectorImpl<ISD::InputArg> &Ins,
3334 SelectionDAG &DAG) const {
3335 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3338 // If -tailcallopt is specified, make fastcc functions tail-callable.
3339 const MachineFunction &MF = DAG.getMachineFunction();
3340 const Function *CallerF = MF.getFunction();
3342 // If the function return type is x86_fp80 and the callee return type is not,
3343 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3344 // perform a tailcall optimization here.
3345 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3348 CallingConv::ID CallerCC = CallerF->getCallingConv();
3349 bool CCMatch = CallerCC == CalleeCC;
3350 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3351 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3353 // Win64 functions have extra shadow space for argument homing. Don't do the
3354 // sibcall if the caller and callee have mismatched expectations for this
3356 if (IsCalleeWin64 != IsCallerWin64)
3359 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3360 if (IsTailCallConvention(CalleeCC) && CCMatch)
3365 // Look for obvious safe cases to perform tail call optimization that do not
3366 // require ABI changes. This is what gcc calls sibcall.
3368 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3369 // emit a special epilogue.
3370 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3371 if (RegInfo->needsStackRealignment(MF))
3374 // Also avoid sibcall optimization if either caller or callee uses struct
3375 // return semantics.
3376 if (isCalleeStructRet || isCallerStructRet)
3379 // An stdcall/thiscall caller is expected to clean up its arguments; the
3380 // callee isn't going to do that.
3381 // FIXME: this is more restrictive than needed. We could produce a tailcall
3382 // when the stack adjustment matches. For example, with a thiscall that takes
3383 // only one argument.
3384 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3385 CallerCC == CallingConv::X86_ThisCall))
3388 // Do not sibcall optimize vararg calls unless all arguments are passed via
3390 if (isVarArg && !Outs.empty()) {
3392 // Optimizing for varargs on Win64 is unlikely to be safe without
3393 // additional testing.
3394 if (IsCalleeWin64 || IsCallerWin64)
3397 SmallVector<CCValAssign, 16> ArgLocs;
3398 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3401 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3402 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3403 if (!ArgLocs[i].isRegLoc())
3407 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3408 // stack. Therefore, if it's not used by the call it is not safe to optimize
3409 // this into a sibcall.
3410 bool Unused = false;
3411 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3418 SmallVector<CCValAssign, 16> RVLocs;
3419 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3421 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3422 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3423 CCValAssign &VA = RVLocs[i];
3424 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3429 // If the calling conventions do not match, then we'd better make sure the
3430 // results are returned in the same way as what the caller expects.
3432 SmallVector<CCValAssign, 16> RVLocs1;
3433 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3435 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3437 SmallVector<CCValAssign, 16> RVLocs2;
3438 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3440 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3442 if (RVLocs1.size() != RVLocs2.size())
3444 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3445 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3447 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3449 if (RVLocs1[i].isRegLoc()) {
3450 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3453 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3459 // If the callee takes no arguments then go on to check the results of the
3461 if (!Outs.empty()) {
3462 // Check if stack adjustment is needed. For now, do not do this if any
3463 // argument is passed on the stack.
3464 SmallVector<CCValAssign, 16> ArgLocs;
3465 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3468 // Allocate shadow area for Win64
3470 CCInfo.AllocateStack(32, 8);
3472 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3473 if (CCInfo.getNextStackOffset()) {
3474 MachineFunction &MF = DAG.getMachineFunction();
3475 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3478 // Check if the arguments are already laid out in the right way as
3479 // the caller's fixed stack objects.
3480 MachineFrameInfo *MFI = MF.getFrameInfo();
3481 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3482 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3483 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3484 CCValAssign &VA = ArgLocs[i];
3485 SDValue Arg = OutVals[i];
3486 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3487 if (VA.getLocInfo() == CCValAssign::Indirect)
3489 if (!VA.isRegLoc()) {
3490 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3497 // If the tailcall address may be in a register, then make sure it's
3498 // possible to register allocate for it. In 32-bit, the call address can
3499 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3500 // callee-saved registers are restored. These happen to be the same
3501 // registers used to pass 'inreg' arguments so watch out for those.
3502 if (!Subtarget->is64Bit() &&
3503 ((!isa<GlobalAddressSDNode>(Callee) &&
3504 !isa<ExternalSymbolSDNode>(Callee)) ||
3505 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3506 unsigned NumInRegs = 0;
3507 // In PIC we need an extra register to formulate the address computation
3509 unsigned MaxInRegs =
3510 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3512 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3513 CCValAssign &VA = ArgLocs[i];
3516 unsigned Reg = VA.getLocReg();
3519 case X86::EAX: case X86::EDX: case X86::ECX:
3520 if (++NumInRegs == MaxInRegs)
3532 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3533 const TargetLibraryInfo *libInfo) const {
3534 return X86::createFastISel(funcInfo, libInfo);
3537 //===----------------------------------------------------------------------===//
3538 // Other Lowering Hooks
3539 //===----------------------------------------------------------------------===//
3541 static bool MayFoldLoad(SDValue Op) {
3542 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3545 static bool MayFoldIntoStore(SDValue Op) {
3546 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3549 static bool isTargetShuffle(unsigned Opcode) {
3551 default: return false;
3552 case X86ISD::BLENDI:
3553 case X86ISD::PSHUFB:
3554 case X86ISD::PSHUFD:
3555 case X86ISD::PSHUFHW:
3556 case X86ISD::PSHUFLW:
3558 case X86ISD::PALIGNR:
3559 case X86ISD::MOVLHPS:
3560 case X86ISD::MOVLHPD:
3561 case X86ISD::MOVHLPS:
3562 case X86ISD::MOVLPS:
3563 case X86ISD::MOVLPD:
3564 case X86ISD::MOVSHDUP:
3565 case X86ISD::MOVSLDUP:
3566 case X86ISD::MOVDDUP:
3569 case X86ISD::UNPCKL:
3570 case X86ISD::UNPCKH:
3571 case X86ISD::VPERMILPI:
3572 case X86ISD::VPERM2X128:
3573 case X86ISD::VPERMI:
3578 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3579 SDValue V1, unsigned TargetMask,
3580 SelectionDAG &DAG) {
3582 default: llvm_unreachable("Unknown x86 shuffle node");
3583 case X86ISD::PSHUFD:
3584 case X86ISD::PSHUFHW:
3585 case X86ISD::PSHUFLW:
3586 case X86ISD::VPERMILPI:
3587 case X86ISD::VPERMI:
3588 return DAG.getNode(Opc, dl, VT, V1,
3589 DAG.getConstant(TargetMask, dl, MVT::i8));
3593 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3594 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3596 default: llvm_unreachable("Unknown x86 shuffle node");
3597 case X86ISD::MOVLHPS:
3598 case X86ISD::MOVLHPD:
3599 case X86ISD::MOVHLPS:
3600 case X86ISD::MOVLPS:
3601 case X86ISD::MOVLPD:
3604 case X86ISD::UNPCKL:
3605 case X86ISD::UNPCKH:
3606 return DAG.getNode(Opc, dl, VT, V1, V2);
3610 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3611 MachineFunction &MF = DAG.getMachineFunction();
3612 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3613 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3614 int ReturnAddrIndex = FuncInfo->getRAIndex();
3616 if (ReturnAddrIndex == 0) {
3617 // Set up a frame object for the return address.
3618 unsigned SlotSize = RegInfo->getSlotSize();
3619 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3622 FuncInfo->setRAIndex(ReturnAddrIndex);
3625 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3628 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3629 bool hasSymbolicDisplacement) {
3630 // Offset should fit into 32 bit immediate field.
3631 if (!isInt<32>(Offset))
3634 // If we don't have a symbolic displacement - we don't have any extra
3636 if (!hasSymbolicDisplacement)
3639 // FIXME: Some tweaks might be needed for medium code model.
3640 if (M != CodeModel::Small && M != CodeModel::Kernel)
3643 // For small code model we assume that latest object is 16MB before end of 31
3644 // bits boundary. We may also accept pretty large negative constants knowing
3645 // that all objects are in the positive half of address space.
3646 if (M == CodeModel::Small && Offset < 16*1024*1024)
3649 // For kernel code model we know that all object resist in the negative half
3650 // of 32bits address space. We may not accept negative offsets, since they may
3651 // be just off and we may accept pretty large positive ones.
3652 if (M == CodeModel::Kernel && Offset >= 0)
3658 /// isCalleePop - Determines whether the callee is required to pop its
3659 /// own arguments. Callee pop is necessary to support tail calls.
3660 bool X86::isCalleePop(CallingConv::ID CallingConv,
3661 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3662 switch (CallingConv) {
3665 case CallingConv::X86_StdCall:
3666 case CallingConv::X86_FastCall:
3667 case CallingConv::X86_ThisCall:
3669 case CallingConv::Fast:
3670 case CallingConv::GHC:
3671 case CallingConv::HiPE:
3678 /// \brief Return true if the condition is an unsigned comparison operation.
3679 static bool isX86CCUnsigned(unsigned X86CC) {
3681 default: llvm_unreachable("Invalid integer condition!");
3682 case X86::COND_E: return true;
3683 case X86::COND_G: return false;
3684 case X86::COND_GE: return false;
3685 case X86::COND_L: return false;
3686 case X86::COND_LE: return false;
3687 case X86::COND_NE: return true;
3688 case X86::COND_B: return true;
3689 case X86::COND_A: return true;
3690 case X86::COND_BE: return true;
3691 case X86::COND_AE: return true;
3693 llvm_unreachable("covered switch fell through?!");
3696 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3697 /// specific condition code, returning the condition code and the LHS/RHS of the
3698 /// comparison to make.
3699 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3700 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3702 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3703 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3704 // X > -1 -> X == 0, jump !sign.
3705 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3706 return X86::COND_NS;
3708 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3709 // X < 0 -> X == 0, jump on sign.
3712 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3714 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3715 return X86::COND_LE;
3719 switch (SetCCOpcode) {
3720 default: llvm_unreachable("Invalid integer condition!");
3721 case ISD::SETEQ: return X86::COND_E;
3722 case ISD::SETGT: return X86::COND_G;
3723 case ISD::SETGE: return X86::COND_GE;
3724 case ISD::SETLT: return X86::COND_L;
3725 case ISD::SETLE: return X86::COND_LE;
3726 case ISD::SETNE: return X86::COND_NE;
3727 case ISD::SETULT: return X86::COND_B;
3728 case ISD::SETUGT: return X86::COND_A;
3729 case ISD::SETULE: return X86::COND_BE;
3730 case ISD::SETUGE: return X86::COND_AE;
3734 // First determine if it is required or is profitable to flip the operands.
3736 // If LHS is a foldable load, but RHS is not, flip the condition.
3737 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3738 !ISD::isNON_EXTLoad(RHS.getNode())) {
3739 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3740 std::swap(LHS, RHS);
3743 switch (SetCCOpcode) {
3749 std::swap(LHS, RHS);
3753 // On a floating point condition, the flags are set as follows:
3755 // 0 | 0 | 0 | X > Y
3756 // 0 | 0 | 1 | X < Y
3757 // 1 | 0 | 0 | X == Y
3758 // 1 | 1 | 1 | unordered
3759 switch (SetCCOpcode) {
3760 default: llvm_unreachable("Condcode should be pre-legalized away");
3762 case ISD::SETEQ: return X86::COND_E;
3763 case ISD::SETOLT: // flipped
3765 case ISD::SETGT: return X86::COND_A;
3766 case ISD::SETOLE: // flipped
3768 case ISD::SETGE: return X86::COND_AE;
3769 case ISD::SETUGT: // flipped
3771 case ISD::SETLT: return X86::COND_B;
3772 case ISD::SETUGE: // flipped
3774 case ISD::SETLE: return X86::COND_BE;
3776 case ISD::SETNE: return X86::COND_NE;
3777 case ISD::SETUO: return X86::COND_P;
3778 case ISD::SETO: return X86::COND_NP;
3780 case ISD::SETUNE: return X86::COND_INVALID;
3784 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3785 /// code. Current x86 isa includes the following FP cmov instructions:
3786 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3787 static bool hasFPCMov(unsigned X86CC) {
3803 /// isFPImmLegal - Returns true if the target can instruction select the
3804 /// specified FP immediate natively. If false, the legalizer will
3805 /// materialize the FP immediate as a load from a constant pool.
3806 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3807 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3808 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3814 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3815 ISD::LoadExtType ExtTy,
3817 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3818 // relocation target a movq or addq instruction: don't let the load shrink.
3819 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3820 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3821 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3822 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3826 /// \brief Returns true if it is beneficial to convert a load of a constant
3827 /// to just the constant itself.
3828 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3830 assert(Ty->isIntegerTy());
3832 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3833 if (BitSize == 0 || BitSize > 64)
3838 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3839 unsigned Index) const {
3840 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3843 return (Index == 0 || Index == ResVT.getVectorNumElements());
3846 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3847 // Speculate cttz only if we can directly use TZCNT.
3848 return Subtarget->hasBMI();
3851 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3852 // Speculate ctlz only if we can directly use LZCNT.
3853 return Subtarget->hasLZCNT();
3856 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3857 /// the specified range (L, H].
3858 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3859 return (Val < 0) || (Val >= Low && Val < Hi);
3862 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3863 /// specified value.
3864 static bool isUndefOrEqual(int Val, int CmpVal) {
3865 return (Val < 0 || Val == CmpVal);
3868 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3869 /// from position Pos and ending in Pos+Size, falls within the specified
3870 /// sequential range (Low, Low+Size]. or is undef.
3871 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3872 unsigned Pos, unsigned Size, int Low) {
3873 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3874 if (!isUndefOrEqual(Mask[i], Low))
3879 /// isVEXTRACTIndex - Return true if the specified
3880 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3881 /// suitable for instruction that extract 128 or 256 bit vectors
3882 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3883 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3884 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3887 // The index should be aligned on a vecWidth-bit boundary.
3889 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3891 MVT VT = N->getSimpleValueType(0);
3892 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3893 bool Result = (Index * ElSize) % vecWidth == 0;
3898 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3899 /// operand specifies a subvector insert that is suitable for input to
3900 /// insertion of 128 or 256-bit subvectors
3901 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3902 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3903 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3905 // The index should be aligned on a vecWidth-bit boundary.
3907 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3909 MVT VT = N->getSimpleValueType(0);
3910 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3911 bool Result = (Index * ElSize) % vecWidth == 0;
3916 bool X86::isVINSERT128Index(SDNode *N) {
3917 return isVINSERTIndex(N, 128);
3920 bool X86::isVINSERT256Index(SDNode *N) {
3921 return isVINSERTIndex(N, 256);
3924 bool X86::isVEXTRACT128Index(SDNode *N) {
3925 return isVEXTRACTIndex(N, 128);
3928 bool X86::isVEXTRACT256Index(SDNode *N) {
3929 return isVEXTRACTIndex(N, 256);
3932 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3933 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3934 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3935 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3938 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3940 MVT VecVT = N->getOperand(0).getSimpleValueType();
3941 MVT ElVT = VecVT.getVectorElementType();
3943 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3944 return Index / NumElemsPerChunk;
3947 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3948 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3949 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3950 llvm_unreachable("Illegal insert subvector for VINSERT");
3953 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3955 MVT VecVT = N->getSimpleValueType(0);
3956 MVT ElVT = VecVT.getVectorElementType();
3958 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3959 return Index / NumElemsPerChunk;
3962 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
3963 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3964 /// and VINSERTI128 instructions.
3965 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
3966 return getExtractVEXTRACTImmediate(N, 128);
3969 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
3970 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
3971 /// and VINSERTI64x4 instructions.
3972 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
3973 return getExtractVEXTRACTImmediate(N, 256);
3976 /// getInsertVINSERT128Immediate - Return the appropriate immediate
3977 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3978 /// and VINSERTI128 instructions.
3979 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
3980 return getInsertVINSERTImmediate(N, 128);
3983 /// getInsertVINSERT256Immediate - Return the appropriate immediate
3984 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
3985 /// and VINSERTI64x4 instructions.
3986 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
3987 return getInsertVINSERTImmediate(N, 256);
3990 /// isZero - Returns true if Elt is a constant integer zero
3991 static bool isZero(SDValue V) {
3992 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
3993 return C && C->isNullValue();
3996 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3998 bool X86::isZeroNode(SDValue Elt) {
4001 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4002 return CFP->getValueAPF().isPosZero();
4006 /// getZeroVector - Returns a vector of specified type with all zero elements.
4008 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4009 SelectionDAG &DAG, SDLoc dl) {
4010 assert(VT.isVector() && "Expected a vector type");
4012 // Always build SSE zero vectors as <4 x i32> bitcasted
4013 // to their dest type. This ensures they get CSE'd.
4015 if (VT.is128BitVector()) { // SSE
4016 if (Subtarget->hasSSE2()) { // SSE2
4017 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4018 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4020 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4021 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4023 } else if (VT.is256BitVector()) { // AVX
4024 if (Subtarget->hasInt256()) { // AVX2
4025 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4026 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4027 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4029 // 256-bit logic and arithmetic instructions in AVX are all
4030 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4031 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4032 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4033 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4035 } else if (VT.is512BitVector()) { // AVX-512
4036 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4037 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4038 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4039 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4040 } else if (VT.getScalarType() == MVT::i1) {
4042 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4043 && "Unexpected vector type");
4044 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4045 && "Unexpected vector type");
4046 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4047 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4048 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4050 llvm_unreachable("Unexpected vector type");
4052 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4055 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4056 SelectionDAG &DAG, SDLoc dl,
4057 unsigned vectorWidth) {
4058 assert((vectorWidth == 128 || vectorWidth == 256) &&
4059 "Unsupported vector width");
4060 EVT VT = Vec.getValueType();
4061 EVT ElVT = VT.getVectorElementType();
4062 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4063 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4064 VT.getVectorNumElements()/Factor);
4066 // Extract from UNDEF is UNDEF.
4067 if (Vec.getOpcode() == ISD::UNDEF)
4068 return DAG.getUNDEF(ResultVT);
4070 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4071 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4073 // This is the index of the first element of the vectorWidth-bit chunk
4075 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4078 // If the input is a buildvector just emit a smaller one.
4079 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4080 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4081 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4084 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4085 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4088 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4089 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4090 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4091 /// instructions or a simple subregister reference. Idx is an index in the
4092 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4093 /// lowering EXTRACT_VECTOR_ELT operations easier.
4094 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4095 SelectionDAG &DAG, SDLoc dl) {
4096 assert((Vec.getValueType().is256BitVector() ||
4097 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4098 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4101 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4102 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4103 SelectionDAG &DAG, SDLoc dl) {
4104 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4105 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4108 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4109 unsigned IdxVal, SelectionDAG &DAG,
4110 SDLoc dl, unsigned vectorWidth) {
4111 assert((vectorWidth == 128 || vectorWidth == 256) &&
4112 "Unsupported vector width");
4113 // Inserting UNDEF is Result
4114 if (Vec.getOpcode() == ISD::UNDEF)
4116 EVT VT = Vec.getValueType();
4117 EVT ElVT = VT.getVectorElementType();
4118 EVT ResultVT = Result.getValueType();
4120 // Insert the relevant vectorWidth bits.
4121 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4123 // This is the index of the first element of the vectorWidth-bit chunk
4125 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4128 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4129 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4132 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4133 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4134 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4135 /// simple superregister reference. Idx is an index in the 128 bits
4136 /// we want. It need not be aligned to a 128-bit boundary. That makes
4137 /// lowering INSERT_VECTOR_ELT operations easier.
4138 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4139 SelectionDAG &DAG, SDLoc dl) {
4140 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4142 // For insertion into the zero index (low half) of a 256-bit vector, it is
4143 // more efficient to generate a blend with immediate instead of an insert*128.
4144 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4145 // extend the subvector to the size of the result vector. Make sure that
4146 // we are not recursing on that node by checking for undef here.
4147 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4148 Result.getOpcode() != ISD::UNDEF) {
4149 EVT ResultVT = Result.getValueType();
4150 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4151 SDValue Undef = DAG.getUNDEF(ResultVT);
4152 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4155 // The blend instruction, and therefore its mask, depend on the data type.
4156 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4157 if (ScalarType.isFloatingPoint()) {
4158 // Choose either vblendps (float) or vblendpd (double).
4159 unsigned ScalarSize = ScalarType.getSizeInBits();
4160 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4161 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4162 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4163 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4166 const X86Subtarget &Subtarget =
4167 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4169 // AVX2 is needed for 256-bit integer blend support.
4170 // Integers must be cast to 32-bit because there is only vpblendd;
4171 // vpblendw can't be used for this because it has a handicapped mask.
4173 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4174 // is still more efficient than using the wrong domain vinsertf128 that
4175 // will be created by InsertSubVector().
4176 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4178 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4179 Vec256 = DAG.getNode(ISD::BITCAST, dl, CastVT, Vec256);
4180 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4181 return DAG.getNode(ISD::BITCAST, dl, ResultVT, Vec256);
4184 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4187 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4188 SelectionDAG &DAG, SDLoc dl) {
4189 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4190 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4193 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4194 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4195 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4196 /// large BUILD_VECTORS.
4197 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4198 unsigned NumElems, SelectionDAG &DAG,
4200 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4201 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4204 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4205 unsigned NumElems, SelectionDAG &DAG,
4207 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4208 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4211 /// getOnesVector - Returns a vector of specified type with all bits set.
4212 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4213 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4214 /// Then bitcast to their original type, ensuring they get CSE'd.
4215 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4217 assert(VT.isVector() && "Expected a vector type");
4219 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4221 if (VT.is256BitVector()) {
4222 if (HasInt256) { // AVX2
4223 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4224 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4226 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4227 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4229 } else if (VT.is128BitVector()) {
4230 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4232 llvm_unreachable("Unexpected vector type");
4234 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4237 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4238 /// operation of specified width.
4239 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4241 unsigned NumElems = VT.getVectorNumElements();
4242 SmallVector<int, 8> Mask;
4243 Mask.push_back(NumElems);
4244 for (unsigned i = 1; i != NumElems; ++i)
4246 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4249 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4250 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4252 unsigned NumElems = VT.getVectorNumElements();
4253 SmallVector<int, 8> Mask;
4254 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4256 Mask.push_back(i + NumElems);
4258 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4261 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4262 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4264 unsigned NumElems = VT.getVectorNumElements();
4265 SmallVector<int, 8> Mask;
4266 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4267 Mask.push_back(i + Half);
4268 Mask.push_back(i + NumElems + Half);
4270 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4273 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4274 /// vector of zero or undef vector. This produces a shuffle where the low
4275 /// element of V2 is swizzled into the zero/undef vector, landing at element
4276 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4277 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4279 const X86Subtarget *Subtarget,
4280 SelectionDAG &DAG) {
4281 MVT VT = V2.getSimpleValueType();
4283 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4284 unsigned NumElems = VT.getVectorNumElements();
4285 SmallVector<int, 16> MaskVec;
4286 for (unsigned i = 0; i != NumElems; ++i)
4287 // If this is the insertion idx, put the low elt of V2 here.
4288 MaskVec.push_back(i == Idx ? NumElems : i);
4289 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4292 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4293 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4294 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4295 /// shuffles which use a single input multiple times, and in those cases it will
4296 /// adjust the mask to only have indices within that single input.
4297 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4298 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4299 unsigned NumElems = VT.getVectorNumElements();
4303 bool IsFakeUnary = false;
4304 switch(N->getOpcode()) {
4305 case X86ISD::BLENDI:
4306 ImmN = N->getOperand(N->getNumOperands()-1);
4307 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4310 ImmN = N->getOperand(N->getNumOperands()-1);
4311 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4312 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4314 case X86ISD::UNPCKH:
4315 DecodeUNPCKHMask(VT, Mask);
4316 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4318 case X86ISD::UNPCKL:
4319 DecodeUNPCKLMask(VT, Mask);
4320 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4322 case X86ISD::MOVHLPS:
4323 DecodeMOVHLPSMask(NumElems, Mask);
4324 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4326 case X86ISD::MOVLHPS:
4327 DecodeMOVLHPSMask(NumElems, Mask);
4328 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4330 case X86ISD::PALIGNR:
4331 ImmN = N->getOperand(N->getNumOperands()-1);
4332 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4334 case X86ISD::PSHUFD:
4335 case X86ISD::VPERMILPI:
4336 ImmN = N->getOperand(N->getNumOperands()-1);
4337 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4340 case X86ISD::PSHUFHW:
4341 ImmN = N->getOperand(N->getNumOperands()-1);
4342 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4345 case X86ISD::PSHUFLW:
4346 ImmN = N->getOperand(N->getNumOperands()-1);
4347 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4350 case X86ISD::PSHUFB: {
4352 SDValue MaskNode = N->getOperand(1);
4353 while (MaskNode->getOpcode() == ISD::BITCAST)
4354 MaskNode = MaskNode->getOperand(0);
4356 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4357 // If we have a build-vector, then things are easy.
4358 EVT VT = MaskNode.getValueType();
4359 assert(VT.isVector() &&
4360 "Can't produce a non-vector with a build_vector!");
4361 if (!VT.isInteger())
4364 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4366 SmallVector<uint64_t, 32> RawMask;
4367 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4368 SDValue Op = MaskNode->getOperand(i);
4369 if (Op->getOpcode() == ISD::UNDEF) {
4370 RawMask.push_back((uint64_t)SM_SentinelUndef);
4373 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4376 APInt MaskElement = CN->getAPIntValue();
4378 // We now have to decode the element which could be any integer size and
4379 // extract each byte of it.
4380 for (int j = 0; j < NumBytesPerElement; ++j) {
4381 // Note that this is x86 and so always little endian: the low byte is
4382 // the first byte of the mask.
4383 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4384 MaskElement = MaskElement.lshr(8);
4387 DecodePSHUFBMask(RawMask, Mask);
4391 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4395 SDValue Ptr = MaskLoad->getBasePtr();
4396 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4397 Ptr->getOpcode() == X86ISD::WrapperRIP)
4398 Ptr = Ptr->getOperand(0);
4400 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4401 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4404 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4405 DecodePSHUFBMask(C, Mask);
4413 case X86ISD::VPERMI:
4414 ImmN = N->getOperand(N->getNumOperands()-1);
4415 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4420 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4422 case X86ISD::VPERM2X128:
4423 ImmN = N->getOperand(N->getNumOperands()-1);
4424 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4425 if (Mask.empty()) return false;
4427 case X86ISD::MOVSLDUP:
4428 DecodeMOVSLDUPMask(VT, Mask);
4431 case X86ISD::MOVSHDUP:
4432 DecodeMOVSHDUPMask(VT, Mask);
4435 case X86ISD::MOVDDUP:
4436 DecodeMOVDDUPMask(VT, Mask);
4439 case X86ISD::MOVLHPD:
4440 case X86ISD::MOVLPD:
4441 case X86ISD::MOVLPS:
4442 // Not yet implemented
4444 default: llvm_unreachable("unknown target shuffle node");
4447 // If we have a fake unary shuffle, the shuffle mask is spread across two
4448 // inputs that are actually the same node. Re-map the mask to always point
4449 // into the first input.
4452 if (M >= (int)Mask.size())
4458 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4459 /// element of the result of the vector shuffle.
4460 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4463 return SDValue(); // Limit search depth.
4465 SDValue V = SDValue(N, 0);
4466 EVT VT = V.getValueType();
4467 unsigned Opcode = V.getOpcode();
4469 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4470 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4471 int Elt = SV->getMaskElt(Index);
4474 return DAG.getUNDEF(VT.getVectorElementType());
4476 unsigned NumElems = VT.getVectorNumElements();
4477 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4478 : SV->getOperand(1);
4479 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4482 // Recurse into target specific vector shuffles to find scalars.
4483 if (isTargetShuffle(Opcode)) {
4484 MVT ShufVT = V.getSimpleValueType();
4485 unsigned NumElems = ShufVT.getVectorNumElements();
4486 SmallVector<int, 16> ShuffleMask;
4489 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4492 int Elt = ShuffleMask[Index];
4494 return DAG.getUNDEF(ShufVT.getVectorElementType());
4496 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4498 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4502 // Actual nodes that may contain scalar elements
4503 if (Opcode == ISD::BITCAST) {
4504 V = V.getOperand(0);
4505 EVT SrcVT = V.getValueType();
4506 unsigned NumElems = VT.getVectorNumElements();
4508 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4512 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4513 return (Index == 0) ? V.getOperand(0)
4514 : DAG.getUNDEF(VT.getVectorElementType());
4516 if (V.getOpcode() == ISD::BUILD_VECTOR)
4517 return V.getOperand(Index);
4522 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4524 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4525 unsigned NumNonZero, unsigned NumZero,
4527 const X86Subtarget* Subtarget,
4528 const TargetLowering &TLI) {
4536 // SSE4.1 - use PINSRB to insert each byte directly.
4537 if (Subtarget->hasSSE41()) {
4538 for (unsigned i = 0; i < 16; ++i) {
4539 bool isNonZero = (NonZeros & (1 << i)) != 0;
4543 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4545 V = DAG.getUNDEF(MVT::v16i8);
4548 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4549 MVT::v16i8, V, Op.getOperand(i),
4550 DAG.getIntPtrConstant(i, dl));
4557 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4558 for (unsigned i = 0; i < 16; ++i) {
4559 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4560 if (ThisIsNonZero && First) {
4562 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4564 V = DAG.getUNDEF(MVT::v8i16);
4569 SDValue ThisElt, LastElt;
4570 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4571 if (LastIsNonZero) {
4572 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4573 MVT::i16, Op.getOperand(i-1));
4575 if (ThisIsNonZero) {
4576 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4577 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4578 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4580 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4584 if (ThisElt.getNode())
4585 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4586 DAG.getIntPtrConstant(i/2, dl));
4590 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4593 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4595 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4596 unsigned NumNonZero, unsigned NumZero,
4598 const X86Subtarget* Subtarget,
4599 const TargetLowering &TLI) {
4606 for (unsigned i = 0; i < 8; ++i) {
4607 bool isNonZero = (NonZeros & (1 << i)) != 0;
4611 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4613 V = DAG.getUNDEF(MVT::v8i16);
4616 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4617 MVT::v8i16, V, Op.getOperand(i),
4618 DAG.getIntPtrConstant(i, dl));
4625 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4626 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4627 const X86Subtarget *Subtarget,
4628 const TargetLowering &TLI) {
4629 // Find all zeroable elements.
4630 std::bitset<4> Zeroable;
4631 for (int i=0; i < 4; ++i) {
4632 SDValue Elt = Op->getOperand(i);
4633 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4635 assert(Zeroable.size() - Zeroable.count() > 1 &&
4636 "We expect at least two non-zero elements!");
4638 // We only know how to deal with build_vector nodes where elements are either
4639 // zeroable or extract_vector_elt with constant index.
4640 SDValue FirstNonZero;
4641 unsigned FirstNonZeroIdx;
4642 for (unsigned i=0; i < 4; ++i) {
4645 SDValue Elt = Op->getOperand(i);
4646 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4647 !isa<ConstantSDNode>(Elt.getOperand(1)))
4649 // Make sure that this node is extracting from a 128-bit vector.
4650 MVT VT = Elt.getOperand(0).getSimpleValueType();
4651 if (!VT.is128BitVector())
4653 if (!FirstNonZero.getNode()) {
4655 FirstNonZeroIdx = i;
4659 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4660 SDValue V1 = FirstNonZero.getOperand(0);
4661 MVT VT = V1.getSimpleValueType();
4663 // See if this build_vector can be lowered as a blend with zero.
4665 unsigned EltMaskIdx, EltIdx;
4667 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4668 if (Zeroable[EltIdx]) {
4669 // The zero vector will be on the right hand side.
4670 Mask[EltIdx] = EltIdx+4;
4674 Elt = Op->getOperand(EltIdx);
4675 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4676 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4677 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4679 Mask[EltIdx] = EltIdx;
4683 // Let the shuffle legalizer deal with blend operations.
4684 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4685 if (V1.getSimpleValueType() != VT)
4686 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4687 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4690 // See if we can lower this build_vector to a INSERTPS.
4691 if (!Subtarget->hasSSE41())
4694 SDValue V2 = Elt.getOperand(0);
4695 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4698 bool CanFold = true;
4699 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4703 SDValue Current = Op->getOperand(i);
4704 SDValue SrcVector = Current->getOperand(0);
4707 CanFold = SrcVector == V1 &&
4708 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4714 assert(V1.getNode() && "Expected at least two non-zero elements!");
4715 if (V1.getSimpleValueType() != MVT::v4f32)
4716 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4717 if (V2.getSimpleValueType() != MVT::v4f32)
4718 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4720 // Ok, we can emit an INSERTPS instruction.
4721 unsigned ZMask = Zeroable.to_ulong();
4723 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4724 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4726 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4727 DAG.getIntPtrConstant(InsertPSMask, DL));
4728 return DAG.getNode(ISD::BITCAST, DL, VT, Result);
4731 /// Return a vector logical shift node.
4732 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4733 unsigned NumBits, SelectionDAG &DAG,
4734 const TargetLowering &TLI, SDLoc dl) {
4735 assert(VT.is128BitVector() && "Unknown type for VShift");
4736 MVT ShVT = MVT::v2i64;
4737 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4738 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4739 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4740 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4741 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4742 return DAG.getNode(ISD::BITCAST, dl, VT,
4743 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4747 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4749 // Check if the scalar load can be widened into a vector load. And if
4750 // the address is "base + cst" see if the cst can be "absorbed" into
4751 // the shuffle mask.
4752 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4753 SDValue Ptr = LD->getBasePtr();
4754 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4756 EVT PVT = LD->getValueType(0);
4757 if (PVT != MVT::i32 && PVT != MVT::f32)
4762 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4763 FI = FINode->getIndex();
4765 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4766 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4767 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4768 Offset = Ptr.getConstantOperandVal(1);
4769 Ptr = Ptr.getOperand(0);
4774 // FIXME: 256-bit vector instructions don't require a strict alignment,
4775 // improve this code to support it better.
4776 unsigned RequiredAlign = VT.getSizeInBits()/8;
4777 SDValue Chain = LD->getChain();
4778 // Make sure the stack object alignment is at least 16 or 32.
4779 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4780 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4781 if (MFI->isFixedObjectIndex(FI)) {
4782 // Can't change the alignment. FIXME: It's possible to compute
4783 // the exact stack offset and reference FI + adjust offset instead.
4784 // If someone *really* cares about this. That's the way to implement it.
4787 MFI->setObjectAlignment(FI, RequiredAlign);
4791 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4792 // Ptr + (Offset & ~15).
4795 if ((Offset % RequiredAlign) & 3)
4797 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4800 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4801 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4804 int EltNo = (Offset - StartOffset) >> 2;
4805 unsigned NumElems = VT.getVectorNumElements();
4807 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4808 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4809 LD->getPointerInfo().getWithOffset(StartOffset),
4810 false, false, false, 0);
4812 SmallVector<int, 8> Mask(NumElems, EltNo);
4814 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4820 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4821 /// elements can be replaced by a single large load which has the same value as
4822 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4824 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4826 /// FIXME: we'd also like to handle the case where the last elements are zero
4827 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4828 /// There's even a handy isZeroNode for that purpose.
4829 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4830 SDLoc &DL, SelectionDAG &DAG,
4831 bool isAfterLegalize) {
4832 unsigned NumElems = Elts.size();
4834 LoadSDNode *LDBase = nullptr;
4835 unsigned LastLoadedElt = -1U;
4837 // For each element in the initializer, see if we've found a load or an undef.
4838 // If we don't find an initial load element, or later load elements are
4839 // non-consecutive, bail out.
4840 for (unsigned i = 0; i < NumElems; ++i) {
4841 SDValue Elt = Elts[i];
4842 // Look through a bitcast.
4843 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4844 Elt = Elt.getOperand(0);
4845 if (!Elt.getNode() ||
4846 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4849 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4851 LDBase = cast<LoadSDNode>(Elt.getNode());
4855 if (Elt.getOpcode() == ISD::UNDEF)
4858 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4859 EVT LdVT = Elt.getValueType();
4860 // Each loaded element must be the correct fractional portion of the
4861 // requested vector load.
4862 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4864 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4869 // If we have found an entire vector of loads and undefs, then return a large
4870 // load of the entire vector width starting at the base pointer. If we found
4871 // consecutive loads for the low half, generate a vzext_load node.
4872 if (LastLoadedElt == NumElems - 1) {
4873 assert(LDBase && "Did not find base load for merging consecutive loads");
4874 EVT EltVT = LDBase->getValueType(0);
4875 // Ensure that the input vector size for the merged loads matches the
4876 // cumulative size of the input elements.
4877 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4880 if (isAfterLegalize &&
4881 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4884 SDValue NewLd = SDValue();
4886 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4887 LDBase->getPointerInfo(), LDBase->isVolatile(),
4888 LDBase->isNonTemporal(), LDBase->isInvariant(),
4889 LDBase->getAlignment());
4891 if (LDBase->hasAnyUseOfValue(1)) {
4892 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4894 SDValue(NewLd.getNode(), 1));
4895 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4896 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4897 SDValue(NewLd.getNode(), 1));
4903 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4904 //of a v4i32 / v4f32. It's probably worth generalizing.
4905 EVT EltVT = VT.getVectorElementType();
4906 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4907 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4908 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4909 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4911 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4912 LDBase->getPointerInfo(),
4913 LDBase->getAlignment(),
4914 false/*isVolatile*/, true/*ReadMem*/,
4917 // Make sure the newly-created LOAD is in the same position as LDBase in
4918 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4919 // update uses of LDBase's output chain to use the TokenFactor.
4920 if (LDBase->hasAnyUseOfValue(1)) {
4921 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4922 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4923 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4924 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4925 SDValue(ResNode.getNode(), 1));
4928 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4933 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4934 /// to generate a splat value for the following cases:
4935 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4936 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4937 /// a scalar load, or a constant.
4938 /// The VBROADCAST node is returned when a pattern is found,
4939 /// or SDValue() otherwise.
4940 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4941 SelectionDAG &DAG) {
4942 // VBROADCAST requires AVX.
4943 // TODO: Splats could be generated for non-AVX CPUs using SSE
4944 // instructions, but there's less potential gain for only 128-bit vectors.
4945 if (!Subtarget->hasAVX())
4948 MVT VT = Op.getSimpleValueType();
4951 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4952 "Unsupported vector type for broadcast.");
4957 switch (Op.getOpcode()) {
4959 // Unknown pattern found.
4962 case ISD::BUILD_VECTOR: {
4963 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4964 BitVector UndefElements;
4965 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4967 // We need a splat of a single value to use broadcast, and it doesn't
4968 // make any sense if the value is only in one element of the vector.
4969 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4973 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4974 Ld.getOpcode() == ISD::ConstantFP);
4976 // Make sure that all of the users of a non-constant load are from the
4977 // BUILD_VECTOR node.
4978 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4983 case ISD::VECTOR_SHUFFLE: {
4984 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4986 // Shuffles must have a splat mask where the first element is
4988 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4991 SDValue Sc = Op.getOperand(0);
4992 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4993 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4995 if (!Subtarget->hasInt256())
4998 // Use the register form of the broadcast instruction available on AVX2.
4999 if (VT.getSizeInBits() >= 256)
5000 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5001 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5004 Ld = Sc.getOperand(0);
5005 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5006 Ld.getOpcode() == ISD::ConstantFP);
5008 // The scalar_to_vector node and the suspected
5009 // load node must have exactly one user.
5010 // Constants may have multiple users.
5012 // AVX-512 has register version of the broadcast
5013 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5014 Ld.getValueType().getSizeInBits() >= 32;
5015 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5022 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5023 bool IsGE256 = (VT.getSizeInBits() >= 256);
5025 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5026 // instruction to save 8 or more bytes of constant pool data.
5027 // TODO: If multiple splats are generated to load the same constant,
5028 // it may be detrimental to overall size. There needs to be a way to detect
5029 // that condition to know if this is truly a size win.
5030 const Function *F = DAG.getMachineFunction().getFunction();
5031 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
5033 // Handle broadcasting a single constant scalar from the constant pool
5035 // On Sandybridge (no AVX2), it is still better to load a constant vector
5036 // from the constant pool and not to broadcast it from a scalar.
5037 // But override that restriction when optimizing for size.
5038 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5039 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5040 EVT CVT = Ld.getValueType();
5041 assert(!CVT.isVector() && "Must not broadcast a vector type");
5043 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5044 // For size optimization, also splat v2f64 and v2i64, and for size opt
5045 // with AVX2, also splat i8 and i16.
5046 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5047 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5048 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5049 const Constant *C = nullptr;
5050 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5051 C = CI->getConstantIntValue();
5052 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5053 C = CF->getConstantFPValue();
5055 assert(C && "Invalid constant type");
5057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5058 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
5059 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5060 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5061 MachinePointerInfo::getConstantPool(),
5062 false, false, false, Alignment);
5064 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5068 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5070 // Handle AVX2 in-register broadcasts.
5071 if (!IsLoad && Subtarget->hasInt256() &&
5072 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5073 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5075 // The scalar source must be a normal load.
5079 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5080 (Subtarget->hasVLX() && ScalarSize == 64))
5081 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5083 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5084 // double since there is no vbroadcastsd xmm
5085 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5086 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5087 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5090 // Unsupported broadcast.
5094 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5095 /// underlying vector and index.
5097 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5099 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5101 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5102 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5105 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5107 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5109 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5110 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5113 // In this case the vector is the extract_subvector expression and the index
5114 // is 2, as specified by the shuffle.
5115 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5116 SDValue ShuffleVec = SVOp->getOperand(0);
5117 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5118 assert(ShuffleVecVT.getVectorElementType() ==
5119 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5121 int ShuffleIdx = SVOp->getMaskElt(Idx);
5122 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5123 ExtractedFromVec = ShuffleVec;
5129 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5130 MVT VT = Op.getSimpleValueType();
5132 // Skip if insert_vec_elt is not supported.
5133 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5134 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5138 unsigned NumElems = Op.getNumOperands();
5142 SmallVector<unsigned, 4> InsertIndices;
5143 SmallVector<int, 8> Mask(NumElems, -1);
5145 for (unsigned i = 0; i != NumElems; ++i) {
5146 unsigned Opc = Op.getOperand(i).getOpcode();
5148 if (Opc == ISD::UNDEF)
5151 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5152 // Quit if more than 1 elements need inserting.
5153 if (InsertIndices.size() > 1)
5156 InsertIndices.push_back(i);
5160 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5161 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5162 // Quit if non-constant index.
5163 if (!isa<ConstantSDNode>(ExtIdx))
5165 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5167 // Quit if extracted from vector of different type.
5168 if (ExtractedFromVec.getValueType() != VT)
5171 if (!VecIn1.getNode())
5172 VecIn1 = ExtractedFromVec;
5173 else if (VecIn1 != ExtractedFromVec) {
5174 if (!VecIn2.getNode())
5175 VecIn2 = ExtractedFromVec;
5176 else if (VecIn2 != ExtractedFromVec)
5177 // Quit if more than 2 vectors to shuffle
5181 if (ExtractedFromVec == VecIn1)
5183 else if (ExtractedFromVec == VecIn2)
5184 Mask[i] = Idx + NumElems;
5187 if (!VecIn1.getNode())
5190 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5191 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5192 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5193 unsigned Idx = InsertIndices[i];
5194 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5195 DAG.getIntPtrConstant(Idx, DL));
5201 static SDValue ConvertI1VectorToInterger(SDValue Op, SelectionDAG &DAG) {
5202 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5203 Op.getScalarValueSizeInBits() == 1 &&
5204 "Can not convert non-constant vector");
5205 uint64_t Immediate = 0;
5206 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5207 SDValue In = Op.getOperand(idx);
5208 if (In.getOpcode() != ISD::UNDEF)
5209 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5213 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5214 return DAG.getConstant(Immediate, dl, VT);
5216 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5218 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5220 MVT VT = Op.getSimpleValueType();
5221 assert((VT.getVectorElementType() == MVT::i1) &&
5222 "Unexpected type in LowerBUILD_VECTORvXi1!");
5225 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5226 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5227 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5228 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5231 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5232 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5233 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5234 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5237 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5238 SDValue Imm = ConvertI1VectorToInterger(Op, DAG);
5239 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5240 return DAG.getNode(ISD::BITCAST, dl, VT, Imm);
5241 SDValue ExtVec = DAG.getNode(ISD::BITCAST, dl, MVT::v8i1, Imm);
5242 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5243 DAG.getIntPtrConstant(0, dl));
5246 // Vector has one or more non-const elements
5247 uint64_t Immediate = 0;
5248 SmallVector<unsigned, 16> NonConstIdx;
5249 bool IsSplat = true;
5250 bool HasConstElts = false;
5252 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5253 SDValue In = Op.getOperand(idx);
5254 if (In.getOpcode() == ISD::UNDEF)
5256 if (!isa<ConstantSDNode>(In))
5257 NonConstIdx.push_back(idx);
5259 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5260 HasConstElts = true;
5264 else if (In != Op.getOperand(SplatIdx))
5268 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5270 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5271 DAG.getConstant(1, dl, VT),
5272 DAG.getConstant(0, dl, VT));
5274 // insert elements one by one
5278 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5279 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5281 else if (HasConstElts)
5282 Imm = DAG.getConstant(0, dl, VT);
5284 Imm = DAG.getUNDEF(VT);
5285 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5286 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, Imm);
5288 SDValue ExtVec = DAG.getNode(ISD::BITCAST, dl, MVT::v8i1, Imm);
5289 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5290 DAG.getIntPtrConstant(0, dl));
5293 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5294 unsigned InsertIdx = NonConstIdx[i];
5295 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5296 Op.getOperand(InsertIdx),
5297 DAG.getIntPtrConstant(InsertIdx, dl));
5302 /// \brief Return true if \p N implements a horizontal binop and return the
5303 /// operands for the horizontal binop into V0 and V1.
5305 /// This is a helper function of LowerToHorizontalOp().
5306 /// This function checks that the build_vector \p N in input implements a
5307 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5308 /// operation to match.
5309 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5310 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5311 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5314 /// This function only analyzes elements of \p N whose indices are
5315 /// in range [BaseIdx, LastIdx).
5316 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5318 unsigned BaseIdx, unsigned LastIdx,
5319 SDValue &V0, SDValue &V1) {
5320 EVT VT = N->getValueType(0);
5322 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5323 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5324 "Invalid Vector in input!");
5326 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5327 bool CanFold = true;
5328 unsigned ExpectedVExtractIdx = BaseIdx;
5329 unsigned NumElts = LastIdx - BaseIdx;
5330 V0 = DAG.getUNDEF(VT);
5331 V1 = DAG.getUNDEF(VT);
5333 // Check if N implements a horizontal binop.
5334 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5335 SDValue Op = N->getOperand(i + BaseIdx);
5338 if (Op->getOpcode() == ISD::UNDEF) {
5339 // Update the expected vector extract index.
5340 if (i * 2 == NumElts)
5341 ExpectedVExtractIdx = BaseIdx;
5342 ExpectedVExtractIdx += 2;
5346 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5351 SDValue Op0 = Op.getOperand(0);
5352 SDValue Op1 = Op.getOperand(1);
5354 // Try to match the following pattern:
5355 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5356 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5357 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5358 Op0.getOperand(0) == Op1.getOperand(0) &&
5359 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5360 isa<ConstantSDNode>(Op1.getOperand(1)));
5364 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5365 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5367 if (i * 2 < NumElts) {
5368 if (V0.getOpcode() == ISD::UNDEF) {
5369 V0 = Op0.getOperand(0);
5370 if (V0.getValueType() != VT)
5374 if (V1.getOpcode() == ISD::UNDEF) {
5375 V1 = Op0.getOperand(0);
5376 if (V1.getValueType() != VT)
5379 if (i * 2 == NumElts)
5380 ExpectedVExtractIdx = BaseIdx;
5383 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5384 if (I0 == ExpectedVExtractIdx)
5385 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5386 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5387 // Try to match the following dag sequence:
5388 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5389 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5393 ExpectedVExtractIdx += 2;
5399 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5400 /// a concat_vector.
5402 /// This is a helper function of LowerToHorizontalOp().
5403 /// This function expects two 256-bit vectors called V0 and V1.
5404 /// At first, each vector is split into two separate 128-bit vectors.
5405 /// Then, the resulting 128-bit vectors are used to implement two
5406 /// horizontal binary operations.
5408 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5410 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5411 /// the two new horizontal binop.
5412 /// When Mode is set, the first horizontal binop dag node would take as input
5413 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5414 /// horizontal binop dag node would take as input the lower 128-bit of V1
5415 /// and the upper 128-bit of V1.
5417 /// HADD V0_LO, V0_HI
5418 /// HADD V1_LO, V1_HI
5420 /// Otherwise, the first horizontal binop dag node takes as input the lower
5421 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5422 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5424 /// HADD V0_LO, V1_LO
5425 /// HADD V0_HI, V1_HI
5427 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5428 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5429 /// the upper 128-bits of the result.
5430 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5431 SDLoc DL, SelectionDAG &DAG,
5432 unsigned X86Opcode, bool Mode,
5433 bool isUndefLO, bool isUndefHI) {
5434 EVT VT = V0.getValueType();
5435 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5436 "Invalid nodes in input!");
5438 unsigned NumElts = VT.getVectorNumElements();
5439 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5440 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5441 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5442 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5443 EVT NewVT = V0_LO.getValueType();
5445 SDValue LO = DAG.getUNDEF(NewVT);
5446 SDValue HI = DAG.getUNDEF(NewVT);
5449 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5450 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5451 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5452 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5453 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5455 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5456 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5457 V1_LO->getOpcode() != ISD::UNDEF))
5458 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5460 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5461 V1_HI->getOpcode() != ISD::UNDEF))
5462 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5465 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5468 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5470 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5471 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5472 EVT VT = BV->getValueType(0);
5473 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5474 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5478 unsigned NumElts = VT.getVectorNumElements();
5479 SDValue InVec0 = DAG.getUNDEF(VT);
5480 SDValue InVec1 = DAG.getUNDEF(VT);
5482 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5483 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5485 // Odd-numbered elements in the input build vector are obtained from
5486 // adding two integer/float elements.
5487 // Even-numbered elements in the input build vector are obtained from
5488 // subtracting two integer/float elements.
5489 unsigned ExpectedOpcode = ISD::FSUB;
5490 unsigned NextExpectedOpcode = ISD::FADD;
5491 bool AddFound = false;
5492 bool SubFound = false;
5494 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5495 SDValue Op = BV->getOperand(i);
5497 // Skip 'undef' values.
5498 unsigned Opcode = Op.getOpcode();
5499 if (Opcode == ISD::UNDEF) {
5500 std::swap(ExpectedOpcode, NextExpectedOpcode);
5504 // Early exit if we found an unexpected opcode.
5505 if (Opcode != ExpectedOpcode)
5508 SDValue Op0 = Op.getOperand(0);
5509 SDValue Op1 = Op.getOperand(1);
5511 // Try to match the following pattern:
5512 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5513 // Early exit if we cannot match that sequence.
5514 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5515 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5516 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5517 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5518 Op0.getOperand(1) != Op1.getOperand(1))
5521 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5525 // We found a valid add/sub node. Update the information accordingly.
5531 // Update InVec0 and InVec1.
5532 if (InVec0.getOpcode() == ISD::UNDEF) {
5533 InVec0 = Op0.getOperand(0);
5534 if (InVec0.getValueType() != VT)
5537 if (InVec1.getOpcode() == ISD::UNDEF) {
5538 InVec1 = Op1.getOperand(0);
5539 if (InVec1.getValueType() != VT)
5543 // Make sure that operands in input to each add/sub node always
5544 // come from a same pair of vectors.
5545 if (InVec0 != Op0.getOperand(0)) {
5546 if (ExpectedOpcode == ISD::FSUB)
5549 // FADD is commutable. Try to commute the operands
5550 // and then test again.
5551 std::swap(Op0, Op1);
5552 if (InVec0 != Op0.getOperand(0))
5556 if (InVec1 != Op1.getOperand(0))
5559 // Update the pair of expected opcodes.
5560 std::swap(ExpectedOpcode, NextExpectedOpcode);
5563 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5564 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5565 InVec1.getOpcode() != ISD::UNDEF)
5566 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5571 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5572 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5573 const X86Subtarget *Subtarget,
5574 SelectionDAG &DAG) {
5575 EVT VT = BV->getValueType(0);
5576 unsigned NumElts = VT.getVectorNumElements();
5577 unsigned NumUndefsLO = 0;
5578 unsigned NumUndefsHI = 0;
5579 unsigned Half = NumElts/2;
5581 // Count the number of UNDEF operands in the build_vector in input.
5582 for (unsigned i = 0, e = Half; i != e; ++i)
5583 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5586 for (unsigned i = Half, e = NumElts; i != e; ++i)
5587 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5590 // Early exit if this is either a build_vector of all UNDEFs or all the
5591 // operands but one are UNDEF.
5592 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5596 SDValue InVec0, InVec1;
5597 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5598 // Try to match an SSE3 float HADD/HSUB.
5599 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5600 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5602 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5603 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5604 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5605 // Try to match an SSSE3 integer HADD/HSUB.
5606 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5607 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5609 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5610 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5613 if (!Subtarget->hasAVX())
5616 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5617 // Try to match an AVX horizontal add/sub of packed single/double
5618 // precision floating point values from 256-bit vectors.
5619 SDValue InVec2, InVec3;
5620 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5621 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5622 ((InVec0.getOpcode() == ISD::UNDEF ||
5623 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5624 ((InVec1.getOpcode() == ISD::UNDEF ||
5625 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5626 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5628 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5629 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5630 ((InVec0.getOpcode() == ISD::UNDEF ||
5631 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5632 ((InVec1.getOpcode() == ISD::UNDEF ||
5633 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5634 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5635 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5636 // Try to match an AVX2 horizontal add/sub of signed integers.
5637 SDValue InVec2, InVec3;
5639 bool CanFold = true;
5641 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5642 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5643 ((InVec0.getOpcode() == ISD::UNDEF ||
5644 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5645 ((InVec1.getOpcode() == ISD::UNDEF ||
5646 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5647 X86Opcode = X86ISD::HADD;
5648 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5649 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5650 ((InVec0.getOpcode() == ISD::UNDEF ||
5651 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5652 ((InVec1.getOpcode() == ISD::UNDEF ||
5653 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5654 X86Opcode = X86ISD::HSUB;
5659 // Fold this build_vector into a single horizontal add/sub.
5660 // Do this only if the target has AVX2.
5661 if (Subtarget->hasAVX2())
5662 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5664 // Do not try to expand this build_vector into a pair of horizontal
5665 // add/sub if we can emit a pair of scalar add/sub.
5666 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5669 // Convert this build_vector into a pair of horizontal binop followed by
5671 bool isUndefLO = NumUndefsLO == Half;
5672 bool isUndefHI = NumUndefsHI == Half;
5673 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5674 isUndefLO, isUndefHI);
5678 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5679 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5681 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5682 X86Opcode = X86ISD::HADD;
5683 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5684 X86Opcode = X86ISD::HSUB;
5685 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5686 X86Opcode = X86ISD::FHADD;
5687 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5688 X86Opcode = X86ISD::FHSUB;
5692 // Don't try to expand this build_vector into a pair of horizontal add/sub
5693 // if we can simply emit a pair of scalar add/sub.
5694 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5697 // Convert this build_vector into two horizontal add/sub followed by
5699 bool isUndefLO = NumUndefsLO == Half;
5700 bool isUndefHI = NumUndefsHI == Half;
5701 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5702 isUndefLO, isUndefHI);
5709 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5712 MVT VT = Op.getSimpleValueType();
5713 MVT ExtVT = VT.getVectorElementType();
5714 unsigned NumElems = Op.getNumOperands();
5716 // Generate vectors for predicate vectors.
5717 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5718 return LowerBUILD_VECTORvXi1(Op, DAG);
5720 // Vectors containing all zeros can be matched by pxor and xorps later
5721 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5722 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5723 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5724 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5727 return getZeroVector(VT, Subtarget, DAG, dl);
5730 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5731 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5732 // vpcmpeqd on 256-bit vectors.
5733 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5734 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5737 if (!VT.is512BitVector())
5738 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5741 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5742 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5744 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5745 return HorizontalOp;
5746 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5749 unsigned EVTBits = ExtVT.getSizeInBits();
5751 unsigned NumZero = 0;
5752 unsigned NumNonZero = 0;
5753 unsigned NonZeros = 0;
5754 bool IsAllConstants = true;
5755 SmallSet<SDValue, 8> Values;
5756 for (unsigned i = 0; i < NumElems; ++i) {
5757 SDValue Elt = Op.getOperand(i);
5758 if (Elt.getOpcode() == ISD::UNDEF)
5761 if (Elt.getOpcode() != ISD::Constant &&
5762 Elt.getOpcode() != ISD::ConstantFP)
5763 IsAllConstants = false;
5764 if (X86::isZeroNode(Elt))
5767 NonZeros |= (1 << i);
5772 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5773 if (NumNonZero == 0)
5774 return DAG.getUNDEF(VT);
5776 // Special case for single non-zero, non-undef, element.
5777 if (NumNonZero == 1) {
5778 unsigned Idx = countTrailingZeros(NonZeros);
5779 SDValue Item = Op.getOperand(Idx);
5781 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5782 // the value are obviously zero, truncate the value to i32 and do the
5783 // insertion that way. Only do this if the value is non-constant or if the
5784 // value is a constant being inserted into element 0. It is cheaper to do
5785 // a constant pool load than it is to do a movd + shuffle.
5786 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5787 (!IsAllConstants || Idx == 0)) {
5788 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5790 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5791 EVT VecVT = MVT::v4i32;
5793 // Truncate the value (which may itself be a constant) to i32, and
5794 // convert it to a vector with movd (S2V+shuffle to zero extend).
5795 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5796 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5798 ISD::BITCAST, dl, VT,
5799 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5803 // If we have a constant or non-constant insertion into the low element of
5804 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5805 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5806 // depending on what the source datatype is.
5809 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5811 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5812 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5813 if (VT.is512BitVector()) {
5814 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5815 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5816 Item, DAG.getIntPtrConstant(0, dl));
5818 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5819 "Expected an SSE value type!");
5820 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5821 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5822 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5825 // We can't directly insert an i8 or i16 into a vector, so zero extend
5827 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5828 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5829 if (VT.is256BitVector()) {
5830 if (Subtarget->hasAVX()) {
5831 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5832 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5834 // Without AVX, we need to extend to a 128-bit vector and then
5835 // insert into the 256-bit vector.
5836 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5837 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5838 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5841 assert(VT.is128BitVector() && "Expected an SSE value type!");
5842 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5843 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5845 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5849 // Is it a vector logical left shift?
5850 if (NumElems == 2 && Idx == 1 &&
5851 X86::isZeroNode(Op.getOperand(0)) &&
5852 !X86::isZeroNode(Op.getOperand(1))) {
5853 unsigned NumBits = VT.getSizeInBits();
5854 return getVShift(true, VT,
5855 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5856 VT, Op.getOperand(1)),
5857 NumBits/2, DAG, *this, dl);
5860 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5863 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5864 // is a non-constant being inserted into an element other than the low one,
5865 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5866 // movd/movss) to move this into the low element, then shuffle it into
5868 if (EVTBits == 32) {
5869 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5870 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5874 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5875 if (Values.size() == 1) {
5876 if (EVTBits == 32) {
5877 // Instead of a shuffle like this:
5878 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5879 // Check if it's possible to issue this instead.
5880 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5881 unsigned Idx = countTrailingZeros(NonZeros);
5882 SDValue Item = Op.getOperand(Idx);
5883 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5884 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5889 // A vector full of immediates; various special cases are already
5890 // handled, so this is best done with a single constant-pool load.
5894 // For AVX-length vectors, see if we can use a vector load to get all of the
5895 // elements, otherwise build the individual 128-bit pieces and use
5896 // shuffles to put them in place.
5897 if (VT.is256BitVector() || VT.is512BitVector()) {
5898 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5900 // Check for a build vector of consecutive loads.
5901 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5904 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5906 // Build both the lower and upper subvector.
5907 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5908 makeArrayRef(&V[0], NumElems/2));
5909 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5910 makeArrayRef(&V[NumElems / 2], NumElems/2));
5912 // Recreate the wider vector with the lower and upper part.
5913 if (VT.is256BitVector())
5914 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5915 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5918 // Let legalizer expand 2-wide build_vectors.
5919 if (EVTBits == 64) {
5920 if (NumNonZero == 1) {
5921 // One half is zero or undef.
5922 unsigned Idx = countTrailingZeros(NonZeros);
5923 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5924 Op.getOperand(Idx));
5925 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5930 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5931 if (EVTBits == 8 && NumElems == 16)
5932 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5936 if (EVTBits == 16 && NumElems == 8)
5937 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5941 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5942 if (EVTBits == 32 && NumElems == 4)
5943 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
5946 // If element VT is == 32 bits, turn it into a number of shuffles.
5947 SmallVector<SDValue, 8> V(NumElems);
5948 if (NumElems == 4 && NumZero > 0) {
5949 for (unsigned i = 0; i < 4; ++i) {
5950 bool isZero = !(NonZeros & (1 << i));
5952 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5954 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5957 for (unsigned i = 0; i < 2; ++i) {
5958 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5961 V[i] = V[i*2]; // Must be a zero vector.
5964 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5967 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5970 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5975 bool Reverse1 = (NonZeros & 0x3) == 2;
5976 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5980 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5981 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5983 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5986 if (Values.size() > 1 && VT.is128BitVector()) {
5987 // Check for a build vector of consecutive loads.
5988 for (unsigned i = 0; i < NumElems; ++i)
5989 V[i] = Op.getOperand(i);
5991 // Check for elements which are consecutive loads.
5992 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5995 // Check for a build vector from mostly shuffle plus few inserting.
5996 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
5999 // For SSE 4.1, use insertps to put the high elements into the low element.
6000 if (Subtarget->hasSSE41()) {
6002 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6003 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6005 Result = DAG.getUNDEF(VT);
6007 for (unsigned i = 1; i < NumElems; ++i) {
6008 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6009 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6010 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6015 // Otherwise, expand into a number of unpckl*, start by extending each of
6016 // our (non-undef) elements to the full vector width with the element in the
6017 // bottom slot of the vector (which generates no code for SSE).
6018 for (unsigned i = 0; i < NumElems; ++i) {
6019 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6020 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6022 V[i] = DAG.getUNDEF(VT);
6025 // Next, we iteratively mix elements, e.g. for v4f32:
6026 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6027 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6028 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6029 unsigned EltStride = NumElems >> 1;
6030 while (EltStride != 0) {
6031 for (unsigned i = 0; i < EltStride; ++i) {
6032 // If V[i+EltStride] is undef and this is the first round of mixing,
6033 // then it is safe to just drop this shuffle: V[i] is already in the
6034 // right place, the one element (since it's the first round) being
6035 // inserted as undef can be dropped. This isn't safe for successive
6036 // rounds because they will permute elements within both vectors.
6037 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6038 EltStride == NumElems/2)
6041 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6050 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
6051 // to create 256-bit vectors from two other 128-bit ones.
6052 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6054 MVT ResVT = Op.getSimpleValueType();
6056 assert((ResVT.is256BitVector() ||
6057 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6059 SDValue V1 = Op.getOperand(0);
6060 SDValue V2 = Op.getOperand(1);
6061 unsigned NumElems = ResVT.getVectorNumElements();
6062 if (ResVT.is256BitVector())
6063 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6065 if (Op.getNumOperands() == 4) {
6066 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6067 ResVT.getVectorNumElements()/2);
6068 SDValue V3 = Op.getOperand(2);
6069 SDValue V4 = Op.getOperand(3);
6070 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6071 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6073 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6076 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6077 const X86Subtarget *Subtarget,
6078 SelectionDAG & DAG) {
6080 MVT ResVT = Op.getSimpleValueType();
6081 unsigned NumOfOperands = Op.getNumOperands();
6083 assert(isPowerOf2_32(NumOfOperands) &&
6084 "Unexpected number of operands in CONCAT_VECTORS");
6086 if (NumOfOperands > 2) {
6087 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6088 ResVT.getVectorNumElements()/2);
6089 SmallVector<SDValue, 2> Ops;
6090 for (unsigned i = 0; i < NumOfOperands/2; i++)
6091 Ops.push_back(Op.getOperand(i));
6092 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6094 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6095 Ops.push_back(Op.getOperand(i));
6096 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6097 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6100 SDValue V1 = Op.getOperand(0);
6101 SDValue V2 = Op.getOperand(1);
6102 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6103 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6105 if (IsZeroV1 && IsZeroV2)
6106 return getZeroVector(ResVT, Subtarget, DAG, dl);
6108 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6109 SDValue Undef = DAG.getUNDEF(ResVT);
6110 unsigned NumElems = ResVT.getVectorNumElements();
6111 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6113 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6114 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6118 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6119 // Zero the upper bits of V1
6120 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6121 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6124 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6127 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6128 const X86Subtarget *Subtarget,
6129 SelectionDAG &DAG) {
6130 MVT VT = Op.getSimpleValueType();
6131 if (VT.getVectorElementType() == MVT::i1)
6132 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6134 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6135 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6136 Op.getNumOperands() == 4)));
6138 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6139 // from two other 128-bit ones.
6141 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6142 return LowerAVXCONCAT_VECTORS(Op, DAG);
6146 //===----------------------------------------------------------------------===//
6147 // Vector shuffle lowering
6149 // This is an experimental code path for lowering vector shuffles on x86. It is
6150 // designed to handle arbitrary vector shuffles and blends, gracefully
6151 // degrading performance as necessary. It works hard to recognize idiomatic
6152 // shuffles and lower them to optimal instruction patterns without leaving
6153 // a framework that allows reasonably efficient handling of all vector shuffle
6155 //===----------------------------------------------------------------------===//
6157 /// \brief Tiny helper function to identify a no-op mask.
6159 /// This is a somewhat boring predicate function. It checks whether the mask
6160 /// array input, which is assumed to be a single-input shuffle mask of the kind
6161 /// used by the X86 shuffle instructions (not a fully general
6162 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6163 /// in-place shuffle are 'no-op's.
6164 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6165 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6166 if (Mask[i] != -1 && Mask[i] != i)
6171 /// \brief Helper function to classify a mask as a single-input mask.
6173 /// This isn't a generic single-input test because in the vector shuffle
6174 /// lowering we canonicalize single inputs to be the first input operand. This
6175 /// means we can more quickly test for a single input by only checking whether
6176 /// an input from the second operand exists. We also assume that the size of
6177 /// mask corresponds to the size of the input vectors which isn't true in the
6178 /// fully general case.
6179 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6181 if (M >= (int)Mask.size())
6186 /// \brief Test whether there are elements crossing 128-bit lanes in this
6189 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6190 /// and we routinely test for these.
6191 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6192 int LaneSize = 128 / VT.getScalarSizeInBits();
6193 int Size = Mask.size();
6194 for (int i = 0; i < Size; ++i)
6195 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6200 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6202 /// This checks a shuffle mask to see if it is performing the same
6203 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6204 /// that it is also not lane-crossing. It may however involve a blend from the
6205 /// same lane of a second vector.
6207 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6208 /// non-trivial to compute in the face of undef lanes. The representation is
6209 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6210 /// entries from both V1 and V2 inputs to the wider mask.
6212 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6213 SmallVectorImpl<int> &RepeatedMask) {
6214 int LaneSize = 128 / VT.getScalarSizeInBits();
6215 RepeatedMask.resize(LaneSize, -1);
6216 int Size = Mask.size();
6217 for (int i = 0; i < Size; ++i) {
6220 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6221 // This entry crosses lanes, so there is no way to model this shuffle.
6224 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6225 if (RepeatedMask[i % LaneSize] == -1)
6226 // This is the first non-undef entry in this slot of a 128-bit lane.
6227 RepeatedMask[i % LaneSize] =
6228 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6229 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6230 // Found a mismatch with the repeated mask.
6236 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6239 /// This is a fast way to test a shuffle mask against a fixed pattern:
6241 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6243 /// It returns true if the mask is exactly as wide as the argument list, and
6244 /// each element of the mask is either -1 (signifying undef) or the value given
6245 /// in the argument.
6246 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6247 ArrayRef<int> ExpectedMask) {
6248 if (Mask.size() != ExpectedMask.size())
6251 int Size = Mask.size();
6253 // If the values are build vectors, we can look through them to find
6254 // equivalent inputs that make the shuffles equivalent.
6255 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6256 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6258 for (int i = 0; i < Size; ++i)
6259 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6260 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6261 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6262 if (!MaskBV || !ExpectedBV ||
6263 MaskBV->getOperand(Mask[i] % Size) !=
6264 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6271 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6273 /// This helper function produces an 8-bit shuffle immediate corresponding to
6274 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6275 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6278 /// NB: We rely heavily on "undef" masks preserving the input lane.
6279 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6280 SelectionDAG &DAG) {
6281 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6282 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6283 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6284 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6285 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6288 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6289 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6290 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6291 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6292 return DAG.getConstant(Imm, DL, MVT::i8);
6295 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6297 /// This is used as a fallback approach when first class blend instructions are
6298 /// unavailable. Currently it is only suitable for integer vectors, but could
6299 /// be generalized for floating point vectors if desirable.
6300 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6301 SDValue V2, ArrayRef<int> Mask,
6302 SelectionDAG &DAG) {
6303 assert(VT.isInteger() && "Only supports integer vector types!");
6304 MVT EltVT = VT.getScalarType();
6305 int NumEltBits = EltVT.getSizeInBits();
6306 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6307 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6309 SmallVector<SDValue, 16> MaskOps;
6310 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6311 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6312 return SDValue(); // Shuffled input!
6313 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6316 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6317 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6318 // We have to cast V2 around.
6319 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6320 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6321 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6322 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6323 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6324 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6327 /// \brief Try to emit a blend instruction for a shuffle.
6329 /// This doesn't do any checks for the availability of instructions for blending
6330 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6331 /// be matched in the backend with the type given. What it does check for is
6332 /// that the shuffle mask is in fact a blend.
6333 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6334 SDValue V2, ArrayRef<int> Mask,
6335 const X86Subtarget *Subtarget,
6336 SelectionDAG &DAG) {
6337 unsigned BlendMask = 0;
6338 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6339 if (Mask[i] >= Size) {
6340 if (Mask[i] != i + Size)
6341 return SDValue(); // Shuffled V2 input!
6342 BlendMask |= 1u << i;
6345 if (Mask[i] >= 0 && Mask[i] != i)
6346 return SDValue(); // Shuffled V1 input!
6348 switch (VT.SimpleTy) {
6353 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6354 DAG.getConstant(BlendMask, DL, MVT::i8));
6358 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6362 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6363 // that instruction.
6364 if (Subtarget->hasAVX2()) {
6365 // Scale the blend by the number of 32-bit dwords per element.
6366 int Scale = VT.getScalarSizeInBits() / 32;
6368 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6369 if (Mask[i] >= Size)
6370 for (int j = 0; j < Scale; ++j)
6371 BlendMask |= 1u << (i * Scale + j);
6373 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6374 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6375 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6376 return DAG.getNode(ISD::BITCAST, DL, VT,
6377 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6378 DAG.getConstant(BlendMask, DL, MVT::i8)));
6382 // For integer shuffles we need to expand the mask and cast the inputs to
6383 // v8i16s prior to blending.
6384 int Scale = 8 / VT.getVectorNumElements();
6386 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6387 if (Mask[i] >= Size)
6388 for (int j = 0; j < Scale; ++j)
6389 BlendMask |= 1u << (i * Scale + j);
6391 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6392 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6393 return DAG.getNode(ISD::BITCAST, DL, VT,
6394 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6395 DAG.getConstant(BlendMask, DL, MVT::i8)));
6399 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6400 SmallVector<int, 8> RepeatedMask;
6401 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6402 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6403 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6405 for (int i = 0; i < 8; ++i)
6406 if (RepeatedMask[i] >= 16)
6407 BlendMask |= 1u << i;
6408 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6409 DAG.getConstant(BlendMask, DL, MVT::i8));
6415 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6416 "256-bit byte-blends require AVX2 support!");
6418 // Scale the blend by the number of bytes per element.
6419 int Scale = VT.getScalarSizeInBits() / 8;
6421 // This form of blend is always done on bytes. Compute the byte vector
6423 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6425 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6426 // mix of LLVM's code generator and the x86 backend. We tell the code
6427 // generator that boolean values in the elements of an x86 vector register
6428 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6429 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6430 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6431 // of the element (the remaining are ignored) and 0 in that high bit would
6432 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6433 // the LLVM model for boolean values in vector elements gets the relevant
6434 // bit set, it is set backwards and over constrained relative to x86's
6436 SmallVector<SDValue, 32> VSELECTMask;
6437 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6438 for (int j = 0; j < Scale; ++j)
6439 VSELECTMask.push_back(
6440 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6441 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6444 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6445 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6447 ISD::BITCAST, DL, VT,
6448 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6449 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6454 llvm_unreachable("Not a supported integer vector type!");
6458 /// \brief Try to lower as a blend of elements from two inputs followed by
6459 /// a single-input permutation.
6461 /// This matches the pattern where we can blend elements from two inputs and
6462 /// then reduce the shuffle to a single-input permutation.
6463 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6466 SelectionDAG &DAG) {
6467 // We build up the blend mask while checking whether a blend is a viable way
6468 // to reduce the shuffle.
6469 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6470 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6472 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6476 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6478 if (BlendMask[Mask[i] % Size] == -1)
6479 BlendMask[Mask[i] % Size] = Mask[i];
6480 else if (BlendMask[Mask[i] % Size] != Mask[i])
6481 return SDValue(); // Can't blend in the needed input!
6483 PermuteMask[i] = Mask[i] % Size;
6486 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6487 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6490 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6491 /// blends and permutes.
6493 /// This matches the extremely common pattern for handling combined
6494 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6495 /// operations. It will try to pick the best arrangement of shuffles and
6497 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6501 SelectionDAG &DAG) {
6502 // Shuffle the input elements into the desired positions in V1 and V2 and
6503 // blend them together.
6504 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6505 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6506 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6507 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6508 if (Mask[i] >= 0 && Mask[i] < Size) {
6509 V1Mask[i] = Mask[i];
6511 } else if (Mask[i] >= Size) {
6512 V2Mask[i] = Mask[i] - Size;
6513 BlendMask[i] = i + Size;
6516 // Try to lower with the simpler initial blend strategy unless one of the
6517 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6518 // shuffle may be able to fold with a load or other benefit. However, when
6519 // we'll have to do 2x as many shuffles in order to achieve this, blending
6520 // first is a better strategy.
6521 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6522 if (SDValue BlendPerm =
6523 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6526 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6527 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6528 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6531 /// \brief Try to lower a vector shuffle as a byte rotation.
6533 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6534 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6535 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6536 /// try to generically lower a vector shuffle through such an pattern. It
6537 /// does not check for the profitability of lowering either as PALIGNR or
6538 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6539 /// This matches shuffle vectors that look like:
6541 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6543 /// Essentially it concatenates V1 and V2, shifts right by some number of
6544 /// elements, and takes the low elements as the result. Note that while this is
6545 /// specified as a *right shift* because x86 is little-endian, it is a *left
6546 /// rotate* of the vector lanes.
6547 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6550 const X86Subtarget *Subtarget,
6551 SelectionDAG &DAG) {
6552 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6554 int NumElts = Mask.size();
6555 int NumLanes = VT.getSizeInBits() / 128;
6556 int NumLaneElts = NumElts / NumLanes;
6558 // We need to detect various ways of spelling a rotation:
6559 // [11, 12, 13, 14, 15, 0, 1, 2]
6560 // [-1, 12, 13, 14, -1, -1, 1, -1]
6561 // [-1, -1, -1, -1, -1, -1, 1, 2]
6562 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6563 // [-1, 4, 5, 6, -1, -1, 9, -1]
6564 // [-1, 4, 5, 6, -1, -1, -1, -1]
6567 for (int l = 0; l < NumElts; l += NumLaneElts) {
6568 for (int i = 0; i < NumLaneElts; ++i) {
6569 if (Mask[l + i] == -1)
6571 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6573 // Get the mod-Size index and lane correct it.
6574 int LaneIdx = (Mask[l + i] % NumElts) - l;
6575 // Make sure it was in this lane.
6576 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6579 // Determine where a rotated vector would have started.
6580 int StartIdx = i - LaneIdx;
6582 // The identity rotation isn't interesting, stop.
6585 // If we found the tail of a vector the rotation must be the missing
6586 // front. If we found the head of a vector, it must be how much of the
6588 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6591 Rotation = CandidateRotation;
6592 else if (Rotation != CandidateRotation)
6593 // The rotations don't match, so we can't match this mask.
6596 // Compute which value this mask is pointing at.
6597 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6599 // Compute which of the two target values this index should be assigned
6600 // to. This reflects whether the high elements are remaining or the low
6601 // elements are remaining.
6602 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6604 // Either set up this value if we've not encountered it before, or check
6605 // that it remains consistent.
6608 else if (TargetV != MaskV)
6609 // This may be a rotation, but it pulls from the inputs in some
6610 // unsupported interleaving.
6615 // Check that we successfully analyzed the mask, and normalize the results.
6616 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6617 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6623 // The actual rotate instruction rotates bytes, so we need to scale the
6624 // rotation based on how many bytes are in the vector lane.
6625 int Scale = 16 / NumLaneElts;
6627 // SSSE3 targets can use the palignr instruction.
6628 if (Subtarget->hasSSSE3()) {
6629 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6630 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6631 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6632 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6634 return DAG.getNode(ISD::BITCAST, DL, VT,
6635 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6636 DAG.getConstant(Rotation * Scale, DL,
6640 assert(VT.getSizeInBits() == 128 &&
6641 "Rotate-based lowering only supports 128-bit lowering!");
6642 assert(Mask.size() <= 16 &&
6643 "Can shuffle at most 16 bytes in a 128-bit vector!");
6645 // Default SSE2 implementation
6646 int LoByteShift = 16 - Rotation * Scale;
6647 int HiByteShift = Rotation * Scale;
6649 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6650 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6651 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6653 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6654 DAG.getConstant(LoByteShift, DL, MVT::i8));
6655 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6656 DAG.getConstant(HiByteShift, DL, MVT::i8));
6657 return DAG.getNode(ISD::BITCAST, DL, VT,
6658 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6661 /// \brief Compute whether each element of a shuffle is zeroable.
6663 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6664 /// Either it is an undef element in the shuffle mask, the element of the input
6665 /// referenced is undef, or the element of the input referenced is known to be
6666 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6667 /// as many lanes with this technique as possible to simplify the remaining
6669 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6670 SDValue V1, SDValue V2) {
6671 SmallBitVector Zeroable(Mask.size(), false);
6673 while (V1.getOpcode() == ISD::BITCAST)
6674 V1 = V1->getOperand(0);
6675 while (V2.getOpcode() == ISD::BITCAST)
6676 V2 = V2->getOperand(0);
6678 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6679 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6681 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6683 // Handle the easy cases.
6684 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6689 // If this is an index into a build_vector node (which has the same number
6690 // of elements), dig out the input value and use it.
6691 SDValue V = M < Size ? V1 : V2;
6692 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6695 SDValue Input = V.getOperand(M % Size);
6696 // The UNDEF opcode check really should be dead code here, but not quite
6697 // worth asserting on (it isn't invalid, just unexpected).
6698 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6705 /// \brief Try to emit a bitmask instruction for a shuffle.
6707 /// This handles cases where we can model a blend exactly as a bitmask due to
6708 /// one of the inputs being zeroable.
6709 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6710 SDValue V2, ArrayRef<int> Mask,
6711 SelectionDAG &DAG) {
6712 MVT EltVT = VT.getScalarType();
6713 int NumEltBits = EltVT.getSizeInBits();
6714 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6715 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6716 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6718 if (EltVT.isFloatingPoint()) {
6719 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6720 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6722 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6723 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6725 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6728 if (Mask[i] % Size != i)
6729 return SDValue(); // Not a blend.
6731 V = Mask[i] < Size ? V1 : V2;
6732 else if (V != (Mask[i] < Size ? V1 : V2))
6733 return SDValue(); // Can only let one input through the mask.
6735 VMaskOps[i] = AllOnes;
6738 return SDValue(); // No non-zeroable elements!
6740 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6741 V = DAG.getNode(VT.isFloatingPoint()
6742 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6747 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6749 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6750 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6751 /// matches elements from one of the input vectors shuffled to the left or
6752 /// right with zeroable elements 'shifted in'. It handles both the strictly
6753 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6756 /// PSHL : (little-endian) left bit shift.
6757 /// [ zz, 0, zz, 2 ]
6758 /// [ -1, 4, zz, -1 ]
6759 /// PSRL : (little-endian) right bit shift.
6761 /// [ -1, -1, 7, zz]
6762 /// PSLLDQ : (little-endian) left byte shift
6763 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6764 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6765 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6766 /// PSRLDQ : (little-endian) right byte shift
6767 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6768 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6769 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6770 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6771 SDValue V2, ArrayRef<int> Mask,
6772 SelectionDAG &DAG) {
6773 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6775 int Size = Mask.size();
6776 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6778 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6779 for (int i = 0; i < Size; i += Scale)
6780 for (int j = 0; j < Shift; ++j)
6781 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6787 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6788 for (int i = 0; i != Size; i += Scale) {
6789 unsigned Pos = Left ? i + Shift : i;
6790 unsigned Low = Left ? i : i + Shift;
6791 unsigned Len = Scale - Shift;
6792 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6793 Low + (V == V1 ? 0 : Size)))
6797 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6798 bool ByteShift = ShiftEltBits > 64;
6799 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6800 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6801 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6803 // Normalize the scale for byte shifts to still produce an i64 element
6805 Scale = ByteShift ? Scale / 2 : Scale;
6807 // We need to round trip through the appropriate type for the shift.
6808 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6809 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6810 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6811 "Illegal integer vector type");
6812 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6814 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6815 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6816 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6819 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6820 // keep doubling the size of the integer elements up to that. We can
6821 // then shift the elements of the integer vector by whole multiples of
6822 // their width within the elements of the larger integer vector. Test each
6823 // multiple to see if we can find a match with the moved element indices
6824 // and that the shifted in elements are all zeroable.
6825 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6826 for (int Shift = 1; Shift != Scale; ++Shift)
6827 for (bool Left : {true, false})
6828 if (CheckZeros(Shift, Scale, Left))
6829 for (SDValue V : {V1, V2})
6830 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6837 /// \brief Lower a vector shuffle as a zero or any extension.
6839 /// Given a specific number of elements, element bit width, and extension
6840 /// stride, produce either a zero or any extension based on the available
6841 /// features of the subtarget.
6842 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6843 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6844 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6845 assert(Scale > 1 && "Need a scale to extend.");
6846 int NumElements = VT.getVectorNumElements();
6847 int EltBits = VT.getScalarSizeInBits();
6848 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6849 "Only 8, 16, and 32 bit elements can be extended.");
6850 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6852 // Found a valid zext mask! Try various lowering strategies based on the
6853 // input type and available ISA extensions.
6854 if (Subtarget->hasSSE41()) {
6855 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6856 NumElements / Scale);
6857 return DAG.getNode(ISD::BITCAST, DL, VT,
6858 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6861 // For any extends we can cheat for larger element sizes and use shuffle
6862 // instructions that can fold with a load and/or copy.
6863 if (AnyExt && EltBits == 32) {
6864 int PSHUFDMask[4] = {0, -1, 1, -1};
6866 ISD::BITCAST, DL, VT,
6867 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6868 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6869 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
6871 if (AnyExt && EltBits == 16 && Scale > 2) {
6872 int PSHUFDMask[4] = {0, -1, 0, -1};
6873 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6874 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6875 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
6876 int PSHUFHWMask[4] = {1, -1, -1, -1};
6878 ISD::BITCAST, DL, VT,
6879 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6880 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6881 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
6884 // If this would require more than 2 unpack instructions to expand, use
6885 // pshufb when available. We can only use more than 2 unpack instructions
6886 // when zero extending i8 elements which also makes it easier to use pshufb.
6887 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6888 assert(NumElements == 16 && "Unexpected byte vector width!");
6889 SDValue PSHUFBMask[16];
6890 for (int i = 0; i < 16; ++i)
6892 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
6893 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6894 return DAG.getNode(ISD::BITCAST, DL, VT,
6895 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6896 DAG.getNode(ISD::BUILD_VECTOR, DL,
6897 MVT::v16i8, PSHUFBMask)));
6900 // Otherwise emit a sequence of unpacks.
6902 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6903 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6904 : getZeroVector(InputVT, Subtarget, DAG, DL);
6905 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6906 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6910 } while (Scale > 1);
6911 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6914 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6916 /// This routine will try to do everything in its power to cleverly lower
6917 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6918 /// check for the profitability of this lowering, it tries to aggressively
6919 /// match this pattern. It will use all of the micro-architectural details it
6920 /// can to emit an efficient lowering. It handles both blends with all-zero
6921 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6922 /// masking out later).
6924 /// The reason we have dedicated lowering for zext-style shuffles is that they
6925 /// are both incredibly common and often quite performance sensitive.
6926 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6927 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6928 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6929 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6931 int Bits = VT.getSizeInBits();
6932 int NumElements = VT.getVectorNumElements();
6933 assert(VT.getScalarSizeInBits() <= 32 &&
6934 "Exceeds 32-bit integer zero extension limit");
6935 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6937 // Define a helper function to check a particular ext-scale and lower to it if
6939 auto Lower = [&](int Scale) -> SDValue {
6942 for (int i = 0; i < NumElements; ++i) {
6944 continue; // Valid anywhere but doesn't tell us anything.
6945 if (i % Scale != 0) {
6946 // Each of the extended elements need to be zeroable.
6950 // We no longer are in the anyext case.
6955 // Each of the base elements needs to be consecutive indices into the
6956 // same input vector.
6957 SDValue V = Mask[i] < NumElements ? V1 : V2;
6960 else if (InputV != V)
6961 return SDValue(); // Flip-flopping inputs.
6963 if (Mask[i] % NumElements != i / Scale)
6964 return SDValue(); // Non-consecutive strided elements.
6967 // If we fail to find an input, we have a zero-shuffle which should always
6968 // have already been handled.
6969 // FIXME: Maybe handle this here in case during blending we end up with one?
6973 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6974 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6977 // The widest scale possible for extending is to a 64-bit integer.
6978 assert(Bits % 64 == 0 &&
6979 "The number of bits in a vector must be divisible by 64 on x86!");
6980 int NumExtElements = Bits / 64;
6982 // Each iteration, try extending the elements half as much, but into twice as
6984 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6985 assert(NumElements % NumExtElements == 0 &&
6986 "The input vector size must be divisible by the extended size.");
6987 if (SDValue V = Lower(NumElements / NumExtElements))
6991 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6995 // Returns one of the source operands if the shuffle can be reduced to a
6996 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6997 auto CanZExtLowHalf = [&]() {
6998 for (int i = NumElements / 2; i != NumElements; ++i)
7001 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7003 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7008 if (SDValue V = CanZExtLowHalf()) {
7009 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
7010 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7011 return DAG.getNode(ISD::BITCAST, DL, VT, V);
7014 // No viable ext lowering found.
7018 /// \brief Try to get a scalar value for a specific element of a vector.
7020 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7021 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7022 SelectionDAG &DAG) {
7023 MVT VT = V.getSimpleValueType();
7024 MVT EltVT = VT.getVectorElementType();
7025 while (V.getOpcode() == ISD::BITCAST)
7026 V = V.getOperand(0);
7027 // If the bitcasts shift the element size, we can't extract an equivalent
7029 MVT NewVT = V.getSimpleValueType();
7030 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7033 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7034 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7035 // Ensure the scalar operand is the same size as the destination.
7036 // FIXME: Add support for scalar truncation where possible.
7037 SDValue S = V.getOperand(Idx);
7038 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7039 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7045 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7047 /// This is particularly important because the set of instructions varies
7048 /// significantly based on whether the operand is a load or not.
7049 static bool isShuffleFoldableLoad(SDValue V) {
7050 while (V.getOpcode() == ISD::BITCAST)
7051 V = V.getOperand(0);
7053 return ISD::isNON_EXTLoad(V.getNode());
7056 /// \brief Try to lower insertion of a single element into a zero vector.
7058 /// This is a common pattern that we have especially efficient patterns to lower
7059 /// across all subtarget feature sets.
7060 static SDValue lowerVectorShuffleAsElementInsertion(
7061 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7062 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7063 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7065 MVT EltVT = VT.getVectorElementType();
7067 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7068 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7070 bool IsV1Zeroable = true;
7071 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7072 if (i != V2Index && !Zeroable[i]) {
7073 IsV1Zeroable = false;
7077 // Check for a single input from a SCALAR_TO_VECTOR node.
7078 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7079 // all the smarts here sunk into that routine. However, the current
7080 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7081 // vector shuffle lowering is dead.
7082 if (SDValue V2S = getScalarValueForVectorElement(
7083 V2, Mask[V2Index] - Mask.size(), DAG)) {
7084 // We need to zext the scalar if it is smaller than an i32.
7085 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
7086 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7087 // Using zext to expand a narrow element won't work for non-zero
7092 // Zero-extend directly to i32.
7094 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7096 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7097 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7098 EltVT == MVT::i16) {
7099 // Either not inserting from the low element of the input or the input
7100 // element size is too small to use VZEXT_MOVL to clear the high bits.
7104 if (!IsV1Zeroable) {
7105 // If V1 can't be treated as a zero vector we have fewer options to lower
7106 // this. We can't support integer vectors or non-zero targets cheaply, and
7107 // the V1 elements can't be permuted in any way.
7108 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7109 if (!VT.isFloatingPoint() || V2Index != 0)
7111 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7112 V1Mask[V2Index] = -1;
7113 if (!isNoopShuffleMask(V1Mask))
7115 // This is essentially a special case blend operation, but if we have
7116 // general purpose blend operations, they are always faster. Bail and let
7117 // the rest of the lowering handle these as blends.
7118 if (Subtarget->hasSSE41())
7121 // Otherwise, use MOVSD or MOVSS.
7122 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7123 "Only two types of floating point element types to handle!");
7124 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7128 // This lowering only works for the low element with floating point vectors.
7129 if (VT.isFloatingPoint() && V2Index != 0)
7132 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7134 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7137 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7138 // the desired position. Otherwise it is more efficient to do a vector
7139 // shift left. We know that we can do a vector shift left because all
7140 // the inputs are zero.
7141 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7142 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7143 V2Shuffle[V2Index] = 0;
7144 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7146 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
7148 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7150 V2Index * EltVT.getSizeInBits()/8, DL,
7151 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
7152 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
7158 /// \brief Try to lower broadcast of a single element.
7160 /// For convenience, this code also bundles all of the subtarget feature set
7161 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7162 /// a convenient way to factor it out.
7163 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7165 const X86Subtarget *Subtarget,
7166 SelectionDAG &DAG) {
7167 if (!Subtarget->hasAVX())
7169 if (VT.isInteger() && !Subtarget->hasAVX2())
7172 // Check that the mask is a broadcast.
7173 int BroadcastIdx = -1;
7175 if (M >= 0 && BroadcastIdx == -1)
7177 else if (M >= 0 && M != BroadcastIdx)
7180 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7181 "a sorted mask where the broadcast "
7184 // Go up the chain of (vector) values to find a scalar load that we can
7185 // combine with the broadcast.
7187 switch (V.getOpcode()) {
7188 case ISD::CONCAT_VECTORS: {
7189 int OperandSize = Mask.size() / V.getNumOperands();
7190 V = V.getOperand(BroadcastIdx / OperandSize);
7191 BroadcastIdx %= OperandSize;
7195 case ISD::INSERT_SUBVECTOR: {
7196 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7197 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7201 int BeginIdx = (int)ConstantIdx->getZExtValue();
7203 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7204 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7205 BroadcastIdx -= BeginIdx;
7216 // Check if this is a broadcast of a scalar. We special case lowering
7217 // for scalars so that we can more effectively fold with loads.
7218 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7219 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7220 V = V.getOperand(BroadcastIdx);
7222 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7223 // Only AVX2 has register broadcasts.
7224 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7226 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7227 // We can't broadcast from a vector register without AVX2, and we can only
7228 // broadcast from the zero-element of a vector register.
7232 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7235 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7236 // INSERTPS when the V1 elements are already in the correct locations
7237 // because otherwise we can just always use two SHUFPS instructions which
7238 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7239 // perform INSERTPS if a single V1 element is out of place and all V2
7240 // elements are zeroable.
7241 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7243 SelectionDAG &DAG) {
7244 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7245 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7246 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7247 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7249 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7252 int V1DstIndex = -1;
7253 int V2DstIndex = -1;
7254 bool V1UsedInPlace = false;
7256 for (int i = 0; i < 4; ++i) {
7257 // Synthesize a zero mask from the zeroable elements (includes undefs).
7263 // Flag if we use any V1 inputs in place.
7265 V1UsedInPlace = true;
7269 // We can only insert a single non-zeroable element.
7270 if (V1DstIndex != -1 || V2DstIndex != -1)
7274 // V1 input out of place for insertion.
7277 // V2 input for insertion.
7282 // Don't bother if we have no (non-zeroable) element for insertion.
7283 if (V1DstIndex == -1 && V2DstIndex == -1)
7286 // Determine element insertion src/dst indices. The src index is from the
7287 // start of the inserted vector, not the start of the concatenated vector.
7288 unsigned V2SrcIndex = 0;
7289 if (V1DstIndex != -1) {
7290 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7291 // and don't use the original V2 at all.
7292 V2SrcIndex = Mask[V1DstIndex];
7293 V2DstIndex = V1DstIndex;
7296 V2SrcIndex = Mask[V2DstIndex] - 4;
7299 // If no V1 inputs are used in place, then the result is created only from
7300 // the zero mask and the V2 insertion - so remove V1 dependency.
7302 V1 = DAG.getUNDEF(MVT::v4f32);
7304 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7305 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7307 // Insert the V2 element into the desired position.
7309 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7310 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7313 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7314 /// UNPCK instruction.
7316 /// This specifically targets cases where we end up with alternating between
7317 /// the two inputs, and so can permute them into something that feeds a single
7318 /// UNPCK instruction. Note that this routine only targets integer vectors
7319 /// because for floating point vectors we have a generalized SHUFPS lowering
7320 /// strategy that handles everything that doesn't *exactly* match an unpack,
7321 /// making this clever lowering unnecessary.
7322 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7323 SDValue V2, ArrayRef<int> Mask,
7324 SelectionDAG &DAG) {
7325 assert(!VT.isFloatingPoint() &&
7326 "This routine only supports integer vectors.");
7327 assert(!isSingleInputShuffleMask(Mask) &&
7328 "This routine should only be used when blending two inputs.");
7329 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7331 int Size = Mask.size();
7333 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7334 return M >= 0 && M % Size < Size / 2;
7336 int NumHiInputs = std::count_if(
7337 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7339 bool UnpackLo = NumLoInputs >= NumHiInputs;
7341 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7342 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7343 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7345 for (int i = 0; i < Size; ++i) {
7349 // Each element of the unpack contains Scale elements from this mask.
7350 int UnpackIdx = i / Scale;
7352 // We only handle the case where V1 feeds the first slots of the unpack.
7353 // We rely on canonicalization to ensure this is the case.
7354 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7357 // Setup the mask for this input. The indexing is tricky as we have to
7358 // handle the unpack stride.
7359 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7360 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7364 // If we will have to shuffle both inputs to use the unpack, check whether
7365 // we can just unpack first and shuffle the result. If so, skip this unpack.
7366 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7367 !isNoopShuffleMask(V2Mask))
7370 // Shuffle the inputs into place.
7371 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7372 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7374 // Cast the inputs to the type we will use to unpack them.
7375 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7376 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7378 // Unpack the inputs and cast the result back to the desired type.
7379 return DAG.getNode(ISD::BITCAST, DL, VT,
7380 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7381 DL, UnpackVT, V1, V2));
7384 // We try each unpack from the largest to the smallest to try and find one
7385 // that fits this mask.
7386 int OrigNumElements = VT.getVectorNumElements();
7387 int OrigScalarSize = VT.getScalarSizeInBits();
7388 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7389 int Scale = ScalarSize / OrigScalarSize;
7390 int NumElements = OrigNumElements / Scale;
7391 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7392 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7396 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7398 if (NumLoInputs == 0 || NumHiInputs == 0) {
7399 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7400 "We have to have *some* inputs!");
7401 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7403 // FIXME: We could consider the total complexity of the permute of each
7404 // possible unpacking. Or at the least we should consider how many
7405 // half-crossings are created.
7406 // FIXME: We could consider commuting the unpacks.
7408 SmallVector<int, 32> PermMask;
7409 PermMask.assign(Size, -1);
7410 for (int i = 0; i < Size; ++i) {
7414 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7417 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7419 return DAG.getVectorShuffle(
7420 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7422 DAG.getUNDEF(VT), PermMask);
7428 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7430 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7431 /// support for floating point shuffles but not integer shuffles. These
7432 /// instructions will incur a domain crossing penalty on some chips though so
7433 /// it is better to avoid lowering through this for integer vectors where
7435 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7436 const X86Subtarget *Subtarget,
7437 SelectionDAG &DAG) {
7439 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7440 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7441 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7442 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7443 ArrayRef<int> Mask = SVOp->getMask();
7444 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7446 if (isSingleInputShuffleMask(Mask)) {
7447 // Use low duplicate instructions for masks that match their pattern.
7448 if (Subtarget->hasSSE3())
7449 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7450 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7452 // Straight shuffle of a single input vector. Simulate this by using the
7453 // single input as both of the "inputs" to this instruction..
7454 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7456 if (Subtarget->hasAVX()) {
7457 // If we have AVX, we can use VPERMILPS which will allow folding a load
7458 // into the shuffle.
7459 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7460 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7463 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7464 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7466 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7467 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7469 // If we have a single input, insert that into V1 if we can do so cheaply.
7470 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7471 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7472 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7474 // Try inverting the insertion since for v2 masks it is easy to do and we
7475 // can't reliably sort the mask one way or the other.
7476 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7477 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7478 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7479 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7483 // Try to use one of the special instruction patterns to handle two common
7484 // blend patterns if a zero-blend above didn't work.
7485 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7486 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7487 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7488 // We can either use a special instruction to load over the low double or
7489 // to move just the low double.
7491 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7493 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7495 if (Subtarget->hasSSE41())
7496 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7500 // Use dedicated unpack instructions for masks that match their pattern.
7501 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7502 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7503 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7504 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7506 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7507 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7508 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7511 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7513 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7514 /// the integer unit to minimize domain crossing penalties. However, for blends
7515 /// it falls back to the floating point shuffle operation with appropriate bit
7517 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7518 const X86Subtarget *Subtarget,
7519 SelectionDAG &DAG) {
7521 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7522 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7523 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7524 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7525 ArrayRef<int> Mask = SVOp->getMask();
7526 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7528 if (isSingleInputShuffleMask(Mask)) {
7529 // Check for being able to broadcast a single element.
7530 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7531 Mask, Subtarget, DAG))
7534 // Straight shuffle of a single input vector. For everything from SSE2
7535 // onward this has a single fast instruction with no scary immediates.
7536 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7537 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7538 int WidenedMask[4] = {
7539 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7540 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7542 ISD::BITCAST, DL, MVT::v2i64,
7543 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7544 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7546 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7547 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7548 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7549 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7551 // If we have a blend of two PACKUS operations an the blend aligns with the
7552 // low and half halves, we can just merge the PACKUS operations. This is
7553 // particularly important as it lets us merge shuffles that this routine itself
7555 auto GetPackNode = [](SDValue V) {
7556 while (V.getOpcode() == ISD::BITCAST)
7557 V = V.getOperand(0);
7559 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7561 if (SDValue V1Pack = GetPackNode(V1))
7562 if (SDValue V2Pack = GetPackNode(V2))
7563 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7564 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7565 Mask[0] == 0 ? V1Pack.getOperand(0)
7566 : V1Pack.getOperand(1),
7567 Mask[1] == 2 ? V2Pack.getOperand(0)
7568 : V2Pack.getOperand(1)));
7570 // Try to use shift instructions.
7572 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7575 // When loading a scalar and then shuffling it into a vector we can often do
7576 // the insertion cheaply.
7577 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7578 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7580 // Try inverting the insertion since for v2 masks it is easy to do and we
7581 // can't reliably sort the mask one way or the other.
7582 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7583 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7584 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7587 // We have different paths for blend lowering, but they all must use the
7588 // *exact* same predicate.
7589 bool IsBlendSupported = Subtarget->hasSSE41();
7590 if (IsBlendSupported)
7591 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7595 // Use dedicated unpack instructions for masks that match their pattern.
7596 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7597 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7598 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7599 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7601 // Try to use byte rotation instructions.
7602 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7603 if (Subtarget->hasSSSE3())
7604 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7605 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7608 // If we have direct support for blends, we should lower by decomposing into
7609 // a permute. That will be faster than the domain cross.
7610 if (IsBlendSupported)
7611 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7614 // We implement this with SHUFPD which is pretty lame because it will likely
7615 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7616 // However, all the alternatives are still more cycles and newer chips don't
7617 // have this problem. It would be really nice if x86 had better shuffles here.
7618 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7619 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7620 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7621 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7624 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7626 /// This is used to disable more specialized lowerings when the shufps lowering
7627 /// will happen to be efficient.
7628 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7629 // This routine only handles 128-bit shufps.
7630 assert(Mask.size() == 4 && "Unsupported mask size!");
7632 // To lower with a single SHUFPS we need to have the low half and high half
7633 // each requiring a single input.
7634 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7636 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7642 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7644 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7645 /// It makes no assumptions about whether this is the *best* lowering, it simply
7647 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7648 ArrayRef<int> Mask, SDValue V1,
7649 SDValue V2, SelectionDAG &DAG) {
7650 SDValue LowV = V1, HighV = V2;
7651 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7654 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7656 if (NumV2Elements == 1) {
7658 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7661 // Compute the index adjacent to V2Index and in the same half by toggling
7663 int V2AdjIndex = V2Index ^ 1;
7665 if (Mask[V2AdjIndex] == -1) {
7666 // Handles all the cases where we have a single V2 element and an undef.
7667 // This will only ever happen in the high lanes because we commute the
7668 // vector otherwise.
7670 std::swap(LowV, HighV);
7671 NewMask[V2Index] -= 4;
7673 // Handle the case where the V2 element ends up adjacent to a V1 element.
7674 // To make this work, blend them together as the first step.
7675 int V1Index = V2AdjIndex;
7676 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7677 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7678 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7680 // Now proceed to reconstruct the final blend as we have the necessary
7681 // high or low half formed.
7688 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7689 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7691 } else if (NumV2Elements == 2) {
7692 if (Mask[0] < 4 && Mask[1] < 4) {
7693 // Handle the easy case where we have V1 in the low lanes and V2 in the
7697 } else if (Mask[2] < 4 && Mask[3] < 4) {
7698 // We also handle the reversed case because this utility may get called
7699 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7700 // arrange things in the right direction.
7706 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7707 // trying to place elements directly, just blend them and set up the final
7708 // shuffle to place them.
7710 // The first two blend mask elements are for V1, the second two are for
7712 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7713 Mask[2] < 4 ? Mask[2] : Mask[3],
7714 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7715 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7716 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7717 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7719 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7722 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7723 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7724 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7725 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7728 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7729 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
7732 /// \brief Lower 4-lane 32-bit floating point shuffles.
7734 /// Uses instructions exclusively from the floating point unit to minimize
7735 /// domain crossing penalties, as these are sufficient to implement all v4f32
7737 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7738 const X86Subtarget *Subtarget,
7739 SelectionDAG &DAG) {
7741 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7742 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7743 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7744 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7745 ArrayRef<int> Mask = SVOp->getMask();
7746 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7749 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7751 if (NumV2Elements == 0) {
7752 // Check for being able to broadcast a single element.
7753 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7754 Mask, Subtarget, DAG))
7757 // Use even/odd duplicate instructions for masks that match their pattern.
7758 if (Subtarget->hasSSE3()) {
7759 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7760 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7761 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7762 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7765 if (Subtarget->hasAVX()) {
7766 // If we have AVX, we can use VPERMILPS which will allow folding a load
7767 // into the shuffle.
7768 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7769 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7772 // Otherwise, use a straight shuffle of a single input vector. We pass the
7773 // input vector to both operands to simulate this with a SHUFPS.
7774 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7775 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7778 // There are special ways we can lower some single-element blends. However, we
7779 // have custom ways we can lower more complex single-element blends below that
7780 // we defer to if both this and BLENDPS fail to match, so restrict this to
7781 // when the V2 input is targeting element 0 of the mask -- that is the fast
7783 if (NumV2Elements == 1 && Mask[0] >= 4)
7784 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7785 Mask, Subtarget, DAG))
7788 if (Subtarget->hasSSE41()) {
7789 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7793 // Use INSERTPS if we can complete the shuffle efficiently.
7794 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7797 if (!isSingleSHUFPSMask(Mask))
7798 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7799 DL, MVT::v4f32, V1, V2, Mask, DAG))
7803 // Use dedicated unpack instructions for masks that match their pattern.
7804 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7805 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7806 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7807 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7808 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7809 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7810 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7811 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7813 // Otherwise fall back to a SHUFPS lowering strategy.
7814 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7817 /// \brief Lower 4-lane i32 vector shuffles.
7819 /// We try to handle these with integer-domain shuffles where we can, but for
7820 /// blends we use the floating point domain blend instructions.
7821 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7822 const X86Subtarget *Subtarget,
7823 SelectionDAG &DAG) {
7825 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7826 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7827 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7828 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7829 ArrayRef<int> Mask = SVOp->getMask();
7830 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7832 // Whenever we can lower this as a zext, that instruction is strictly faster
7833 // than any alternative. It also allows us to fold memory operands into the
7834 // shuffle in many cases.
7835 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7836 Mask, Subtarget, DAG))
7840 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7842 if (NumV2Elements == 0) {
7843 // Check for being able to broadcast a single element.
7844 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7845 Mask, Subtarget, DAG))
7848 // Straight shuffle of a single input vector. For everything from SSE2
7849 // onward this has a single fast instruction with no scary immediates.
7850 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7851 // but we aren't actually going to use the UNPCK instruction because doing
7852 // so prevents folding a load into this instruction or making a copy.
7853 const int UnpackLoMask[] = {0, 0, 1, 1};
7854 const int UnpackHiMask[] = {2, 2, 3, 3};
7855 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7856 Mask = UnpackLoMask;
7857 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7858 Mask = UnpackHiMask;
7860 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7861 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
7864 // Try to use shift instructions.
7866 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7869 // There are special ways we can lower some single-element blends.
7870 if (NumV2Elements == 1)
7871 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7872 Mask, Subtarget, DAG))
7875 // We have different paths for blend lowering, but they all must use the
7876 // *exact* same predicate.
7877 bool IsBlendSupported = Subtarget->hasSSE41();
7878 if (IsBlendSupported)
7879 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7883 if (SDValue Masked =
7884 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7887 // Use dedicated unpack instructions for masks that match their pattern.
7888 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7889 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7890 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7891 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7892 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7893 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7894 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7895 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7897 // Try to use byte rotation instructions.
7898 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7899 if (Subtarget->hasSSSE3())
7900 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7901 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7904 // If we have direct support for blends, we should lower by decomposing into
7905 // a permute. That will be faster than the domain cross.
7906 if (IsBlendSupported)
7907 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7910 // Try to lower by permuting the inputs into an unpack instruction.
7911 if (SDValue Unpack =
7912 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7915 // We implement this with SHUFPS because it can blend from two vectors.
7916 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7917 // up the inputs, bypassing domain shift penalties that we would encur if we
7918 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7920 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7921 DAG.getVectorShuffle(
7923 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7924 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7927 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7928 /// shuffle lowering, and the most complex part.
7930 /// The lowering strategy is to try to form pairs of input lanes which are
7931 /// targeted at the same half of the final vector, and then use a dword shuffle
7932 /// to place them onto the right half, and finally unpack the paired lanes into
7933 /// their final position.
7935 /// The exact breakdown of how to form these dword pairs and align them on the
7936 /// correct sides is really tricky. See the comments within the function for
7937 /// more of the details.
7939 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
7940 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
7941 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
7942 /// vector, form the analogous 128-bit 8-element Mask.
7943 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7944 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
7945 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7946 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
7947 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
7949 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
7950 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7951 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7953 SmallVector<int, 4> LoInputs;
7954 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7955 [](int M) { return M >= 0; });
7956 std::sort(LoInputs.begin(), LoInputs.end());
7957 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7958 SmallVector<int, 4> HiInputs;
7959 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7960 [](int M) { return M >= 0; });
7961 std::sort(HiInputs.begin(), HiInputs.end());
7962 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7964 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7965 int NumHToL = LoInputs.size() - NumLToL;
7967 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7968 int NumHToH = HiInputs.size() - NumLToH;
7969 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7970 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7971 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7972 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7974 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7975 // such inputs we can swap two of the dwords across the half mark and end up
7976 // with <=2 inputs to each half in each half. Once there, we can fall through
7977 // to the generic code below. For example:
7979 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7980 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7982 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7983 // and an existing 2-into-2 on the other half. In this case we may have to
7984 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7985 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7986 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7987 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7988 // half than the one we target for fixing) will be fixed when we re-enter this
7989 // path. We will also combine away any sequence of PSHUFD instructions that
7990 // result into a single instruction. Here is an example of the tricky case:
7992 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7993 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7995 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7997 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7998 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8000 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8001 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8003 // The result is fine to be handled by the generic logic.
8004 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8005 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8006 int AOffset, int BOffset) {
8007 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8008 "Must call this with A having 3 or 1 inputs from the A half.");
8009 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8010 "Must call this with B having 1 or 3 inputs from the B half.");
8011 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8012 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8014 // Compute the index of dword with only one word among the three inputs in
8015 // a half by taking the sum of the half with three inputs and subtracting
8016 // the sum of the actual three inputs. The difference is the remaining
8019 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8020 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8021 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8022 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8023 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8024 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8025 int TripleNonInputIdx =
8026 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8027 TripleDWord = TripleNonInputIdx / 2;
8029 // We use xor with one to compute the adjacent DWord to whichever one the
8031 OneInputDWord = (OneInput / 2) ^ 1;
8033 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8034 // and BToA inputs. If there is also such a problem with the BToB and AToB
8035 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8036 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8037 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8038 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8039 // Compute how many inputs will be flipped by swapping these DWords. We
8041 // to balance this to ensure we don't form a 3-1 shuffle in the other
8043 int NumFlippedAToBInputs =
8044 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8045 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8046 int NumFlippedBToBInputs =
8047 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8048 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8049 if ((NumFlippedAToBInputs == 1 &&
8050 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8051 (NumFlippedBToBInputs == 1 &&
8052 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8053 // We choose whether to fix the A half or B half based on whether that
8054 // half has zero flipped inputs. At zero, we may not be able to fix it
8055 // with that half. We also bias towards fixing the B half because that
8056 // will more commonly be the high half, and we have to bias one way.
8057 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8058 ArrayRef<int> Inputs) {
8059 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8060 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8061 PinnedIdx ^ 1) != Inputs.end();
8062 // Determine whether the free index is in the flipped dword or the
8063 // unflipped dword based on where the pinned index is. We use this bit
8064 // in an xor to conditionally select the adjacent dword.
8065 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8066 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8067 FixFreeIdx) != Inputs.end();
8068 if (IsFixIdxInput == IsFixFreeIdxInput)
8070 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8071 FixFreeIdx) != Inputs.end();
8072 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8073 "We need to be changing the number of flipped inputs!");
8074 int PSHUFHalfMask[] = {0, 1, 2, 3};
8075 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8076 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8078 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8081 if (M != -1 && M == FixIdx)
8083 else if (M != -1 && M == FixFreeIdx)
8086 if (NumFlippedBToBInputs != 0) {
8088 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8089 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8091 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8093 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8094 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8099 int PSHUFDMask[] = {0, 1, 2, 3};
8100 PSHUFDMask[ADWord] = BDWord;
8101 PSHUFDMask[BDWord] = ADWord;
8102 V = DAG.getNode(ISD::BITCAST, DL, VT,
8103 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8104 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8105 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8108 // Adjust the mask to match the new locations of A and B.
8110 if (M != -1 && M/2 == ADWord)
8111 M = 2 * BDWord + M % 2;
8112 else if (M != -1 && M/2 == BDWord)
8113 M = 2 * ADWord + M % 2;
8115 // Recurse back into this routine to re-compute state now that this isn't
8116 // a 3 and 1 problem.
8117 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8120 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8121 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8122 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8123 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8125 // At this point there are at most two inputs to the low and high halves from
8126 // each half. That means the inputs can always be grouped into dwords and
8127 // those dwords can then be moved to the correct half with a dword shuffle.
8128 // We use at most one low and one high word shuffle to collect these paired
8129 // inputs into dwords, and finally a dword shuffle to place them.
8130 int PSHUFLMask[4] = {-1, -1, -1, -1};
8131 int PSHUFHMask[4] = {-1, -1, -1, -1};
8132 int PSHUFDMask[4] = {-1, -1, -1, -1};
8134 // First fix the masks for all the inputs that are staying in their
8135 // original halves. This will then dictate the targets of the cross-half
8137 auto fixInPlaceInputs =
8138 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8139 MutableArrayRef<int> SourceHalfMask,
8140 MutableArrayRef<int> HalfMask, int HalfOffset) {
8141 if (InPlaceInputs.empty())
8143 if (InPlaceInputs.size() == 1) {
8144 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8145 InPlaceInputs[0] - HalfOffset;
8146 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8149 if (IncomingInputs.empty()) {
8150 // Just fix all of the in place inputs.
8151 for (int Input : InPlaceInputs) {
8152 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8153 PSHUFDMask[Input / 2] = Input / 2;
8158 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8159 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8160 InPlaceInputs[0] - HalfOffset;
8161 // Put the second input next to the first so that they are packed into
8162 // a dword. We find the adjacent index by toggling the low bit.
8163 int AdjIndex = InPlaceInputs[0] ^ 1;
8164 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8165 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8166 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8168 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8169 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8171 // Now gather the cross-half inputs and place them into a free dword of
8172 // their target half.
8173 // FIXME: This operation could almost certainly be simplified dramatically to
8174 // look more like the 3-1 fixing operation.
8175 auto moveInputsToRightHalf = [&PSHUFDMask](
8176 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8177 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8178 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8180 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8181 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8183 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8185 int LowWord = Word & ~1;
8186 int HighWord = Word | 1;
8187 return isWordClobbered(SourceHalfMask, LowWord) ||
8188 isWordClobbered(SourceHalfMask, HighWord);
8191 if (IncomingInputs.empty())
8194 if (ExistingInputs.empty()) {
8195 // Map any dwords with inputs from them into the right half.
8196 for (int Input : IncomingInputs) {
8197 // If the source half mask maps over the inputs, turn those into
8198 // swaps and use the swapped lane.
8199 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8200 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8201 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8202 Input - SourceOffset;
8203 // We have to swap the uses in our half mask in one sweep.
8204 for (int &M : HalfMask)
8205 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8207 else if (M == Input)
8208 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8210 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8211 Input - SourceOffset &&
8212 "Previous placement doesn't match!");
8214 // Note that this correctly re-maps both when we do a swap and when
8215 // we observe the other side of the swap above. We rely on that to
8216 // avoid swapping the members of the input list directly.
8217 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8220 // Map the input's dword into the correct half.
8221 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8222 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8224 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8226 "Previous placement doesn't match!");
8229 // And just directly shift any other-half mask elements to be same-half
8230 // as we will have mirrored the dword containing the element into the
8231 // same position within that half.
8232 for (int &M : HalfMask)
8233 if (M >= SourceOffset && M < SourceOffset + 4) {
8234 M = M - SourceOffset + DestOffset;
8235 assert(M >= 0 && "This should never wrap below zero!");
8240 // Ensure we have the input in a viable dword of its current half. This
8241 // is particularly tricky because the original position may be clobbered
8242 // by inputs being moved and *staying* in that half.
8243 if (IncomingInputs.size() == 1) {
8244 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8245 int InputFixed = std::find(std::begin(SourceHalfMask),
8246 std::end(SourceHalfMask), -1) -
8247 std::begin(SourceHalfMask) + SourceOffset;
8248 SourceHalfMask[InputFixed - SourceOffset] =
8249 IncomingInputs[0] - SourceOffset;
8250 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8252 IncomingInputs[0] = InputFixed;
8254 } else if (IncomingInputs.size() == 2) {
8255 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8256 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8257 // We have two non-adjacent or clobbered inputs we need to extract from
8258 // the source half. To do this, we need to map them into some adjacent
8259 // dword slot in the source mask.
8260 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8261 IncomingInputs[1] - SourceOffset};
8263 // If there is a free slot in the source half mask adjacent to one of
8264 // the inputs, place the other input in it. We use (Index XOR 1) to
8265 // compute an adjacent index.
8266 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8267 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8268 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8269 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8270 InputsFixed[1] = InputsFixed[0] ^ 1;
8271 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8272 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8273 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8274 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8275 InputsFixed[0] = InputsFixed[1] ^ 1;
8276 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8277 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8278 // The two inputs are in the same DWord but it is clobbered and the
8279 // adjacent DWord isn't used at all. Move both inputs to the free
8281 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8282 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8283 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8284 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8286 // The only way we hit this point is if there is no clobbering
8287 // (because there are no off-half inputs to this half) and there is no
8288 // free slot adjacent to one of the inputs. In this case, we have to
8289 // swap an input with a non-input.
8290 for (int i = 0; i < 4; ++i)
8291 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8292 "We can't handle any clobbers here!");
8293 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8294 "Cannot have adjacent inputs here!");
8296 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8297 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8299 // We also have to update the final source mask in this case because
8300 // it may need to undo the above swap.
8301 for (int &M : FinalSourceHalfMask)
8302 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8303 M = InputsFixed[1] + SourceOffset;
8304 else if (M == InputsFixed[1] + SourceOffset)
8305 M = (InputsFixed[0] ^ 1) + SourceOffset;
8307 InputsFixed[1] = InputsFixed[0] ^ 1;
8310 // Point everything at the fixed inputs.
8311 for (int &M : HalfMask)
8312 if (M == IncomingInputs[0])
8313 M = InputsFixed[0] + SourceOffset;
8314 else if (M == IncomingInputs[1])
8315 M = InputsFixed[1] + SourceOffset;
8317 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8318 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8321 llvm_unreachable("Unhandled input size!");
8324 // Now hoist the DWord down to the right half.
8325 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8326 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8327 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8328 for (int &M : HalfMask)
8329 for (int Input : IncomingInputs)
8331 M = FreeDWord * 2 + Input % 2;
8333 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8334 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8335 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8336 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8338 // Now enact all the shuffles we've computed to move the inputs into their
8340 if (!isNoopShuffleMask(PSHUFLMask))
8341 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8342 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8343 if (!isNoopShuffleMask(PSHUFHMask))
8344 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8345 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8346 if (!isNoopShuffleMask(PSHUFDMask))
8347 V = DAG.getNode(ISD::BITCAST, DL, VT,
8348 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
8349 DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
8350 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL,
8353 // At this point, each half should contain all its inputs, and we can then
8354 // just shuffle them into their final position.
8355 assert(std::count_if(LoMask.begin(), LoMask.end(),
8356 [](int M) { return M >= 4; }) == 0 &&
8357 "Failed to lift all the high half inputs to the low mask!");
8358 assert(std::count_if(HiMask.begin(), HiMask.end(),
8359 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8360 "Failed to lift all the low half inputs to the high mask!");
8362 // Do a half shuffle for the low mask.
8363 if (!isNoopShuffleMask(LoMask))
8364 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8365 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8367 // Do a half shuffle with the high mask after shifting its values down.
8368 for (int &M : HiMask)
8371 if (!isNoopShuffleMask(HiMask))
8372 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8373 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8378 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8379 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8380 SDValue V2, ArrayRef<int> Mask,
8381 SelectionDAG &DAG, bool &V1InUse,
8383 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8389 int Size = Mask.size();
8390 int Scale = 16 / Size;
8391 for (int i = 0; i < 16; ++i) {
8392 if (Mask[i / Scale] == -1) {
8393 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8395 const int ZeroMask = 0x80;
8396 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8398 int V2Idx = Mask[i / Scale] < Size
8400 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8401 if (Zeroable[i / Scale])
8402 V1Idx = V2Idx = ZeroMask;
8403 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8404 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8405 V1InUse |= (ZeroMask != V1Idx);
8406 V2InUse |= (ZeroMask != V2Idx);
8411 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8412 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8413 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8415 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8416 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8417 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8419 // If we need shuffled inputs from both, blend the two.
8421 if (V1InUse && V2InUse)
8422 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8424 V = V1InUse ? V1 : V2;
8426 // Cast the result back to the correct type.
8427 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8430 /// \brief Generic lowering of 8-lane i16 shuffles.
8432 /// This handles both single-input shuffles and combined shuffle/blends with
8433 /// two inputs. The single input shuffles are immediately delegated to
8434 /// a dedicated lowering routine.
8436 /// The blends are lowered in one of three fundamental ways. If there are few
8437 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8438 /// of the input is significantly cheaper when lowered as an interleaving of
8439 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8440 /// halves of the inputs separately (making them have relatively few inputs)
8441 /// and then concatenate them.
8442 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8443 const X86Subtarget *Subtarget,
8444 SelectionDAG &DAG) {
8446 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8447 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8448 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8449 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8450 ArrayRef<int> OrigMask = SVOp->getMask();
8451 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8452 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8453 MutableArrayRef<int> Mask(MaskStorage);
8455 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8457 // Whenever we can lower this as a zext, that instruction is strictly faster
8458 // than any alternative.
8459 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8460 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8463 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8465 auto isV2 = [](int M) { return M >= 8; };
8467 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8469 if (NumV2Inputs == 0) {
8470 // Check for being able to broadcast a single element.
8471 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8472 Mask, Subtarget, DAG))
8475 // Try to use shift instructions.
8477 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8480 // Use dedicated unpack instructions for masks that match their pattern.
8481 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8482 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8483 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8484 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8486 // Try to use byte rotation instructions.
8487 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8488 Mask, Subtarget, DAG))
8491 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8495 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8496 "All single-input shuffles should be canonicalized to be V1-input "
8499 // Try to use shift instructions.
8501 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8504 // There are special ways we can lower some single-element blends.
8505 if (NumV2Inputs == 1)
8506 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8507 Mask, Subtarget, DAG))
8510 // We have different paths for blend lowering, but they all must use the
8511 // *exact* same predicate.
8512 bool IsBlendSupported = Subtarget->hasSSE41();
8513 if (IsBlendSupported)
8514 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8518 if (SDValue Masked =
8519 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8522 // Use dedicated unpack instructions for masks that match their pattern.
8523 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8524 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8525 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8526 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8528 // Try to use byte rotation instructions.
8529 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8530 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8533 if (SDValue BitBlend =
8534 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8537 if (SDValue Unpack =
8538 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8541 // If we can't directly blend but can use PSHUFB, that will be better as it
8542 // can both shuffle and set up the inefficient blend.
8543 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8544 bool V1InUse, V2InUse;
8545 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8549 // We can always bit-blend if we have to so the fallback strategy is to
8550 // decompose into single-input permutes and blends.
8551 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8555 /// \brief Check whether a compaction lowering can be done by dropping even
8556 /// elements and compute how many times even elements must be dropped.
8558 /// This handles shuffles which take every Nth element where N is a power of
8559 /// two. Example shuffle masks:
8561 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8562 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8563 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8564 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8565 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8566 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8568 /// Any of these lanes can of course be undef.
8570 /// This routine only supports N <= 3.
8571 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8574 /// \returns N above, or the number of times even elements must be dropped if
8575 /// there is such a number. Otherwise returns zero.
8576 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8577 // Figure out whether we're looping over two inputs or just one.
8578 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8580 // The modulus for the shuffle vector entries is based on whether this is
8581 // a single input or not.
8582 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8583 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8584 "We should only be called with masks with a power-of-2 size!");
8586 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8588 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8589 // and 2^3 simultaneously. This is because we may have ambiguity with
8590 // partially undef inputs.
8591 bool ViableForN[3] = {true, true, true};
8593 for (int i = 0, e = Mask.size(); i < e; ++i) {
8594 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8599 bool IsAnyViable = false;
8600 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8601 if (ViableForN[j]) {
8604 // The shuffle mask must be equal to (i * 2^N) % M.
8605 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8608 ViableForN[j] = false;
8610 // Early exit if we exhaust the possible powers of two.
8615 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8619 // Return 0 as there is no viable power of two.
8623 /// \brief Generic lowering of v16i8 shuffles.
8625 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8626 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8627 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8628 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8630 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8631 const X86Subtarget *Subtarget,
8632 SelectionDAG &DAG) {
8634 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8635 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8636 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8637 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8638 ArrayRef<int> Mask = SVOp->getMask();
8639 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8641 // Try to use shift instructions.
8643 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8646 // Try to use byte rotation instructions.
8647 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8648 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8651 // Try to use a zext lowering.
8652 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8653 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8657 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8659 // For single-input shuffles, there are some nicer lowering tricks we can use.
8660 if (NumV2Elements == 0) {
8661 // Check for being able to broadcast a single element.
8662 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8663 Mask, Subtarget, DAG))
8666 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8667 // Notably, this handles splat and partial-splat shuffles more efficiently.
8668 // However, it only makes sense if the pre-duplication shuffle simplifies
8669 // things significantly. Currently, this means we need to be able to
8670 // express the pre-duplication shuffle as an i16 shuffle.
8672 // FIXME: We should check for other patterns which can be widened into an
8673 // i16 shuffle as well.
8674 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8675 for (int i = 0; i < 16; i += 2)
8676 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8681 auto tryToWidenViaDuplication = [&]() -> SDValue {
8682 if (!canWidenViaDuplication(Mask))
8684 SmallVector<int, 4> LoInputs;
8685 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8686 [](int M) { return M >= 0 && M < 8; });
8687 std::sort(LoInputs.begin(), LoInputs.end());
8688 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8690 SmallVector<int, 4> HiInputs;
8691 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8692 [](int M) { return M >= 8; });
8693 std::sort(HiInputs.begin(), HiInputs.end());
8694 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8697 bool TargetLo = LoInputs.size() >= HiInputs.size();
8698 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8699 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8701 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8702 SmallDenseMap<int, int, 8> LaneMap;
8703 for (int I : InPlaceInputs) {
8704 PreDupI16Shuffle[I/2] = I/2;
8707 int j = TargetLo ? 0 : 4, je = j + 4;
8708 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8709 // Check if j is already a shuffle of this input. This happens when
8710 // there are two adjacent bytes after we move the low one.
8711 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8712 // If we haven't yet mapped the input, search for a slot into which
8714 while (j < je && PreDupI16Shuffle[j] != -1)
8718 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8721 // Map this input with the i16 shuffle.
8722 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8725 // Update the lane map based on the mapping we ended up with.
8726 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8729 ISD::BITCAST, DL, MVT::v16i8,
8730 DAG.getVectorShuffle(MVT::v8i16, DL,
8731 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8732 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8734 // Unpack the bytes to form the i16s that will be shuffled into place.
8735 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8736 MVT::v16i8, V1, V1);
8738 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8739 for (int i = 0; i < 16; ++i)
8740 if (Mask[i] != -1) {
8741 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8742 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8743 if (PostDupI16Shuffle[i / 2] == -1)
8744 PostDupI16Shuffle[i / 2] = MappedMask;
8746 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8747 "Conflicting entrties in the original shuffle!");
8750 ISD::BITCAST, DL, MVT::v16i8,
8751 DAG.getVectorShuffle(MVT::v8i16, DL,
8752 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8753 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8755 if (SDValue V = tryToWidenViaDuplication())
8759 // Use dedicated unpack instructions for masks that match their pattern.
8760 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8761 0, 16, 1, 17, 2, 18, 3, 19,
8763 4, 20, 5, 21, 6, 22, 7, 23}))
8764 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8765 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8766 8, 24, 9, 25, 10, 26, 11, 27,
8768 12, 28, 13, 29, 14, 30, 15, 31}))
8769 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8771 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8772 // with PSHUFB. It is important to do this before we attempt to generate any
8773 // blends but after all of the single-input lowerings. If the single input
8774 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8775 // want to preserve that and we can DAG combine any longer sequences into
8776 // a PSHUFB in the end. But once we start blending from multiple inputs,
8777 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8778 // and there are *very* few patterns that would actually be faster than the
8779 // PSHUFB approach because of its ability to zero lanes.
8781 // FIXME: The only exceptions to the above are blends which are exact
8782 // interleavings with direct instructions supporting them. We currently don't
8783 // handle those well here.
8784 if (Subtarget->hasSSSE3()) {
8785 bool V1InUse = false;
8786 bool V2InUse = false;
8788 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8789 DAG, V1InUse, V2InUse);
8791 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8792 // do so. This avoids using them to handle blends-with-zero which is
8793 // important as a single pshufb is significantly faster for that.
8794 if (V1InUse && V2InUse) {
8795 if (Subtarget->hasSSE41())
8796 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8797 Mask, Subtarget, DAG))
8800 // We can use an unpack to do the blending rather than an or in some
8801 // cases. Even though the or may be (very minorly) more efficient, we
8802 // preference this lowering because there are common cases where part of
8803 // the complexity of the shuffles goes away when we do the final blend as
8805 // FIXME: It might be worth trying to detect if the unpack-feeding
8806 // shuffles will both be pshufb, in which case we shouldn't bother with
8808 if (SDValue Unpack =
8809 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8816 // There are special ways we can lower some single-element blends.
8817 if (NumV2Elements == 1)
8818 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8819 Mask, Subtarget, DAG))
8822 if (SDValue BitBlend =
8823 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8826 // Check whether a compaction lowering can be done. This handles shuffles
8827 // which take every Nth element for some even N. See the helper function for
8830 // We special case these as they can be particularly efficiently handled with
8831 // the PACKUSB instruction on x86 and they show up in common patterns of
8832 // rearranging bytes to truncate wide elements.
8833 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8834 // NumEvenDrops is the power of two stride of the elements. Another way of
8835 // thinking about it is that we need to drop the even elements this many
8836 // times to get the original input.
8837 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8839 // First we need to zero all the dropped bytes.
8840 assert(NumEvenDrops <= 3 &&
8841 "No support for dropping even elements more than 3 times.");
8842 // We use the mask type to pick which bytes are preserved based on how many
8843 // elements are dropped.
8844 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8845 SDValue ByteClearMask =
8846 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8847 DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
8848 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8850 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8852 // Now pack things back together.
8853 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8854 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8855 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8856 for (int i = 1; i < NumEvenDrops; ++i) {
8857 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8858 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8864 // Handle multi-input cases by blending single-input shuffles.
8865 if (NumV2Elements > 0)
8866 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8869 // The fallback path for single-input shuffles widens this into two v8i16
8870 // vectors with unpacks, shuffles those, and then pulls them back together
8874 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8875 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8876 for (int i = 0; i < 16; ++i)
8878 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8880 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8882 SDValue VLoHalf, VHiHalf;
8883 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8884 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8886 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8887 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8888 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8889 [](int M) { return M >= 0 && M % 2 == 1; })) {
8890 // Use a mask to drop the high bytes.
8891 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8892 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8893 DAG.getConstant(0x00FF, DL, MVT::v8i16));
8895 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8896 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8898 // Squash the masks to point directly into VLoHalf.
8899 for (int &M : LoBlendMask)
8902 for (int &M : HiBlendMask)
8906 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8907 // VHiHalf so that we can blend them as i16s.
8908 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8909 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8910 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8911 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8914 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8915 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8917 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8920 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8922 /// This routine breaks down the specific type of 128-bit shuffle and
8923 /// dispatches to the lowering routines accordingly.
8924 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8925 MVT VT, const X86Subtarget *Subtarget,
8926 SelectionDAG &DAG) {
8927 switch (VT.SimpleTy) {
8929 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8931 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8933 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8935 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8937 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8939 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8942 llvm_unreachable("Unimplemented!");
8946 /// \brief Helper function to test whether a shuffle mask could be
8947 /// simplified by widening the elements being shuffled.
8949 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8950 /// leaves it in an unspecified state.
8952 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8953 /// shuffle masks. The latter have the special property of a '-2' representing
8954 /// a zero-ed lane of a vector.
8955 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8956 SmallVectorImpl<int> &WidenedMask) {
8957 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8958 // If both elements are undef, its trivial.
8959 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8960 WidenedMask.push_back(SM_SentinelUndef);
8964 // Check for an undef mask and a mask value properly aligned to fit with
8965 // a pair of values. If we find such a case, use the non-undef mask's value.
8966 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8967 WidenedMask.push_back(Mask[i + 1] / 2);
8970 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8971 WidenedMask.push_back(Mask[i] / 2);
8975 // When zeroing, we need to spread the zeroing across both lanes to widen.
8976 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8977 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8978 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8979 WidenedMask.push_back(SM_SentinelZero);
8985 // Finally check if the two mask values are adjacent and aligned with
8987 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8988 WidenedMask.push_back(Mask[i] / 2);
8992 // Otherwise we can't safely widen the elements used in this shuffle.
8995 assert(WidenedMask.size() == Mask.size() / 2 &&
8996 "Incorrect size of mask after widening the elements!");
9001 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9003 /// This routine just extracts two subvectors, shuffles them independently, and
9004 /// then concatenates them back together. This should work effectively with all
9005 /// AVX vector shuffle types.
9006 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9007 SDValue V2, ArrayRef<int> Mask,
9008 SelectionDAG &DAG) {
9009 assert(VT.getSizeInBits() >= 256 &&
9010 "Only for 256-bit or wider vector shuffles!");
9011 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9012 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9014 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9015 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9017 int NumElements = VT.getVectorNumElements();
9018 int SplitNumElements = NumElements / 2;
9019 MVT ScalarVT = VT.getScalarType();
9020 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9022 // Rather than splitting build-vectors, just build two narrower build
9023 // vectors. This helps shuffling with splats and zeros.
9024 auto SplitVector = [&](SDValue V) {
9025 while (V.getOpcode() == ISD::BITCAST)
9026 V = V->getOperand(0);
9028 MVT OrigVT = V.getSimpleValueType();
9029 int OrigNumElements = OrigVT.getVectorNumElements();
9030 int OrigSplitNumElements = OrigNumElements / 2;
9031 MVT OrigScalarVT = OrigVT.getScalarType();
9032 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9036 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9038 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9039 DAG.getIntPtrConstant(0, DL));
9040 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9041 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9044 SmallVector<SDValue, 16> LoOps, HiOps;
9045 for (int i = 0; i < OrigSplitNumElements; ++i) {
9046 LoOps.push_back(BV->getOperand(i));
9047 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9049 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9050 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9052 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
9053 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
9056 SDValue LoV1, HiV1, LoV2, HiV2;
9057 std::tie(LoV1, HiV1) = SplitVector(V1);
9058 std::tie(LoV2, HiV2) = SplitVector(V2);
9060 // Now create two 4-way blends of these half-width vectors.
9061 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9062 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9063 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9064 for (int i = 0; i < SplitNumElements; ++i) {
9065 int M = HalfMask[i];
9066 if (M >= NumElements) {
9067 if (M >= NumElements + SplitNumElements)
9071 V2BlendMask.push_back(M - NumElements);
9072 V1BlendMask.push_back(-1);
9073 BlendMask.push_back(SplitNumElements + i);
9074 } else if (M >= 0) {
9075 if (M >= SplitNumElements)
9079 V2BlendMask.push_back(-1);
9080 V1BlendMask.push_back(M);
9081 BlendMask.push_back(i);
9083 V2BlendMask.push_back(-1);
9084 V1BlendMask.push_back(-1);
9085 BlendMask.push_back(-1);
9089 // Because the lowering happens after all combining takes place, we need to
9090 // manually combine these blend masks as much as possible so that we create
9091 // a minimal number of high-level vector shuffle nodes.
9093 // First try just blending the halves of V1 or V2.
9094 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9095 return DAG.getUNDEF(SplitVT);
9096 if (!UseLoV2 && !UseHiV2)
9097 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9098 if (!UseLoV1 && !UseHiV1)
9099 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9101 SDValue V1Blend, V2Blend;
9102 if (UseLoV1 && UseHiV1) {
9104 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9106 // We only use half of V1 so map the usage down into the final blend mask.
9107 V1Blend = UseLoV1 ? LoV1 : HiV1;
9108 for (int i = 0; i < SplitNumElements; ++i)
9109 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9110 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9112 if (UseLoV2 && UseHiV2) {
9114 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9116 // We only use half of V2 so map the usage down into the final blend mask.
9117 V2Blend = UseLoV2 ? LoV2 : HiV2;
9118 for (int i = 0; i < SplitNumElements; ++i)
9119 if (BlendMask[i] >= SplitNumElements)
9120 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9122 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9124 SDValue Lo = HalfBlend(LoMask);
9125 SDValue Hi = HalfBlend(HiMask);
9126 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9129 /// \brief Either split a vector in halves or decompose the shuffles and the
9132 /// This is provided as a good fallback for many lowerings of non-single-input
9133 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9134 /// between splitting the shuffle into 128-bit components and stitching those
9135 /// back together vs. extracting the single-input shuffles and blending those
9137 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9138 SDValue V2, ArrayRef<int> Mask,
9139 SelectionDAG &DAG) {
9140 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9141 "lower single-input shuffles as it "
9142 "could then recurse on itself.");
9143 int Size = Mask.size();
9145 // If this can be modeled as a broadcast of two elements followed by a blend,
9146 // prefer that lowering. This is especially important because broadcasts can
9147 // often fold with memory operands.
9148 auto DoBothBroadcast = [&] {
9149 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9152 if (V2BroadcastIdx == -1)
9153 V2BroadcastIdx = M - Size;
9154 else if (M - Size != V2BroadcastIdx)
9156 } else if (M >= 0) {
9157 if (V1BroadcastIdx == -1)
9159 else if (M != V1BroadcastIdx)
9164 if (DoBothBroadcast())
9165 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9168 // If the inputs all stem from a single 128-bit lane of each input, then we
9169 // split them rather than blending because the split will decompose to
9170 // unusually few instructions.
9171 int LaneCount = VT.getSizeInBits() / 128;
9172 int LaneSize = Size / LaneCount;
9173 SmallBitVector LaneInputs[2];
9174 LaneInputs[0].resize(LaneCount, false);
9175 LaneInputs[1].resize(LaneCount, false);
9176 for (int i = 0; i < Size; ++i)
9178 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9179 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9180 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9182 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9183 // that the decomposed single-input shuffles don't end up here.
9184 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9187 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9188 /// a permutation and blend of those lanes.
9190 /// This essentially blends the out-of-lane inputs to each lane into the lane
9191 /// from a permuted copy of the vector. This lowering strategy results in four
9192 /// instructions in the worst case for a single-input cross lane shuffle which
9193 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9194 /// of. Special cases for each particular shuffle pattern should be handled
9195 /// prior to trying this lowering.
9196 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9197 SDValue V1, SDValue V2,
9199 SelectionDAG &DAG) {
9200 // FIXME: This should probably be generalized for 512-bit vectors as well.
9201 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9202 int LaneSize = Mask.size() / 2;
9204 // If there are only inputs from one 128-bit lane, splitting will in fact be
9205 // less expensive. The flags track whether the given lane contains an element
9206 // that crosses to another lane.
9207 bool LaneCrossing[2] = {false, false};
9208 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9209 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9210 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9211 if (!LaneCrossing[0] || !LaneCrossing[1])
9212 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9214 if (isSingleInputShuffleMask(Mask)) {
9215 SmallVector<int, 32> FlippedBlendMask;
9216 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9217 FlippedBlendMask.push_back(
9218 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9220 : Mask[i] % LaneSize +
9221 (i / LaneSize) * LaneSize + Size));
9223 // Flip the vector, and blend the results which should now be in-lane. The
9224 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9225 // 5 for the high source. The value 3 selects the high half of source 2 and
9226 // the value 2 selects the low half of source 2. We only use source 2 to
9227 // allow folding it into a memory operand.
9228 unsigned PERMMask = 3 | 2 << 4;
9229 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9230 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9231 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9234 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9235 // will be handled by the above logic and a blend of the results, much like
9236 // other patterns in AVX.
9237 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9240 /// \brief Handle lowering 2-lane 128-bit shuffles.
9241 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9242 SDValue V2, ArrayRef<int> Mask,
9243 const X86Subtarget *Subtarget,
9244 SelectionDAG &DAG) {
9245 // TODO: If minimizing size and one of the inputs is a zero vector and the
9246 // the zero vector has only one use, we could use a VPERM2X128 to save the
9247 // instruction bytes needed to explicitly generate the zero vector.
9249 // Blends are faster and handle all the non-lane-crossing cases.
9250 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9254 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9255 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9257 // If either input operand is a zero vector, use VPERM2X128 because its mask
9258 // allows us to replace the zero input with an implicit zero.
9259 if (!IsV1Zero && !IsV2Zero) {
9260 // Check for patterns which can be matched with a single insert of a 128-bit
9262 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9263 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9264 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9265 VT.getVectorNumElements() / 2);
9266 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9267 DAG.getIntPtrConstant(0, DL));
9268 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9269 OnlyUsesV1 ? V1 : V2,
9270 DAG.getIntPtrConstant(0, DL));
9271 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9275 // Otherwise form a 128-bit permutation. After accounting for undefs,
9276 // convert the 64-bit shuffle mask selection values into 128-bit
9277 // selection bits by dividing the indexes by 2 and shifting into positions
9278 // defined by a vperm2*128 instruction's immediate control byte.
9280 // The immediate permute control byte looks like this:
9281 // [1:0] - select 128 bits from sources for low half of destination
9283 // [3] - zero low half of destination
9284 // [5:4] - select 128 bits from sources for high half of destination
9286 // [7] - zero high half of destination
9288 int MaskLO = Mask[0];
9289 if (MaskLO == SM_SentinelUndef)
9290 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9292 int MaskHI = Mask[2];
9293 if (MaskHI == SM_SentinelUndef)
9294 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9296 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9298 // If either input is a zero vector, replace it with an undef input.
9299 // Shuffle mask values < 4 are selecting elements of V1.
9300 // Shuffle mask values >= 4 are selecting elements of V2.
9301 // Adjust each half of the permute mask by clearing the half that was
9302 // selecting the zero vector and setting the zero mask bit.
9304 V1 = DAG.getUNDEF(VT);
9306 PermMask = (PermMask & 0xf0) | 0x08;
9308 PermMask = (PermMask & 0x0f) | 0x80;
9311 V2 = DAG.getUNDEF(VT);
9313 PermMask = (PermMask & 0xf0) | 0x08;
9315 PermMask = (PermMask & 0x0f) | 0x80;
9318 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9319 DAG.getConstant(PermMask, DL, MVT::i8));
9322 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9323 /// shuffling each lane.
9325 /// This will only succeed when the result of fixing the 128-bit lanes results
9326 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9327 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9328 /// the lane crosses early and then use simpler shuffles within each lane.
9330 /// FIXME: It might be worthwhile at some point to support this without
9331 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9332 /// in x86 only floating point has interesting non-repeating shuffles, and even
9333 /// those are still *marginally* more expensive.
9334 static SDValue lowerVectorShuffleByMerging128BitLanes(
9335 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9336 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9337 assert(!isSingleInputShuffleMask(Mask) &&
9338 "This is only useful with multiple inputs.");
9340 int Size = Mask.size();
9341 int LaneSize = 128 / VT.getScalarSizeInBits();
9342 int NumLanes = Size / LaneSize;
9343 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9345 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9346 // check whether the in-128-bit lane shuffles share a repeating pattern.
9347 SmallVector<int, 4> Lanes;
9348 Lanes.resize(NumLanes, -1);
9349 SmallVector<int, 4> InLaneMask;
9350 InLaneMask.resize(LaneSize, -1);
9351 for (int i = 0; i < Size; ++i) {
9355 int j = i / LaneSize;
9358 // First entry we've seen for this lane.
9359 Lanes[j] = Mask[i] / LaneSize;
9360 } else if (Lanes[j] != Mask[i] / LaneSize) {
9361 // This doesn't match the lane selected previously!
9365 // Check that within each lane we have a consistent shuffle mask.
9366 int k = i % LaneSize;
9367 if (InLaneMask[k] < 0) {
9368 InLaneMask[k] = Mask[i] % LaneSize;
9369 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9370 // This doesn't fit a repeating in-lane mask.
9375 // First shuffle the lanes into place.
9376 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9377 VT.getSizeInBits() / 64);
9378 SmallVector<int, 8> LaneMask;
9379 LaneMask.resize(NumLanes * 2, -1);
9380 for (int i = 0; i < NumLanes; ++i)
9381 if (Lanes[i] >= 0) {
9382 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9383 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9386 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9387 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9388 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9390 // Cast it back to the type we actually want.
9391 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9393 // Now do a simple shuffle that isn't lane crossing.
9394 SmallVector<int, 8> NewMask;
9395 NewMask.resize(Size, -1);
9396 for (int i = 0; i < Size; ++i)
9398 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9399 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9400 "Must not introduce lane crosses at this point!");
9402 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9405 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9408 /// This returns true if the elements from a particular input are already in the
9409 /// slot required by the given mask and require no permutation.
9410 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9411 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9412 int Size = Mask.size();
9413 for (int i = 0; i < Size; ++i)
9414 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9420 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9422 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9423 /// isn't available.
9424 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9425 const X86Subtarget *Subtarget,
9426 SelectionDAG &DAG) {
9428 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9429 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9430 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9431 ArrayRef<int> Mask = SVOp->getMask();
9432 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9434 SmallVector<int, 4> WidenedMask;
9435 if (canWidenShuffleElements(Mask, WidenedMask))
9436 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9439 if (isSingleInputShuffleMask(Mask)) {
9440 // Check for being able to broadcast a single element.
9441 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9442 Mask, Subtarget, DAG))
9445 // Use low duplicate instructions for masks that match their pattern.
9446 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9447 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9449 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9450 // Non-half-crossing single input shuffles can be lowerid with an
9451 // interleaved permutation.
9452 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9453 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9454 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9455 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9458 // With AVX2 we have direct support for this permutation.
9459 if (Subtarget->hasAVX2())
9460 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9461 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9463 // Otherwise, fall back.
9464 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9468 // X86 has dedicated unpack instructions that can handle specific blend
9469 // operations: UNPCKH and UNPCKL.
9470 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9471 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9472 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9473 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9474 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9475 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9476 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9477 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9479 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9483 // Check if the blend happens to exactly fit that of SHUFPD.
9484 if ((Mask[0] == -1 || Mask[0] < 2) &&
9485 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9486 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9487 (Mask[3] == -1 || Mask[3] >= 6)) {
9488 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9489 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9490 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9491 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9493 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9494 (Mask[1] == -1 || Mask[1] < 2) &&
9495 (Mask[2] == -1 || Mask[2] >= 6) &&
9496 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9497 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9498 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9499 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9500 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
9503 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9504 // shuffle. However, if we have AVX2 and either inputs are already in place,
9505 // we will be able to shuffle even across lanes the other input in a single
9506 // instruction so skip this pattern.
9507 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9508 isShuffleMaskInputInPlace(1, Mask))))
9509 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9510 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9513 // If we have AVX2 then we always want to lower with a blend because an v4 we
9514 // can fully permute the elements.
9515 if (Subtarget->hasAVX2())
9516 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9519 // Otherwise fall back on generic lowering.
9520 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9523 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9525 /// This routine is only called when we have AVX2 and thus a reasonable
9526 /// instruction set for v4i64 shuffling..
9527 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9528 const X86Subtarget *Subtarget,
9529 SelectionDAG &DAG) {
9531 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9532 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9533 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9534 ArrayRef<int> Mask = SVOp->getMask();
9535 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9536 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9538 SmallVector<int, 4> WidenedMask;
9539 if (canWidenShuffleElements(Mask, WidenedMask))
9540 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9543 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9547 // Check for being able to broadcast a single element.
9548 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9549 Mask, Subtarget, DAG))
9552 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9553 // use lower latency instructions that will operate on both 128-bit lanes.
9554 SmallVector<int, 2> RepeatedMask;
9555 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9556 if (isSingleInputShuffleMask(Mask)) {
9557 int PSHUFDMask[] = {-1, -1, -1, -1};
9558 for (int i = 0; i < 2; ++i)
9559 if (RepeatedMask[i] >= 0) {
9560 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9561 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9564 ISD::BITCAST, DL, MVT::v4i64,
9565 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9566 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9567 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9571 // AVX2 provides a direct instruction for permuting a single input across
9573 if (isSingleInputShuffleMask(Mask))
9574 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9575 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9577 // Try to use shift instructions.
9579 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9582 // Use dedicated unpack instructions for masks that match their pattern.
9583 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9584 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9585 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9586 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9587 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9588 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9589 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9590 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9592 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9593 // shuffle. However, if we have AVX2 and either inputs are already in place,
9594 // we will be able to shuffle even across lanes the other input in a single
9595 // instruction so skip this pattern.
9596 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9597 isShuffleMaskInputInPlace(1, Mask))))
9598 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9599 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9602 // Otherwise fall back on generic blend lowering.
9603 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9607 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9609 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9610 /// isn't available.
9611 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9612 const X86Subtarget *Subtarget,
9613 SelectionDAG &DAG) {
9615 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9616 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9617 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9618 ArrayRef<int> Mask = SVOp->getMask();
9619 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9621 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9625 // Check for being able to broadcast a single element.
9626 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9627 Mask, Subtarget, DAG))
9630 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9631 // options to efficiently lower the shuffle.
9632 SmallVector<int, 4> RepeatedMask;
9633 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9634 assert(RepeatedMask.size() == 4 &&
9635 "Repeated masks must be half the mask width!");
9637 // Use even/odd duplicate instructions for masks that match their pattern.
9638 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9639 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9640 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9641 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9643 if (isSingleInputShuffleMask(Mask))
9644 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9645 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9647 // Use dedicated unpack instructions for masks that match their pattern.
9648 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9649 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9650 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9651 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9652 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9653 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9654 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9655 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9657 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9658 // have already handled any direct blends. We also need to squash the
9659 // repeated mask into a simulated v4f32 mask.
9660 for (int i = 0; i < 4; ++i)
9661 if (RepeatedMask[i] >= 8)
9662 RepeatedMask[i] -= 4;
9663 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9666 // If we have a single input shuffle with different shuffle patterns in the
9667 // two 128-bit lanes use the variable mask to VPERMILPS.
9668 if (isSingleInputShuffleMask(Mask)) {
9669 SDValue VPermMask[8];
9670 for (int i = 0; i < 8; ++i)
9671 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9672 : DAG.getConstant(Mask[i], DL, MVT::i32);
9673 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9675 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9676 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9678 if (Subtarget->hasAVX2())
9679 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9680 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9681 DAG.getNode(ISD::BUILD_VECTOR, DL,
9682 MVT::v8i32, VPermMask)),
9685 // Otherwise, fall back.
9686 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9690 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9692 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9693 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9696 // If we have AVX2 then we always want to lower with a blend because at v8 we
9697 // can fully permute the elements.
9698 if (Subtarget->hasAVX2())
9699 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9702 // Otherwise fall back on generic lowering.
9703 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9706 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9708 /// This routine is only called when we have AVX2 and thus a reasonable
9709 /// instruction set for v8i32 shuffling..
9710 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9711 const X86Subtarget *Subtarget,
9712 SelectionDAG &DAG) {
9714 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9715 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9716 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9717 ArrayRef<int> Mask = SVOp->getMask();
9718 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9719 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9721 // Whenever we can lower this as a zext, that instruction is strictly faster
9722 // than any alternative. It also allows us to fold memory operands into the
9723 // shuffle in many cases.
9724 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9725 Mask, Subtarget, DAG))
9728 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9732 // Check for being able to broadcast a single element.
9733 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9734 Mask, Subtarget, DAG))
9737 // If the shuffle mask is repeated in each 128-bit lane we can use more
9738 // efficient instructions that mirror the shuffles across the two 128-bit
9740 SmallVector<int, 4> RepeatedMask;
9741 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9742 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9743 if (isSingleInputShuffleMask(Mask))
9744 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9745 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9747 // Use dedicated unpack instructions for masks that match their pattern.
9748 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9749 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9750 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9751 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9752 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9753 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9754 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9755 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9758 // Try to use shift instructions.
9760 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9763 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9764 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9767 // If the shuffle patterns aren't repeated but it is a single input, directly
9768 // generate a cross-lane VPERMD instruction.
9769 if (isSingleInputShuffleMask(Mask)) {
9770 SDValue VPermMask[8];
9771 for (int i = 0; i < 8; ++i)
9772 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9773 : DAG.getConstant(Mask[i], DL, MVT::i32);
9775 X86ISD::VPERMV, DL, MVT::v8i32,
9776 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9779 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9781 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9782 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9785 // Otherwise fall back on generic blend lowering.
9786 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9790 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9792 /// This routine is only called when we have AVX2 and thus a reasonable
9793 /// instruction set for v16i16 shuffling..
9794 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9795 const X86Subtarget *Subtarget,
9796 SelectionDAG &DAG) {
9798 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9799 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9800 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9801 ArrayRef<int> Mask = SVOp->getMask();
9802 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9803 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9805 // Whenever we can lower this as a zext, that instruction is strictly faster
9806 // than any alternative. It also allows us to fold memory operands into the
9807 // shuffle in many cases.
9808 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9809 Mask, Subtarget, DAG))
9812 // Check for being able to broadcast a single element.
9813 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9814 Mask, Subtarget, DAG))
9817 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9821 // Use dedicated unpack instructions for masks that match their pattern.
9822 if (isShuffleEquivalent(V1, V2, Mask,
9823 {// First 128-bit lane:
9824 0, 16, 1, 17, 2, 18, 3, 19,
9825 // Second 128-bit lane:
9826 8, 24, 9, 25, 10, 26, 11, 27}))
9827 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9828 if (isShuffleEquivalent(V1, V2, Mask,
9829 {// First 128-bit lane:
9830 4, 20, 5, 21, 6, 22, 7, 23,
9831 // Second 128-bit lane:
9832 12, 28, 13, 29, 14, 30, 15, 31}))
9833 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9835 // Try to use shift instructions.
9837 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9840 // Try to use byte rotation instructions.
9841 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9842 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9845 if (isSingleInputShuffleMask(Mask)) {
9846 // There are no generalized cross-lane shuffle operations available on i16
9848 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9849 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9852 SmallVector<int, 8> RepeatedMask;
9853 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
9854 // As this is a single-input shuffle, the repeated mask should be
9855 // a strictly valid v8i16 mask that we can pass through to the v8i16
9856 // lowering to handle even the v16 case.
9857 return lowerV8I16GeneralSingleInputVectorShuffle(
9858 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
9861 SDValue PSHUFBMask[32];
9862 for (int i = 0; i < 16; ++i) {
9863 if (Mask[i] == -1) {
9864 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9868 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9869 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9870 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
9871 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
9874 ISD::BITCAST, DL, MVT::v16i16,
9876 X86ISD::PSHUFB, DL, MVT::v32i8,
9877 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9878 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9881 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9883 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9884 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9887 // Otherwise fall back on generic lowering.
9888 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9891 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9893 /// This routine is only called when we have AVX2 and thus a reasonable
9894 /// instruction set for v32i8 shuffling..
9895 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9896 const X86Subtarget *Subtarget,
9897 SelectionDAG &DAG) {
9899 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9900 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9901 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9902 ArrayRef<int> Mask = SVOp->getMask();
9903 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9904 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9906 // Whenever we can lower this as a zext, that instruction is strictly faster
9907 // than any alternative. It also allows us to fold memory operands into the
9908 // shuffle in many cases.
9909 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9910 Mask, Subtarget, DAG))
9913 // Check for being able to broadcast a single element.
9914 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9915 Mask, Subtarget, DAG))
9918 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9922 // Use dedicated unpack instructions for masks that match their pattern.
9923 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9925 if (isShuffleEquivalent(
9927 {// First 128-bit lane:
9928 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9929 // Second 128-bit lane:
9930 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9931 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9932 if (isShuffleEquivalent(
9934 {// First 128-bit lane:
9935 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9936 // Second 128-bit lane:
9937 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9938 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9940 // Try to use shift instructions.
9942 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9945 // Try to use byte rotation instructions.
9946 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9947 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9950 if (isSingleInputShuffleMask(Mask)) {
9951 // There are no generalized cross-lane shuffle operations available on i8
9953 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9954 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9957 SDValue PSHUFBMask[32];
9958 for (int i = 0; i < 32; ++i)
9961 ? DAG.getUNDEF(MVT::i8)
9962 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
9966 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9967 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9970 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9972 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9973 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9976 // Otherwise fall back on generic lowering.
9977 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9980 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9982 /// This routine either breaks down the specific type of a 256-bit x86 vector
9983 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9984 /// together based on the available instructions.
9985 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9986 MVT VT, const X86Subtarget *Subtarget,
9987 SelectionDAG &DAG) {
9989 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9990 ArrayRef<int> Mask = SVOp->getMask();
9992 // If we have a single input to the zero element, insert that into V1 if we
9993 // can do so cheaply.
9994 int NumElts = VT.getVectorNumElements();
9995 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
9996 return M >= NumElts;
9999 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10000 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10001 DL, VT, V1, V2, Mask, Subtarget, DAG))
10004 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10005 // check for those subtargets here and avoid much of the subtarget querying in
10006 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10007 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10008 // floating point types there eventually, just immediately cast everything to
10009 // a float and operate entirely in that domain.
10010 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10011 int ElementBits = VT.getScalarSizeInBits();
10012 if (ElementBits < 32)
10013 // No floating point type available, decompose into 128-bit vectors.
10014 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10016 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10017 VT.getVectorNumElements());
10018 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
10019 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
10020 return DAG.getNode(ISD::BITCAST, DL, VT,
10021 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10024 switch (VT.SimpleTy) {
10026 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10028 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10030 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10032 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10034 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10036 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10039 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10043 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10044 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10045 const X86Subtarget *Subtarget,
10046 SelectionDAG &DAG) {
10048 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10049 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10050 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10051 ArrayRef<int> Mask = SVOp->getMask();
10052 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10054 // X86 has dedicated unpack instructions that can handle specific blend
10055 // operations: UNPCKH and UNPCKL.
10056 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10057 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10058 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10059 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10061 // FIXME: Implement direct support for this type!
10062 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10065 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10066 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10067 const X86Subtarget *Subtarget,
10068 SelectionDAG &DAG) {
10070 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10071 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10072 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10073 ArrayRef<int> Mask = SVOp->getMask();
10074 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10076 // Use dedicated unpack instructions for masks that match their pattern.
10077 if (isShuffleEquivalent(V1, V2, Mask,
10078 {// First 128-bit lane.
10079 0, 16, 1, 17, 4, 20, 5, 21,
10080 // Second 128-bit lane.
10081 8, 24, 9, 25, 12, 28, 13, 29}))
10082 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10083 if (isShuffleEquivalent(V1, V2, Mask,
10084 {// First 128-bit lane.
10085 2, 18, 3, 19, 6, 22, 7, 23,
10086 // Second 128-bit lane.
10087 10, 26, 11, 27, 14, 30, 15, 31}))
10088 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10090 // FIXME: Implement direct support for this type!
10091 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10094 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10095 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10096 const X86Subtarget *Subtarget,
10097 SelectionDAG &DAG) {
10099 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10100 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10101 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10102 ArrayRef<int> Mask = SVOp->getMask();
10103 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10105 // X86 has dedicated unpack instructions that can handle specific blend
10106 // operations: UNPCKH and UNPCKL.
10107 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10108 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10109 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10110 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10112 // FIXME: Implement direct support for this type!
10113 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10116 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10117 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10118 const X86Subtarget *Subtarget,
10119 SelectionDAG &DAG) {
10121 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10122 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10123 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10124 ArrayRef<int> Mask = SVOp->getMask();
10125 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10127 // Use dedicated unpack instructions for masks that match their pattern.
10128 if (isShuffleEquivalent(V1, V2, Mask,
10129 {// First 128-bit lane.
10130 0, 16, 1, 17, 4, 20, 5, 21,
10131 // Second 128-bit lane.
10132 8, 24, 9, 25, 12, 28, 13, 29}))
10133 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10134 if (isShuffleEquivalent(V1, V2, Mask,
10135 {// First 128-bit lane.
10136 2, 18, 3, 19, 6, 22, 7, 23,
10137 // Second 128-bit lane.
10138 10, 26, 11, 27, 14, 30, 15, 31}))
10139 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10141 // FIXME: Implement direct support for this type!
10142 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10145 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10146 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10147 const X86Subtarget *Subtarget,
10148 SelectionDAG &DAG) {
10150 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10151 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10152 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10153 ArrayRef<int> Mask = SVOp->getMask();
10154 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10155 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10157 // FIXME: Implement direct support for this type!
10158 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10161 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10162 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10163 const X86Subtarget *Subtarget,
10164 SelectionDAG &DAG) {
10166 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10167 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10168 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10169 ArrayRef<int> Mask = SVOp->getMask();
10170 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10171 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10173 // FIXME: Implement direct support for this type!
10174 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10177 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10179 /// This routine either breaks down the specific type of a 512-bit x86 vector
10180 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10181 /// together based on the available instructions.
10182 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10183 MVT VT, const X86Subtarget *Subtarget,
10184 SelectionDAG &DAG) {
10186 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10187 ArrayRef<int> Mask = SVOp->getMask();
10188 assert(Subtarget->hasAVX512() &&
10189 "Cannot lower 512-bit vectors w/ basic ISA!");
10191 // Check for being able to broadcast a single element.
10192 if (SDValue Broadcast =
10193 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10196 // Dispatch to each element type for lowering. If we don't have supprot for
10197 // specific element type shuffles at 512 bits, immediately split them and
10198 // lower them. Each lowering routine of a given type is allowed to assume that
10199 // the requisite ISA extensions for that element type are available.
10200 switch (VT.SimpleTy) {
10202 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10204 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10206 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10208 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10210 if (Subtarget->hasBWI())
10211 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10214 if (Subtarget->hasBWI())
10215 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10219 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10222 // Otherwise fall back on splitting.
10223 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10226 /// \brief Top-level lowering for x86 vector shuffles.
10228 /// This handles decomposition, canonicalization, and lowering of all x86
10229 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10230 /// above in helper routines. The canonicalization attempts to widen shuffles
10231 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10232 /// s.t. only one of the two inputs needs to be tested, etc.
10233 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10234 SelectionDAG &DAG) {
10235 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10236 ArrayRef<int> Mask = SVOp->getMask();
10237 SDValue V1 = Op.getOperand(0);
10238 SDValue V2 = Op.getOperand(1);
10239 MVT VT = Op.getSimpleValueType();
10240 int NumElements = VT.getVectorNumElements();
10243 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10245 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10246 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10247 if (V1IsUndef && V2IsUndef)
10248 return DAG.getUNDEF(VT);
10250 // When we create a shuffle node we put the UNDEF node to second operand,
10251 // but in some cases the first operand may be transformed to UNDEF.
10252 // In this case we should just commute the node.
10254 return DAG.getCommutedVectorShuffle(*SVOp);
10256 // Check for non-undef masks pointing at an undef vector and make the masks
10257 // undef as well. This makes it easier to match the shuffle based solely on
10261 if (M >= NumElements) {
10262 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10263 for (int &M : NewMask)
10264 if (M >= NumElements)
10266 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10269 // We actually see shuffles that are entirely re-arrangements of a set of
10270 // zero inputs. This mostly happens while decomposing complex shuffles into
10271 // simple ones. Directly lower these as a buildvector of zeros.
10272 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10273 if (Zeroable.all())
10274 return getZeroVector(VT, Subtarget, DAG, dl);
10276 // Try to collapse shuffles into using a vector type with fewer elements but
10277 // wider element types. We cap this to not form integers or floating point
10278 // elements wider than 64 bits, but it might be interesting to form i128
10279 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10280 SmallVector<int, 16> WidenedMask;
10281 if (VT.getScalarSizeInBits() < 64 &&
10282 canWidenShuffleElements(Mask, WidenedMask)) {
10283 MVT NewEltVT = VT.isFloatingPoint()
10284 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10285 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10286 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10287 // Make sure that the new vector type is legal. For example, v2f64 isn't
10289 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10290 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
10291 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
10292 return DAG.getNode(ISD::BITCAST, dl, VT,
10293 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10297 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10298 for (int M : SVOp->getMask())
10300 ++NumUndefElements;
10301 else if (M < NumElements)
10306 // Commute the shuffle as needed such that more elements come from V1 than
10307 // V2. This allows us to match the shuffle pattern strictly on how many
10308 // elements come from V1 without handling the symmetric cases.
10309 if (NumV2Elements > NumV1Elements)
10310 return DAG.getCommutedVectorShuffle(*SVOp);
10312 // When the number of V1 and V2 elements are the same, try to minimize the
10313 // number of uses of V2 in the low half of the vector. When that is tied,
10314 // ensure that the sum of indices for V1 is equal to or lower than the sum
10315 // indices for V2. When those are equal, try to ensure that the number of odd
10316 // indices for V1 is lower than the number of odd indices for V2.
10317 if (NumV1Elements == NumV2Elements) {
10318 int LowV1Elements = 0, LowV2Elements = 0;
10319 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10320 if (M >= NumElements)
10324 if (LowV2Elements > LowV1Elements) {
10325 return DAG.getCommutedVectorShuffle(*SVOp);
10326 } else if (LowV2Elements == LowV1Elements) {
10327 int SumV1Indices = 0, SumV2Indices = 0;
10328 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10329 if (SVOp->getMask()[i] >= NumElements)
10331 else if (SVOp->getMask()[i] >= 0)
10333 if (SumV2Indices < SumV1Indices) {
10334 return DAG.getCommutedVectorShuffle(*SVOp);
10335 } else if (SumV2Indices == SumV1Indices) {
10336 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10337 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10338 if (SVOp->getMask()[i] >= NumElements)
10339 NumV2OddIndices += i % 2;
10340 else if (SVOp->getMask()[i] >= 0)
10341 NumV1OddIndices += i % 2;
10342 if (NumV2OddIndices < NumV1OddIndices)
10343 return DAG.getCommutedVectorShuffle(*SVOp);
10348 // For each vector width, delegate to a specialized lowering routine.
10349 if (VT.getSizeInBits() == 128)
10350 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10352 if (VT.getSizeInBits() == 256)
10353 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10355 // Force AVX-512 vectors to be scalarized for now.
10356 // FIXME: Implement AVX-512 support!
10357 if (VT.getSizeInBits() == 512)
10358 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10360 llvm_unreachable("Unimplemented!");
10363 // This function assumes its argument is a BUILD_VECTOR of constants or
10364 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10366 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10367 unsigned &MaskValue) {
10369 unsigned NumElems = BuildVector->getNumOperands();
10370 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10371 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10372 unsigned NumElemsInLane = NumElems / NumLanes;
10374 // Blend for v16i16 should be symetric for the both lanes.
10375 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10376 SDValue EltCond = BuildVector->getOperand(i);
10377 SDValue SndLaneEltCond =
10378 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10380 int Lane1Cond = -1, Lane2Cond = -1;
10381 if (isa<ConstantSDNode>(EltCond))
10382 Lane1Cond = !isZero(EltCond);
10383 if (isa<ConstantSDNode>(SndLaneEltCond))
10384 Lane2Cond = !isZero(SndLaneEltCond);
10386 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10387 // Lane1Cond != 0, means we want the first argument.
10388 // Lane1Cond == 0, means we want the second argument.
10389 // The encoding of this argument is 0 for the first argument, 1
10390 // for the second. Therefore, invert the condition.
10391 MaskValue |= !Lane1Cond << i;
10392 else if (Lane1Cond < 0)
10393 MaskValue |= !Lane2Cond << i;
10400 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10401 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10402 const X86Subtarget *Subtarget,
10403 SelectionDAG &DAG) {
10404 SDValue Cond = Op.getOperand(0);
10405 SDValue LHS = Op.getOperand(1);
10406 SDValue RHS = Op.getOperand(2);
10408 MVT VT = Op.getSimpleValueType();
10410 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10412 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10414 // Only non-legal VSELECTs reach this lowering, convert those into generic
10415 // shuffles and re-use the shuffle lowering path for blends.
10416 SmallVector<int, 32> Mask;
10417 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10418 SDValue CondElt = CondBV->getOperand(i);
10420 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10422 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10425 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10426 // A vselect where all conditions and data are constants can be optimized into
10427 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10428 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10429 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10430 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10433 // Try to lower this to a blend-style vector shuffle. This can handle all
10434 // constant condition cases.
10435 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10438 // Variable blends are only legal from SSE4.1 onward.
10439 if (!Subtarget->hasSSE41())
10442 // Only some types will be legal on some subtargets. If we can emit a legal
10443 // VSELECT-matching blend, return Op, and but if we need to expand, return
10445 switch (Op.getSimpleValueType().SimpleTy) {
10447 // Most of the vector types have blends past SSE4.1.
10451 // The byte blends for AVX vectors were introduced only in AVX2.
10452 if (Subtarget->hasAVX2())
10459 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10460 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10463 // FIXME: We should custom lower this by fixing the condition and using i8
10469 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10470 MVT VT = Op.getSimpleValueType();
10473 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10476 if (VT.getSizeInBits() == 8) {
10477 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10478 Op.getOperand(0), Op.getOperand(1));
10479 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10480 DAG.getValueType(VT));
10481 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10484 if (VT.getSizeInBits() == 16) {
10485 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10486 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10488 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10489 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10490 DAG.getNode(ISD::BITCAST, dl,
10493 Op.getOperand(1)));
10494 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10495 Op.getOperand(0), Op.getOperand(1));
10496 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10497 DAG.getValueType(VT));
10498 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10501 if (VT == MVT::f32) {
10502 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10503 // the result back to FR32 register. It's only worth matching if the
10504 // result has a single use which is a store or a bitcast to i32. And in
10505 // the case of a store, it's not worth it if the index is a constant 0,
10506 // because a MOVSSmr can be used instead, which is smaller and faster.
10507 if (!Op.hasOneUse())
10509 SDNode *User = *Op.getNode()->use_begin();
10510 if ((User->getOpcode() != ISD::STORE ||
10511 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10512 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10513 (User->getOpcode() != ISD::BITCAST ||
10514 User->getValueType(0) != MVT::i32))
10516 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10517 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10520 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10523 if (VT == MVT::i32 || VT == MVT::i64) {
10524 // ExtractPS/pextrq works with constant index.
10525 if (isa<ConstantSDNode>(Op.getOperand(1)))
10531 /// Extract one bit from mask vector, like v16i1 or v8i1.
10532 /// AVX-512 feature.
10534 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10535 SDValue Vec = Op.getOperand(0);
10537 MVT VecVT = Vec.getSimpleValueType();
10538 SDValue Idx = Op.getOperand(1);
10539 MVT EltVT = Op.getSimpleValueType();
10541 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10542 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10543 "Unexpected vector type in ExtractBitFromMaskVector");
10545 // variable index can't be handled in mask registers,
10546 // extend vector to VR512
10547 if (!isa<ConstantSDNode>(Idx)) {
10548 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10549 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10550 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10551 ExtVT.getVectorElementType(), Ext, Idx);
10552 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10555 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10556 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10557 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10558 rc = getRegClassFor(MVT::v16i1);
10559 unsigned MaxSift = rc->getSize()*8 - 1;
10560 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10561 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10562 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10563 DAG.getConstant(MaxSift, dl, MVT::i8));
10564 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10565 DAG.getIntPtrConstant(0, dl));
10569 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10570 SelectionDAG &DAG) const {
10572 SDValue Vec = Op.getOperand(0);
10573 MVT VecVT = Vec.getSimpleValueType();
10574 SDValue Idx = Op.getOperand(1);
10576 if (Op.getSimpleValueType() == MVT::i1)
10577 return ExtractBitFromMaskVector(Op, DAG);
10579 if (!isa<ConstantSDNode>(Idx)) {
10580 if (VecVT.is512BitVector() ||
10581 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10582 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10585 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10586 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10587 MaskEltVT.getSizeInBits());
10589 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10590 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10591 getZeroVector(MaskVT, Subtarget, DAG, dl),
10592 Idx, DAG.getConstant(0, dl, getPointerTy()));
10593 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10594 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10595 Perm, DAG.getConstant(0, dl, getPointerTy()));
10600 // If this is a 256-bit vector result, first extract the 128-bit vector and
10601 // then extract the element from the 128-bit vector.
10602 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10604 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10605 // Get the 128-bit vector.
10606 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10607 MVT EltVT = VecVT.getVectorElementType();
10609 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10611 //if (IdxVal >= NumElems/2)
10612 // IdxVal -= NumElems/2;
10613 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10614 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10615 DAG.getConstant(IdxVal, dl, MVT::i32));
10618 assert(VecVT.is128BitVector() && "Unexpected vector length");
10620 if (Subtarget->hasSSE41()) {
10621 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10626 MVT VT = Op.getSimpleValueType();
10627 // TODO: handle v16i8.
10628 if (VT.getSizeInBits() == 16) {
10629 SDValue Vec = Op.getOperand(0);
10630 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10632 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10633 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10634 DAG.getNode(ISD::BITCAST, dl,
10636 Op.getOperand(1)));
10637 // Transform it so it match pextrw which produces a 32-bit result.
10638 MVT EltVT = MVT::i32;
10639 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10640 Op.getOperand(0), Op.getOperand(1));
10641 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10642 DAG.getValueType(VT));
10643 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10646 if (VT.getSizeInBits() == 32) {
10647 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10651 // SHUFPS the element to the lowest double word, then movss.
10652 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10653 MVT VVT = Op.getOperand(0).getSimpleValueType();
10654 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10655 DAG.getUNDEF(VVT), Mask);
10656 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10657 DAG.getIntPtrConstant(0, dl));
10660 if (VT.getSizeInBits() == 64) {
10661 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10662 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10663 // to match extract_elt for f64.
10664 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10668 // UNPCKHPD the element to the lowest double word, then movsd.
10669 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10670 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10671 int Mask[2] = { 1, -1 };
10672 MVT VVT = Op.getOperand(0).getSimpleValueType();
10673 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10674 DAG.getUNDEF(VVT), Mask);
10675 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10676 DAG.getIntPtrConstant(0, dl));
10682 /// Insert one bit to mask vector, like v16i1 or v8i1.
10683 /// AVX-512 feature.
10685 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10687 SDValue Vec = Op.getOperand(0);
10688 SDValue Elt = Op.getOperand(1);
10689 SDValue Idx = Op.getOperand(2);
10690 MVT VecVT = Vec.getSimpleValueType();
10692 if (!isa<ConstantSDNode>(Idx)) {
10693 // Non constant index. Extend source and destination,
10694 // insert element and then truncate the result.
10695 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10696 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10697 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10698 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10699 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10700 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10703 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10704 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10706 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10707 DAG.getConstant(IdxVal, dl, MVT::i8));
10708 if (Vec.getOpcode() == ISD::UNDEF)
10710 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10713 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10714 SelectionDAG &DAG) const {
10715 MVT VT = Op.getSimpleValueType();
10716 MVT EltVT = VT.getVectorElementType();
10718 if (EltVT == MVT::i1)
10719 return InsertBitToMaskVector(Op, DAG);
10722 SDValue N0 = Op.getOperand(0);
10723 SDValue N1 = Op.getOperand(1);
10724 SDValue N2 = Op.getOperand(2);
10725 if (!isa<ConstantSDNode>(N2))
10727 auto *N2C = cast<ConstantSDNode>(N2);
10728 unsigned IdxVal = N2C->getZExtValue();
10730 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10731 // into that, and then insert the subvector back into the result.
10732 if (VT.is256BitVector() || VT.is512BitVector()) {
10733 // With a 256-bit vector, we can insert into the zero element efficiently
10734 // using a blend if we have AVX or AVX2 and the right data type.
10735 if (VT.is256BitVector() && IdxVal == 0) {
10736 // TODO: It is worthwhile to cast integer to floating point and back
10737 // and incur a domain crossing penalty if that's what we'll end up
10738 // doing anyway after extracting to a 128-bit vector.
10739 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
10740 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
10741 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
10742 N2 = DAG.getIntPtrConstant(1, dl);
10743 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
10747 // Get the desired 128-bit vector chunk.
10748 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10750 // Insert the element into the desired chunk.
10751 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10752 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10754 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10755 DAG.getConstant(IdxIn128, dl, MVT::i32));
10757 // Insert the changed part back into the bigger vector
10758 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10760 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10762 if (Subtarget->hasSSE41()) {
10763 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10765 if (VT == MVT::v8i16) {
10766 Opc = X86ISD::PINSRW;
10768 assert(VT == MVT::v16i8);
10769 Opc = X86ISD::PINSRB;
10772 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10774 if (N1.getValueType() != MVT::i32)
10775 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10776 if (N2.getValueType() != MVT::i32)
10777 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10778 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10781 if (EltVT == MVT::f32) {
10782 // Bits [7:6] of the constant are the source select. This will always be
10783 // zero here. The DAG Combiner may combine an extract_elt index into
10784 // these bits. For example (insert (extract, 3), 2) could be matched by
10785 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
10786 // Bits [5:4] of the constant are the destination select. This is the
10787 // value of the incoming immediate.
10788 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10789 // combine either bitwise AND or insert of float 0.0 to set these bits.
10791 const Function *F = DAG.getMachineFunction().getFunction();
10792 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
10793 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
10794 // If this is an insertion of 32-bits into the low 32-bits of
10795 // a vector, we prefer to generate a blend with immediate rather
10796 // than an insertps. Blends are simpler operations in hardware and so
10797 // will always have equal or better performance than insertps.
10798 // But if optimizing for size and there's a load folding opportunity,
10799 // generate insertps because blendps does not have a 32-bit memory
10801 N2 = DAG.getIntPtrConstant(1, dl);
10802 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10803 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
10805 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
10806 // Create this as a scalar to vector..
10807 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10808 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10811 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10812 // PINSR* works with constant index.
10817 if (EltVT == MVT::i8)
10820 if (EltVT.getSizeInBits() == 16) {
10821 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10822 // as its second argument.
10823 if (N1.getValueType() != MVT::i32)
10824 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10825 if (N2.getValueType() != MVT::i32)
10826 N2 = DAG.getIntPtrConstant(IdxVal, dl);
10827 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10832 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10834 MVT OpVT = Op.getSimpleValueType();
10836 // If this is a 256-bit vector result, first insert into a 128-bit
10837 // vector and then insert into the 256-bit vector.
10838 if (!OpVT.is128BitVector()) {
10839 // Insert into a 128-bit vector.
10840 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10841 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10842 OpVT.getVectorNumElements() / SizeFactor);
10844 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10846 // Insert the 128-bit vector.
10847 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10850 if (OpVT == MVT::v1i64 &&
10851 Op.getOperand(0).getValueType() == MVT::i64)
10852 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10854 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10855 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10856 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10857 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10860 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10861 // a simple subregister reference or explicit instructions to grab
10862 // upper bits of a vector.
10863 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10864 SelectionDAG &DAG) {
10866 SDValue In = Op.getOperand(0);
10867 SDValue Idx = Op.getOperand(1);
10868 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10869 MVT ResVT = Op.getSimpleValueType();
10870 MVT InVT = In.getSimpleValueType();
10872 if (Subtarget->hasFp256()) {
10873 if (ResVT.is128BitVector() &&
10874 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10875 isa<ConstantSDNode>(Idx)) {
10876 return Extract128BitVector(In, IdxVal, DAG, dl);
10878 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10879 isa<ConstantSDNode>(Idx)) {
10880 return Extract256BitVector(In, IdxVal, DAG, dl);
10886 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10887 // simple superregister reference or explicit instructions to insert
10888 // the upper bits of a vector.
10889 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10890 SelectionDAG &DAG) {
10891 if (!Subtarget->hasAVX())
10895 SDValue Vec = Op.getOperand(0);
10896 SDValue SubVec = Op.getOperand(1);
10897 SDValue Idx = Op.getOperand(2);
10899 if (!isa<ConstantSDNode>(Idx))
10902 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10903 MVT OpVT = Op.getSimpleValueType();
10904 MVT SubVecVT = SubVec.getSimpleValueType();
10906 // Fold two 16-byte subvector loads into one 32-byte load:
10907 // (insert_subvector (insert_subvector undef, (load addr), 0),
10908 // (load addr + 16), Elts/2)
10910 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10911 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10912 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10913 !Subtarget->isUnalignedMem32Slow()) {
10914 SDValue SubVec2 = Vec.getOperand(1);
10915 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10916 if (Idx2->getZExtValue() == 0) {
10917 SDValue Ops[] = { SubVec2, SubVec };
10918 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10925 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10926 SubVecVT.is128BitVector())
10927 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10929 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10930 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10932 if (OpVT.getVectorElementType() == MVT::i1) {
10933 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
10935 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
10936 SDValue Undef = DAG.getUNDEF(OpVT);
10937 unsigned NumElems = OpVT.getVectorNumElements();
10938 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
10940 if (IdxVal == OpVT.getVectorNumElements() / 2) {
10941 // Zero upper bits of the Vec
10942 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10943 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10945 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10947 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10948 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10951 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
10953 // Zero upper bits of the Vec2
10954 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
10955 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
10956 // Zero lower bits of the Vec
10957 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
10958 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
10959 // Merge them together
10960 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
10966 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10967 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10968 // one of the above mentioned nodes. It has to be wrapped because otherwise
10969 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10970 // be used to form addressing mode. These wrapped nodes will be selected
10973 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10974 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10976 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10977 // global base reg.
10978 unsigned char OpFlag = 0;
10979 unsigned WrapperKind = X86ISD::Wrapper;
10980 CodeModel::Model M = DAG.getTarget().getCodeModel();
10982 if (Subtarget->isPICStyleRIPRel() &&
10983 (M == CodeModel::Small || M == CodeModel::Kernel))
10984 WrapperKind = X86ISD::WrapperRIP;
10985 else if (Subtarget->isPICStyleGOT())
10986 OpFlag = X86II::MO_GOTOFF;
10987 else if (Subtarget->isPICStyleStubPIC())
10988 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10990 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10991 CP->getAlignment(),
10992 CP->getOffset(), OpFlag);
10994 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10995 // With PIC, the address is actually $g + Offset.
10997 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10998 DAG.getNode(X86ISD::GlobalBaseReg,
10999 SDLoc(), getPointerTy()),
11006 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11007 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11009 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11010 // global base reg.
11011 unsigned char OpFlag = 0;
11012 unsigned WrapperKind = X86ISD::Wrapper;
11013 CodeModel::Model M = DAG.getTarget().getCodeModel();
11015 if (Subtarget->isPICStyleRIPRel() &&
11016 (M == CodeModel::Small || M == CodeModel::Kernel))
11017 WrapperKind = X86ISD::WrapperRIP;
11018 else if (Subtarget->isPICStyleGOT())
11019 OpFlag = X86II::MO_GOTOFF;
11020 else if (Subtarget->isPICStyleStubPIC())
11021 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11023 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
11026 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11028 // With PIC, the address is actually $g + Offset.
11030 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11031 DAG.getNode(X86ISD::GlobalBaseReg,
11032 SDLoc(), getPointerTy()),
11039 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11040 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11042 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11043 // global base reg.
11044 unsigned char OpFlag = 0;
11045 unsigned WrapperKind = X86ISD::Wrapper;
11046 CodeModel::Model M = DAG.getTarget().getCodeModel();
11048 if (Subtarget->isPICStyleRIPRel() &&
11049 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11050 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11051 OpFlag = X86II::MO_GOTPCREL;
11052 WrapperKind = X86ISD::WrapperRIP;
11053 } else if (Subtarget->isPICStyleGOT()) {
11054 OpFlag = X86II::MO_GOT;
11055 } else if (Subtarget->isPICStyleStubPIC()) {
11056 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11057 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11058 OpFlag = X86II::MO_DARWIN_NONLAZY;
11061 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
11064 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11066 // With PIC, the address is actually $g + Offset.
11067 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11068 !Subtarget->is64Bit()) {
11069 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11070 DAG.getNode(X86ISD::GlobalBaseReg,
11071 SDLoc(), getPointerTy()),
11075 // For symbols that require a load from a stub to get the address, emit the
11077 if (isGlobalStubReference(OpFlag))
11078 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
11079 MachinePointerInfo::getGOT(), false, false, false, 0);
11085 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11086 // Create the TargetBlockAddressAddress node.
11087 unsigned char OpFlags =
11088 Subtarget->ClassifyBlockAddressReference();
11089 CodeModel::Model M = DAG.getTarget().getCodeModel();
11090 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11091 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11093 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
11096 if (Subtarget->isPICStyleRIPRel() &&
11097 (M == CodeModel::Small || M == CodeModel::Kernel))
11098 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11100 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11102 // With PIC, the address is actually $g + Offset.
11103 if (isGlobalRelativeToPICBase(OpFlags)) {
11104 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11105 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11113 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11114 int64_t Offset, SelectionDAG &DAG) const {
11115 // Create the TargetGlobalAddress node, folding in the constant
11116 // offset if it is legal.
11117 unsigned char OpFlags =
11118 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11119 CodeModel::Model M = DAG.getTarget().getCodeModel();
11121 if (OpFlags == X86II::MO_NO_FLAG &&
11122 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11123 // A direct static reference to a global.
11124 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
11127 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
11130 if (Subtarget->isPICStyleRIPRel() &&
11131 (M == CodeModel::Small || M == CodeModel::Kernel))
11132 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
11134 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
11136 // With PIC, the address is actually $g + Offset.
11137 if (isGlobalRelativeToPICBase(OpFlags)) {
11138 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
11139 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
11143 // For globals that require a load from a stub to get the address, emit the
11145 if (isGlobalStubReference(OpFlags))
11146 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
11147 MachinePointerInfo::getGOT(), false, false, false, 0);
11149 // If there was a non-zero offset that we didn't fold, create an explicit
11150 // addition for it.
11152 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
11153 DAG.getConstant(Offset, dl, getPointerTy()));
11159 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11160 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11161 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11162 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11166 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11167 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11168 unsigned char OperandFlags, bool LocalDynamic = false) {
11169 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11170 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11172 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11173 GA->getValueType(0),
11177 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11181 SDValue Ops[] = { Chain, TGA, *InFlag };
11182 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11184 SDValue Ops[] = { Chain, TGA };
11185 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11188 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11189 MFI->setAdjustsStack(true);
11190 MFI->setHasCalls(true);
11192 SDValue Flag = Chain.getValue(1);
11193 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11196 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11198 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11201 SDLoc dl(GA); // ? function entry point might be better
11202 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11203 DAG.getNode(X86ISD::GlobalBaseReg,
11204 SDLoc(), PtrVT), InFlag);
11205 InFlag = Chain.getValue(1);
11207 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11210 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11212 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11214 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11215 X86::RAX, X86II::MO_TLSGD);
11218 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11224 // Get the start address of the TLS block for this module.
11225 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11226 .getInfo<X86MachineFunctionInfo>();
11227 MFI->incNumLocalDynamicTLSAccesses();
11231 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11232 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11235 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11236 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11237 InFlag = Chain.getValue(1);
11238 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11239 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11242 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11246 unsigned char OperandFlags = X86II::MO_DTPOFF;
11247 unsigned WrapperKind = X86ISD::Wrapper;
11248 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11249 GA->getValueType(0),
11250 GA->getOffset(), OperandFlags);
11251 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11253 // Add x@dtpoff with the base.
11254 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11257 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11258 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11259 const EVT PtrVT, TLSModel::Model model,
11260 bool is64Bit, bool isPIC) {
11263 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11264 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11265 is64Bit ? 257 : 256));
11267 SDValue ThreadPointer =
11268 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11269 MachinePointerInfo(Ptr), false, false, false, 0);
11271 unsigned char OperandFlags = 0;
11272 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11274 unsigned WrapperKind = X86ISD::Wrapper;
11275 if (model == TLSModel::LocalExec) {
11276 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11277 } else if (model == TLSModel::InitialExec) {
11279 OperandFlags = X86II::MO_GOTTPOFF;
11280 WrapperKind = X86ISD::WrapperRIP;
11282 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11285 llvm_unreachable("Unexpected model");
11288 // emit "addl x@ntpoff,%eax" (local exec)
11289 // or "addl x@indntpoff,%eax" (initial exec)
11290 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11292 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11293 GA->getOffset(), OperandFlags);
11294 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11296 if (model == TLSModel::InitialExec) {
11297 if (isPIC && !is64Bit) {
11298 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11299 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11303 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11304 MachinePointerInfo::getGOT(), false, false, false, 0);
11307 // The address of the thread local variable is the add of the thread
11308 // pointer with the offset of the variable.
11309 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11313 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11315 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11316 const GlobalValue *GV = GA->getGlobal();
11318 if (Subtarget->isTargetELF()) {
11319 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11321 case TLSModel::GeneralDynamic:
11322 if (Subtarget->is64Bit())
11323 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
11324 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
11325 case TLSModel::LocalDynamic:
11326 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
11327 Subtarget->is64Bit());
11328 case TLSModel::InitialExec:
11329 case TLSModel::LocalExec:
11330 return LowerToTLSExecModel(
11331 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
11332 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
11334 llvm_unreachable("Unknown TLS model.");
11337 if (Subtarget->isTargetDarwin()) {
11338 // Darwin only has one model of TLS. Lower to that.
11339 unsigned char OpFlag = 0;
11340 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11341 X86ISD::WrapperRIP : X86ISD::Wrapper;
11343 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11344 // global base reg.
11345 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11346 !Subtarget->is64Bit();
11348 OpFlag = X86II::MO_TLVP_PIC_BASE;
11350 OpFlag = X86II::MO_TLVP;
11352 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11353 GA->getValueType(0),
11354 GA->getOffset(), OpFlag);
11355 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
11357 // With PIC32, the address is actually $g + Offset.
11359 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
11360 DAG.getNode(X86ISD::GlobalBaseReg,
11361 SDLoc(), getPointerTy()),
11364 // Lowering the machine isd will make sure everything is in the right
11366 SDValue Chain = DAG.getEntryNode();
11367 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11368 SDValue Args[] = { Chain, Offset };
11369 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11371 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11372 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11373 MFI->setAdjustsStack(true);
11375 // And our return value (tls address) is in the standard call return value
11377 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11378 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11379 Chain.getValue(1));
11382 if (Subtarget->isTargetKnownWindowsMSVC() ||
11383 Subtarget->isTargetWindowsGNU()) {
11384 // Just use the implicit TLS architecture
11385 // Need to generate someting similar to:
11386 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11388 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11389 // mov rcx, qword [rdx+rcx*8]
11390 // mov eax, .tls$:tlsvar
11391 // [rax+rcx] contains the address
11392 // Windows 64bit: gs:0x58
11393 // Windows 32bit: fs:__tls_array
11396 SDValue Chain = DAG.getEntryNode();
11398 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11399 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11400 // use its literal value of 0x2C.
11401 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11402 ? Type::getInt8PtrTy(*DAG.getContext(),
11404 : Type::getInt32PtrTy(*DAG.getContext(),
11408 Subtarget->is64Bit()
11409 ? DAG.getIntPtrConstant(0x58, dl)
11410 : (Subtarget->isTargetWindowsGNU()
11411 ? DAG.getIntPtrConstant(0x2C, dl)
11412 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11414 SDValue ThreadPointer =
11415 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11416 MachinePointerInfo(Ptr), false, false, false, 0);
11419 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11420 res = ThreadPointer;
11422 // Load the _tls_index variable
11423 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11424 if (Subtarget->is64Bit())
11425 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain, IDX,
11426 MachinePointerInfo(), MVT::i32, false, false,
11429 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11430 false, false, false, 0);
11432 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()), dl,
11434 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11436 res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11439 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11440 false, false, false, 0);
11442 // Get the offset of start of .tls section
11443 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11444 GA->getValueType(0),
11445 GA->getOffset(), X86II::MO_SECREL);
11446 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11448 // The address of the thread local variable is the add of the thread
11449 // pointer with the offset of the variable.
11450 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11453 llvm_unreachable("TLS not implemented for this target.");
11456 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11457 /// and take a 2 x i32 value to shift plus a shift amount.
11458 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11459 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11460 MVT VT = Op.getSimpleValueType();
11461 unsigned VTBits = VT.getSizeInBits();
11463 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11464 SDValue ShOpLo = Op.getOperand(0);
11465 SDValue ShOpHi = Op.getOperand(1);
11466 SDValue ShAmt = Op.getOperand(2);
11467 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11468 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11470 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11471 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11472 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11473 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11474 : DAG.getConstant(0, dl, VT);
11476 SDValue Tmp2, Tmp3;
11477 if (Op.getOpcode() == ISD::SHL_PARTS) {
11478 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11479 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11481 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11482 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11485 // If the shift amount is larger or equal than the width of a part we can't
11486 // rely on the results of shld/shrd. Insert a test and select the appropriate
11487 // values for large shift amounts.
11488 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11489 DAG.getConstant(VTBits, dl, MVT::i8));
11490 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11491 AndNode, DAG.getConstant(0, dl, MVT::i8));
11494 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11495 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11496 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11498 if (Op.getOpcode() == ISD::SHL_PARTS) {
11499 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11500 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11502 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11503 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11506 SDValue Ops[2] = { Lo, Hi };
11507 return DAG.getMergeValues(Ops, dl);
11510 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11511 SelectionDAG &DAG) const {
11512 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11515 if (SrcVT.isVector()) {
11516 if (SrcVT.getVectorElementType() == MVT::i1) {
11517 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11518 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11519 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11520 Op.getOperand(0)));
11525 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11526 "Unknown SINT_TO_FP to lower!");
11528 // These are really Legal; return the operand so the caller accepts it as
11530 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11532 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11533 Subtarget->is64Bit()) {
11537 unsigned Size = SrcVT.getSizeInBits()/8;
11538 MachineFunction &MF = DAG.getMachineFunction();
11539 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11540 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11541 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11543 MachinePointerInfo::getFixedStack(SSFI),
11545 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11548 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11550 SelectionDAG &DAG) const {
11554 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11556 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11558 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11560 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11562 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11563 MachineMemOperand *MMO;
11565 int SSFI = FI->getIndex();
11567 DAG.getMachineFunction()
11568 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11569 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11571 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11572 StackSlot = StackSlot.getOperand(1);
11574 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11575 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11577 Tys, Ops, SrcVT, MMO);
11580 Chain = Result.getValue(1);
11581 SDValue InFlag = Result.getValue(2);
11583 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11584 // shouldn't be necessary except that RFP cannot be live across
11585 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11586 MachineFunction &MF = DAG.getMachineFunction();
11587 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11588 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11589 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11590 Tys = DAG.getVTList(MVT::Other);
11592 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11594 MachineMemOperand *MMO =
11595 DAG.getMachineFunction()
11596 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11597 MachineMemOperand::MOStore, SSFISize, SSFISize);
11599 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11600 Ops, Op.getValueType(), MMO);
11601 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11602 MachinePointerInfo::getFixedStack(SSFI),
11603 false, false, false, 0);
11609 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11610 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11611 SelectionDAG &DAG) const {
11612 // This algorithm is not obvious. Here it is what we're trying to output:
11615 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11616 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11618 haddpd %xmm0, %xmm0
11620 pshufd $0x4e, %xmm0, %xmm1
11626 LLVMContext *Context = DAG.getContext();
11628 // Build some magic constants.
11629 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11630 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11631 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11633 SmallVector<Constant*,2> CV1;
11635 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11636 APInt(64, 0x4330000000000000ULL))));
11638 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11639 APInt(64, 0x4530000000000000ULL))));
11640 Constant *C1 = ConstantVector::get(CV1);
11641 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11643 // Load the 64-bit value into an XMM register.
11644 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11646 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11647 MachinePointerInfo::getConstantPool(),
11648 false, false, false, 16);
11649 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11650 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11653 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11654 MachinePointerInfo::getConstantPool(),
11655 false, false, false, 16);
11656 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11657 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11660 if (Subtarget->hasSSE3()) {
11661 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11662 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11664 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11665 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11667 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11668 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11672 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11673 DAG.getIntPtrConstant(0, dl));
11676 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11677 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11678 SelectionDAG &DAG) const {
11680 // FP constant to bias correct the final result.
11681 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
11684 // Load the 32-bit value into an XMM register.
11685 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11688 // Zero out the upper parts of the register.
11689 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11691 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11692 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11693 DAG.getIntPtrConstant(0, dl));
11695 // Or the load with the bias.
11696 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11697 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11698 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11699 MVT::v2f64, Load)),
11700 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11701 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11702 MVT::v2f64, Bias)));
11703 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11704 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11705 DAG.getIntPtrConstant(0, dl));
11707 // Subtract the bias.
11708 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11710 // Handle final rounding.
11711 EVT DestVT = Op.getValueType();
11713 if (DestVT.bitsLT(MVT::f64))
11714 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11715 DAG.getIntPtrConstant(0, dl));
11716 if (DestVT.bitsGT(MVT::f64))
11717 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11719 // Handle final rounding.
11723 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11724 const X86Subtarget &Subtarget) {
11725 // The algorithm is the following:
11726 // #ifdef __SSE4_1__
11727 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11728 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11729 // (uint4) 0x53000000, 0xaa);
11731 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11732 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11734 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11735 // return (float4) lo + fhi;
11738 SDValue V = Op->getOperand(0);
11739 EVT VecIntVT = V.getValueType();
11740 bool Is128 = VecIntVT == MVT::v4i32;
11741 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11742 // If we convert to something else than the supported type, e.g., to v4f64,
11744 if (VecFloatVT != Op->getValueType(0))
11747 unsigned NumElts = VecIntVT.getVectorNumElements();
11748 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11749 "Unsupported custom type");
11750 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11752 // In the #idef/#else code, we have in common:
11753 // - The vector of constants:
11759 // Create the splat vector for 0x4b000000.
11760 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
11761 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11762 CstLow, CstLow, CstLow, CstLow};
11763 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11764 makeArrayRef(&CstLowArray[0], NumElts));
11765 // Create the splat vector for 0x53000000.
11766 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
11767 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11768 CstHigh, CstHigh, CstHigh, CstHigh};
11769 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11770 makeArrayRef(&CstHighArray[0], NumElts));
11772 // Create the right shift.
11773 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
11774 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11775 CstShift, CstShift, CstShift, CstShift};
11776 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11777 makeArrayRef(&CstShiftArray[0], NumElts));
11778 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11781 if (Subtarget.hasSSE41()) {
11782 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11783 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11784 SDValue VecCstLowBitcast =
11785 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11786 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11787 // Low will be bitcasted right away, so do not bother bitcasting back to its
11789 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11790 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11791 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11792 // (uint4) 0x53000000, 0xaa);
11793 SDValue VecCstHighBitcast =
11794 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11795 SDValue VecShiftBitcast =
11796 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11797 // High will be bitcasted right away, so do not bother bitcasting back to
11798 // its original type.
11799 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11800 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
11802 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
11803 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11804 CstMask, CstMask, CstMask);
11805 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11806 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11807 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11809 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11810 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11813 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11814 SDValue CstFAdd = DAG.getConstantFP(
11815 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
11816 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11817 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11818 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11819 makeArrayRef(&CstFAddArray[0], NumElts));
11821 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11822 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11824 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11825 // return (float4) lo + fhi;
11826 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11827 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11830 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11831 SelectionDAG &DAG) const {
11832 SDValue N0 = Op.getOperand(0);
11833 MVT SVT = N0.getSimpleValueType();
11836 switch (SVT.SimpleTy) {
11838 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11843 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11844 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11845 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11849 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11852 if (Subtarget->hasAVX512())
11853 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
11854 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
11856 llvm_unreachable(nullptr);
11859 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11860 SelectionDAG &DAG) const {
11861 SDValue N0 = Op.getOperand(0);
11864 if (Op.getValueType().isVector())
11865 return lowerUINT_TO_FP_vec(Op, DAG);
11867 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11868 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11869 // the optimization here.
11870 if (DAG.SignBitIsZero(N0))
11871 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11873 MVT SrcVT = N0.getSimpleValueType();
11874 MVT DstVT = Op.getSimpleValueType();
11875 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11876 return LowerUINT_TO_FP_i64(Op, DAG);
11877 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11878 return LowerUINT_TO_FP_i32(Op, DAG);
11879 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11882 // Make a 64-bit buffer, and use it to build an FILD.
11883 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11884 if (SrcVT == MVT::i32) {
11885 SDValue WordOff = DAG.getConstant(4, dl, getPointerTy());
11886 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11887 getPointerTy(), StackSlot, WordOff);
11888 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11889 StackSlot, MachinePointerInfo(),
11891 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
11892 OffsetSlot, MachinePointerInfo(),
11894 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11898 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11899 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11900 StackSlot, MachinePointerInfo(),
11902 // For i64 source, we need to add the appropriate power of 2 if the input
11903 // was negative. This is the same as the optimization in
11904 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11905 // we must be careful to do the computation in x87 extended precision, not
11906 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11907 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11908 MachineMemOperand *MMO =
11909 DAG.getMachineFunction()
11910 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11911 MachineMemOperand::MOLoad, 8, 8);
11913 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11914 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11915 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11918 APInt FF(32, 0x5F800000ULL);
11920 // Check whether the sign bit is set.
11921 SDValue SignSet = DAG.getSetCC(dl,
11922 getSetCCResultType(*DAG.getContext(), MVT::i64),
11924 DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
11926 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11927 SDValue FudgePtr = DAG.getConstantPool(
11928 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11931 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11932 SDValue Zero = DAG.getIntPtrConstant(0, dl);
11933 SDValue Four = DAG.getIntPtrConstant(4, dl);
11934 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11936 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11938 // Load the value out, extending it from f32 to f80.
11939 // FIXME: Avoid the extend by constructing the right constant pool?
11940 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11941 FudgePtr, MachinePointerInfo::getConstantPool(),
11942 MVT::f32, false, false, false, 4);
11943 // Extend everything to 80 bits to force it to be done on x87.
11944 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11945 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
11946 DAG.getIntPtrConstant(0, dl));
11949 std::pair<SDValue,SDValue>
11950 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11951 bool IsSigned, bool IsReplace) const {
11954 EVT DstTy = Op.getValueType();
11956 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11957 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11961 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11962 DstTy.getSimpleVT() >= MVT::i16 &&
11963 "Unknown FP_TO_INT to lower!");
11965 // These are really Legal.
11966 if (DstTy == MVT::i32 &&
11967 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11968 return std::make_pair(SDValue(), SDValue());
11969 if (Subtarget->is64Bit() &&
11970 DstTy == MVT::i64 &&
11971 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11972 return std::make_pair(SDValue(), SDValue());
11974 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11975 // stack slot, or into the FTOL runtime function.
11976 MachineFunction &MF = DAG.getMachineFunction();
11977 unsigned MemSize = DstTy.getSizeInBits()/8;
11978 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11979 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11982 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11983 Opc = X86ISD::WIN_FTOL;
11985 switch (DstTy.getSimpleVT().SimpleTy) {
11986 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11987 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11988 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11989 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11992 SDValue Chain = DAG.getEntryNode();
11993 SDValue Value = Op.getOperand(0);
11994 EVT TheVT = Op.getOperand(0).getValueType();
11995 // FIXME This causes a redundant load/store if the SSE-class value is already
11996 // in memory, such as if it is on the callstack.
11997 if (isScalarFPTypeInSSEReg(TheVT)) {
11998 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11999 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12000 MachinePointerInfo::getFixedStack(SSFI),
12002 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12004 Chain, StackSlot, DAG.getValueType(TheVT)
12007 MachineMemOperand *MMO =
12008 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12009 MachineMemOperand::MOLoad, MemSize, MemSize);
12010 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12011 Chain = Value.getValue(1);
12012 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12013 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
12016 MachineMemOperand *MMO =
12017 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12018 MachineMemOperand::MOStore, MemSize, MemSize);
12020 if (Opc != X86ISD::WIN_FTOL) {
12021 // Build the FP_TO_INT*_IN_MEM
12022 SDValue Ops[] = { Chain, Value, StackSlot };
12023 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12025 return std::make_pair(FIST, StackSlot);
12027 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12028 DAG.getVTList(MVT::Other, MVT::Glue),
12030 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12031 MVT::i32, ftol.getValue(1));
12032 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12033 MVT::i32, eax.getValue(2));
12034 SDValue Ops[] = { eax, edx };
12035 SDValue pair = IsReplace
12036 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12037 : DAG.getMergeValues(Ops, DL);
12038 return std::make_pair(pair, SDValue());
12042 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12043 const X86Subtarget *Subtarget) {
12044 MVT VT = Op->getSimpleValueType(0);
12045 SDValue In = Op->getOperand(0);
12046 MVT InVT = In.getSimpleValueType();
12049 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12050 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12052 // Optimize vectors in AVX mode:
12055 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12056 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12057 // Concat upper and lower parts.
12060 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12061 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12062 // Concat upper and lower parts.
12065 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12066 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12067 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12070 if (Subtarget->hasInt256())
12071 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12073 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12074 SDValue Undef = DAG.getUNDEF(InVT);
12075 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12076 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12077 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12079 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12080 VT.getVectorNumElements()/2);
12082 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
12083 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
12085 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12088 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12089 SelectionDAG &DAG) {
12090 MVT VT = Op->getSimpleValueType(0);
12091 SDValue In = Op->getOperand(0);
12092 MVT InVT = In.getSimpleValueType();
12094 unsigned int NumElts = VT.getVectorNumElements();
12095 if (NumElts != 8 && NumElts != 16)
12098 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12099 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12101 assert(InVT.getVectorElementType() == MVT::i1);
12102 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12104 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12106 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12108 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12109 if (VT.is512BitVector())
12111 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12114 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12115 SelectionDAG &DAG) {
12116 if (Subtarget->hasFp256()) {
12117 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12125 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12126 SelectionDAG &DAG) {
12128 MVT VT = Op.getSimpleValueType();
12129 SDValue In = Op.getOperand(0);
12130 MVT SVT = In.getSimpleValueType();
12132 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12133 return LowerZERO_EXTEND_AVX512(Op, DAG);
12135 if (Subtarget->hasFp256()) {
12136 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
12141 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12142 VT.getVectorNumElements() != SVT.getVectorNumElements());
12146 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12148 MVT VT = Op.getSimpleValueType();
12149 SDValue In = Op.getOperand(0);
12150 MVT InVT = In.getSimpleValueType();
12152 if (VT == MVT::i1) {
12153 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12154 "Invalid scalar TRUNCATE operation");
12155 if (InVT.getSizeInBits() >= 32)
12157 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12158 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12160 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12161 "Invalid TRUNCATE operation");
12163 // move vector to mask - truncate solution for SKX
12164 if (VT.getVectorElementType() == MVT::i1) {
12165 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12166 Subtarget->hasBWI())
12167 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12168 if ((InVT.is256BitVector() || InVT.is128BitVector())
12169 && InVT.getScalarSizeInBits() <= 16 &&
12170 Subtarget->hasBWI() && Subtarget->hasVLX())
12171 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12172 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12173 Subtarget->hasDQI())
12174 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12175 if ((InVT.is256BitVector() || InVT.is128BitVector())
12176 && InVT.getScalarSizeInBits() >= 32 &&
12177 Subtarget->hasDQI() && Subtarget->hasVLX())
12178 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12180 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
12181 if (VT.getVectorElementType().getSizeInBits() >=8)
12182 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12184 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12185 unsigned NumElts = InVT.getVectorNumElements();
12186 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12187 if (InVT.getSizeInBits() < 512) {
12188 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12189 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12194 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12195 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12196 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12199 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12200 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12201 if (Subtarget->hasInt256()) {
12202 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12203 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
12204 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12206 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12207 DAG.getIntPtrConstant(0, DL));
12210 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12211 DAG.getIntPtrConstant(0, DL));
12212 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12213 DAG.getIntPtrConstant(2, DL));
12214 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12215 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12216 static const int ShufMask[] = {0, 2, 4, 6};
12217 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12220 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12221 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12222 if (Subtarget->hasInt256()) {
12223 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
12225 SmallVector<SDValue,32> pshufbMask;
12226 for (unsigned i = 0; i < 2; ++i) {
12227 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12228 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12229 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12230 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12231 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12232 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12233 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12234 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12235 for (unsigned j = 0; j < 8; ++j)
12236 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12238 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12239 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12240 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
12242 static const int ShufMask[] = {0, 2, -1, -1};
12243 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12245 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12246 DAG.getIntPtrConstant(0, DL));
12247 return DAG.getNode(ISD::BITCAST, DL, VT, In);
12250 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12251 DAG.getIntPtrConstant(0, DL));
12253 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12254 DAG.getIntPtrConstant(4, DL));
12256 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
12257 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
12259 // The PSHUFB mask:
12260 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12261 -1, -1, -1, -1, -1, -1, -1, -1};
12263 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12264 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12265 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12267 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
12268 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
12270 // The MOVLHPS Mask:
12271 static const int ShufMask2[] = {0, 1, 4, 5};
12272 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12273 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
12276 // Handle truncation of V256 to V128 using shuffles.
12277 if (!VT.is128BitVector() || !InVT.is256BitVector())
12280 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12282 unsigned NumElems = VT.getVectorNumElements();
12283 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12285 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12286 // Prepare truncation shuffle mask
12287 for (unsigned i = 0; i != NumElems; ++i)
12288 MaskVec[i] = i * 2;
12289 SDValue V = DAG.getVectorShuffle(NVT, DL,
12290 DAG.getNode(ISD::BITCAST, DL, NVT, In),
12291 DAG.getUNDEF(NVT), &MaskVec[0]);
12292 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12293 DAG.getIntPtrConstant(0, DL));
12296 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12297 SelectionDAG &DAG) const {
12298 assert(!Op.getSimpleValueType().isVector());
12300 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12301 /*IsSigned=*/ true, /*IsReplace=*/ false);
12302 SDValue FIST = Vals.first, StackSlot = Vals.second;
12303 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12304 if (!FIST.getNode()) return Op;
12306 if (StackSlot.getNode())
12307 // Load the result.
12308 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12309 FIST, StackSlot, MachinePointerInfo(),
12310 false, false, false, 0);
12312 // The node is the result.
12316 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12317 SelectionDAG &DAG) const {
12318 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12319 /*IsSigned=*/ false, /*IsReplace=*/ false);
12320 SDValue FIST = Vals.first, StackSlot = Vals.second;
12321 assert(FIST.getNode() && "Unexpected failure");
12323 if (StackSlot.getNode())
12324 // Load the result.
12325 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12326 FIST, StackSlot, MachinePointerInfo(),
12327 false, false, false, 0);
12329 // The node is the result.
12333 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12335 MVT VT = Op.getSimpleValueType();
12336 SDValue In = Op.getOperand(0);
12337 MVT SVT = In.getSimpleValueType();
12339 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12341 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12342 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12343 In, DAG.getUNDEF(SVT)));
12346 /// The only differences between FABS and FNEG are the mask and the logic op.
12347 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12348 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12349 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12350 "Wrong opcode for lowering FABS or FNEG.");
12352 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12354 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12355 // into an FNABS. We'll lower the FABS after that if it is still in use.
12357 for (SDNode *User : Op->uses())
12358 if (User->getOpcode() == ISD::FNEG)
12361 SDValue Op0 = Op.getOperand(0);
12362 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12365 MVT VT = Op.getSimpleValueType();
12366 // Assume scalar op for initialization; update for vector if needed.
12367 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
12368 // generate a 16-byte vector constant and logic op even for the scalar case.
12369 // Using a 16-byte mask allows folding the load of the mask with
12370 // the logic op, so it can save (~4 bytes) on code size.
12372 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
12373 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12374 // decide if we should generate a 16-byte constant mask when we only need 4 or
12375 // 8 bytes for the scalar case.
12376 if (VT.isVector()) {
12377 EltVT = VT.getVectorElementType();
12378 NumElts = VT.getVectorNumElements();
12381 unsigned EltBits = EltVT.getSizeInBits();
12382 LLVMContext *Context = DAG.getContext();
12383 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12385 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12386 Constant *C = ConstantInt::get(*Context, MaskElt);
12387 C = ConstantVector::getSplat(NumElts, C);
12388 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12389 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
12390 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12391 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12392 MachinePointerInfo::getConstantPool(),
12393 false, false, false, Alignment);
12395 if (VT.isVector()) {
12396 // For a vector, cast operands to a vector type, perform the logic op,
12397 // and cast the result back to the original value type.
12398 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12399 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12400 SDValue Operand = IsFNABS ?
12401 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12402 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12403 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12404 return DAG.getNode(ISD::BITCAST, dl, VT,
12405 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12408 // If not vector, then scalar.
12409 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12410 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12411 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12414 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12415 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12416 LLVMContext *Context = DAG.getContext();
12417 SDValue Op0 = Op.getOperand(0);
12418 SDValue Op1 = Op.getOperand(1);
12420 MVT VT = Op.getSimpleValueType();
12421 MVT SrcVT = Op1.getSimpleValueType();
12423 // If second operand is smaller, extend it first.
12424 if (SrcVT.bitsLT(VT)) {
12425 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12428 // And if it is bigger, shrink it first.
12429 if (SrcVT.bitsGT(VT)) {
12430 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12434 // At this point the operands and the result should have the same
12435 // type, and that won't be f80 since that is not custom lowered.
12437 const fltSemantics &Sem =
12438 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12439 const unsigned SizeInBits = VT.getSizeInBits();
12441 SmallVector<Constant *, 4> CV(
12442 VT == MVT::f64 ? 2 : 4,
12443 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12445 // First, clear all bits but the sign bit from the second operand (sign).
12446 CV[0] = ConstantFP::get(*Context,
12447 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12448 Constant *C = ConstantVector::get(CV);
12449 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12450 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12451 MachinePointerInfo::getConstantPool(),
12452 false, false, false, 16);
12453 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12455 // Next, clear the sign bit from the first operand (magnitude).
12456 // If it's a constant, we can clear it here.
12457 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12458 APFloat APF = Op0CN->getValueAPF();
12459 // If the magnitude is a positive zero, the sign bit alone is enough.
12460 if (APF.isPosZero())
12463 CV[0] = ConstantFP::get(*Context, APF);
12465 CV[0] = ConstantFP::get(
12467 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12469 C = ConstantVector::get(CV);
12470 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12471 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12472 MachinePointerInfo::getConstantPool(),
12473 false, false, false, 16);
12474 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12475 if (!isa<ConstantFPSDNode>(Op0))
12476 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12478 // OR the magnitude value with the sign bit.
12479 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12482 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12483 SDValue N0 = Op.getOperand(0);
12485 MVT VT = Op.getSimpleValueType();
12487 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12488 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12489 DAG.getConstant(1, dl, VT));
12490 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12493 // Check whether an OR'd tree is PTEST-able.
12494 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12495 SelectionDAG &DAG) {
12496 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12498 if (!Subtarget->hasSSE41())
12501 if (!Op->hasOneUse())
12504 SDNode *N = Op.getNode();
12507 SmallVector<SDValue, 8> Opnds;
12508 DenseMap<SDValue, unsigned> VecInMap;
12509 SmallVector<SDValue, 8> VecIns;
12510 EVT VT = MVT::Other;
12512 // Recognize a special case where a vector is casted into wide integer to
12514 Opnds.push_back(N->getOperand(0));
12515 Opnds.push_back(N->getOperand(1));
12517 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12518 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12519 // BFS traverse all OR'd operands.
12520 if (I->getOpcode() == ISD::OR) {
12521 Opnds.push_back(I->getOperand(0));
12522 Opnds.push_back(I->getOperand(1));
12523 // Re-evaluate the number of nodes to be traversed.
12524 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12528 // Quit if a non-EXTRACT_VECTOR_ELT
12529 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12532 // Quit if without a constant index.
12533 SDValue Idx = I->getOperand(1);
12534 if (!isa<ConstantSDNode>(Idx))
12537 SDValue ExtractedFromVec = I->getOperand(0);
12538 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12539 if (M == VecInMap.end()) {
12540 VT = ExtractedFromVec.getValueType();
12541 // Quit if not 128/256-bit vector.
12542 if (!VT.is128BitVector() && !VT.is256BitVector())
12544 // Quit if not the same type.
12545 if (VecInMap.begin() != VecInMap.end() &&
12546 VT != VecInMap.begin()->first.getValueType())
12548 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12549 VecIns.push_back(ExtractedFromVec);
12551 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12554 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12555 "Not extracted from 128-/256-bit vector.");
12557 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12559 for (DenseMap<SDValue, unsigned>::const_iterator
12560 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12561 // Quit if not all elements are used.
12562 if (I->second != FullMask)
12566 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12568 // Cast all vectors into TestVT for PTEST.
12569 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12570 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12572 // If more than one full vectors are evaluated, OR them first before PTEST.
12573 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12574 // Each iteration will OR 2 nodes and append the result until there is only
12575 // 1 node left, i.e. the final OR'd value of all vectors.
12576 SDValue LHS = VecIns[Slot];
12577 SDValue RHS = VecIns[Slot + 1];
12578 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12581 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12582 VecIns.back(), VecIns.back());
12585 /// \brief return true if \c Op has a use that doesn't just read flags.
12586 static bool hasNonFlagsUse(SDValue Op) {
12587 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12589 SDNode *User = *UI;
12590 unsigned UOpNo = UI.getOperandNo();
12591 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12592 // Look pass truncate.
12593 UOpNo = User->use_begin().getOperandNo();
12594 User = *User->use_begin();
12597 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12598 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12604 /// Emit nodes that will be selected as "test Op0,Op0", or something
12606 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12607 SelectionDAG &DAG) const {
12608 if (Op.getValueType() == MVT::i1) {
12609 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12610 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12611 DAG.getConstant(0, dl, MVT::i8));
12613 // CF and OF aren't always set the way we want. Determine which
12614 // of these we need.
12615 bool NeedCF = false;
12616 bool NeedOF = false;
12619 case X86::COND_A: case X86::COND_AE:
12620 case X86::COND_B: case X86::COND_BE:
12623 case X86::COND_G: case X86::COND_GE:
12624 case X86::COND_L: case X86::COND_LE:
12625 case X86::COND_O: case X86::COND_NO: {
12626 // Check if we really need to set the
12627 // Overflow flag. If NoSignedWrap is present
12628 // that is not actually needed.
12629 switch (Op->getOpcode()) {
12634 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
12635 if (BinNode->Flags.hasNoSignedWrap())
12645 // See if we can use the EFLAGS value from the operand instead of
12646 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12647 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12648 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12649 // Emit a CMP with 0, which is the TEST pattern.
12650 //if (Op.getValueType() == MVT::i1)
12651 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12652 // DAG.getConstant(0, MVT::i1));
12653 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12654 DAG.getConstant(0, dl, Op.getValueType()));
12656 unsigned Opcode = 0;
12657 unsigned NumOperands = 0;
12659 // Truncate operations may prevent the merge of the SETCC instruction
12660 // and the arithmetic instruction before it. Attempt to truncate the operands
12661 // of the arithmetic instruction and use a reduced bit-width instruction.
12662 bool NeedTruncation = false;
12663 SDValue ArithOp = Op;
12664 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12665 SDValue Arith = Op->getOperand(0);
12666 // Both the trunc and the arithmetic op need to have one user each.
12667 if (Arith->hasOneUse())
12668 switch (Arith.getOpcode()) {
12675 NeedTruncation = true;
12681 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12682 // which may be the result of a CAST. We use the variable 'Op', which is the
12683 // non-casted variable when we check for possible users.
12684 switch (ArithOp.getOpcode()) {
12686 // Due to an isel shortcoming, be conservative if this add is likely to be
12687 // selected as part of a load-modify-store instruction. When the root node
12688 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12689 // uses of other nodes in the match, such as the ADD in this case. This
12690 // leads to the ADD being left around and reselected, with the result being
12691 // two adds in the output. Alas, even if none our users are stores, that
12692 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12693 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12694 // climbing the DAG back to the root, and it doesn't seem to be worth the
12696 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12697 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12698 if (UI->getOpcode() != ISD::CopyToReg &&
12699 UI->getOpcode() != ISD::SETCC &&
12700 UI->getOpcode() != ISD::STORE)
12703 if (ConstantSDNode *C =
12704 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12705 // An add of one will be selected as an INC.
12706 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12707 Opcode = X86ISD::INC;
12712 // An add of negative one (subtract of one) will be selected as a DEC.
12713 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12714 Opcode = X86ISD::DEC;
12720 // Otherwise use a regular EFLAGS-setting add.
12721 Opcode = X86ISD::ADD;
12726 // If we have a constant logical shift that's only used in a comparison
12727 // against zero turn it into an equivalent AND. This allows turning it into
12728 // a TEST instruction later.
12729 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12730 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12731 EVT VT = Op.getValueType();
12732 unsigned BitWidth = VT.getSizeInBits();
12733 unsigned ShAmt = Op->getConstantOperandVal(1);
12734 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12736 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12737 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12738 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12739 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12741 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12742 DAG.getConstant(Mask, dl, VT));
12743 DAG.ReplaceAllUsesWith(Op, New);
12749 // If the primary and result isn't used, don't bother using X86ISD::AND,
12750 // because a TEST instruction will be better.
12751 if (!hasNonFlagsUse(Op))
12757 // Due to the ISEL shortcoming noted above, be conservative if this op is
12758 // likely to be selected as part of a load-modify-store instruction.
12759 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12760 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12761 if (UI->getOpcode() == ISD::STORE)
12764 // Otherwise use a regular EFLAGS-setting instruction.
12765 switch (ArithOp.getOpcode()) {
12766 default: llvm_unreachable("unexpected operator!");
12767 case ISD::SUB: Opcode = X86ISD::SUB; break;
12768 case ISD::XOR: Opcode = X86ISD::XOR; break;
12769 case ISD::AND: Opcode = X86ISD::AND; break;
12771 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12772 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12773 if (EFLAGS.getNode())
12776 Opcode = X86ISD::OR;
12790 return SDValue(Op.getNode(), 1);
12796 // If we found that truncation is beneficial, perform the truncation and
12798 if (NeedTruncation) {
12799 EVT VT = Op.getValueType();
12800 SDValue WideVal = Op->getOperand(0);
12801 EVT WideVT = WideVal.getValueType();
12802 unsigned ConvertedOp = 0;
12803 // Use a target machine opcode to prevent further DAGCombine
12804 // optimizations that may separate the arithmetic operations
12805 // from the setcc node.
12806 switch (WideVal.getOpcode()) {
12808 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12809 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12810 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12811 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12812 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12816 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12817 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12818 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12819 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12820 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12826 // Emit a CMP with 0, which is the TEST pattern.
12827 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12828 DAG.getConstant(0, dl, Op.getValueType()));
12830 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12831 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12833 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12834 DAG.ReplaceAllUsesWith(Op, New);
12835 return SDValue(New.getNode(), 1);
12838 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12840 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12841 SDLoc dl, SelectionDAG &DAG) const {
12842 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12843 if (C->getAPIntValue() == 0)
12844 return EmitTest(Op0, X86CC, dl, DAG);
12846 if (Op0.getValueType() == MVT::i1)
12847 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12850 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12851 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12852 // Do the comparison at i32 if it's smaller, besides the Atom case.
12853 // This avoids subregister aliasing issues. Keep the smaller reference
12854 // if we're optimizing for size, however, as that'll allow better folding
12855 // of memory operations.
12856 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12857 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12858 Attribute::MinSize) &&
12859 !Subtarget->isAtom()) {
12860 unsigned ExtendOp =
12861 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12862 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12863 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12865 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12866 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12867 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12869 return SDValue(Sub.getNode(), 1);
12871 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12874 /// Convert a comparison if required by the subtarget.
12875 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12876 SelectionDAG &DAG) const {
12877 // If the subtarget does not support the FUCOMI instruction, floating-point
12878 // comparisons have to be converted.
12879 if (Subtarget->hasCMov() ||
12880 Cmp.getOpcode() != X86ISD::CMP ||
12881 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12882 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12885 // The instruction selector will select an FUCOM instruction instead of
12886 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12887 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12888 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12890 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12891 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12892 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12893 DAG.getConstant(8, dl, MVT::i8));
12894 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12895 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12898 /// The minimum architected relative accuracy is 2^-12. We need one
12899 /// Newton-Raphson step to have a good float result (24 bits of precision).
12900 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12901 DAGCombinerInfo &DCI,
12902 unsigned &RefinementSteps,
12903 bool &UseOneConstNR) const {
12904 // FIXME: We should use instruction latency models to calculate the cost of
12905 // each potential sequence, but this is very hard to do reliably because
12906 // at least Intel's Core* chips have variable timing based on the number of
12907 // significant digits in the divisor and/or sqrt operand.
12908 if (!Subtarget->useSqrtEst())
12911 EVT VT = Op.getValueType();
12913 // SSE1 has rsqrtss and rsqrtps.
12914 // TODO: Add support for AVX512 (v16f32).
12915 // It is likely not profitable to do this for f64 because a double-precision
12916 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12917 // instructions: convert to single, rsqrtss, convert back to double, refine
12918 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12919 // along with FMA, this could be a throughput win.
12920 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12921 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12922 RefinementSteps = 1;
12923 UseOneConstNR = false;
12924 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12929 /// The minimum architected relative accuracy is 2^-12. We need one
12930 /// Newton-Raphson step to have a good float result (24 bits of precision).
12931 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12932 DAGCombinerInfo &DCI,
12933 unsigned &RefinementSteps) const {
12934 // FIXME: We should use instruction latency models to calculate the cost of
12935 // each potential sequence, but this is very hard to do reliably because
12936 // at least Intel's Core* chips have variable timing based on the number of
12937 // significant digits in the divisor.
12938 if (!Subtarget->useReciprocalEst())
12941 EVT VT = Op.getValueType();
12943 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12944 // TODO: Add support for AVX512 (v16f32).
12945 // It is likely not profitable to do this for f64 because a double-precision
12946 // reciprocal estimate with refinement on x86 prior to FMA requires
12947 // 15 instructions: convert to single, rcpss, convert back to double, refine
12948 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12949 // along with FMA, this could be a throughput win.
12950 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12951 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12952 RefinementSteps = ReciprocalEstimateRefinementSteps;
12953 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12958 /// If we have at least two divisions that use the same divisor, convert to
12959 /// multplication by a reciprocal. This may need to be adjusted for a given
12960 /// CPU if a division's cost is not at least twice the cost of a multiplication.
12961 /// This is because we still need one division to calculate the reciprocal and
12962 /// then we need two multiplies by that reciprocal as replacements for the
12963 /// original divisions.
12964 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
12965 return NumUsers > 1;
12968 static bool isAllOnes(SDValue V) {
12969 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12970 return C && C->isAllOnesValue();
12973 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12974 /// if it's possible.
12975 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12976 SDLoc dl, SelectionDAG &DAG) const {
12977 SDValue Op0 = And.getOperand(0);
12978 SDValue Op1 = And.getOperand(1);
12979 if (Op0.getOpcode() == ISD::TRUNCATE)
12980 Op0 = Op0.getOperand(0);
12981 if (Op1.getOpcode() == ISD::TRUNCATE)
12982 Op1 = Op1.getOperand(0);
12985 if (Op1.getOpcode() == ISD::SHL)
12986 std::swap(Op0, Op1);
12987 if (Op0.getOpcode() == ISD::SHL) {
12988 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12989 if (And00C->getZExtValue() == 1) {
12990 // If we looked past a truncate, check that it's only truncating away
12992 unsigned BitWidth = Op0.getValueSizeInBits();
12993 unsigned AndBitWidth = And.getValueSizeInBits();
12994 if (BitWidth > AndBitWidth) {
12996 DAG.computeKnownBits(Op0, Zeros, Ones);
12997 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13001 RHS = Op0.getOperand(1);
13003 } else if (Op1.getOpcode() == ISD::Constant) {
13004 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13005 uint64_t AndRHSVal = AndRHS->getZExtValue();
13006 SDValue AndLHS = Op0;
13008 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13009 LHS = AndLHS.getOperand(0);
13010 RHS = AndLHS.getOperand(1);
13013 // Use BT if the immediate can't be encoded in a TEST instruction.
13014 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13016 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13020 if (LHS.getNode()) {
13021 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13022 // instruction. Since the shift amount is in-range-or-undefined, we know
13023 // that doing a bittest on the i32 value is ok. We extend to i32 because
13024 // the encoding for the i16 version is larger than the i32 version.
13025 // Also promote i16 to i32 for performance / code size reason.
13026 if (LHS.getValueType() == MVT::i8 ||
13027 LHS.getValueType() == MVT::i16)
13028 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13030 // If the operand types disagree, extend the shift amount to match. Since
13031 // BT ignores high bits (like shifts) we can use anyextend.
13032 if (LHS.getValueType() != RHS.getValueType())
13033 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13035 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13036 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13037 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13038 DAG.getConstant(Cond, dl, MVT::i8), BT);
13044 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13046 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13051 // SSE Condition code mapping:
13060 switch (SetCCOpcode) {
13061 default: llvm_unreachable("Unexpected SETCC condition");
13063 case ISD::SETEQ: SSECC = 0; break;
13065 case ISD::SETGT: Swap = true; // Fallthrough
13067 case ISD::SETOLT: SSECC = 1; break;
13069 case ISD::SETGE: Swap = true; // Fallthrough
13071 case ISD::SETOLE: SSECC = 2; break;
13072 case ISD::SETUO: SSECC = 3; break;
13074 case ISD::SETNE: SSECC = 4; break;
13075 case ISD::SETULE: Swap = true; // Fallthrough
13076 case ISD::SETUGE: SSECC = 5; break;
13077 case ISD::SETULT: Swap = true; // Fallthrough
13078 case ISD::SETUGT: SSECC = 6; break;
13079 case ISD::SETO: SSECC = 7; break;
13081 case ISD::SETONE: SSECC = 8; break;
13084 std::swap(Op0, Op1);
13089 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13090 // ones, and then concatenate the result back.
13091 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13092 MVT VT = Op.getSimpleValueType();
13094 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13095 "Unsupported value type for operation");
13097 unsigned NumElems = VT.getVectorNumElements();
13099 SDValue CC = Op.getOperand(2);
13101 // Extract the LHS vectors
13102 SDValue LHS = Op.getOperand(0);
13103 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13104 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13106 // Extract the RHS vectors
13107 SDValue RHS = Op.getOperand(1);
13108 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13109 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13111 // Issue the operation on the smaller types and concatenate the result back
13112 MVT EltVT = VT.getVectorElementType();
13113 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13114 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13115 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13116 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13119 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13120 SDValue Op0 = Op.getOperand(0);
13121 SDValue Op1 = Op.getOperand(1);
13122 SDValue CC = Op.getOperand(2);
13123 MVT VT = Op.getSimpleValueType();
13126 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13127 "Unexpected type for boolean compare operation");
13128 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13129 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13130 DAG.getConstant(-1, dl, VT));
13131 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13132 DAG.getConstant(-1, dl, VT));
13133 switch (SetCCOpcode) {
13134 default: llvm_unreachable("Unexpected SETCC condition");
13136 // (x != y) -> ~(x ^ y)
13137 return DAG.getNode(ISD::XOR, dl, VT,
13138 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13139 DAG.getConstant(-1, dl, VT));
13141 // (x == y) -> (x ^ y)
13142 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13145 // (x > y) -> (x & ~y)
13146 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13149 // (x < y) -> (~x & y)
13150 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13153 // (x <= y) -> (~x | y)
13154 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13157 // (x >=y) -> (x | ~y)
13158 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13162 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13163 const X86Subtarget *Subtarget) {
13164 SDValue Op0 = Op.getOperand(0);
13165 SDValue Op1 = Op.getOperand(1);
13166 SDValue CC = Op.getOperand(2);
13167 MVT VT = Op.getSimpleValueType();
13170 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13171 Op.getValueType().getScalarType() == MVT::i1 &&
13172 "Cannot set masked compare for this operation");
13174 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13176 bool Unsigned = false;
13179 switch (SetCCOpcode) {
13180 default: llvm_unreachable("Unexpected SETCC condition");
13181 case ISD::SETNE: SSECC = 4; break;
13182 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13183 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13184 case ISD::SETLT: Swap = true; //fall-through
13185 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13186 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13187 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13188 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13189 case ISD::SETULE: Unsigned = true; //fall-through
13190 case ISD::SETLE: SSECC = 2; break;
13194 std::swap(Op0, Op1);
13196 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13197 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13198 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13199 DAG.getConstant(SSECC, dl, MVT::i8));
13202 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13203 /// operand \p Op1. If non-trivial (for example because it's not constant)
13204 /// return an empty value.
13205 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13207 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13211 MVT VT = Op1.getSimpleValueType();
13212 MVT EVT = VT.getVectorElementType();
13213 unsigned n = VT.getVectorNumElements();
13214 SmallVector<SDValue, 8> ULTOp1;
13216 for (unsigned i = 0; i < n; ++i) {
13217 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13218 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13221 // Avoid underflow.
13222 APInt Val = Elt->getAPIntValue();
13226 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13229 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13232 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13233 SelectionDAG &DAG) {
13234 SDValue Op0 = Op.getOperand(0);
13235 SDValue Op1 = Op.getOperand(1);
13236 SDValue CC = Op.getOperand(2);
13237 MVT VT = Op.getSimpleValueType();
13238 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13239 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13244 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13245 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13248 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13249 unsigned Opc = X86ISD::CMPP;
13250 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13251 assert(VT.getVectorNumElements() <= 16);
13252 Opc = X86ISD::CMPM;
13254 // In the two special cases we can't handle, emit two comparisons.
13257 unsigned CombineOpc;
13258 if (SetCCOpcode == ISD::SETUEQ) {
13259 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13261 assert(SetCCOpcode == ISD::SETONE);
13262 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13265 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13266 DAG.getConstant(CC0, dl, MVT::i8));
13267 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13268 DAG.getConstant(CC1, dl, MVT::i8));
13269 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13271 // Handle all other FP comparisons here.
13272 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13273 DAG.getConstant(SSECC, dl, MVT::i8));
13276 // Break 256-bit integer vector compare into smaller ones.
13277 if (VT.is256BitVector() && !Subtarget->hasInt256())
13278 return Lower256IntVSETCC(Op, DAG);
13280 EVT OpVT = Op1.getValueType();
13281 if (OpVT.getVectorElementType() == MVT::i1)
13282 return LowerBoolVSETCC_AVX512(Op, DAG);
13284 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13285 if (Subtarget->hasAVX512()) {
13286 if (Op1.getValueType().is512BitVector() ||
13287 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13288 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13289 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13291 // In AVX-512 architecture setcc returns mask with i1 elements,
13292 // But there is no compare instruction for i8 and i16 elements in KNL.
13293 // We are not talking about 512-bit operands in this case, these
13294 // types are illegal.
13296 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13297 OpVT.getVectorElementType().getSizeInBits() >= 8))
13298 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13299 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13302 // We are handling one of the integer comparisons here. Since SSE only has
13303 // GT and EQ comparisons for integer, swapping operands and multiple
13304 // operations may be required for some comparisons.
13306 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13307 bool Subus = false;
13309 switch (SetCCOpcode) {
13310 default: llvm_unreachable("Unexpected SETCC condition");
13311 case ISD::SETNE: Invert = true;
13312 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13313 case ISD::SETLT: Swap = true;
13314 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13315 case ISD::SETGE: Swap = true;
13316 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13317 Invert = true; break;
13318 case ISD::SETULT: Swap = true;
13319 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13320 FlipSigns = true; break;
13321 case ISD::SETUGE: Swap = true;
13322 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13323 FlipSigns = true; Invert = true; break;
13326 // Special case: Use min/max operations for SETULE/SETUGE
13327 MVT VET = VT.getVectorElementType();
13329 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13330 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13333 switch (SetCCOpcode) {
13335 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
13336 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
13339 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13342 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13343 if (!MinMax && hasSubus) {
13344 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13346 // t = psubus Op0, Op1
13347 // pcmpeq t, <0..0>
13348 switch (SetCCOpcode) {
13350 case ISD::SETULT: {
13351 // If the comparison is against a constant we can turn this into a
13352 // setule. With psubus, setule does not require a swap. This is
13353 // beneficial because the constant in the register is no longer
13354 // destructed as the destination so it can be hoisted out of a loop.
13355 // Only do this pre-AVX since vpcmp* is no longer destructive.
13356 if (Subtarget->hasAVX())
13358 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13359 if (ULEOp1.getNode()) {
13361 Subus = true; Invert = false; Swap = false;
13365 // Psubus is better than flip-sign because it requires no inversion.
13366 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13367 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13371 Opc = X86ISD::SUBUS;
13377 std::swap(Op0, Op1);
13379 // Check that the operation in question is available (most are plain SSE2,
13380 // but PCMPGTQ and PCMPEQQ have different requirements).
13381 if (VT == MVT::v2i64) {
13382 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13383 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13385 // First cast everything to the right type.
13386 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13387 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13389 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13390 // bits of the inputs before performing those operations. The lower
13391 // compare is always unsigned.
13394 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13396 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13397 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13398 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13399 Sign, Zero, Sign, Zero);
13401 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13402 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13404 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13405 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13406 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13408 // Create masks for only the low parts/high parts of the 64 bit integers.
13409 static const int MaskHi[] = { 1, 1, 3, 3 };
13410 static const int MaskLo[] = { 0, 0, 2, 2 };
13411 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13412 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13413 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13415 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13416 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13419 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13421 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13424 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13425 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13426 // pcmpeqd + pshufd + pand.
13427 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13429 // First cast everything to the right type.
13430 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
13431 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
13434 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13436 // Make sure the lower and upper halves are both all-ones.
13437 static const int Mask[] = { 1, 0, 3, 2 };
13438 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13439 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13442 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13444 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
13448 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13449 // bits of the inputs before performing those operations.
13451 EVT EltVT = VT.getVectorElementType();
13452 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13454 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13455 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13458 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13460 // If the logical-not of the result is required, perform that now.
13462 Result = DAG.getNOT(dl, Result, VT);
13465 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13468 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13469 getZeroVector(VT, Subtarget, DAG, dl));
13474 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13476 MVT VT = Op.getSimpleValueType();
13478 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13480 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13481 && "SetCC type must be 8-bit or 1-bit integer");
13482 SDValue Op0 = Op.getOperand(0);
13483 SDValue Op1 = Op.getOperand(1);
13485 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13487 // Optimize to BT if possible.
13488 // Lower (X & (1 << N)) == 0 to BT(X, N).
13489 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13490 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13491 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13492 Op1.getOpcode() == ISD::Constant &&
13493 cast<ConstantSDNode>(Op1)->isNullValue() &&
13494 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13495 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13496 if (NewSetCC.getNode()) {
13498 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13503 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13505 if (Op1.getOpcode() == ISD::Constant &&
13506 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13507 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13508 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13510 // If the input is a setcc, then reuse the input setcc or use a new one with
13511 // the inverted condition.
13512 if (Op0.getOpcode() == X86ISD::SETCC) {
13513 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13514 bool Invert = (CC == ISD::SETNE) ^
13515 cast<ConstantSDNode>(Op1)->isNullValue();
13519 CCode = X86::GetOppositeBranchCondition(CCode);
13520 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13521 DAG.getConstant(CCode, dl, MVT::i8),
13522 Op0.getOperand(1));
13524 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13528 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13529 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13530 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13532 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13533 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13536 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13537 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13538 if (X86CC == X86::COND_INVALID)
13541 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13542 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13543 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13544 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13546 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13550 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13551 static bool isX86LogicalCmp(SDValue Op) {
13552 unsigned Opc = Op.getNode()->getOpcode();
13553 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13554 Opc == X86ISD::SAHF)
13556 if (Op.getResNo() == 1 &&
13557 (Opc == X86ISD::ADD ||
13558 Opc == X86ISD::SUB ||
13559 Opc == X86ISD::ADC ||
13560 Opc == X86ISD::SBB ||
13561 Opc == X86ISD::SMUL ||
13562 Opc == X86ISD::UMUL ||
13563 Opc == X86ISD::INC ||
13564 Opc == X86ISD::DEC ||
13565 Opc == X86ISD::OR ||
13566 Opc == X86ISD::XOR ||
13567 Opc == X86ISD::AND))
13570 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13576 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13577 if (V.getOpcode() != ISD::TRUNCATE)
13580 SDValue VOp0 = V.getOperand(0);
13581 unsigned InBits = VOp0.getValueSizeInBits();
13582 unsigned Bits = V.getValueSizeInBits();
13583 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13586 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13587 bool addTest = true;
13588 SDValue Cond = Op.getOperand(0);
13589 SDValue Op1 = Op.getOperand(1);
13590 SDValue Op2 = Op.getOperand(2);
13592 EVT VT = Op1.getValueType();
13595 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13596 // are available or VBLENDV if AVX is available.
13597 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13598 if (Cond.getOpcode() == ISD::SETCC &&
13599 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13600 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13601 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13602 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13603 int SSECC = translateX86FSETCC(
13604 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13607 if (Subtarget->hasAVX512()) {
13608 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13609 DAG.getConstant(SSECC, DL, MVT::i8));
13610 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13613 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13614 DAG.getConstant(SSECC, DL, MVT::i8));
13616 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13617 // of 3 logic instructions for size savings and potentially speed.
13618 // Unfortunately, there is no scalar form of VBLENDV.
13620 // If either operand is a constant, don't try this. We can expect to
13621 // optimize away at least one of the logic instructions later in that
13622 // case, so that sequence would be faster than a variable blend.
13624 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13625 // uses XMM0 as the selection register. That may need just as many
13626 // instructions as the AND/ANDN/OR sequence due to register moves, so
13629 if (Subtarget->hasAVX() &&
13630 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13632 // Convert to vectors, do a VSELECT, and convert back to scalar.
13633 // All of the conversions should be optimized away.
13635 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13636 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13637 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13638 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13640 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13641 VCmp = DAG.getNode(ISD::BITCAST, DL, VCmpVT, VCmp);
13643 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13645 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13646 VSel, DAG.getIntPtrConstant(0, DL));
13648 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13649 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13650 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13654 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
13656 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
13657 Op1Scalar = ConvertI1VectorToInterger(Op1, DAG);
13658 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
13659 Op1Scalar = Op1.getOperand(0);
13661 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
13662 Op2Scalar = ConvertI1VectorToInterger(Op2, DAG);
13663 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
13664 Op2Scalar = Op2.getOperand(0);
13665 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
13666 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
13667 Op1Scalar.getValueType(),
13668 Cond, Op1Scalar, Op2Scalar);
13669 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
13670 return DAG.getNode(ISD::BITCAST, DL, VT, newSelect);
13671 SDValue ExtVec = DAG.getNode(ISD::BITCAST, DL, MVT::v8i1, newSelect);
13672 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
13673 DAG.getIntPtrConstant(0, DL));
13677 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
13678 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
13679 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13680 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
13681 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
13682 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
13683 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
13685 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
13688 if (Cond.getOpcode() == ISD::SETCC) {
13689 SDValue NewCond = LowerSETCC(Cond, DAG);
13690 if (NewCond.getNode())
13694 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13695 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13696 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13697 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13698 if (Cond.getOpcode() == X86ISD::SETCC &&
13699 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13700 isZero(Cond.getOperand(1).getOperand(1))) {
13701 SDValue Cmp = Cond.getOperand(1);
13703 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13705 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13706 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13707 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13709 SDValue CmpOp0 = Cmp.getOperand(0);
13710 // Apply further optimizations for special cases
13711 // (select (x != 0), -1, 0) -> neg & sbb
13712 // (select (x == 0), 0, -1) -> neg & sbb
13713 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13714 if (YC->isNullValue() &&
13715 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13716 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13717 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13718 DAG.getConstant(0, DL,
13719 CmpOp0.getValueType()),
13721 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13722 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13723 SDValue(Neg.getNode(), 1));
13727 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13728 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
13729 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13731 SDValue Res = // Res = 0 or -1.
13732 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13733 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
13735 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13736 Res = DAG.getNOT(DL, Res, Res.getValueType());
13738 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13739 if (!N2C || !N2C->isNullValue())
13740 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13745 // Look past (and (setcc_carry (cmp ...)), 1).
13746 if (Cond.getOpcode() == ISD::AND &&
13747 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13748 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13749 if (C && C->getAPIntValue() == 1)
13750 Cond = Cond.getOperand(0);
13753 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13754 // setting operand in place of the X86ISD::SETCC.
13755 unsigned CondOpcode = Cond.getOpcode();
13756 if (CondOpcode == X86ISD::SETCC ||
13757 CondOpcode == X86ISD::SETCC_CARRY) {
13758 CC = Cond.getOperand(0);
13760 SDValue Cmp = Cond.getOperand(1);
13761 unsigned Opc = Cmp.getOpcode();
13762 MVT VT = Op.getSimpleValueType();
13764 bool IllegalFPCMov = false;
13765 if (VT.isFloatingPoint() && !VT.isVector() &&
13766 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13767 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13769 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13770 Opc == X86ISD::BT) { // FIXME
13774 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13775 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13776 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13777 Cond.getOperand(0).getValueType() != MVT::i8)) {
13778 SDValue LHS = Cond.getOperand(0);
13779 SDValue RHS = Cond.getOperand(1);
13780 unsigned X86Opcode;
13783 switch (CondOpcode) {
13784 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13785 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13786 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13787 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13788 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13789 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13790 default: llvm_unreachable("unexpected overflowing operator");
13792 if (CondOpcode == ISD::UMULO)
13793 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13796 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13798 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13800 if (CondOpcode == ISD::UMULO)
13801 Cond = X86Op.getValue(2);
13803 Cond = X86Op.getValue(1);
13805 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
13810 // Look pass the truncate if the high bits are known zero.
13811 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13812 Cond = Cond.getOperand(0);
13814 // We know the result of AND is compared against zero. Try to match
13816 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13817 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13818 if (NewSetCC.getNode()) {
13819 CC = NewSetCC.getOperand(0);
13820 Cond = NewSetCC.getOperand(1);
13827 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
13828 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13831 // a < b ? -1 : 0 -> RES = ~setcc_carry
13832 // a < b ? 0 : -1 -> RES = setcc_carry
13833 // a >= b ? -1 : 0 -> RES = setcc_carry
13834 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13835 if (Cond.getOpcode() == X86ISD::SUB) {
13836 Cond = ConvertCmpIfNecessary(Cond, DAG);
13837 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13839 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13840 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13841 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13842 DAG.getConstant(X86::COND_B, DL, MVT::i8),
13844 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13845 return DAG.getNOT(DL, Res, Res.getValueType());
13850 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13851 // widen the cmov and push the truncate through. This avoids introducing a new
13852 // branch during isel and doesn't add any extensions.
13853 if (Op.getValueType() == MVT::i8 &&
13854 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13855 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13856 if (T1.getValueType() == T2.getValueType() &&
13857 // Blacklist CopyFromReg to avoid partial register stalls.
13858 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13859 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13860 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13861 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13865 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13866 // condition is true.
13867 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13868 SDValue Ops[] = { Op2, Op1, CC, Cond };
13869 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13872 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13873 SelectionDAG &DAG) {
13874 MVT VT = Op->getSimpleValueType(0);
13875 SDValue In = Op->getOperand(0);
13876 MVT InVT = In.getSimpleValueType();
13877 MVT VTElt = VT.getVectorElementType();
13878 MVT InVTElt = InVT.getVectorElementType();
13882 if ((InVTElt == MVT::i1) &&
13883 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13884 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13886 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13887 VTElt.getSizeInBits() <= 16)) ||
13889 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13890 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13892 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13893 VTElt.getSizeInBits() >= 32))))
13894 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13896 unsigned int NumElts = VT.getVectorNumElements();
13898 if (NumElts != 8 && NumElts != 16)
13901 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13902 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13903 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13904 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13907 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13908 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13910 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
13913 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
13915 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
13916 if (VT.is512BitVector())
13918 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
13921 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
13922 const X86Subtarget *Subtarget,
13923 SelectionDAG &DAG) {
13924 SDValue In = Op->getOperand(0);
13925 MVT VT = Op->getSimpleValueType(0);
13926 MVT InVT = In.getSimpleValueType();
13927 assert(VT.getSizeInBits() == InVT.getSizeInBits());
13929 MVT InSVT = InVT.getScalarType();
13930 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
13932 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
13934 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
13939 // SSE41 targets can use the pmovsx* instructions directly.
13940 if (Subtarget->hasSSE41())
13941 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13943 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
13947 // As SRAI is only available on i16/i32 types, we expand only up to i32
13948 // and handle i64 separately.
13949 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
13950 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
13951 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
13952 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
13953 Curr = DAG.getNode(ISD::BITCAST, dl, CurrVT, Curr);
13956 SDValue SignExt = Curr;
13957 if (CurrVT != InVT) {
13958 unsigned SignExtShift =
13959 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
13960 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
13961 DAG.getConstant(SignExtShift, dl, MVT::i8));
13967 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
13968 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
13969 DAG.getConstant(31, dl, MVT::i8));
13970 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
13971 return DAG.getNode(ISD::BITCAST, dl, VT, Ext);
13977 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13978 SelectionDAG &DAG) {
13979 MVT VT = Op->getSimpleValueType(0);
13980 SDValue In = Op->getOperand(0);
13981 MVT InVT = In.getSimpleValueType();
13984 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13985 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13987 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13988 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13989 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13992 if (Subtarget->hasInt256())
13993 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13995 // Optimize vectors in AVX mode
13996 // Sign extend v8i16 to v8i32 and
13999 // Divide input vector into two parts
14000 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14001 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14002 // concat the vectors to original VT
14004 unsigned NumElems = InVT.getVectorNumElements();
14005 SDValue Undef = DAG.getUNDEF(InVT);
14007 SmallVector<int,8> ShufMask1(NumElems, -1);
14008 for (unsigned i = 0; i != NumElems/2; ++i)
14011 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14013 SmallVector<int,8> ShufMask2(NumElems, -1);
14014 for (unsigned i = 0; i != NumElems/2; ++i)
14015 ShufMask2[i] = i + NumElems/2;
14017 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14019 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14020 VT.getVectorNumElements()/2);
14022 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14023 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14025 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14028 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14029 // may emit an illegal shuffle but the expansion is still better than scalar
14030 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14031 // we'll emit a shuffle and a arithmetic shift.
14032 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14033 // TODO: It is possible to support ZExt by zeroing the undef values during
14034 // the shuffle phase or after the shuffle.
14035 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14036 SelectionDAG &DAG) {
14037 MVT RegVT = Op.getSimpleValueType();
14038 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14039 assert(RegVT.isInteger() &&
14040 "We only custom lower integer vector sext loads.");
14042 // Nothing useful we can do without SSE2 shuffles.
14043 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14045 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14047 EVT MemVT = Ld->getMemoryVT();
14048 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14049 unsigned RegSz = RegVT.getSizeInBits();
14051 ISD::LoadExtType Ext = Ld->getExtensionType();
14053 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14054 && "Only anyext and sext are currently implemented.");
14055 assert(MemVT != RegVT && "Cannot extend to the same type");
14056 assert(MemVT.isVector() && "Must load a vector from memory");
14058 unsigned NumElems = RegVT.getVectorNumElements();
14059 unsigned MemSz = MemVT.getSizeInBits();
14060 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14062 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14063 // The only way in which we have a legal 256-bit vector result but not the
14064 // integer 256-bit operations needed to directly lower a sextload is if we
14065 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14066 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14067 // correctly legalized. We do this late to allow the canonical form of
14068 // sextload to persist throughout the rest of the DAG combiner -- it wants
14069 // to fold together any extensions it can, and so will fuse a sign_extend
14070 // of an sextload into a sextload targeting a wider value.
14072 if (MemSz == 128) {
14073 // Just switch this to a normal load.
14074 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14075 "it must be a legal 128-bit vector "
14077 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14078 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14079 Ld->isInvariant(), Ld->getAlignment());
14081 assert(MemSz < 128 &&
14082 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14083 // Do an sext load to a 128-bit vector type. We want to use the same
14084 // number of elements, but elements half as wide. This will end up being
14085 // recursively lowered by this routine, but will succeed as we definitely
14086 // have all the necessary features if we're using AVX1.
14088 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14089 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14091 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14092 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14093 Ld->isNonTemporal(), Ld->isInvariant(),
14094 Ld->getAlignment());
14097 // Replace chain users with the new chain.
14098 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14099 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14101 // Finally, do a normal sign-extend to the desired register.
14102 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14105 // All sizes must be a power of two.
14106 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14107 "Non-power-of-two elements are not custom lowered!");
14109 // Attempt to load the original value using scalar loads.
14110 // Find the largest scalar type that divides the total loaded size.
14111 MVT SclrLoadTy = MVT::i8;
14112 for (MVT Tp : MVT::integer_valuetypes()) {
14113 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14118 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14119 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14121 SclrLoadTy = MVT::f64;
14123 // Calculate the number of scalar loads that we need to perform
14124 // in order to load our vector from memory.
14125 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14127 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14128 "Can only lower sext loads with a single scalar load!");
14130 unsigned loadRegZize = RegSz;
14131 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14134 // Represent our vector as a sequence of elements which are the
14135 // largest scalar that we can load.
14136 EVT LoadUnitVecVT = EVT::getVectorVT(
14137 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14139 // Represent the data using the same element type that is stored in
14140 // memory. In practice, we ''widen'' MemVT.
14142 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14143 loadRegZize / MemVT.getScalarType().getSizeInBits());
14145 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14146 "Invalid vector type");
14148 // We can't shuffle using an illegal type.
14149 assert(TLI.isTypeLegal(WideVecVT) &&
14150 "We only lower types that form legal widened vector types");
14152 SmallVector<SDValue, 8> Chains;
14153 SDValue Ptr = Ld->getBasePtr();
14154 SDValue Increment =
14155 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl, TLI.getPointerTy());
14156 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14158 for (unsigned i = 0; i < NumLoads; ++i) {
14159 // Perform a single load.
14160 SDValue ScalarLoad =
14161 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14162 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14163 Ld->getAlignment());
14164 Chains.push_back(ScalarLoad.getValue(1));
14165 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14166 // another round of DAGCombining.
14168 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14170 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14171 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14173 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14176 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14178 // Bitcast the loaded value to a vector of the original element type, in
14179 // the size of the target vector type.
14180 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
14181 unsigned SizeRatio = RegSz / MemSz;
14183 if (Ext == ISD::SEXTLOAD) {
14184 // If we have SSE4.1, we can directly emit a VSEXT node.
14185 if (Subtarget->hasSSE41()) {
14186 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14187 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14191 // Otherwise we'll shuffle the small elements in the high bits of the
14192 // larger type and perform an arithmetic shift. If the shift is not legal
14193 // it's better to scalarize.
14194 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14195 "We can't implement a sext load without an arithmetic right shift!");
14197 // Redistribute the loaded elements into the different locations.
14198 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14199 for (unsigned i = 0; i != NumElems; ++i)
14200 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14202 SDValue Shuff = DAG.getVectorShuffle(
14203 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14205 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14207 // Build the arithmetic shift.
14208 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14209 MemVT.getVectorElementType().getSizeInBits();
14211 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14212 DAG.getConstant(Amt, dl, RegVT));
14214 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14218 // Redistribute the loaded elements into the different locations.
14219 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14220 for (unsigned i = 0; i != NumElems; ++i)
14221 ShuffleVec[i * SizeRatio] = i;
14223 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14224 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14226 // Bitcast to the requested type.
14227 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
14228 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14232 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14233 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14234 // from the AND / OR.
14235 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14236 Opc = Op.getOpcode();
14237 if (Opc != ISD::OR && Opc != ISD::AND)
14239 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14240 Op.getOperand(0).hasOneUse() &&
14241 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14242 Op.getOperand(1).hasOneUse());
14245 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14246 // 1 and that the SETCC node has a single use.
14247 static bool isXor1OfSetCC(SDValue Op) {
14248 if (Op.getOpcode() != ISD::XOR)
14250 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14251 if (N1C && N1C->getAPIntValue() == 1) {
14252 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14253 Op.getOperand(0).hasOneUse();
14258 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14259 bool addTest = true;
14260 SDValue Chain = Op.getOperand(0);
14261 SDValue Cond = Op.getOperand(1);
14262 SDValue Dest = Op.getOperand(2);
14265 bool Inverted = false;
14267 if (Cond.getOpcode() == ISD::SETCC) {
14268 // Check for setcc([su]{add,sub,mul}o == 0).
14269 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14270 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14271 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14272 Cond.getOperand(0).getResNo() == 1 &&
14273 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14274 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14275 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14276 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14277 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14278 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14280 Cond = Cond.getOperand(0);
14282 SDValue NewCond = LowerSETCC(Cond, DAG);
14283 if (NewCond.getNode())
14288 // FIXME: LowerXALUO doesn't handle these!!
14289 else if (Cond.getOpcode() == X86ISD::ADD ||
14290 Cond.getOpcode() == X86ISD::SUB ||
14291 Cond.getOpcode() == X86ISD::SMUL ||
14292 Cond.getOpcode() == X86ISD::UMUL)
14293 Cond = LowerXALUO(Cond, DAG);
14296 // Look pass (and (setcc_carry (cmp ...)), 1).
14297 if (Cond.getOpcode() == ISD::AND &&
14298 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14299 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14300 if (C && C->getAPIntValue() == 1)
14301 Cond = Cond.getOperand(0);
14304 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14305 // setting operand in place of the X86ISD::SETCC.
14306 unsigned CondOpcode = Cond.getOpcode();
14307 if (CondOpcode == X86ISD::SETCC ||
14308 CondOpcode == X86ISD::SETCC_CARRY) {
14309 CC = Cond.getOperand(0);
14311 SDValue Cmp = Cond.getOperand(1);
14312 unsigned Opc = Cmp.getOpcode();
14313 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14314 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14318 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14322 // These can only come from an arithmetic instruction with overflow,
14323 // e.g. SADDO, UADDO.
14324 Cond = Cond.getNode()->getOperand(1);
14330 CondOpcode = Cond.getOpcode();
14331 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14332 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14333 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14334 Cond.getOperand(0).getValueType() != MVT::i8)) {
14335 SDValue LHS = Cond.getOperand(0);
14336 SDValue RHS = Cond.getOperand(1);
14337 unsigned X86Opcode;
14340 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14341 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14343 switch (CondOpcode) {
14344 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14346 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14348 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14351 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14352 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14354 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14356 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14359 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14360 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14361 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14362 default: llvm_unreachable("unexpected overflowing operator");
14365 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14366 if (CondOpcode == ISD::UMULO)
14367 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14370 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14372 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14374 if (CondOpcode == ISD::UMULO)
14375 Cond = X86Op.getValue(2);
14377 Cond = X86Op.getValue(1);
14379 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14383 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14384 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14385 if (CondOpc == ISD::OR) {
14386 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14387 // two branches instead of an explicit OR instruction with a
14389 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14390 isX86LogicalCmp(Cmp)) {
14391 CC = Cond.getOperand(0).getOperand(0);
14392 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14393 Chain, Dest, CC, Cmp);
14394 CC = Cond.getOperand(1).getOperand(0);
14398 } else { // ISD::AND
14399 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14400 // two branches instead of an explicit AND instruction with a
14401 // separate test. However, we only do this if this block doesn't
14402 // have a fall-through edge, because this requires an explicit
14403 // jmp when the condition is false.
14404 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14405 isX86LogicalCmp(Cmp) &&
14406 Op.getNode()->hasOneUse()) {
14407 X86::CondCode CCode =
14408 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14409 CCode = X86::GetOppositeBranchCondition(CCode);
14410 CC = DAG.getConstant(CCode, dl, MVT::i8);
14411 SDNode *User = *Op.getNode()->use_begin();
14412 // Look for an unconditional branch following this conditional branch.
14413 // We need this because we need to reverse the successors in order
14414 // to implement FCMP_OEQ.
14415 if (User->getOpcode() == ISD::BR) {
14416 SDValue FalseBB = User->getOperand(1);
14418 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14419 assert(NewBR == User);
14423 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14424 Chain, Dest, CC, Cmp);
14425 X86::CondCode CCode =
14426 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14427 CCode = X86::GetOppositeBranchCondition(CCode);
14428 CC = DAG.getConstant(CCode, dl, MVT::i8);
14434 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14435 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14436 // It should be transformed during dag combiner except when the condition
14437 // is set by a arithmetics with overflow node.
14438 X86::CondCode CCode =
14439 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14440 CCode = X86::GetOppositeBranchCondition(CCode);
14441 CC = DAG.getConstant(CCode, dl, MVT::i8);
14442 Cond = Cond.getOperand(0).getOperand(1);
14444 } else if (Cond.getOpcode() == ISD::SETCC &&
14445 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14446 // For FCMP_OEQ, we can emit
14447 // two branches instead of an explicit AND instruction with a
14448 // separate test. However, we only do this if this block doesn't
14449 // have a fall-through edge, because this requires an explicit
14450 // jmp when the condition is false.
14451 if (Op.getNode()->hasOneUse()) {
14452 SDNode *User = *Op.getNode()->use_begin();
14453 // Look for an unconditional branch following this conditional branch.
14454 // We need this because we need to reverse the successors in order
14455 // to implement FCMP_OEQ.
14456 if (User->getOpcode() == ISD::BR) {
14457 SDValue FalseBB = User->getOperand(1);
14459 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14460 assert(NewBR == User);
14464 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14465 Cond.getOperand(0), Cond.getOperand(1));
14466 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14467 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14468 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14469 Chain, Dest, CC, Cmp);
14470 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14475 } else if (Cond.getOpcode() == ISD::SETCC &&
14476 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14477 // For FCMP_UNE, we can emit
14478 // two branches instead of an explicit AND instruction with a
14479 // separate test. However, we only do this if this block doesn't
14480 // have a fall-through edge, because this requires an explicit
14481 // jmp when the condition is false.
14482 if (Op.getNode()->hasOneUse()) {
14483 SDNode *User = *Op.getNode()->use_begin();
14484 // Look for an unconditional branch following this conditional branch.
14485 // We need this because we need to reverse the successors in order
14486 // to implement FCMP_UNE.
14487 if (User->getOpcode() == ISD::BR) {
14488 SDValue FalseBB = User->getOperand(1);
14490 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14491 assert(NewBR == User);
14494 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14495 Cond.getOperand(0), Cond.getOperand(1));
14496 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14497 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14498 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14499 Chain, Dest, CC, Cmp);
14500 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14510 // Look pass the truncate if the high bits are known zero.
14511 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14512 Cond = Cond.getOperand(0);
14514 // We know the result of AND is compared against zero. Try to match
14516 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14517 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14518 if (NewSetCC.getNode()) {
14519 CC = NewSetCC.getOperand(0);
14520 Cond = NewSetCC.getOperand(1);
14527 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14528 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14529 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14531 Cond = ConvertCmpIfNecessary(Cond, DAG);
14532 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14533 Chain, Dest, CC, Cond);
14536 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14537 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14538 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14539 // that the guard pages used by the OS virtual memory manager are allocated in
14540 // correct sequence.
14542 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14543 SelectionDAG &DAG) const {
14544 MachineFunction &MF = DAG.getMachineFunction();
14545 bool SplitStack = MF.shouldSplitStack();
14546 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14551 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14552 SDNode* Node = Op.getNode();
14554 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14555 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14556 " not tell us which reg is the stack pointer!");
14557 EVT VT = Node->getValueType(0);
14558 SDValue Tmp1 = SDValue(Node, 0);
14559 SDValue Tmp2 = SDValue(Node, 1);
14560 SDValue Tmp3 = Node->getOperand(2);
14561 SDValue Chain = Tmp1.getOperand(0);
14563 // Chain the dynamic stack allocation so that it doesn't modify the stack
14564 // pointer when other instructions are using the stack.
14565 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14568 SDValue Size = Tmp2.getOperand(1);
14569 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14570 Chain = SP.getValue(1);
14571 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14572 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14573 unsigned StackAlign = TFI.getStackAlignment();
14574 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14575 if (Align > StackAlign)
14576 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14577 DAG.getConstant(-(uint64_t)Align, dl, VT));
14578 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14580 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14581 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14584 SDValue Ops[2] = { Tmp1, Tmp2 };
14585 return DAG.getMergeValues(Ops, dl);
14589 SDValue Chain = Op.getOperand(0);
14590 SDValue Size = Op.getOperand(1);
14591 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14592 EVT VT = Op.getNode()->getValueType(0);
14594 bool Is64Bit = Subtarget->is64Bit();
14595 EVT SPTy = getPointerTy();
14598 MachineRegisterInfo &MRI = MF.getRegInfo();
14601 // The 64 bit implementation of segmented stacks needs to clobber both r10
14602 // r11. This makes it impossible to use it along with nested parameters.
14603 const Function *F = MF.getFunction();
14605 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14607 if (I->hasNestAttr())
14608 report_fatal_error("Cannot use segmented stacks with functions that "
14609 "have nested arguments.");
14612 const TargetRegisterClass *AddrRegClass =
14613 getRegClassFor(getPointerTy());
14614 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14615 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14616 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14617 DAG.getRegister(Vreg, SPTy));
14618 SDValue Ops1[2] = { Value, Chain };
14619 return DAG.getMergeValues(Ops1, dl);
14622 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14624 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14625 Flag = Chain.getValue(1);
14626 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14628 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14630 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14631 unsigned SPReg = RegInfo->getStackRegister();
14632 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14633 Chain = SP.getValue(1);
14636 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14637 DAG.getConstant(-(uint64_t)Align, dl, VT));
14638 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14641 SDValue Ops1[2] = { SP, Chain };
14642 return DAG.getMergeValues(Ops1, dl);
14646 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14647 MachineFunction &MF = DAG.getMachineFunction();
14648 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14650 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14653 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14654 // vastart just stores the address of the VarArgsFrameIndex slot into the
14655 // memory location argument.
14656 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14658 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14659 MachinePointerInfo(SV), false, false, 0);
14663 // gp_offset (0 - 6 * 8)
14664 // fp_offset (48 - 48 + 8 * 16)
14665 // overflow_arg_area (point to parameters coming in memory).
14667 SmallVector<SDValue, 8> MemOps;
14668 SDValue FIN = Op.getOperand(1);
14670 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14671 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14673 FIN, MachinePointerInfo(SV), false, false, 0);
14674 MemOps.push_back(Store);
14677 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14678 FIN, DAG.getIntPtrConstant(4, DL));
14679 Store = DAG.getStore(Op.getOperand(0), DL,
14680 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
14682 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14683 MemOps.push_back(Store);
14685 // Store ptr to overflow_arg_area
14686 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14687 FIN, DAG.getIntPtrConstant(4, DL));
14688 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14690 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14691 MachinePointerInfo(SV, 8),
14693 MemOps.push_back(Store);
14695 // Store ptr to reg_save_area.
14696 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14697 FIN, DAG.getIntPtrConstant(8, DL));
14698 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14700 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14701 MachinePointerInfo(SV, 16), false, false, 0);
14702 MemOps.push_back(Store);
14703 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14706 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14707 assert(Subtarget->is64Bit() &&
14708 "LowerVAARG only handles 64-bit va_arg!");
14709 assert((Subtarget->isTargetLinux() ||
14710 Subtarget->isTargetDarwin()) &&
14711 "Unhandled target in LowerVAARG");
14712 assert(Op.getNode()->getNumOperands() == 4);
14713 SDValue Chain = Op.getOperand(0);
14714 SDValue SrcPtr = Op.getOperand(1);
14715 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14716 unsigned Align = Op.getConstantOperandVal(3);
14719 EVT ArgVT = Op.getNode()->getValueType(0);
14720 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14721 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14724 // Decide which area this value should be read from.
14725 // TODO: Implement the AMD64 ABI in its entirety. This simple
14726 // selection mechanism works only for the basic types.
14727 if (ArgVT == MVT::f80) {
14728 llvm_unreachable("va_arg for f80 not yet implemented");
14729 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14730 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14731 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14732 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14734 llvm_unreachable("Unhandled argument type in LowerVAARG");
14737 if (ArgMode == 2) {
14738 // Sanity Check: Make sure using fp_offset makes sense.
14739 assert(!Subtarget->useSoftFloat() &&
14740 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14741 Attribute::NoImplicitFloat)) &&
14742 Subtarget->hasSSE1());
14745 // Insert VAARG_64 node into the DAG
14746 // VAARG_64 returns two values: Variable Argument Address, Chain
14747 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
14748 DAG.getConstant(ArgMode, dl, MVT::i8),
14749 DAG.getConstant(Align, dl, MVT::i32)};
14750 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14751 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14752 VTs, InstOps, MVT::i64,
14753 MachinePointerInfo(SV),
14755 /*Volatile=*/false,
14757 /*WriteMem=*/true);
14758 Chain = VAARG.getValue(1);
14760 // Load the next argument and return it
14761 return DAG.getLoad(ArgVT, dl,
14764 MachinePointerInfo(),
14765 false, false, false, 0);
14768 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14769 SelectionDAG &DAG) {
14770 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14771 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14772 SDValue Chain = Op.getOperand(0);
14773 SDValue DstPtr = Op.getOperand(1);
14774 SDValue SrcPtr = Op.getOperand(2);
14775 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14776 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14779 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14780 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
14782 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14785 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14786 // amount is a constant. Takes immediate version of shift as input.
14787 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14788 SDValue SrcOp, uint64_t ShiftAmt,
14789 SelectionDAG &DAG) {
14790 MVT ElementType = VT.getVectorElementType();
14792 // Fold this packed shift into its first operand if ShiftAmt is 0.
14796 // Check for ShiftAmt >= element width
14797 if (ShiftAmt >= ElementType.getSizeInBits()) {
14798 if (Opc == X86ISD::VSRAI)
14799 ShiftAmt = ElementType.getSizeInBits() - 1;
14801 return DAG.getConstant(0, dl, VT);
14804 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14805 && "Unknown target vector shift-by-constant node");
14807 // Fold this packed vector shift into a build vector if SrcOp is a
14808 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14809 if (VT == SrcOp.getSimpleValueType() &&
14810 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14811 SmallVector<SDValue, 8> Elts;
14812 unsigned NumElts = SrcOp->getNumOperands();
14813 ConstantSDNode *ND;
14816 default: llvm_unreachable(nullptr);
14817 case X86ISD::VSHLI:
14818 for (unsigned i=0; i!=NumElts; ++i) {
14819 SDValue CurrentOp = SrcOp->getOperand(i);
14820 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14821 Elts.push_back(CurrentOp);
14824 ND = cast<ConstantSDNode>(CurrentOp);
14825 const APInt &C = ND->getAPIntValue();
14826 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
14829 case X86ISD::VSRLI:
14830 for (unsigned i=0; i!=NumElts; ++i) {
14831 SDValue CurrentOp = SrcOp->getOperand(i);
14832 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14833 Elts.push_back(CurrentOp);
14836 ND = cast<ConstantSDNode>(CurrentOp);
14837 const APInt &C = ND->getAPIntValue();
14838 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
14841 case X86ISD::VSRAI:
14842 for (unsigned i=0; i!=NumElts; ++i) {
14843 SDValue CurrentOp = SrcOp->getOperand(i);
14844 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14845 Elts.push_back(CurrentOp);
14848 ND = cast<ConstantSDNode>(CurrentOp);
14849 const APInt &C = ND->getAPIntValue();
14850 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
14855 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14858 return DAG.getNode(Opc, dl, VT, SrcOp,
14859 DAG.getConstant(ShiftAmt, dl, MVT::i8));
14862 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14863 // may or may not be a constant. Takes immediate version of shift as input.
14864 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14865 SDValue SrcOp, SDValue ShAmt,
14866 SelectionDAG &DAG) {
14867 MVT SVT = ShAmt.getSimpleValueType();
14868 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14870 // Catch shift-by-constant.
14871 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14872 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14873 CShAmt->getZExtValue(), DAG);
14875 // Change opcode to non-immediate version
14877 default: llvm_unreachable("Unknown target vector shift node");
14878 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14879 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14880 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14883 const X86Subtarget &Subtarget =
14884 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14885 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14886 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14887 // Let the shuffle legalizer expand this shift amount node.
14888 SDValue Op0 = ShAmt.getOperand(0);
14889 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14890 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14892 // Need to build a vector containing shift amount.
14893 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14894 SmallVector<SDValue, 4> ShOps;
14895 ShOps.push_back(ShAmt);
14896 if (SVT == MVT::i32) {
14897 ShOps.push_back(DAG.getConstant(0, dl, SVT));
14898 ShOps.push_back(DAG.getUNDEF(SVT));
14900 ShOps.push_back(DAG.getUNDEF(SVT));
14902 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14903 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14906 // The return type has to be a 128-bit type with the same element
14907 // type as the input type.
14908 MVT EltVT = VT.getVectorElementType();
14909 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14911 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14912 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14915 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14916 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14917 /// necessary casting for \p Mask when lowering masking intrinsics.
14918 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14919 SDValue PreservedSrc,
14920 const X86Subtarget *Subtarget,
14921 SelectionDAG &DAG) {
14922 EVT VT = Op.getValueType();
14923 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14924 MVT::i1, VT.getVectorNumElements());
14925 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14926 Mask.getValueType().getSizeInBits());
14929 assert(MaskVT.isSimple() && "invalid mask type");
14931 if (isAllOnes(Mask))
14934 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14935 // are extracted by EXTRACT_SUBVECTOR.
14936 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14937 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14938 DAG.getIntPtrConstant(0, dl));
14940 switch (Op.getOpcode()) {
14942 case X86ISD::PCMPEQM:
14943 case X86ISD::PCMPGTM:
14945 case X86ISD::CMPMU:
14946 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14948 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14949 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14950 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14953 /// \brief Creates an SDNode for a predicated scalar operation.
14954 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14955 /// The mask is comming as MVT::i8 and it should be truncated
14956 /// to MVT::i1 while lowering masking intrinsics.
14957 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14958 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14959 /// a scalar instruction.
14960 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14961 SDValue PreservedSrc,
14962 const X86Subtarget *Subtarget,
14963 SelectionDAG &DAG) {
14964 if (isAllOnes(Mask))
14967 EVT VT = Op.getValueType();
14969 // The mask should be of type MVT::i1
14970 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14972 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14973 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14974 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14977 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14978 SelectionDAG &DAG) {
14980 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14981 EVT VT = Op.getValueType();
14982 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14984 switch(IntrData->Type) {
14985 case INTR_TYPE_1OP:
14986 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14987 case INTR_TYPE_2OP:
14988 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14990 case INTR_TYPE_3OP:
14991 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14992 Op.getOperand(2), Op.getOperand(3));
14993 case INTR_TYPE_1OP_MASK_RM: {
14994 SDValue Src = Op.getOperand(1);
14995 SDValue Src0 = Op.getOperand(2);
14996 SDValue Mask = Op.getOperand(3);
14997 SDValue RoundingMode = Op.getOperand(4);
14998 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15000 Mask, Src0, Subtarget, DAG);
15002 case INTR_TYPE_SCALAR_MASK_RM: {
15003 SDValue Src1 = Op.getOperand(1);
15004 SDValue Src2 = Op.getOperand(2);
15005 SDValue Src0 = Op.getOperand(3);
15006 SDValue Mask = Op.getOperand(4);
15007 // There are 2 kinds of intrinsics in this group:
15008 // (1) With supress-all-exceptions (sae) or rounding mode- 6 operands
15009 // (2) With rounding mode and sae - 7 operands.
15010 if (Op.getNumOperands() == 6) {
15011 SDValue Sae = Op.getOperand(5);
15012 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15013 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15015 Mask, Src0, Subtarget, DAG);
15017 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15018 SDValue RoundingMode = Op.getOperand(5);
15019 SDValue Sae = Op.getOperand(6);
15020 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15021 RoundingMode, Sae),
15022 Mask, Src0, Subtarget, DAG);
15024 case INTR_TYPE_2OP_MASK: {
15025 SDValue Src1 = Op.getOperand(1);
15026 SDValue Src2 = Op.getOperand(2);
15027 SDValue PassThru = Op.getOperand(3);
15028 SDValue Mask = Op.getOperand(4);
15029 // We specify 2 possible opcodes for intrinsics with rounding modes.
15030 // First, we check if the intrinsic may have non-default rounding mode,
15031 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15032 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15033 if (IntrWithRoundingModeOpcode != 0) {
15034 SDValue Rnd = Op.getOperand(5);
15035 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15036 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15037 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15038 dl, Op.getValueType(),
15040 Mask, PassThru, Subtarget, DAG);
15043 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15045 Mask, PassThru, Subtarget, DAG);
15047 case FMA_OP_MASK: {
15048 SDValue Src1 = Op.getOperand(1);
15049 SDValue Src2 = Op.getOperand(2);
15050 SDValue Src3 = Op.getOperand(3);
15051 SDValue Mask = Op.getOperand(4);
15052 // We specify 2 possible opcodes for intrinsics with rounding modes.
15053 // First, we check if the intrinsic may have non-default rounding mode,
15054 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15055 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15056 if (IntrWithRoundingModeOpcode != 0) {
15057 SDValue Rnd = Op.getOperand(5);
15058 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15059 X86::STATIC_ROUNDING::CUR_DIRECTION)
15060 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15061 dl, Op.getValueType(),
15062 Src1, Src2, Src3, Rnd),
15063 Mask, Src1, Subtarget, DAG);
15065 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15066 dl, Op.getValueType(),
15068 Mask, Src1, Subtarget, DAG);
15071 case CMP_MASK_CC: {
15072 // Comparison intrinsics with masks.
15073 // Example of transformation:
15074 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15075 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15077 // (v8i1 (insert_subvector undef,
15078 // (v2i1 (and (PCMPEQM %a, %b),
15079 // (extract_subvector
15080 // (v8i1 (bitcast %mask)), 0))), 0))))
15081 EVT VT = Op.getOperand(1).getValueType();
15082 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15083 VT.getVectorNumElements());
15084 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15085 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15086 Mask.getValueType().getSizeInBits());
15088 if (IntrData->Type == CMP_MASK_CC) {
15089 SDValue CC = Op.getOperand(3);
15090 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15091 // We specify 2 possible opcodes for intrinsics with rounding modes.
15092 // First, we check if the intrinsic may have non-default rounding mode,
15093 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15094 if (IntrData->Opc1 != 0) {
15095 SDValue Rnd = Op.getOperand(5);
15096 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15097 X86::STATIC_ROUNDING::CUR_DIRECTION)
15098 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15099 Op.getOperand(2), CC, Rnd);
15101 //default rounding mode
15103 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15104 Op.getOperand(2), CC);
15107 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15108 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15111 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15112 DAG.getTargetConstant(0, dl,
15115 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15116 DAG.getUNDEF(BitcastVT), CmpMask,
15117 DAG.getIntPtrConstant(0, dl));
15118 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
15120 case COMI: { // Comparison intrinsics
15121 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15122 SDValue LHS = Op.getOperand(1);
15123 SDValue RHS = Op.getOperand(2);
15124 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15125 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15126 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15127 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15128 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15129 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15132 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15133 Op.getOperand(1), Op.getOperand(2), DAG);
15135 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15136 Op.getSimpleValueType(),
15138 Op.getOperand(2), DAG),
15139 Op.getOperand(4), Op.getOperand(3), Subtarget,
15141 case COMPRESS_EXPAND_IN_REG: {
15142 SDValue Mask = Op.getOperand(3);
15143 SDValue DataToCompress = Op.getOperand(1);
15144 SDValue PassThru = Op.getOperand(2);
15145 if (isAllOnes(Mask)) // return data as is
15146 return Op.getOperand(1);
15147 EVT VT = Op.getValueType();
15148 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15149 VT.getVectorNumElements());
15150 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15151 Mask.getValueType().getSizeInBits());
15153 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15154 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15155 DAG.getIntPtrConstant(0, dl));
15157 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
15161 SDValue Mask = Op.getOperand(3);
15162 EVT VT = Op.getValueType();
15163 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15164 VT.getVectorNumElements());
15165 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15166 Mask.getValueType().getSizeInBits());
15168 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15169 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15170 DAG.getIntPtrConstant(0, dl));
15171 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15180 default: return SDValue(); // Don't custom lower most intrinsics.
15182 case Intrinsic::x86_avx2_permd:
15183 case Intrinsic::x86_avx2_permps:
15184 // Operands intentionally swapped. Mask is last operand to intrinsic,
15185 // but second operand for node/instruction.
15186 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15187 Op.getOperand(2), Op.getOperand(1));
15189 case Intrinsic::x86_avx512_mask_valign_q_512:
15190 case Intrinsic::x86_avx512_mask_valign_d_512:
15191 // Vector source operands are swapped.
15192 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
15193 Op.getValueType(), Op.getOperand(2),
15196 Op.getOperand(5), Op.getOperand(4),
15199 // ptest and testp intrinsics. The intrinsic these come from are designed to
15200 // return an integer value, not just an instruction so lower it to the ptest
15201 // or testp pattern and a setcc for the result.
15202 case Intrinsic::x86_sse41_ptestz:
15203 case Intrinsic::x86_sse41_ptestc:
15204 case Intrinsic::x86_sse41_ptestnzc:
15205 case Intrinsic::x86_avx_ptestz_256:
15206 case Intrinsic::x86_avx_ptestc_256:
15207 case Intrinsic::x86_avx_ptestnzc_256:
15208 case Intrinsic::x86_avx_vtestz_ps:
15209 case Intrinsic::x86_avx_vtestc_ps:
15210 case Intrinsic::x86_avx_vtestnzc_ps:
15211 case Intrinsic::x86_avx_vtestz_pd:
15212 case Intrinsic::x86_avx_vtestc_pd:
15213 case Intrinsic::x86_avx_vtestnzc_pd:
15214 case Intrinsic::x86_avx_vtestz_ps_256:
15215 case Intrinsic::x86_avx_vtestc_ps_256:
15216 case Intrinsic::x86_avx_vtestnzc_ps_256:
15217 case Intrinsic::x86_avx_vtestz_pd_256:
15218 case Intrinsic::x86_avx_vtestc_pd_256:
15219 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15220 bool IsTestPacked = false;
15223 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15224 case Intrinsic::x86_avx_vtestz_ps:
15225 case Intrinsic::x86_avx_vtestz_pd:
15226 case Intrinsic::x86_avx_vtestz_ps_256:
15227 case Intrinsic::x86_avx_vtestz_pd_256:
15228 IsTestPacked = true; // Fallthrough
15229 case Intrinsic::x86_sse41_ptestz:
15230 case Intrinsic::x86_avx_ptestz_256:
15232 X86CC = X86::COND_E;
15234 case Intrinsic::x86_avx_vtestc_ps:
15235 case Intrinsic::x86_avx_vtestc_pd:
15236 case Intrinsic::x86_avx_vtestc_ps_256:
15237 case Intrinsic::x86_avx_vtestc_pd_256:
15238 IsTestPacked = true; // Fallthrough
15239 case Intrinsic::x86_sse41_ptestc:
15240 case Intrinsic::x86_avx_ptestc_256:
15242 X86CC = X86::COND_B;
15244 case Intrinsic::x86_avx_vtestnzc_ps:
15245 case Intrinsic::x86_avx_vtestnzc_pd:
15246 case Intrinsic::x86_avx_vtestnzc_ps_256:
15247 case Intrinsic::x86_avx_vtestnzc_pd_256:
15248 IsTestPacked = true; // Fallthrough
15249 case Intrinsic::x86_sse41_ptestnzc:
15250 case Intrinsic::x86_avx_ptestnzc_256:
15252 X86CC = X86::COND_A;
15256 SDValue LHS = Op.getOperand(1);
15257 SDValue RHS = Op.getOperand(2);
15258 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15259 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15260 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15261 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15262 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15264 case Intrinsic::x86_avx512_kortestz_w:
15265 case Intrinsic::x86_avx512_kortestc_w: {
15266 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15267 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
15268 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
15269 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15270 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15271 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15272 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15275 case Intrinsic::x86_sse42_pcmpistria128:
15276 case Intrinsic::x86_sse42_pcmpestria128:
15277 case Intrinsic::x86_sse42_pcmpistric128:
15278 case Intrinsic::x86_sse42_pcmpestric128:
15279 case Intrinsic::x86_sse42_pcmpistrio128:
15280 case Intrinsic::x86_sse42_pcmpestrio128:
15281 case Intrinsic::x86_sse42_pcmpistris128:
15282 case Intrinsic::x86_sse42_pcmpestris128:
15283 case Intrinsic::x86_sse42_pcmpistriz128:
15284 case Intrinsic::x86_sse42_pcmpestriz128: {
15288 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15289 case Intrinsic::x86_sse42_pcmpistria128:
15290 Opcode = X86ISD::PCMPISTRI;
15291 X86CC = X86::COND_A;
15293 case Intrinsic::x86_sse42_pcmpestria128:
15294 Opcode = X86ISD::PCMPESTRI;
15295 X86CC = X86::COND_A;
15297 case Intrinsic::x86_sse42_pcmpistric128:
15298 Opcode = X86ISD::PCMPISTRI;
15299 X86CC = X86::COND_B;
15301 case Intrinsic::x86_sse42_pcmpestric128:
15302 Opcode = X86ISD::PCMPESTRI;
15303 X86CC = X86::COND_B;
15305 case Intrinsic::x86_sse42_pcmpistrio128:
15306 Opcode = X86ISD::PCMPISTRI;
15307 X86CC = X86::COND_O;
15309 case Intrinsic::x86_sse42_pcmpestrio128:
15310 Opcode = X86ISD::PCMPESTRI;
15311 X86CC = X86::COND_O;
15313 case Intrinsic::x86_sse42_pcmpistris128:
15314 Opcode = X86ISD::PCMPISTRI;
15315 X86CC = X86::COND_S;
15317 case Intrinsic::x86_sse42_pcmpestris128:
15318 Opcode = X86ISD::PCMPESTRI;
15319 X86CC = X86::COND_S;
15321 case Intrinsic::x86_sse42_pcmpistriz128:
15322 Opcode = X86ISD::PCMPISTRI;
15323 X86CC = X86::COND_E;
15325 case Intrinsic::x86_sse42_pcmpestriz128:
15326 Opcode = X86ISD::PCMPESTRI;
15327 X86CC = X86::COND_E;
15330 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15331 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15332 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15333 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15334 DAG.getConstant(X86CC, dl, MVT::i8),
15335 SDValue(PCMP.getNode(), 1));
15336 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15339 case Intrinsic::x86_sse42_pcmpistri128:
15340 case Intrinsic::x86_sse42_pcmpestri128: {
15342 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15343 Opcode = X86ISD::PCMPISTRI;
15345 Opcode = X86ISD::PCMPESTRI;
15347 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15348 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15349 return DAG.getNode(Opcode, dl, VTs, NewOps);
15352 case Intrinsic::x86_seh_lsda: {
15353 // Compute the symbol for the LSDA. We know it'll get emitted later.
15354 MachineFunction &MF = DAG.getMachineFunction();
15355 SDValue Op1 = Op.getOperand(1);
15357 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15358 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15359 GlobalValue::getRealLinkageName(Fn->getName()));
15360 StringRef Name = LSDASym->getName();
15361 assert(Name.data()[Name.size()] == '\0' && "not null terminated");
15363 // Generate a simple absolute symbol reference. This intrinsic is only
15364 // supported on 32-bit Windows, which isn't PIC.
15366 DAG.getTargetExternalSymbol(Name.data(), VT, X86II::MO_NOPREFIX);
15367 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15372 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15373 SDValue Src, SDValue Mask, SDValue Base,
15374 SDValue Index, SDValue ScaleOp, SDValue Chain,
15375 const X86Subtarget * Subtarget) {
15377 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15378 assert(C && "Invalid scale type");
15379 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15380 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15381 Index.getSimpleValueType().getVectorNumElements());
15383 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15385 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15387 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15388 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15389 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15390 SDValue Segment = DAG.getRegister(0, MVT::i32);
15391 if (Src.getOpcode() == ISD::UNDEF)
15392 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15393 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15394 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15395 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15396 return DAG.getMergeValues(RetOps, dl);
15399 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15400 SDValue Src, SDValue Mask, SDValue Base,
15401 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15403 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15404 assert(C && "Invalid scale type");
15405 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15406 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15407 SDValue Segment = DAG.getRegister(0, MVT::i32);
15408 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15409 Index.getSimpleValueType().getVectorNumElements());
15411 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15413 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15415 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15416 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15417 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15418 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15419 return SDValue(Res, 1);
15422 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15423 SDValue Mask, SDValue Base, SDValue Index,
15424 SDValue ScaleOp, SDValue Chain) {
15426 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15427 assert(C && "Invalid scale type");
15428 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15429 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15430 SDValue Segment = DAG.getRegister(0, MVT::i32);
15432 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15434 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15436 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15438 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
15439 //SDVTList VTs = DAG.getVTList(MVT::Other);
15440 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15441 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15442 return SDValue(Res, 0);
15445 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15446 // read performance monitor counters (x86_rdpmc).
15447 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15448 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15449 SmallVectorImpl<SDValue> &Results) {
15450 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15451 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15454 // The ECX register is used to select the index of the performance counter
15456 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
15458 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
15460 // Reads the content of a 64-bit performance counter and returns it in the
15461 // registers EDX:EAX.
15462 if (Subtarget->is64Bit()) {
15463 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15464 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15467 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15468 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15471 Chain = HI.getValue(1);
15473 if (Subtarget->is64Bit()) {
15474 // The EAX register is loaded with the low-order 32 bits. The EDX register
15475 // is loaded with the supported high-order bits of the counter.
15476 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15477 DAG.getConstant(32, DL, MVT::i8));
15478 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15479 Results.push_back(Chain);
15483 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15484 SDValue Ops[] = { LO, HI };
15485 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15486 Results.push_back(Pair);
15487 Results.push_back(Chain);
15490 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
15491 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
15492 // also used to custom lower READCYCLECOUNTER nodes.
15493 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
15494 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15495 SmallVectorImpl<SDValue> &Results) {
15496 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15497 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
15500 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
15501 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
15502 // and the EAX register is loaded with the low-order 32 bits.
15503 if (Subtarget->is64Bit()) {
15504 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
15505 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
15508 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
15509 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
15512 SDValue Chain = HI.getValue(1);
15514 if (Opcode == X86ISD::RDTSCP_DAG) {
15515 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15517 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
15518 // the ECX register. Add 'ecx' explicitly to the chain.
15519 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
15521 // Explicitly store the content of ECX at the location passed in input
15522 // to the 'rdtscp' intrinsic.
15523 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
15524 MachinePointerInfo(), false, false, 0);
15527 if (Subtarget->is64Bit()) {
15528 // The EDX register is loaded with the high-order 32 bits of the MSR, and
15529 // the EAX register is loaded with the low-order 32 bits.
15530 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
15531 DAG.getConstant(32, DL, MVT::i8));
15532 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
15533 Results.push_back(Chain);
15537 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
15538 SDValue Ops[] = { LO, HI };
15539 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
15540 Results.push_back(Pair);
15541 Results.push_back(Chain);
15544 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
15545 SelectionDAG &DAG) {
15546 SmallVector<SDValue, 2> Results;
15548 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
15550 return DAG.getMergeValues(Results, DL);
15554 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15555 SelectionDAG &DAG) {
15556 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
15558 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
15563 switch(IntrData->Type) {
15565 llvm_unreachable("Unknown Intrinsic Type");
15569 // Emit the node with the right value type.
15570 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
15571 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15573 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
15574 // Otherwise return the value from Rand, which is always 0, casted to i32.
15575 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
15576 DAG.getConstant(1, dl, Op->getValueType(1)),
15577 DAG.getConstant(X86::COND_B, dl, MVT::i32),
15578 SDValue(Result.getNode(), 1) };
15579 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
15580 DAG.getVTList(Op->getValueType(1), MVT::Glue),
15583 // Return { result, isValid, chain }.
15584 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
15585 SDValue(Result.getNode(), 2));
15588 //gather(v1, mask, index, base, scale);
15589 SDValue Chain = Op.getOperand(0);
15590 SDValue Src = Op.getOperand(2);
15591 SDValue Base = Op.getOperand(3);
15592 SDValue Index = Op.getOperand(4);
15593 SDValue Mask = Op.getOperand(5);
15594 SDValue Scale = Op.getOperand(6);
15595 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
15599 //scatter(base, mask, index, v1, scale);
15600 SDValue Chain = Op.getOperand(0);
15601 SDValue Base = Op.getOperand(2);
15602 SDValue Mask = Op.getOperand(3);
15603 SDValue Index = Op.getOperand(4);
15604 SDValue Src = Op.getOperand(5);
15605 SDValue Scale = Op.getOperand(6);
15606 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
15610 SDValue Hint = Op.getOperand(6);
15611 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
15612 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
15613 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
15614 SDValue Chain = Op.getOperand(0);
15615 SDValue Mask = Op.getOperand(2);
15616 SDValue Index = Op.getOperand(3);
15617 SDValue Base = Op.getOperand(4);
15618 SDValue Scale = Op.getOperand(5);
15619 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15621 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15623 SmallVector<SDValue, 2> Results;
15624 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
15626 return DAG.getMergeValues(Results, dl);
15628 // Read Performance Monitoring Counters.
15630 SmallVector<SDValue, 2> Results;
15631 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15632 return DAG.getMergeValues(Results, dl);
15634 // XTEST intrinsics.
15636 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15637 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15638 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15639 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
15641 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15642 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15643 Ret, SDValue(InTrans.getNode(), 1));
15647 SmallVector<SDValue, 2> Results;
15648 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15649 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15650 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15651 DAG.getConstant(-1, dl, MVT::i8));
15652 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15653 Op.getOperand(4), GenCF.getValue(1));
15654 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15655 Op.getOperand(5), MachinePointerInfo(),
15657 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15658 DAG.getConstant(X86::COND_B, dl, MVT::i8),
15660 Results.push_back(SetCC);
15661 Results.push_back(Store);
15662 return DAG.getMergeValues(Results, dl);
15664 case COMPRESS_TO_MEM: {
15666 SDValue Mask = Op.getOperand(4);
15667 SDValue DataToCompress = Op.getOperand(3);
15668 SDValue Addr = Op.getOperand(2);
15669 SDValue Chain = Op.getOperand(0);
15671 if (isAllOnes(Mask)) // return just a store
15672 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15673 MachinePointerInfo(), false, false, 0);
15675 EVT VT = DataToCompress.getValueType();
15676 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15677 VT.getVectorNumElements());
15678 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15679 Mask.getValueType().getSizeInBits());
15680 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15681 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15682 DAG.getIntPtrConstant(0, dl));
15684 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15685 DataToCompress, DAG.getUNDEF(VT));
15686 return DAG.getStore(Chain, dl, Compressed, Addr,
15687 MachinePointerInfo(), false, false, 0);
15689 case EXPAND_FROM_MEM: {
15691 SDValue Mask = Op.getOperand(4);
15692 SDValue PathThru = Op.getOperand(3);
15693 SDValue Addr = Op.getOperand(2);
15694 SDValue Chain = Op.getOperand(0);
15695 EVT VT = Op.getValueType();
15697 if (isAllOnes(Mask)) // return just a load
15698 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15700 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15701 VT.getVectorNumElements());
15702 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15703 Mask.getValueType().getSizeInBits());
15704 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15705 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15706 DAG.getIntPtrConstant(0, dl));
15708 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15709 false, false, false, 0);
15711 SDValue Results[] = {
15712 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15714 return DAG.getMergeValues(Results, dl);
15719 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15720 SelectionDAG &DAG) const {
15721 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15722 MFI->setReturnAddressIsTaken(true);
15724 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15727 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15729 EVT PtrVT = getPointerTy();
15732 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15733 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15734 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
15735 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15736 DAG.getNode(ISD::ADD, dl, PtrVT,
15737 FrameAddr, Offset),
15738 MachinePointerInfo(), false, false, false, 0);
15741 // Just load the return address.
15742 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15743 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15744 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15747 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15748 MachineFunction &MF = DAG.getMachineFunction();
15749 MachineFrameInfo *MFI = MF.getFrameInfo();
15750 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15751 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15752 EVT VT = Op.getValueType();
15754 MFI->setFrameAddressIsTaken(true);
15756 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15757 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15758 // is not possible to crawl up the stack without looking at the unwind codes
15760 int FrameAddrIndex = FuncInfo->getFAIndex();
15761 if (!FrameAddrIndex) {
15762 // Set up a frame object for the return address.
15763 unsigned SlotSize = RegInfo->getSlotSize();
15764 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15765 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
15766 FuncInfo->setFAIndex(FrameAddrIndex);
15768 return DAG.getFrameIndex(FrameAddrIndex, VT);
15771 unsigned FrameReg =
15772 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15773 SDLoc dl(Op); // FIXME probably not meaningful
15774 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15775 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15776 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15777 "Invalid Frame Register!");
15778 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15780 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15781 MachinePointerInfo(),
15782 false, false, false, 0);
15786 // FIXME? Maybe this could be a TableGen attribute on some registers and
15787 // this table could be generated automatically from RegInfo.
15788 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15790 unsigned Reg = StringSwitch<unsigned>(RegName)
15791 .Case("esp", X86::ESP)
15792 .Case("rsp", X86::RSP)
15796 report_fatal_error("Invalid register name global variable");
15799 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15800 SelectionDAG &DAG) const {
15801 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15802 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
15805 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15806 SDValue Chain = Op.getOperand(0);
15807 SDValue Offset = Op.getOperand(1);
15808 SDValue Handler = Op.getOperand(2);
15811 EVT PtrVT = getPointerTy();
15812 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15813 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15814 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15815 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15816 "Invalid Frame Register!");
15817 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15818 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15820 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15821 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
15823 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15824 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15826 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15828 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15829 DAG.getRegister(StoreAddrReg, PtrVT));
15832 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15833 SelectionDAG &DAG) const {
15835 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15836 DAG.getVTList(MVT::i32, MVT::Other),
15837 Op.getOperand(0), Op.getOperand(1));
15840 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15841 SelectionDAG &DAG) const {
15843 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15844 Op.getOperand(0), Op.getOperand(1));
15847 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15848 return Op.getOperand(0);
15851 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15852 SelectionDAG &DAG) const {
15853 SDValue Root = Op.getOperand(0);
15854 SDValue Trmp = Op.getOperand(1); // trampoline
15855 SDValue FPtr = Op.getOperand(2); // nested function
15856 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15859 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15860 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15862 if (Subtarget->is64Bit()) {
15863 SDValue OutChains[6];
15865 // Large code-model.
15866 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15867 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15869 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15870 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15872 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15874 // Load the pointer to the nested function into R11.
15875 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15876 SDValue Addr = Trmp;
15877 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15878 Addr, MachinePointerInfo(TrmpAddr),
15881 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15882 DAG.getConstant(2, dl, MVT::i64));
15883 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15884 MachinePointerInfo(TrmpAddr, 2),
15887 // Load the 'nest' parameter value into R10.
15888 // R10 is specified in X86CallingConv.td
15889 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15890 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15891 DAG.getConstant(10, dl, MVT::i64));
15892 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15893 Addr, MachinePointerInfo(TrmpAddr, 10),
15896 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15897 DAG.getConstant(12, dl, MVT::i64));
15898 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15899 MachinePointerInfo(TrmpAddr, 12),
15902 // Jump to the nested function.
15903 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15904 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15905 DAG.getConstant(20, dl, MVT::i64));
15906 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
15907 Addr, MachinePointerInfo(TrmpAddr, 20),
15910 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15911 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15912 DAG.getConstant(22, dl, MVT::i64));
15913 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
15914 Addr, MachinePointerInfo(TrmpAddr, 22),
15917 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15919 const Function *Func =
15920 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15921 CallingConv::ID CC = Func->getCallingConv();
15926 llvm_unreachable("Unsupported calling convention");
15927 case CallingConv::C:
15928 case CallingConv::X86_StdCall: {
15929 // Pass 'nest' parameter in ECX.
15930 // Must be kept in sync with X86CallingConv.td
15931 NestReg = X86::ECX;
15933 // Check that ECX wasn't needed by an 'inreg' parameter.
15934 FunctionType *FTy = Func->getFunctionType();
15935 const AttributeSet &Attrs = Func->getAttributes();
15937 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15938 unsigned InRegCount = 0;
15941 for (FunctionType::param_iterator I = FTy->param_begin(),
15942 E = FTy->param_end(); I != E; ++I, ++Idx)
15943 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15944 // FIXME: should only count parameters that are lowered to integers.
15945 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15947 if (InRegCount > 2) {
15948 report_fatal_error("Nest register in use - reduce number of inreg"
15954 case CallingConv::X86_FastCall:
15955 case CallingConv::X86_ThisCall:
15956 case CallingConv::Fast:
15957 // Pass 'nest' parameter in EAX.
15958 // Must be kept in sync with X86CallingConv.td
15959 NestReg = X86::EAX;
15963 SDValue OutChains[4];
15964 SDValue Addr, Disp;
15966 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15967 DAG.getConstant(10, dl, MVT::i32));
15968 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15970 // This is storing the opcode for MOV32ri.
15971 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15972 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15973 OutChains[0] = DAG.getStore(Root, dl,
15974 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
15975 Trmp, MachinePointerInfo(TrmpAddr),
15978 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15979 DAG.getConstant(1, dl, MVT::i32));
15980 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15981 MachinePointerInfo(TrmpAddr, 1),
15984 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15985 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15986 DAG.getConstant(5, dl, MVT::i32));
15987 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
15988 Addr, MachinePointerInfo(TrmpAddr, 5),
15991 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15992 DAG.getConstant(6, dl, MVT::i32));
15993 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15994 MachinePointerInfo(TrmpAddr, 6),
15997 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16001 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16002 SelectionDAG &DAG) const {
16004 The rounding mode is in bits 11:10 of FPSR, and has the following
16006 00 Round to nearest
16011 FLT_ROUNDS, on the other hand, expects the following:
16018 To perform the conversion, we do:
16019 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16022 MachineFunction &MF = DAG.getMachineFunction();
16023 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16024 unsigned StackAlignment = TFI.getStackAlignment();
16025 MVT VT = Op.getSimpleValueType();
16028 // Save FP Control Word to stack slot
16029 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16030 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
16032 MachineMemOperand *MMO =
16033 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
16034 MachineMemOperand::MOStore, 2, 2);
16036 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16037 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16038 DAG.getVTList(MVT::Other),
16039 Ops, MVT::i16, MMO);
16041 // Load FP Control Word from stack slot
16042 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16043 MachinePointerInfo(), false, false, false, 0);
16045 // Transform as necessary
16047 DAG.getNode(ISD::SRL, DL, MVT::i16,
16048 DAG.getNode(ISD::AND, DL, MVT::i16,
16049 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16050 DAG.getConstant(11, DL, MVT::i8));
16052 DAG.getNode(ISD::SRL, DL, MVT::i16,
16053 DAG.getNode(ISD::AND, DL, MVT::i16,
16054 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16055 DAG.getConstant(9, DL, MVT::i8));
16058 DAG.getNode(ISD::AND, DL, MVT::i16,
16059 DAG.getNode(ISD::ADD, DL, MVT::i16,
16060 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16061 DAG.getConstant(1, DL, MVT::i16)),
16062 DAG.getConstant(3, DL, MVT::i16));
16064 return DAG.getNode((VT.getSizeInBits() < 16 ?
16065 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16068 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16069 MVT VT = Op.getSimpleValueType();
16071 unsigned NumBits = VT.getSizeInBits();
16074 Op = Op.getOperand(0);
16075 if (VT == MVT::i8) {
16076 // Zero extend to i32 since there is not an i8 bsr.
16078 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16081 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16082 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16083 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16085 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16088 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16089 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16092 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16094 // Finally xor with NumBits-1.
16095 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16096 DAG.getConstant(NumBits - 1, dl, OpVT));
16099 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16103 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16104 MVT VT = Op.getSimpleValueType();
16106 unsigned NumBits = VT.getSizeInBits();
16109 Op = Op.getOperand(0);
16110 if (VT == MVT::i8) {
16111 // Zero extend to i32 since there is not an i8 bsr.
16113 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16116 // Issue a bsr (scan bits in reverse).
16117 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16118 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16120 // And xor with NumBits-1.
16121 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16122 DAG.getConstant(NumBits - 1, dl, OpVT));
16125 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16129 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16130 MVT VT = Op.getSimpleValueType();
16131 unsigned NumBits = VT.getSizeInBits();
16133 Op = Op.getOperand(0);
16135 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16136 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16137 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16139 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16142 DAG.getConstant(NumBits, dl, VT),
16143 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16146 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16149 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16150 // ones, and then concatenate the result back.
16151 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16152 MVT VT = Op.getSimpleValueType();
16154 assert(VT.is256BitVector() && VT.isInteger() &&
16155 "Unsupported value type for operation");
16157 unsigned NumElems = VT.getVectorNumElements();
16160 // Extract the LHS vectors
16161 SDValue LHS = Op.getOperand(0);
16162 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16163 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16165 // Extract the RHS vectors
16166 SDValue RHS = Op.getOperand(1);
16167 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16168 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16170 MVT EltVT = VT.getVectorElementType();
16171 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16173 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16174 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16175 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16178 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16179 assert(Op.getSimpleValueType().is256BitVector() &&
16180 Op.getSimpleValueType().isInteger() &&
16181 "Only handle AVX 256-bit vector integer operation");
16182 return Lower256IntArith(Op, DAG);
16185 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16186 assert(Op.getSimpleValueType().is256BitVector() &&
16187 Op.getSimpleValueType().isInteger() &&
16188 "Only handle AVX 256-bit vector integer operation");
16189 return Lower256IntArith(Op, DAG);
16192 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16193 SelectionDAG &DAG) {
16195 MVT VT = Op.getSimpleValueType();
16197 // Decompose 256-bit ops into smaller 128-bit ops.
16198 if (VT.is256BitVector() && !Subtarget->hasInt256())
16199 return Lower256IntArith(Op, DAG);
16201 SDValue A = Op.getOperand(0);
16202 SDValue B = Op.getOperand(1);
16204 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16205 // pairs, multiply and truncate.
16206 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16207 if (Subtarget->hasInt256()) {
16208 if (VT == MVT::v32i8) {
16209 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16210 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16211 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16212 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16213 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16214 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16215 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16216 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16217 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16218 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16221 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16222 return DAG.getNode(
16223 ISD::TRUNCATE, dl, VT,
16224 DAG.getNode(ISD::MUL, dl, ExVT,
16225 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16226 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16229 assert(VT == MVT::v16i8 &&
16230 "Pre-AVX2 support only supports v16i8 multiplication");
16231 MVT ExVT = MVT::v8i16;
16233 // Extract the lo parts and sign extend to i16
16235 if (Subtarget->hasSSE41()) {
16236 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16237 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16239 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16240 -1, 4, -1, 5, -1, 6, -1, 7};
16241 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16242 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16243 ALo = DAG.getNode(ISD::BITCAST, dl, ExVT, ALo);
16244 BLo = DAG.getNode(ISD::BITCAST, dl, ExVT, BLo);
16245 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16246 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16249 // Extract the hi parts and sign extend to i16
16251 if (Subtarget->hasSSE41()) {
16252 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16253 -1, -1, -1, -1, -1, -1, -1, -1};
16254 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16255 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16256 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16257 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16259 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16260 -1, 12, -1, 13, -1, 14, -1, 15};
16261 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16262 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16263 AHi = DAG.getNode(ISD::BITCAST, dl, ExVT, AHi);
16264 BHi = DAG.getNode(ISD::BITCAST, dl, ExVT, BHi);
16265 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16266 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16269 // Multiply, mask the lower 8bits of the lo/hi results and pack
16270 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16271 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16272 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16273 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16274 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16277 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16278 if (VT == MVT::v4i32) {
16279 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16280 "Should not custom lower when pmuldq is available!");
16282 // Extract the odd parts.
16283 static const int UnpackMask[] = { 1, -1, 3, -1 };
16284 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16285 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16287 // Multiply the even parts.
16288 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16289 // Now multiply odd parts.
16290 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16292 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
16293 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
16295 // Merge the two vectors back together with a shuffle. This expands into 2
16297 static const int ShufMask[] = { 0, 4, 2, 6 };
16298 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16301 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16302 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16304 // Ahi = psrlqi(a, 32);
16305 // Bhi = psrlqi(b, 32);
16307 // AloBlo = pmuludq(a, b);
16308 // AloBhi = pmuludq(a, Bhi);
16309 // AhiBlo = pmuludq(Ahi, b);
16311 // AloBhi = psllqi(AloBhi, 32);
16312 // AhiBlo = psllqi(AhiBlo, 32);
16313 // return AloBlo + AloBhi + AhiBlo;
16315 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16316 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16318 // Bit cast to 32-bit vectors for MULUDQ
16319 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16320 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16321 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
16322 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
16323 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
16324 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
16326 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
16327 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
16328 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
16330 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
16331 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
16333 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
16334 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
16337 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
16338 assert(Subtarget->isTargetWin64() && "Unexpected target");
16339 EVT VT = Op.getValueType();
16340 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
16341 "Unexpected return type for lowering");
16345 switch (Op->getOpcode()) {
16346 default: llvm_unreachable("Unexpected request for libcall!");
16347 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
16348 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
16349 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
16350 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
16351 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
16352 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
16356 SDValue InChain = DAG.getEntryNode();
16358 TargetLowering::ArgListTy Args;
16359 TargetLowering::ArgListEntry Entry;
16360 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
16361 EVT ArgVT = Op->getOperand(i).getValueType();
16362 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
16363 "Unexpected argument type for lowering");
16364 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
16365 Entry.Node = StackPtr;
16366 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
16368 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16369 Entry.Ty = PointerType::get(ArgTy,0);
16370 Entry.isSExt = false;
16371 Entry.isZExt = false;
16372 Args.push_back(Entry);
16375 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
16378 TargetLowering::CallLoweringInfo CLI(DAG);
16379 CLI.setDebugLoc(dl).setChain(InChain)
16380 .setCallee(getLibcallCallingConv(LC),
16381 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
16382 Callee, std::move(Args), 0)
16383 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
16385 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
16386 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
16389 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
16390 SelectionDAG &DAG) {
16391 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
16392 EVT VT = Op0.getValueType();
16395 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
16396 (VT == MVT::v8i32 && Subtarget->hasInt256()));
16398 // PMULxD operations multiply each even value (starting at 0) of LHS with
16399 // the related value of RHS and produce a widen result.
16400 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16401 // => <2 x i64> <ae|cg>
16403 // In other word, to have all the results, we need to perform two PMULxD:
16404 // 1. one with the even values.
16405 // 2. one with the odd values.
16406 // To achieve #2, with need to place the odd values at an even position.
16408 // Place the odd value at an even position (basically, shift all values 1
16409 // step to the left):
16410 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
16411 // <a|b|c|d> => <b|undef|d|undef>
16412 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
16413 // <e|f|g|h> => <f|undef|h|undef>
16414 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
16416 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
16418 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
16419 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
16421 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
16422 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
16423 // => <2 x i64> <ae|cg>
16424 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
16425 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
16426 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
16427 // => <2 x i64> <bf|dh>
16428 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
16429 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
16431 // Shuffle it back into the right order.
16432 SDValue Highs, Lows;
16433 if (VT == MVT::v8i32) {
16434 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
16435 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16436 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
16437 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16439 const int HighMask[] = {1, 5, 3, 7};
16440 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
16441 const int LowMask[] = {0, 4, 2, 6};
16442 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
16445 // If we have a signed multiply but no PMULDQ fix up the high parts of a
16446 // unsigned multiply.
16447 if (IsSigned && !Subtarget->hasSSE41()) {
16449 DAG.getConstant(31, dl,
16450 DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
16451 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
16452 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
16453 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
16454 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
16456 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
16457 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
16460 // The first result of MUL_LOHI is actually the low value, followed by the
16462 SDValue Ops[] = {Lows, Highs};
16463 return DAG.getMergeValues(Ops, dl);
16466 // Return true if the requred (according to Opcode) shift-imm form is natively
16467 // supported by the Subtarget
16468 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
16470 if (VT.getScalarSizeInBits() < 16)
16473 if (VT.is512BitVector() &&
16474 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
16477 bool LShift = VT.is128BitVector() ||
16478 (VT.is256BitVector() && Subtarget->hasInt256());
16480 bool AShift = LShift && (Subtarget->hasVLX() ||
16481 (VT != MVT::v2i64 && VT != MVT::v4i64));
16482 return (Opcode == ISD::SRA) ? AShift : LShift;
16485 // The shift amount is a variable, but it is the same for all vector lanes.
16486 // These instrcutions are defined together with shift-immediate.
16488 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
16490 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
16493 // Return true if the requred (according to Opcode) variable-shift form is
16494 // natively supported by the Subtarget
16495 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
16498 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
16501 // vXi16 supported only on AVX-512, BWI
16502 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
16505 if (VT.is512BitVector() || Subtarget->hasVLX())
16508 bool LShift = VT.is128BitVector() || VT.is256BitVector();
16509 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
16510 return (Opcode == ISD::SRA) ? AShift : LShift;
16513 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
16514 const X86Subtarget *Subtarget) {
16515 MVT VT = Op.getSimpleValueType();
16517 SDValue R = Op.getOperand(0);
16518 SDValue Amt = Op.getOperand(1);
16520 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16521 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16523 // Optimize shl/srl/sra with constant shift amount.
16524 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
16525 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
16526 uint64_t ShiftAmt = ShiftConst->getZExtValue();
16528 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
16529 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16531 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
16532 unsigned NumElts = VT.getVectorNumElements();
16533 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
16535 if (Op.getOpcode() == ISD::SHL) {
16536 // Make a large shift.
16537 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
16539 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
16540 // Zero out the rightmost bits.
16541 SmallVector<SDValue, 32> V(
16542 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
16543 return DAG.getNode(ISD::AND, dl, VT, SHL,
16544 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16546 if (Op.getOpcode() == ISD::SRL) {
16547 // Make a large shift.
16548 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
16550 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
16551 // Zero out the leftmost bits.
16552 SmallVector<SDValue, 32> V(
16553 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
16554 return DAG.getNode(ISD::AND, dl, VT, SRL,
16555 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
16557 if (Op.getOpcode() == ISD::SRA) {
16558 if (ShiftAmt == 7) {
16559 // R s>> 7 === R s< 0
16560 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
16561 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
16564 // R s>> a === ((R u>> a) ^ m) - m
16565 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
16566 SmallVector<SDValue, 32> V(NumElts,
16567 DAG.getConstant(128 >> ShiftAmt, dl,
16569 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
16570 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
16571 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
16574 llvm_unreachable("Unknown shift opcode.");
16579 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16580 if (!Subtarget->is64Bit() &&
16581 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
16582 Amt.getOpcode() == ISD::BITCAST &&
16583 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16584 Amt = Amt.getOperand(0);
16585 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16586 VT.getVectorNumElements();
16587 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
16588 uint64_t ShiftAmt = 0;
16589 for (unsigned i = 0; i != Ratio; ++i) {
16590 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
16594 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
16596 // Check remaining shift amounts.
16597 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16598 uint64_t ShAmt = 0;
16599 for (unsigned j = 0; j != Ratio; ++j) {
16600 ConstantSDNode *C =
16601 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
16605 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
16607 if (ShAmt != ShiftAmt)
16610 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
16616 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
16617 const X86Subtarget* Subtarget) {
16618 MVT VT = Op.getSimpleValueType();
16620 SDValue R = Op.getOperand(0);
16621 SDValue Amt = Op.getOperand(1);
16623 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
16624 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
16626 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
16627 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
16629 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
16631 EVT EltVT = VT.getVectorElementType();
16633 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
16634 // Check if this build_vector node is doing a splat.
16635 // If so, then set BaseShAmt equal to the splat value.
16636 BaseShAmt = BV->getSplatValue();
16637 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
16638 BaseShAmt = SDValue();
16640 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
16641 Amt = Amt.getOperand(0);
16643 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
16644 if (SVN && SVN->isSplat()) {
16645 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
16646 SDValue InVec = Amt.getOperand(0);
16647 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
16648 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
16649 "Unexpected shuffle index found!");
16650 BaseShAmt = InVec.getOperand(SplatIdx);
16651 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
16652 if (ConstantSDNode *C =
16653 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
16654 if (C->getZExtValue() == SplatIdx)
16655 BaseShAmt = InVec.getOperand(1);
16660 // Avoid introducing an extract element from a shuffle.
16661 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
16662 DAG.getIntPtrConstant(SplatIdx, dl));
16666 if (BaseShAmt.getNode()) {
16667 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
16668 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
16669 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
16670 else if (EltVT.bitsLT(MVT::i32))
16671 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
16673 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
16677 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
16678 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
16679 Amt.getOpcode() == ISD::BITCAST &&
16680 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16681 Amt = Amt.getOperand(0);
16682 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16683 VT.getVectorNumElements();
16684 std::vector<SDValue> Vals(Ratio);
16685 for (unsigned i = 0; i != Ratio; ++i)
16686 Vals[i] = Amt.getOperand(i);
16687 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16688 for (unsigned j = 0; j != Ratio; ++j)
16689 if (Vals[j] != Amt.getOperand(i + j))
16692 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
16697 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16698 SelectionDAG &DAG) {
16699 MVT VT = Op.getSimpleValueType();
16701 SDValue R = Op.getOperand(0);
16702 SDValue Amt = Op.getOperand(1);
16704 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16705 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16707 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
16710 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
16713 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
16716 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
16717 // shifts per-lane and then shuffle the partial results back together.
16718 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
16719 // Splat the shift amounts so the scalar shifts above will catch it.
16720 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
16721 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
16722 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
16723 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
16724 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
16727 // If possible, lower this packed shift into a vector multiply instead of
16728 // expanding it into a sequence of scalar shifts.
16729 // Do this only if the vector shift count is a constant build_vector.
16730 if (Op.getOpcode() == ISD::SHL &&
16731 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16732 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16733 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16734 SmallVector<SDValue, 8> Elts;
16735 EVT SVT = VT.getScalarType();
16736 unsigned SVTBits = SVT.getSizeInBits();
16737 const APInt &One = APInt(SVTBits, 1);
16738 unsigned NumElems = VT.getVectorNumElements();
16740 for (unsigned i=0; i !=NumElems; ++i) {
16741 SDValue Op = Amt->getOperand(i);
16742 if (Op->getOpcode() == ISD::UNDEF) {
16743 Elts.push_back(Op);
16747 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16748 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16749 uint64_t ShAmt = C.getZExtValue();
16750 if (ShAmt >= SVTBits) {
16751 Elts.push_back(DAG.getUNDEF(SVT));
16754 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
16756 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16757 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16760 // Lower SHL with variable shift amount.
16761 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16762 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
16764 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
16765 DAG.getConstant(0x3f800000U, dl, VT));
16766 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16767 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16768 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16771 // If possible, lower this shift as a sequence of two shifts by
16772 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16774 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16776 // Could be rewritten as:
16777 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16779 // The advantage is that the two shifts from the example would be
16780 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16781 // the vector shift into four scalar shifts plus four pairs of vector
16783 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16784 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16785 unsigned TargetOpcode = X86ISD::MOVSS;
16786 bool CanBeSimplified;
16787 // The splat value for the first packed shift (the 'X' from the example).
16788 SDValue Amt1 = Amt->getOperand(0);
16789 // The splat value for the second packed shift (the 'Y' from the example).
16790 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16791 Amt->getOperand(2);
16793 // See if it is possible to replace this node with a sequence of
16794 // two shifts followed by a MOVSS/MOVSD
16795 if (VT == MVT::v4i32) {
16796 // Check if it is legal to use a MOVSS.
16797 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16798 Amt2 == Amt->getOperand(3);
16799 if (!CanBeSimplified) {
16800 // Otherwise, check if we can still simplify this node using a MOVSD.
16801 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16802 Amt->getOperand(2) == Amt->getOperand(3);
16803 TargetOpcode = X86ISD::MOVSD;
16804 Amt2 = Amt->getOperand(2);
16807 // Do similar checks for the case where the machine value type
16809 CanBeSimplified = Amt1 == Amt->getOperand(1);
16810 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16811 CanBeSimplified = Amt2 == Amt->getOperand(i);
16813 if (!CanBeSimplified) {
16814 TargetOpcode = X86ISD::MOVSD;
16815 CanBeSimplified = true;
16816 Amt2 = Amt->getOperand(4);
16817 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16818 CanBeSimplified = Amt1 == Amt->getOperand(i);
16819 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16820 CanBeSimplified = Amt2 == Amt->getOperand(j);
16824 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16825 isa<ConstantSDNode>(Amt2)) {
16826 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16827 EVT CastVT = MVT::v4i32;
16829 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
16830 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16832 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
16833 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16834 if (TargetOpcode == X86ISD::MOVSD)
16835 CastVT = MVT::v2i64;
16836 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16837 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16838 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16840 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16844 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16845 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
16846 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, dl, VT));
16848 SDValue VSelM = DAG.getConstant(0x80, dl, VT);
16849 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16850 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16852 // r = VSELECT(r, shl(r, 4), a);
16853 SDValue M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(4, dl, VT));
16854 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16857 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16858 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16859 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16861 // r = VSELECT(r, shl(r, 2), a);
16862 M = DAG.getNode(ISD::SHL, dl, VT, R, DAG.getConstant(2, dl, VT));
16863 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16866 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16867 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16868 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16870 // return VSELECT(r, r+r, a);
16871 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16872 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16876 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16877 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16878 // solution better.
16879 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16880 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16882 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16883 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16884 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16885 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16886 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16889 // Decompose 256-bit shifts into smaller 128-bit shifts.
16890 if (VT.is256BitVector()) {
16891 unsigned NumElems = VT.getVectorNumElements();
16892 MVT EltVT = VT.getVectorElementType();
16893 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16895 // Extract the two vectors
16896 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16897 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16899 // Recreate the shift amount vectors
16900 SDValue Amt1, Amt2;
16901 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16902 // Constant shift amount
16903 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
16904 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
16905 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
16907 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16908 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16910 // Variable shift amount
16911 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16912 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16915 // Issue new vector shifts for the smaller types
16916 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16917 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16919 // Concatenate the result back
16920 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16926 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16927 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16928 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16929 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16930 // has only one use.
16931 SDNode *N = Op.getNode();
16932 SDValue LHS = N->getOperand(0);
16933 SDValue RHS = N->getOperand(1);
16934 unsigned BaseOp = 0;
16937 switch (Op.getOpcode()) {
16938 default: llvm_unreachable("Unknown ovf instruction!");
16940 // A subtract of one will be selected as a INC. Note that INC doesn't
16941 // set CF, so we can't do this for UADDO.
16942 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16944 BaseOp = X86ISD::INC;
16945 Cond = X86::COND_O;
16948 BaseOp = X86ISD::ADD;
16949 Cond = X86::COND_O;
16952 BaseOp = X86ISD::ADD;
16953 Cond = X86::COND_B;
16956 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16957 // set CF, so we can't do this for USUBO.
16958 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16960 BaseOp = X86ISD::DEC;
16961 Cond = X86::COND_O;
16964 BaseOp = X86ISD::SUB;
16965 Cond = X86::COND_O;
16968 BaseOp = X86ISD::SUB;
16969 Cond = X86::COND_B;
16972 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16973 Cond = X86::COND_O;
16975 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16976 if (N->getValueType(0) == MVT::i8) {
16977 BaseOp = X86ISD::UMUL8;
16978 Cond = X86::COND_O;
16981 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16983 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16986 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16987 DAG.getConstant(X86::COND_O, DL, MVT::i32),
16988 SDValue(Sum.getNode(), 2));
16990 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16994 // Also sets EFLAGS.
16995 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16996 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16999 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17000 DAG.getConstant(Cond, DL, MVT::i32),
17001 SDValue(Sum.getNode(), 1));
17003 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17006 /// Returns true if the operand type is exactly twice the native width, and
17007 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
17008 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
17009 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
17010 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
17011 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
17014 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
17015 else if (OpWidth == 128)
17016 return Subtarget->hasCmpxchg16b();
17021 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
17022 return needsCmpXchgNb(SI->getValueOperand()->getType());
17025 // Note: this turns large loads into lock cmpxchg8b/16b.
17026 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
17027 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
17028 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
17029 return needsCmpXchgNb(PTy->getElementType());
17032 TargetLoweringBase::AtomicRMWExpansionKind
17033 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
17034 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17035 const Type *MemType = AI->getType();
17037 // If the operand is too big, we must see if cmpxchg8/16b is available
17038 // and default to library calls otherwise.
17039 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
17040 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
17041 : AtomicRMWExpansionKind::None;
17044 AtomicRMWInst::BinOp Op = AI->getOperation();
17047 llvm_unreachable("Unknown atomic operation");
17048 case AtomicRMWInst::Xchg:
17049 case AtomicRMWInst::Add:
17050 case AtomicRMWInst::Sub:
17051 // It's better to use xadd, xsub or xchg for these in all cases.
17052 return AtomicRMWExpansionKind::None;
17053 case AtomicRMWInst::Or:
17054 case AtomicRMWInst::And:
17055 case AtomicRMWInst::Xor:
17056 // If the atomicrmw's result isn't actually used, we can just add a "lock"
17057 // prefix to a normal instruction for these operations.
17058 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
17059 : AtomicRMWExpansionKind::None;
17060 case AtomicRMWInst::Nand:
17061 case AtomicRMWInst::Max:
17062 case AtomicRMWInst::Min:
17063 case AtomicRMWInst::UMax:
17064 case AtomicRMWInst::UMin:
17065 // These always require a non-trivial set of data operations on x86. We must
17066 // use a cmpxchg loop.
17067 return AtomicRMWExpansionKind::CmpXChg;
17071 static bool hasMFENCE(const X86Subtarget& Subtarget) {
17072 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
17073 // no-sse2). There isn't any reason to disable it if the target processor
17075 return Subtarget.hasSSE2() || Subtarget.is64Bit();
17079 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
17080 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17081 const Type *MemType = AI->getType();
17082 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
17083 // there is no benefit in turning such RMWs into loads, and it is actually
17084 // harmful as it introduces a mfence.
17085 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
17088 auto Builder = IRBuilder<>(AI);
17089 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
17090 auto SynchScope = AI->getSynchScope();
17091 // We must restrict the ordering to avoid generating loads with Release or
17092 // ReleaseAcquire orderings.
17093 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
17094 auto Ptr = AI->getPointerOperand();
17096 // Before the load we need a fence. Here is an example lifted from
17097 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
17100 // x.store(1, relaxed);
17101 // r1 = y.fetch_add(0, release);
17103 // y.fetch_add(42, acquire);
17104 // r2 = x.load(relaxed);
17105 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
17106 // lowered to just a load without a fence. A mfence flushes the store buffer,
17107 // making the optimization clearly correct.
17108 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
17109 // otherwise, we might be able to be more agressive on relaxed idempotent
17110 // rmw. In practice, they do not look useful, so we don't try to be
17111 // especially clever.
17112 if (SynchScope == SingleThread)
17113 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
17114 // the IR level, so we must wrap it in an intrinsic.
17117 if (!hasMFENCE(*Subtarget))
17118 // FIXME: it might make sense to use a locked operation here but on a
17119 // different cache-line to prevent cache-line bouncing. In practice it
17120 // is probably a small win, and x86 processors without mfence are rare
17121 // enough that we do not bother.
17125 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
17126 Builder.CreateCall(MFence, {});
17128 // Finally we can emit the atomic load.
17129 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
17130 AI->getType()->getPrimitiveSizeInBits());
17131 Loaded->setAtomic(Order, SynchScope);
17132 AI->replaceAllUsesWith(Loaded);
17133 AI->eraseFromParent();
17137 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
17138 SelectionDAG &DAG) {
17140 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
17141 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
17142 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
17143 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
17145 // The only fence that needs an instruction is a sequentially-consistent
17146 // cross-thread fence.
17147 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
17148 if (hasMFENCE(*Subtarget))
17149 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
17151 SDValue Chain = Op.getOperand(0);
17152 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
17154 DAG.getRegister(X86::ESP, MVT::i32), // Base
17155 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
17156 DAG.getRegister(0, MVT::i32), // Index
17157 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
17158 DAG.getRegister(0, MVT::i32), // Segment.
17162 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
17163 return SDValue(Res, 0);
17166 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
17167 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
17170 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
17171 SelectionDAG &DAG) {
17172 MVT T = Op.getSimpleValueType();
17176 switch(T.SimpleTy) {
17177 default: llvm_unreachable("Invalid value type!");
17178 case MVT::i8: Reg = X86::AL; size = 1; break;
17179 case MVT::i16: Reg = X86::AX; size = 2; break;
17180 case MVT::i32: Reg = X86::EAX; size = 4; break;
17182 assert(Subtarget->is64Bit() && "Node not type legal!");
17183 Reg = X86::RAX; size = 8;
17186 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
17187 Op.getOperand(2), SDValue());
17188 SDValue Ops[] = { cpIn.getValue(0),
17191 DAG.getTargetConstant(size, DL, MVT::i8),
17192 cpIn.getValue(1) };
17193 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17194 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
17195 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
17199 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
17200 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
17201 MVT::i32, cpOut.getValue(2));
17202 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
17203 DAG.getConstant(X86::COND_E, DL, MVT::i8),
17206 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
17207 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
17208 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
17212 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
17213 SelectionDAG &DAG) {
17214 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
17215 MVT DstVT = Op.getSimpleValueType();
17217 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
17218 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17219 if (DstVT != MVT::f64)
17220 // This conversion needs to be expanded.
17223 SDValue InVec = Op->getOperand(0);
17225 unsigned NumElts = SrcVT.getVectorNumElements();
17226 EVT SVT = SrcVT.getVectorElementType();
17228 // Widen the vector in input in the case of MVT::v2i32.
17229 // Example: from MVT::v2i32 to MVT::v4i32.
17230 SmallVector<SDValue, 16> Elts;
17231 for (unsigned i = 0, e = NumElts; i != e; ++i)
17232 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
17233 DAG.getIntPtrConstant(i, dl)));
17235 // Explicitly mark the extra elements as Undef.
17236 Elts.append(NumElts, DAG.getUNDEF(SVT));
17238 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17239 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
17240 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
17241 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
17242 DAG.getIntPtrConstant(0, dl));
17245 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
17246 Subtarget->hasMMX() && "Unexpected custom BITCAST");
17247 assert((DstVT == MVT::i64 ||
17248 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
17249 "Unexpected custom BITCAST");
17250 // i64 <=> MMX conversions are Legal.
17251 if (SrcVT==MVT::i64 && DstVT.isVector())
17253 if (DstVT==MVT::i64 && SrcVT.isVector())
17255 // MMX <=> MMX conversions are Legal.
17256 if (SrcVT.isVector() && DstVT.isVector())
17258 // All other conversions need to be expanded.
17262 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
17263 SelectionDAG &DAG) {
17264 SDNode *Node = Op.getNode();
17267 Op = Op.getOperand(0);
17268 EVT VT = Op.getValueType();
17269 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17270 "CTPOP lowering only implemented for 128/256-bit wide vector types");
17272 unsigned NumElts = VT.getVectorNumElements();
17273 EVT EltVT = VT.getVectorElementType();
17274 unsigned Len = EltVT.getSizeInBits();
17276 // This is the vectorized version of the "best" algorithm from
17277 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
17278 // with a minor tweak to use a series of adds + shifts instead of vector
17279 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
17281 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
17282 // v8i32 => Always profitable
17284 // FIXME: There a couple of possible improvements:
17286 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
17287 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
17289 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
17290 "CTPOP not implemented for this vector element type.");
17292 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
17293 // extra legalization.
17294 bool NeedsBitcast = EltVT == MVT::i32;
17295 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
17297 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), dl,
17299 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), dl,
17301 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), dl,
17304 // v = v - ((v >> 1) & 0x55555555...)
17305 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, dl, EltVT));
17306 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
17307 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
17309 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17311 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
17312 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
17314 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
17316 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
17317 if (VT != And.getValueType())
17318 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17319 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
17321 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
17322 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
17323 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
17324 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, dl, EltVT));
17325 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
17327 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
17328 if (NeedsBitcast) {
17329 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
17330 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
17331 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
17334 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
17335 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
17336 if (VT != AndRHS.getValueType()) {
17337 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
17338 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
17340 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
17342 // v = (v + (v >> 4)) & 0x0F0F0F0F...
17343 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, dl, EltVT));
17344 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
17345 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
17346 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17348 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
17349 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
17350 if (NeedsBitcast) {
17351 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17352 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
17354 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
17355 if (VT != And.getValueType())
17356 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17358 // The algorithm mentioned above uses:
17359 // v = (v * 0x01010101...) >> (Len - 8)
17361 // Change it to use vector adds + vector shifts which yield faster results on
17362 // Haswell than using vector integer multiplication.
17364 // For i32 elements:
17365 // v = v + (v >> 8)
17366 // v = v + (v >> 16)
17368 // For i64 elements:
17369 // v = v + (v >> 8)
17370 // v = v + (v >> 16)
17371 // v = v + (v >> 32)
17374 SmallVector<SDValue, 8> Csts;
17375 for (unsigned i = 8; i <= Len/2; i *= 2) {
17376 Csts.assign(NumElts, DAG.getConstant(i, dl, EltVT));
17377 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
17378 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
17379 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
17383 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
17384 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), dl,
17386 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
17387 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
17388 if (NeedsBitcast) {
17389 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
17390 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
17392 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
17393 if (VT != And.getValueType())
17394 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
17399 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
17400 SDNode *Node = Op.getNode();
17402 EVT T = Node->getValueType(0);
17403 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
17404 DAG.getConstant(0, dl, T), Node->getOperand(2));
17405 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
17406 cast<AtomicSDNode>(Node)->getMemoryVT(),
17407 Node->getOperand(0),
17408 Node->getOperand(1), negOp,
17409 cast<AtomicSDNode>(Node)->getMemOperand(),
17410 cast<AtomicSDNode>(Node)->getOrdering(),
17411 cast<AtomicSDNode>(Node)->getSynchScope());
17414 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
17415 SDNode *Node = Op.getNode();
17417 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
17419 // Convert seq_cst store -> xchg
17420 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
17421 // FIXME: On 32-bit, store -> fist or movq would be more efficient
17422 // (The only way to get a 16-byte store is cmpxchg16b)
17423 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
17424 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
17425 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17426 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
17427 cast<AtomicSDNode>(Node)->getMemoryVT(),
17428 Node->getOperand(0),
17429 Node->getOperand(1), Node->getOperand(2),
17430 cast<AtomicSDNode>(Node)->getMemOperand(),
17431 cast<AtomicSDNode>(Node)->getOrdering(),
17432 cast<AtomicSDNode>(Node)->getSynchScope());
17433 return Swap.getValue(1);
17435 // Other atomic stores have a simple pattern.
17439 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
17440 EVT VT = Op.getNode()->getSimpleValueType(0);
17442 // Let legalize expand this if it isn't a legal type yet.
17443 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17446 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17449 bool ExtraOp = false;
17450 switch (Op.getOpcode()) {
17451 default: llvm_unreachable("Invalid code");
17452 case ISD::ADDC: Opc = X86ISD::ADD; break;
17453 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
17454 case ISD::SUBC: Opc = X86ISD::SUB; break;
17455 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
17459 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17461 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
17462 Op.getOperand(1), Op.getOperand(2));
17465 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
17466 SelectionDAG &DAG) {
17467 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
17469 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
17470 // which returns the values as { float, float } (in XMM0) or
17471 // { double, double } (which is returned in XMM0, XMM1).
17473 SDValue Arg = Op.getOperand(0);
17474 EVT ArgVT = Arg.getValueType();
17475 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17477 TargetLowering::ArgListTy Args;
17478 TargetLowering::ArgListEntry Entry;
17482 Entry.isSExt = false;
17483 Entry.isZExt = false;
17484 Args.push_back(Entry);
17486 bool isF64 = ArgVT == MVT::f64;
17487 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
17488 // the small struct {f32, f32} is returned in (eax, edx). For f64,
17489 // the results are returned via SRet in memory.
17490 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
17491 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17492 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
17494 Type *RetTy = isF64
17495 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
17496 : (Type*)VectorType::get(ArgTy, 4);
17498 TargetLowering::CallLoweringInfo CLI(DAG);
17499 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
17500 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
17502 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
17505 // Returned in xmm0 and xmm1.
17506 return CallResult.first;
17508 // Returned in bits 0:31 and 32:64 xmm0.
17509 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17510 CallResult.first, DAG.getIntPtrConstant(0, dl));
17511 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
17512 CallResult.first, DAG.getIntPtrConstant(1, dl));
17513 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
17514 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
17517 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
17518 SelectionDAG &DAG) {
17519 assert(Subtarget->hasAVX512() &&
17520 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17522 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
17523 EVT VT = N->getValue().getValueType();
17524 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
17527 // X86 scatter kills mask register, so its type should be added to
17528 // the list of return values
17529 if (N->getNumValues() == 1) {
17530 SDValue Index = N->getIndex();
17531 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17532 !Index.getValueType().is512BitVector())
17533 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17535 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
17536 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17537 N->getOperand(3), Index };
17539 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
17540 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
17541 return SDValue(NewScatter.getNode(), 0);
17546 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
17547 SelectionDAG &DAG) {
17548 assert(Subtarget->hasAVX512() &&
17549 "MGATHER/MSCATTER are supported on AVX-512 arch only");
17551 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
17552 EVT VT = Op.getValueType();
17553 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
17556 SDValue Index = N->getIndex();
17557 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
17558 !Index.getValueType().is512BitVector()) {
17559 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
17560 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
17561 N->getOperand(3), Index };
17562 DAG.UpdateNodeOperands(N, Ops);
17567 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
17568 SelectionDAG &DAG) const {
17569 // TODO: Eventually, the lowering of these nodes should be informed by or
17570 // deferred to the GC strategy for the function in which they appear. For
17571 // now, however, they must be lowered to something. Since they are logically
17572 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17573 // require special handling for these nodes), lower them as literal NOOPs for
17575 SmallVector<SDValue, 2> Ops;
17577 Ops.push_back(Op.getOperand(0));
17578 if (Op->getGluedNode())
17579 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17582 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17583 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17588 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
17589 SelectionDAG &DAG) const {
17590 // TODO: Eventually, the lowering of these nodes should be informed by or
17591 // deferred to the GC strategy for the function in which they appear. For
17592 // now, however, they must be lowered to something. Since they are logically
17593 // no-ops in the case of a null GC strategy (or a GC strategy which does not
17594 // require special handling for these nodes), lower them as literal NOOPs for
17596 SmallVector<SDValue, 2> Ops;
17598 Ops.push_back(Op.getOperand(0));
17599 if (Op->getGluedNode())
17600 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
17603 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
17604 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
17609 /// LowerOperation - Provide custom lowering hooks for some operations.
17611 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
17612 switch (Op.getOpcode()) {
17613 default: llvm_unreachable("Should not custom lower this!");
17614 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
17615 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
17616 return LowerCMP_SWAP(Op, Subtarget, DAG);
17617 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
17618 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
17619 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
17620 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
17621 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
17622 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
17623 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
17624 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
17625 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
17626 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
17627 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
17628 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
17629 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
17630 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
17631 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
17632 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
17633 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
17634 case ISD::SHL_PARTS:
17635 case ISD::SRA_PARTS:
17636 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
17637 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
17638 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
17639 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
17640 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
17641 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
17642 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
17643 case ISD::SIGN_EXTEND_VECTOR_INREG:
17644 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
17645 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
17646 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
17647 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
17648 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
17650 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
17651 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
17652 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
17653 case ISD::SETCC: return LowerSETCC(Op, DAG);
17654 case ISD::SELECT: return LowerSELECT(Op, DAG);
17655 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
17656 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
17657 case ISD::VASTART: return LowerVASTART(Op, DAG);
17658 case ISD::VAARG: return LowerVAARG(Op, DAG);
17659 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
17660 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
17661 case ISD::INTRINSIC_VOID:
17662 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
17663 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
17664 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
17665 case ISD::FRAME_TO_ARGS_OFFSET:
17666 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
17667 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
17668 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
17669 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
17670 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
17671 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
17672 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
17673 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
17674 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
17675 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
17676 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
17677 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
17678 case ISD::UMUL_LOHI:
17679 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
17682 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
17688 case ISD::UMULO: return LowerXALUO(Op, DAG);
17689 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
17690 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
17694 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
17695 case ISD::ADD: return LowerADD(Op, DAG);
17696 case ISD::SUB: return LowerSUB(Op, DAG);
17697 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
17698 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
17699 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
17700 case ISD::GC_TRANSITION_START:
17701 return LowerGC_TRANSITION_START(Op, DAG);
17702 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
17706 /// ReplaceNodeResults - Replace a node with an illegal result type
17707 /// with a new node built out of custom code.
17708 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
17709 SmallVectorImpl<SDValue>&Results,
17710 SelectionDAG &DAG) const {
17712 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17713 switch (N->getOpcode()) {
17715 llvm_unreachable("Do not know how to custom type legalize this operation!");
17716 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17717 case X86ISD::FMINC:
17719 case X86ISD::FMAXC:
17720 case X86ISD::FMAX: {
17721 EVT VT = N->getValueType(0);
17722 if (VT != MVT::v2f32)
17723 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17724 SDValue UNDEF = DAG.getUNDEF(VT);
17725 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17726 N->getOperand(0), UNDEF);
17727 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17728 N->getOperand(1), UNDEF);
17729 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17732 case ISD::SIGN_EXTEND_INREG:
17737 // We don't want to expand or promote these.
17744 case ISD::UDIVREM: {
17745 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17746 Results.push_back(V);
17749 case ISD::FP_TO_SINT:
17750 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
17751 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
17752 if (N->getOperand(0).getValueType() == MVT::f16)
17755 case ISD::FP_TO_UINT: {
17756 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17758 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17761 std::pair<SDValue,SDValue> Vals =
17762 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17763 SDValue FIST = Vals.first, StackSlot = Vals.second;
17764 if (FIST.getNode()) {
17765 EVT VT = N->getValueType(0);
17766 // Return a load from the stack slot.
17767 if (StackSlot.getNode())
17768 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17769 MachinePointerInfo(),
17770 false, false, false, 0));
17772 Results.push_back(FIST);
17776 case ISD::UINT_TO_FP: {
17777 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17778 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17779 N->getValueType(0) != MVT::v2f32)
17781 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17783 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
17785 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17786 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17787 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17788 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17789 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17790 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17793 case ISD::FP_ROUND: {
17794 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17796 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17797 Results.push_back(V);
17800 case ISD::FP_EXTEND: {
17801 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
17802 // No other ValueType for FP_EXTEND should reach this point.
17803 assert(N->getValueType(0) == MVT::v2f32 &&
17804 "Do not know how to legalize this Node");
17807 case ISD::INTRINSIC_W_CHAIN: {
17808 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17810 default : llvm_unreachable("Do not know how to custom type "
17811 "legalize this intrinsic operation!");
17812 case Intrinsic::x86_rdtsc:
17813 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17815 case Intrinsic::x86_rdtscp:
17816 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17818 case Intrinsic::x86_rdpmc:
17819 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17822 case ISD::READCYCLECOUNTER: {
17823 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17826 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17827 EVT T = N->getValueType(0);
17828 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17829 bool Regs64bit = T == MVT::i128;
17830 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17831 SDValue cpInL, cpInH;
17832 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17833 DAG.getConstant(0, dl, HalfT));
17834 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17835 DAG.getConstant(1, dl, HalfT));
17836 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17837 Regs64bit ? X86::RAX : X86::EAX,
17839 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17840 Regs64bit ? X86::RDX : X86::EDX,
17841 cpInH, cpInL.getValue(1));
17842 SDValue swapInL, swapInH;
17843 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17844 DAG.getConstant(0, dl, HalfT));
17845 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17846 DAG.getConstant(1, dl, HalfT));
17847 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17848 Regs64bit ? X86::RBX : X86::EBX,
17849 swapInL, cpInH.getValue(1));
17850 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17851 Regs64bit ? X86::RCX : X86::ECX,
17852 swapInH, swapInL.getValue(1));
17853 SDValue Ops[] = { swapInH.getValue(0),
17855 swapInH.getValue(1) };
17856 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17857 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17858 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17859 X86ISD::LCMPXCHG8_DAG;
17860 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17861 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17862 Regs64bit ? X86::RAX : X86::EAX,
17863 HalfT, Result.getValue(1));
17864 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17865 Regs64bit ? X86::RDX : X86::EDX,
17866 HalfT, cpOutL.getValue(2));
17867 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17869 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17870 MVT::i32, cpOutH.getValue(2));
17872 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17873 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
17874 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17876 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17877 Results.push_back(Success);
17878 Results.push_back(EFLAGS.getValue(1));
17881 case ISD::ATOMIC_SWAP:
17882 case ISD::ATOMIC_LOAD_ADD:
17883 case ISD::ATOMIC_LOAD_SUB:
17884 case ISD::ATOMIC_LOAD_AND:
17885 case ISD::ATOMIC_LOAD_OR:
17886 case ISD::ATOMIC_LOAD_XOR:
17887 case ISD::ATOMIC_LOAD_NAND:
17888 case ISD::ATOMIC_LOAD_MIN:
17889 case ISD::ATOMIC_LOAD_MAX:
17890 case ISD::ATOMIC_LOAD_UMIN:
17891 case ISD::ATOMIC_LOAD_UMAX:
17892 case ISD::ATOMIC_LOAD: {
17893 // Delegate to generic TypeLegalization. Situations we can really handle
17894 // should have already been dealt with by AtomicExpandPass.cpp.
17897 case ISD::BITCAST: {
17898 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17899 EVT DstVT = N->getValueType(0);
17900 EVT SrcVT = N->getOperand(0)->getValueType(0);
17902 if (SrcVT != MVT::f64 ||
17903 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17906 unsigned NumElts = DstVT.getVectorNumElements();
17907 EVT SVT = DstVT.getVectorElementType();
17908 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17909 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17910 MVT::v2f64, N->getOperand(0));
17911 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17913 if (ExperimentalVectorWideningLegalization) {
17914 // If we are legalizing vectors by widening, we already have the desired
17915 // legal vector type, just return it.
17916 Results.push_back(ToVecInt);
17920 SmallVector<SDValue, 8> Elts;
17921 for (unsigned i = 0, e = NumElts; i != e; ++i)
17922 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17923 ToVecInt, DAG.getIntPtrConstant(i, dl)));
17925 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17930 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17931 switch ((X86ISD::NodeType)Opcode) {
17932 case X86ISD::FIRST_NUMBER: break;
17933 case X86ISD::BSF: return "X86ISD::BSF";
17934 case X86ISD::BSR: return "X86ISD::BSR";
17935 case X86ISD::SHLD: return "X86ISD::SHLD";
17936 case X86ISD::SHRD: return "X86ISD::SHRD";
17937 case X86ISD::FAND: return "X86ISD::FAND";
17938 case X86ISD::FANDN: return "X86ISD::FANDN";
17939 case X86ISD::FOR: return "X86ISD::FOR";
17940 case X86ISD::FXOR: return "X86ISD::FXOR";
17941 case X86ISD::FSRL: return "X86ISD::FSRL";
17942 case X86ISD::FILD: return "X86ISD::FILD";
17943 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17944 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17945 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17946 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17947 case X86ISD::FLD: return "X86ISD::FLD";
17948 case X86ISD::FST: return "X86ISD::FST";
17949 case X86ISD::CALL: return "X86ISD::CALL";
17950 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17951 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17952 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17953 case X86ISD::BT: return "X86ISD::BT";
17954 case X86ISD::CMP: return "X86ISD::CMP";
17955 case X86ISD::COMI: return "X86ISD::COMI";
17956 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17957 case X86ISD::CMPM: return "X86ISD::CMPM";
17958 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17959 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
17960 case X86ISD::SETCC: return "X86ISD::SETCC";
17961 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17962 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17963 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
17964 case X86ISD::CMOV: return "X86ISD::CMOV";
17965 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17966 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17967 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17968 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17969 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17970 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17971 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17972 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
17973 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
17974 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
17975 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17976 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17977 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17978 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17979 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17980 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
17981 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17982 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17983 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17984 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17985 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17986 case X86ISD::ADDUS: return "X86ISD::ADDUS";
17987 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17988 case X86ISD::HADD: return "X86ISD::HADD";
17989 case X86ISD::HSUB: return "X86ISD::HSUB";
17990 case X86ISD::FHADD: return "X86ISD::FHADD";
17991 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17992 case X86ISD::UMAX: return "X86ISD::UMAX";
17993 case X86ISD::UMIN: return "X86ISD::UMIN";
17994 case X86ISD::SMAX: return "X86ISD::SMAX";
17995 case X86ISD::SMIN: return "X86ISD::SMIN";
17996 case X86ISD::FMAX: return "X86ISD::FMAX";
17997 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
17998 case X86ISD::FMIN: return "X86ISD::FMIN";
17999 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
18000 case X86ISD::FMAXC: return "X86ISD::FMAXC";
18001 case X86ISD::FMINC: return "X86ISD::FMINC";
18002 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
18003 case X86ISD::FRCP: return "X86ISD::FRCP";
18004 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
18005 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
18006 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
18007 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
18008 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
18009 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
18010 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
18011 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
18012 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
18013 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
18014 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
18015 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
18016 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
18017 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
18018 case X86ISD::VZEXT: return "X86ISD::VZEXT";
18019 case X86ISD::VSEXT: return "X86ISD::VSEXT";
18020 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
18021 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
18022 case X86ISD::VINSERT: return "X86ISD::VINSERT";
18023 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
18024 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
18025 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
18026 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
18027 case X86ISD::VSHL: return "X86ISD::VSHL";
18028 case X86ISD::VSRL: return "X86ISD::VSRL";
18029 case X86ISD::VSRA: return "X86ISD::VSRA";
18030 case X86ISD::VSHLI: return "X86ISD::VSHLI";
18031 case X86ISD::VSRLI: return "X86ISD::VSRLI";
18032 case X86ISD::VSRAI: return "X86ISD::VSRAI";
18033 case X86ISD::CMPP: return "X86ISD::CMPP";
18034 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
18035 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
18036 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
18037 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
18038 case X86ISD::ADD: return "X86ISD::ADD";
18039 case X86ISD::SUB: return "X86ISD::SUB";
18040 case X86ISD::ADC: return "X86ISD::ADC";
18041 case X86ISD::SBB: return "X86ISD::SBB";
18042 case X86ISD::SMUL: return "X86ISD::SMUL";
18043 case X86ISD::UMUL: return "X86ISD::UMUL";
18044 case X86ISD::SMUL8: return "X86ISD::SMUL8";
18045 case X86ISD::UMUL8: return "X86ISD::UMUL8";
18046 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
18047 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
18048 case X86ISD::INC: return "X86ISD::INC";
18049 case X86ISD::DEC: return "X86ISD::DEC";
18050 case X86ISD::OR: return "X86ISD::OR";
18051 case X86ISD::XOR: return "X86ISD::XOR";
18052 case X86ISD::AND: return "X86ISD::AND";
18053 case X86ISD::BEXTR: return "X86ISD::BEXTR";
18054 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
18055 case X86ISD::PTEST: return "X86ISD::PTEST";
18056 case X86ISD::TESTP: return "X86ISD::TESTP";
18057 case X86ISD::TESTM: return "X86ISD::TESTM";
18058 case X86ISD::TESTNM: return "X86ISD::TESTNM";
18059 case X86ISD::KORTEST: return "X86ISD::KORTEST";
18060 case X86ISD::PACKSS: return "X86ISD::PACKSS";
18061 case X86ISD::PACKUS: return "X86ISD::PACKUS";
18062 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
18063 case X86ISD::VALIGN: return "X86ISD::VALIGN";
18064 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
18065 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
18066 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
18067 case X86ISD::SHUFP: return "X86ISD::SHUFP";
18068 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
18069 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
18070 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
18071 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
18072 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
18073 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
18074 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
18075 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
18076 case X86ISD::MOVSD: return "X86ISD::MOVSD";
18077 case X86ISD::MOVSS: return "X86ISD::MOVSS";
18078 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
18079 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
18080 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
18081 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
18082 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
18083 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
18084 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
18085 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
18086 case X86ISD::VPERMV: return "X86ISD::VPERMV";
18087 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
18088 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
18089 case X86ISD::VPERMI: return "X86ISD::VPERMI";
18090 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
18091 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
18092 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
18093 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
18094 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
18095 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
18096 case X86ISD::MFENCE: return "X86ISD::MFENCE";
18097 case X86ISD::SFENCE: return "X86ISD::SFENCE";
18098 case X86ISD::LFENCE: return "X86ISD::LFENCE";
18099 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
18100 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
18101 case X86ISD::SAHF: return "X86ISD::SAHF";
18102 case X86ISD::RDRAND: return "X86ISD::RDRAND";
18103 case X86ISD::RDSEED: return "X86ISD::RDSEED";
18104 case X86ISD::FMADD: return "X86ISD::FMADD";
18105 case X86ISD::FMSUB: return "X86ISD::FMSUB";
18106 case X86ISD::FNMADD: return "X86ISD::FNMADD";
18107 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
18108 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
18109 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
18110 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
18111 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
18112 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
18113 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
18114 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
18115 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
18116 case X86ISD::RNDSCALE: return "X86ISD::RNDSCALE";
18117 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
18118 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
18119 case X86ISD::XTEST: return "X86ISD::XTEST";
18120 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
18121 case X86ISD::EXPAND: return "X86ISD::EXPAND";
18122 case X86ISD::SELECT: return "X86ISD::SELECT";
18123 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
18124 case X86ISD::RCP28: return "X86ISD::RCP28";
18125 case X86ISD::EXP2: return "X86ISD::EXP2";
18126 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
18127 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
18128 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
18129 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
18130 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
18131 case X86ISD::ADDS: return "X86ISD::ADDS";
18132 case X86ISD::SUBS: return "X86ISD::SUBS";
18137 // isLegalAddressingMode - Return true if the addressing mode represented
18138 // by AM is legal for this target, for a load/store of the specified type.
18139 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
18141 // X86 supports extremely general addressing modes.
18142 CodeModel::Model M = getTargetMachine().getCodeModel();
18143 Reloc::Model R = getTargetMachine().getRelocationModel();
18145 // X86 allows a sign-extended 32-bit immediate field as a displacement.
18146 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
18151 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
18153 // If a reference to this global requires an extra load, we can't fold it.
18154 if (isGlobalStubReference(GVFlags))
18157 // If BaseGV requires a register for the PIC base, we cannot also have a
18158 // BaseReg specified.
18159 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
18162 // If lower 4G is not available, then we must use rip-relative addressing.
18163 if ((M != CodeModel::Small || R != Reloc::Static) &&
18164 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
18168 switch (AM.Scale) {
18174 // These scales always work.
18179 // These scales are formed with basereg+scalereg. Only accept if there is
18184 default: // Other stuff never works.
18191 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
18192 unsigned Bits = Ty->getScalarSizeInBits();
18194 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
18195 // particularly cheaper than those without.
18199 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
18200 // variable shifts just as cheap as scalar ones.
18201 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
18204 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
18205 // fully general vector.
18209 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
18210 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18212 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
18213 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
18214 return NumBits1 > NumBits2;
18217 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
18218 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
18221 if (!isTypeLegal(EVT::getEVT(Ty1)))
18224 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
18226 // Assuming the caller doesn't have a zeroext or signext return parameter,
18227 // truncation all the way down to i1 is valid.
18231 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
18232 return isInt<32>(Imm);
18235 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
18236 // Can also use sub to handle negated immediates.
18237 return isInt<32>(Imm);
18240 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
18241 if (!VT1.isInteger() || !VT2.isInteger())
18243 unsigned NumBits1 = VT1.getSizeInBits();
18244 unsigned NumBits2 = VT2.getSizeInBits();
18245 return NumBits1 > NumBits2;
18248 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
18249 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18250 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
18253 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
18254 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
18255 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
18258 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
18259 EVT VT1 = Val.getValueType();
18260 if (isZExtFree(VT1, VT2))
18263 if (Val.getOpcode() != ISD::LOAD)
18266 if (!VT1.isSimple() || !VT1.isInteger() ||
18267 !VT2.isSimple() || !VT2.isInteger())
18270 switch (VT1.getSimpleVT().SimpleTy) {
18275 // X86 has 8, 16, and 32-bit zero-extending loads.
18282 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
18285 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
18286 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
18289 VT = VT.getScalarType();
18291 if (!VT.isSimple())
18294 switch (VT.getSimpleVT().SimpleTy) {
18305 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
18306 // i16 instructions are longer (0x66 prefix) and potentially slower.
18307 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
18310 /// isShuffleMaskLegal - Targets can use this to indicate that they only
18311 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
18312 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
18313 /// are assumed to be legal.
18315 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
18317 if (!VT.isSimple())
18320 // Not for i1 vectors
18321 if (VT.getScalarType() == MVT::i1)
18324 // Very little shuffling can be done for 64-bit vectors right now.
18325 if (VT.getSizeInBits() == 64)
18328 // We only care that the types being shuffled are legal. The lowering can
18329 // handle any possible shuffle mask that results.
18330 return isTypeLegal(VT.getSimpleVT());
18334 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
18336 // Just delegate to the generic legality, clear masks aren't special.
18337 return isShuffleMaskLegal(Mask, VT);
18340 //===----------------------------------------------------------------------===//
18341 // X86 Scheduler Hooks
18342 //===----------------------------------------------------------------------===//
18344 /// Utility function to emit xbegin specifying the start of an RTM region.
18345 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
18346 const TargetInstrInfo *TII) {
18347 DebugLoc DL = MI->getDebugLoc();
18349 const BasicBlock *BB = MBB->getBasicBlock();
18350 MachineFunction::iterator I = MBB;
18353 // For the v = xbegin(), we generate
18364 MachineBasicBlock *thisMBB = MBB;
18365 MachineFunction *MF = MBB->getParent();
18366 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18367 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18368 MF->insert(I, mainMBB);
18369 MF->insert(I, sinkMBB);
18371 // Transfer the remainder of BB and its successor edges to sinkMBB.
18372 sinkMBB->splice(sinkMBB->begin(), MBB,
18373 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18374 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18378 // # fallthrough to mainMBB
18379 // # abortion to sinkMBB
18380 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
18381 thisMBB->addSuccessor(mainMBB);
18382 thisMBB->addSuccessor(sinkMBB);
18386 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
18387 mainMBB->addSuccessor(sinkMBB);
18390 // EAX is live into the sinkMBB
18391 sinkMBB->addLiveIn(X86::EAX);
18392 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18393 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18396 MI->eraseFromParent();
18400 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
18401 // or XMM0_V32I8 in AVX all of this code can be replaced with that
18402 // in the .td file.
18403 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
18404 const TargetInstrInfo *TII) {
18406 switch (MI->getOpcode()) {
18407 default: llvm_unreachable("illegal opcode!");
18408 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
18409 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
18410 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
18411 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
18412 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
18413 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
18414 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
18415 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
18418 DebugLoc dl = MI->getDebugLoc();
18419 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18421 unsigned NumArgs = MI->getNumOperands();
18422 for (unsigned i = 1; i < NumArgs; ++i) {
18423 MachineOperand &Op = MI->getOperand(i);
18424 if (!(Op.isReg() && Op.isImplicit()))
18425 MIB.addOperand(Op);
18427 if (MI->hasOneMemOperand())
18428 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18430 BuildMI(*BB, MI, dl,
18431 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18432 .addReg(X86::XMM0);
18434 MI->eraseFromParent();
18438 // FIXME: Custom handling because TableGen doesn't support multiple implicit
18439 // defs in an instruction pattern
18440 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
18441 const TargetInstrInfo *TII) {
18443 switch (MI->getOpcode()) {
18444 default: llvm_unreachable("illegal opcode!");
18445 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
18446 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
18447 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
18448 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
18449 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
18450 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
18451 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
18452 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
18455 DebugLoc dl = MI->getDebugLoc();
18456 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
18458 unsigned NumArgs = MI->getNumOperands(); // remove the results
18459 for (unsigned i = 1; i < NumArgs; ++i) {
18460 MachineOperand &Op = MI->getOperand(i);
18461 if (!(Op.isReg() && Op.isImplicit()))
18462 MIB.addOperand(Op);
18464 if (MI->hasOneMemOperand())
18465 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
18467 BuildMI(*BB, MI, dl,
18468 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
18471 MI->eraseFromParent();
18475 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
18476 const X86Subtarget *Subtarget) {
18477 DebugLoc dl = MI->getDebugLoc();
18478 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18479 // Address into RAX/EAX, other two args into ECX, EDX.
18480 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
18481 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
18482 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
18483 for (int i = 0; i < X86::AddrNumOperands; ++i)
18484 MIB.addOperand(MI->getOperand(i));
18486 unsigned ValOps = X86::AddrNumOperands;
18487 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
18488 .addReg(MI->getOperand(ValOps).getReg());
18489 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
18490 .addReg(MI->getOperand(ValOps+1).getReg());
18492 // The instruction doesn't actually take any operands though.
18493 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
18495 MI->eraseFromParent(); // The pseudo is gone now.
18499 MachineBasicBlock *
18500 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
18501 MachineBasicBlock *MBB) const {
18502 // Emit va_arg instruction on X86-64.
18504 // Operands to this pseudo-instruction:
18505 // 0 ) Output : destination address (reg)
18506 // 1-5) Input : va_list address (addr, i64mem)
18507 // 6 ) ArgSize : Size (in bytes) of vararg type
18508 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
18509 // 8 ) Align : Alignment of type
18510 // 9 ) EFLAGS (implicit-def)
18512 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
18513 static_assert(X86::AddrNumOperands == 5,
18514 "VAARG_64 assumes 5 address operands");
18516 unsigned DestReg = MI->getOperand(0).getReg();
18517 MachineOperand &Base = MI->getOperand(1);
18518 MachineOperand &Scale = MI->getOperand(2);
18519 MachineOperand &Index = MI->getOperand(3);
18520 MachineOperand &Disp = MI->getOperand(4);
18521 MachineOperand &Segment = MI->getOperand(5);
18522 unsigned ArgSize = MI->getOperand(6).getImm();
18523 unsigned ArgMode = MI->getOperand(7).getImm();
18524 unsigned Align = MI->getOperand(8).getImm();
18526 // Memory Reference
18527 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
18528 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18529 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18531 // Machine Information
18532 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18533 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
18534 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
18535 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
18536 DebugLoc DL = MI->getDebugLoc();
18538 // struct va_list {
18541 // i64 overflow_area (address)
18542 // i64 reg_save_area (address)
18544 // sizeof(va_list) = 24
18545 // alignment(va_list) = 8
18547 unsigned TotalNumIntRegs = 6;
18548 unsigned TotalNumXMMRegs = 8;
18549 bool UseGPOffset = (ArgMode == 1);
18550 bool UseFPOffset = (ArgMode == 2);
18551 unsigned MaxOffset = TotalNumIntRegs * 8 +
18552 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
18554 /* Align ArgSize to a multiple of 8 */
18555 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
18556 bool NeedsAlign = (Align > 8);
18558 MachineBasicBlock *thisMBB = MBB;
18559 MachineBasicBlock *overflowMBB;
18560 MachineBasicBlock *offsetMBB;
18561 MachineBasicBlock *endMBB;
18563 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
18564 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
18565 unsigned OffsetReg = 0;
18567 if (!UseGPOffset && !UseFPOffset) {
18568 // If we only pull from the overflow region, we don't create a branch.
18569 // We don't need to alter control flow.
18570 OffsetDestReg = 0; // unused
18571 OverflowDestReg = DestReg;
18573 offsetMBB = nullptr;
18574 overflowMBB = thisMBB;
18577 // First emit code to check if gp_offset (or fp_offset) is below the bound.
18578 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
18579 // If not, pull from overflow_area. (branch to overflowMBB)
18584 // offsetMBB overflowMBB
18589 // Registers for the PHI in endMBB
18590 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
18591 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
18593 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18594 MachineFunction *MF = MBB->getParent();
18595 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18596 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18597 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18599 MachineFunction::iterator MBBIter = MBB;
18602 // Insert the new basic blocks
18603 MF->insert(MBBIter, offsetMBB);
18604 MF->insert(MBBIter, overflowMBB);
18605 MF->insert(MBBIter, endMBB);
18607 // Transfer the remainder of MBB and its successor edges to endMBB.
18608 endMBB->splice(endMBB->begin(), thisMBB,
18609 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
18610 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
18612 // Make offsetMBB and overflowMBB successors of thisMBB
18613 thisMBB->addSuccessor(offsetMBB);
18614 thisMBB->addSuccessor(overflowMBB);
18616 // endMBB is a successor of both offsetMBB and overflowMBB
18617 offsetMBB->addSuccessor(endMBB);
18618 overflowMBB->addSuccessor(endMBB);
18620 // Load the offset value into a register
18621 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18622 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
18626 .addDisp(Disp, UseFPOffset ? 4 : 0)
18627 .addOperand(Segment)
18628 .setMemRefs(MMOBegin, MMOEnd);
18630 // Check if there is enough room left to pull this argument.
18631 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
18633 .addImm(MaxOffset + 8 - ArgSizeA8);
18635 // Branch to "overflowMBB" if offset >= max
18636 // Fall through to "offsetMBB" otherwise
18637 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
18638 .addMBB(overflowMBB);
18641 // In offsetMBB, emit code to use the reg_save_area.
18643 assert(OffsetReg != 0);
18645 // Read the reg_save_area address.
18646 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
18647 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
18652 .addOperand(Segment)
18653 .setMemRefs(MMOBegin, MMOEnd);
18655 // Zero-extend the offset
18656 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
18657 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
18660 .addImm(X86::sub_32bit);
18662 // Add the offset to the reg_save_area to get the final address.
18663 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
18664 .addReg(OffsetReg64)
18665 .addReg(RegSaveReg);
18667 // Compute the offset for the next argument
18668 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
18669 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
18671 .addImm(UseFPOffset ? 16 : 8);
18673 // Store it back into the va_list.
18674 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
18678 .addDisp(Disp, UseFPOffset ? 4 : 0)
18679 .addOperand(Segment)
18680 .addReg(NextOffsetReg)
18681 .setMemRefs(MMOBegin, MMOEnd);
18684 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
18689 // Emit code to use overflow area
18692 // Load the overflow_area address into a register.
18693 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
18694 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
18699 .addOperand(Segment)
18700 .setMemRefs(MMOBegin, MMOEnd);
18702 // If we need to align it, do so. Otherwise, just copy the address
18703 // to OverflowDestReg.
18705 // Align the overflow address
18706 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
18707 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
18709 // aligned_addr = (addr + (align-1)) & ~(align-1)
18710 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
18711 .addReg(OverflowAddrReg)
18714 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
18716 .addImm(~(uint64_t)(Align-1));
18718 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
18719 .addReg(OverflowAddrReg);
18722 // Compute the next overflow address after this argument.
18723 // (the overflow address should be kept 8-byte aligned)
18724 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
18725 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
18726 .addReg(OverflowDestReg)
18727 .addImm(ArgSizeA8);
18729 // Store the new overflow address.
18730 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
18735 .addOperand(Segment)
18736 .addReg(NextAddrReg)
18737 .setMemRefs(MMOBegin, MMOEnd);
18739 // If we branched, emit the PHI to the front of endMBB.
18741 BuildMI(*endMBB, endMBB->begin(), DL,
18742 TII->get(X86::PHI), DestReg)
18743 .addReg(OffsetDestReg).addMBB(offsetMBB)
18744 .addReg(OverflowDestReg).addMBB(overflowMBB);
18747 // Erase the pseudo instruction
18748 MI->eraseFromParent();
18753 MachineBasicBlock *
18754 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18756 MachineBasicBlock *MBB) const {
18757 // Emit code to save XMM registers to the stack. The ABI says that the
18758 // number of registers to save is given in %al, so it's theoretically
18759 // possible to do an indirect jump trick to avoid saving all of them,
18760 // however this code takes a simpler approach and just executes all
18761 // of the stores if %al is non-zero. It's less code, and it's probably
18762 // easier on the hardware branch predictor, and stores aren't all that
18763 // expensive anyway.
18765 // Create the new basic blocks. One block contains all the XMM stores,
18766 // and one block is the final destination regardless of whether any
18767 // stores were performed.
18768 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18769 MachineFunction *F = MBB->getParent();
18770 MachineFunction::iterator MBBIter = MBB;
18772 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
18773 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
18774 F->insert(MBBIter, XMMSaveMBB);
18775 F->insert(MBBIter, EndMBB);
18777 // Transfer the remainder of MBB and its successor edges to EndMBB.
18778 EndMBB->splice(EndMBB->begin(), MBB,
18779 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18780 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
18782 // The original block will now fall through to the XMM save block.
18783 MBB->addSuccessor(XMMSaveMBB);
18784 // The XMMSaveMBB will fall through to the end block.
18785 XMMSaveMBB->addSuccessor(EndMBB);
18787 // Now add the instructions.
18788 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18789 DebugLoc DL = MI->getDebugLoc();
18791 unsigned CountReg = MI->getOperand(0).getReg();
18792 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
18793 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
18795 if (!Subtarget->isTargetWin64()) {
18796 // If %al is 0, branch around the XMM save block.
18797 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
18798 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18799 MBB->addSuccessor(EndMBB);
18802 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18803 // that was just emitted, but clearly shouldn't be "saved".
18804 assert((MI->getNumOperands() <= 3 ||
18805 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18806 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18807 && "Expected last argument to be EFLAGS");
18808 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18809 // In the XMM save block, save all the XMM argument registers.
18810 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18811 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18812 MachineMemOperand *MMO =
18813 F->getMachineMemOperand(
18814 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18815 MachineMemOperand::MOStore,
18816 /*Size=*/16, /*Align=*/16);
18817 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18818 .addFrameIndex(RegSaveFrameIndex)
18819 .addImm(/*Scale=*/1)
18820 .addReg(/*IndexReg=*/0)
18821 .addImm(/*Disp=*/Offset)
18822 .addReg(/*Segment=*/0)
18823 .addReg(MI->getOperand(i).getReg())
18824 .addMemOperand(MMO);
18827 MI->eraseFromParent(); // The pseudo instruction is gone now.
18832 // The EFLAGS operand of SelectItr might be missing a kill marker
18833 // because there were multiple uses of EFLAGS, and ISel didn't know
18834 // which to mark. Figure out whether SelectItr should have had a
18835 // kill marker, and set it if it should. Returns the correct kill
18837 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18838 MachineBasicBlock* BB,
18839 const TargetRegisterInfo* TRI) {
18840 // Scan forward through BB for a use/def of EFLAGS.
18841 MachineBasicBlock::iterator miI(std::next(SelectItr));
18842 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18843 const MachineInstr& mi = *miI;
18844 if (mi.readsRegister(X86::EFLAGS))
18846 if (mi.definesRegister(X86::EFLAGS))
18847 break; // Should have kill-flag - update below.
18850 // If we hit the end of the block, check whether EFLAGS is live into a
18852 if (miI == BB->end()) {
18853 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18854 sEnd = BB->succ_end();
18855 sItr != sEnd; ++sItr) {
18856 MachineBasicBlock* succ = *sItr;
18857 if (succ->isLiveIn(X86::EFLAGS))
18862 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18863 // out. SelectMI should have a kill flag on EFLAGS.
18864 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18868 MachineBasicBlock *
18869 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18870 MachineBasicBlock *BB) const {
18871 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18872 DebugLoc DL = MI->getDebugLoc();
18874 // To "insert" a SELECT_CC instruction, we actually have to insert the
18875 // diamond control-flow pattern. The incoming instruction knows the
18876 // destination vreg to set, the condition code register to branch on, the
18877 // true/false values to select between, and a branch opcode to use.
18878 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18879 MachineFunction::iterator It = BB;
18885 // cmpTY ccX, r1, r2
18887 // fallthrough --> copy0MBB
18888 MachineBasicBlock *thisMBB = BB;
18889 MachineFunction *F = BB->getParent();
18891 // We also lower double CMOVs:
18892 // (CMOV (CMOV F, T, cc1), T, cc2)
18893 // to two successives branches. For that, we look for another CMOV as the
18894 // following instruction.
18896 // Without this, we would add a PHI between the two jumps, which ends up
18897 // creating a few copies all around. For instance, for
18899 // (sitofp (zext (fcmp une)))
18901 // we would generate:
18903 // ucomiss %xmm1, %xmm0
18904 // movss <1.0f>, %xmm0
18905 // movaps %xmm0, %xmm1
18907 // xorps %xmm1, %xmm1
18910 // movaps %xmm1, %xmm0
18914 // because this custom-inserter would have generated:
18926 // A: X = ...; Y = ...
18928 // C: Z = PHI [X, A], [Y, B]
18930 // E: PHI [X, C], [Z, D]
18932 // If we lower both CMOVs in a single step, we can instead generate:
18944 // A: X = ...; Y = ...
18946 // E: PHI [X, A], [X, C], [Y, D]
18948 // Which, in our sitofp/fcmp example, gives us something like:
18950 // ucomiss %xmm1, %xmm0
18951 // movss <1.0f>, %xmm0
18954 // xorps %xmm0, %xmm0
18958 MachineInstr *NextCMOV = nullptr;
18959 MachineBasicBlock::iterator NextMIIt =
18960 std::next(MachineBasicBlock::iterator(MI));
18961 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
18962 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
18963 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
18964 NextCMOV = &*NextMIIt;
18966 MachineBasicBlock *jcc1MBB = nullptr;
18968 // If we have a double CMOV, we lower it to two successive branches to
18969 // the same block. EFLAGS is used by both, so mark it as live in the second.
18971 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
18972 F->insert(It, jcc1MBB);
18973 jcc1MBB->addLiveIn(X86::EFLAGS);
18976 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18977 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18978 F->insert(It, copy0MBB);
18979 F->insert(It, sinkMBB);
18981 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18982 // live into the sink and copy blocks.
18983 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18985 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
18986 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
18987 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
18988 copy0MBB->addLiveIn(X86::EFLAGS);
18989 sinkMBB->addLiveIn(X86::EFLAGS);
18992 // Transfer the remainder of BB and its successor edges to sinkMBB.
18993 sinkMBB->splice(sinkMBB->begin(), BB,
18994 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18995 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18997 // Add the true and fallthrough blocks as its successors.
18999 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
19000 BB->addSuccessor(jcc1MBB);
19002 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
19003 // jump to the sinkMBB.
19004 jcc1MBB->addSuccessor(copy0MBB);
19005 jcc1MBB->addSuccessor(sinkMBB);
19007 BB->addSuccessor(copy0MBB);
19010 // The true block target of the first (or only) branch is always sinkMBB.
19011 BB->addSuccessor(sinkMBB);
19013 // Create the conditional branch instruction.
19015 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
19016 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
19019 unsigned Opc2 = X86::GetCondBranchFromCond(
19020 (X86::CondCode)NextCMOV->getOperand(3).getImm());
19021 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
19025 // %FalseValue = ...
19026 // # fallthrough to sinkMBB
19027 copy0MBB->addSuccessor(sinkMBB);
19030 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
19032 MachineInstrBuilder MIB =
19033 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
19034 MI->getOperand(0).getReg())
19035 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
19036 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
19038 // If we have a double CMOV, the second Jcc provides the same incoming
19039 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
19041 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
19042 // Copy the PHI result to the register defined by the second CMOV.
19043 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
19044 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
19045 .addReg(MI->getOperand(0).getReg());
19046 NextCMOV->eraseFromParent();
19049 MI->eraseFromParent(); // The pseudo instruction is gone now.
19053 MachineBasicBlock *
19054 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
19055 MachineBasicBlock *BB) const {
19056 MachineFunction *MF = BB->getParent();
19057 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19058 DebugLoc DL = MI->getDebugLoc();
19059 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19061 assert(MF->shouldSplitStack());
19063 const bool Is64Bit = Subtarget->is64Bit();
19064 const bool IsLP64 = Subtarget->isTarget64BitLP64();
19066 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
19067 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
19070 // ... [Till the alloca]
19071 // If stacklet is not large enough, jump to mallocMBB
19074 // Allocate by subtracting from RSP
19075 // Jump to continueMBB
19078 // Allocate by call to runtime
19082 // [rest of original BB]
19085 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19086 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19087 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19089 MachineRegisterInfo &MRI = MF->getRegInfo();
19090 const TargetRegisterClass *AddrRegClass =
19091 getRegClassFor(getPointerTy());
19093 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19094 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
19095 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
19096 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
19097 sizeVReg = MI->getOperand(1).getReg(),
19098 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
19100 MachineFunction::iterator MBBIter = BB;
19103 MF->insert(MBBIter, bumpMBB);
19104 MF->insert(MBBIter, mallocMBB);
19105 MF->insert(MBBIter, continueMBB);
19107 continueMBB->splice(continueMBB->begin(), BB,
19108 std::next(MachineBasicBlock::iterator(MI)), BB->end());
19109 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
19111 // Add code to the main basic block to check if the stack limit has been hit,
19112 // and if so, jump to mallocMBB otherwise to bumpMBB.
19113 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
19114 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
19115 .addReg(tmpSPVReg).addReg(sizeVReg);
19116 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
19117 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
19118 .addReg(SPLimitVReg);
19119 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
19121 // bumpMBB simply decreases the stack pointer, since we know the current
19122 // stacklet has enough space.
19123 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
19124 .addReg(SPLimitVReg);
19125 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
19126 .addReg(SPLimitVReg);
19127 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19129 // Calls into a routine in libgcc to allocate more space from the heap.
19130 const uint32_t *RegMask =
19131 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
19133 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
19135 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19136 .addExternalSymbol("__morestack_allocate_stack_space")
19137 .addRegMask(RegMask)
19138 .addReg(X86::RDI, RegState::Implicit)
19139 .addReg(X86::RAX, RegState::ImplicitDefine);
19140 } else if (Is64Bit) {
19141 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
19143 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
19144 .addExternalSymbol("__morestack_allocate_stack_space")
19145 .addRegMask(RegMask)
19146 .addReg(X86::EDI, RegState::Implicit)
19147 .addReg(X86::EAX, RegState::ImplicitDefine);
19149 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
19151 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
19152 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
19153 .addExternalSymbol("__morestack_allocate_stack_space")
19154 .addRegMask(RegMask)
19155 .addReg(X86::EAX, RegState::ImplicitDefine);
19159 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
19162 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
19163 .addReg(IsLP64 ? X86::RAX : X86::EAX);
19164 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
19166 // Set up the CFG correctly.
19167 BB->addSuccessor(bumpMBB);
19168 BB->addSuccessor(mallocMBB);
19169 mallocMBB->addSuccessor(continueMBB);
19170 bumpMBB->addSuccessor(continueMBB);
19172 // Take care of the PHI nodes.
19173 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
19174 MI->getOperand(0).getReg())
19175 .addReg(mallocPtrVReg).addMBB(mallocMBB)
19176 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
19178 // Delete the original pseudo instruction.
19179 MI->eraseFromParent();
19182 return continueMBB;
19185 MachineBasicBlock *
19186 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
19187 MachineBasicBlock *BB) const {
19188 DebugLoc DL = MI->getDebugLoc();
19190 assert(!Subtarget->isTargetMachO());
19192 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
19194 MI->eraseFromParent(); // The pseudo instruction is gone now.
19198 MachineBasicBlock *
19199 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
19200 MachineBasicBlock *BB) const {
19201 // This is pretty easy. We're taking the value that we received from
19202 // our load from the relocation, sticking it in either RDI (x86-64)
19203 // or EAX and doing an indirect call. The return value will then
19204 // be in the normal return register.
19205 MachineFunction *F = BB->getParent();
19206 const X86InstrInfo *TII = Subtarget->getInstrInfo();
19207 DebugLoc DL = MI->getDebugLoc();
19209 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
19210 assert(MI->getOperand(3).isGlobal() && "This should be a global");
19212 // Get a register mask for the lowered call.
19213 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
19214 // proper register mask.
19215 const uint32_t *RegMask =
19216 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
19217 if (Subtarget->is64Bit()) {
19218 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19219 TII->get(X86::MOV64rm), X86::RDI)
19221 .addImm(0).addReg(0)
19222 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19223 MI->getOperand(3).getTargetFlags())
19225 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
19226 addDirectMem(MIB, X86::RDI);
19227 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
19228 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
19229 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19230 TII->get(X86::MOV32rm), X86::EAX)
19232 .addImm(0).addReg(0)
19233 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19234 MI->getOperand(3).getTargetFlags())
19236 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19237 addDirectMem(MIB, X86::EAX);
19238 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19240 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
19241 TII->get(X86::MOV32rm), X86::EAX)
19242 .addReg(TII->getGlobalBaseReg(F))
19243 .addImm(0).addReg(0)
19244 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
19245 MI->getOperand(3).getTargetFlags())
19247 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
19248 addDirectMem(MIB, X86::EAX);
19249 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
19252 MI->eraseFromParent(); // The pseudo instruction is gone now.
19256 MachineBasicBlock *
19257 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
19258 MachineBasicBlock *MBB) const {
19259 DebugLoc DL = MI->getDebugLoc();
19260 MachineFunction *MF = MBB->getParent();
19261 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19262 MachineRegisterInfo &MRI = MF->getRegInfo();
19264 const BasicBlock *BB = MBB->getBasicBlock();
19265 MachineFunction::iterator I = MBB;
19268 // Memory Reference
19269 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19270 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19273 unsigned MemOpndSlot = 0;
19275 unsigned CurOp = 0;
19277 DstReg = MI->getOperand(CurOp++).getReg();
19278 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
19279 assert(RC->hasType(MVT::i32) && "Invalid destination!");
19280 unsigned mainDstReg = MRI.createVirtualRegister(RC);
19281 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
19283 MemOpndSlot = CurOp;
19285 MVT PVT = getPointerTy();
19286 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19287 "Invalid Pointer Size!");
19289 // For v = setjmp(buf), we generate
19292 // buf[LabelOffset] = restoreMBB
19293 // SjLjSetup restoreMBB
19299 // v = phi(main, restore)
19302 // if base pointer being used, load it from frame
19305 MachineBasicBlock *thisMBB = MBB;
19306 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19307 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19308 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
19309 MF->insert(I, mainMBB);
19310 MF->insert(I, sinkMBB);
19311 MF->push_back(restoreMBB);
19313 MachineInstrBuilder MIB;
19315 // Transfer the remainder of BB and its successor edges to sinkMBB.
19316 sinkMBB->splice(sinkMBB->begin(), MBB,
19317 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19318 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19321 unsigned PtrStoreOpc = 0;
19322 unsigned LabelReg = 0;
19323 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19324 Reloc::Model RM = MF->getTarget().getRelocationModel();
19325 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
19326 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
19328 // Prepare IP either in reg or imm.
19329 if (!UseImmLabel) {
19330 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
19331 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
19332 LabelReg = MRI.createVirtualRegister(PtrRC);
19333 if (Subtarget->is64Bit()) {
19334 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
19338 .addMBB(restoreMBB)
19341 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
19342 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
19343 .addReg(XII->getGlobalBaseReg(MF))
19346 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
19350 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
19352 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
19353 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19354 if (i == X86::AddrDisp)
19355 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
19357 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
19360 MIB.addReg(LabelReg);
19362 MIB.addMBB(restoreMBB);
19363 MIB.setMemRefs(MMOBegin, MMOEnd);
19365 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
19366 .addMBB(restoreMBB);
19368 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19369 MIB.addRegMask(RegInfo->getNoPreservedMask());
19370 thisMBB->addSuccessor(mainMBB);
19371 thisMBB->addSuccessor(restoreMBB);
19375 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
19376 mainMBB->addSuccessor(sinkMBB);
19379 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19380 TII->get(X86::PHI), DstReg)
19381 .addReg(mainDstReg).addMBB(mainMBB)
19382 .addReg(restoreDstReg).addMBB(restoreMBB);
19385 if (RegInfo->hasBasePointer(*MF)) {
19386 const bool Uses64BitFramePtr =
19387 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
19388 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
19389 X86FI->setRestoreBasePointer(MF);
19390 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
19391 unsigned BasePtr = RegInfo->getBaseRegister();
19392 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
19393 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
19394 FramePtr, true, X86FI->getRestoreBasePointerOffset())
19395 .setMIFlag(MachineInstr::FrameSetup);
19397 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
19398 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
19399 restoreMBB->addSuccessor(sinkMBB);
19401 MI->eraseFromParent();
19405 MachineBasicBlock *
19406 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
19407 MachineBasicBlock *MBB) const {
19408 DebugLoc DL = MI->getDebugLoc();
19409 MachineFunction *MF = MBB->getParent();
19410 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19411 MachineRegisterInfo &MRI = MF->getRegInfo();
19413 // Memory Reference
19414 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19415 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19417 MVT PVT = getPointerTy();
19418 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
19419 "Invalid Pointer Size!");
19421 const TargetRegisterClass *RC =
19422 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
19423 unsigned Tmp = MRI.createVirtualRegister(RC);
19424 // Since FP is only updated here but NOT referenced, it's treated as GPR.
19425 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
19426 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
19427 unsigned SP = RegInfo->getStackRegister();
19429 MachineInstrBuilder MIB;
19431 const int64_t LabelOffset = 1 * PVT.getStoreSize();
19432 const int64_t SPOffset = 2 * PVT.getStoreSize();
19434 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
19435 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
19438 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
19439 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
19440 MIB.addOperand(MI->getOperand(i));
19441 MIB.setMemRefs(MMOBegin, MMOEnd);
19443 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
19444 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19445 if (i == X86::AddrDisp)
19446 MIB.addDisp(MI->getOperand(i), LabelOffset);
19448 MIB.addOperand(MI->getOperand(i));
19450 MIB.setMemRefs(MMOBegin, MMOEnd);
19452 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
19453 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
19454 if (i == X86::AddrDisp)
19455 MIB.addDisp(MI->getOperand(i), SPOffset);
19457 MIB.addOperand(MI->getOperand(i));
19459 MIB.setMemRefs(MMOBegin, MMOEnd);
19461 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
19463 MI->eraseFromParent();
19467 // Replace 213-type (isel default) FMA3 instructions with 231-type for
19468 // accumulator loops. Writing back to the accumulator allows the coalescer
19469 // to remove extra copies in the loop.
19470 MachineBasicBlock *
19471 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
19472 MachineBasicBlock *MBB) const {
19473 MachineOperand &AddendOp = MI->getOperand(3);
19475 // Bail out early if the addend isn't a register - we can't switch these.
19476 if (!AddendOp.isReg())
19479 MachineFunction &MF = *MBB->getParent();
19480 MachineRegisterInfo &MRI = MF.getRegInfo();
19482 // Check whether the addend is defined by a PHI:
19483 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
19484 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
19485 if (!AddendDef.isPHI())
19488 // Look for the following pattern:
19490 // %addend = phi [%entry, 0], [%loop, %result]
19492 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
19496 // %addend = phi [%entry, 0], [%loop, %result]
19498 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
19500 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
19501 assert(AddendDef.getOperand(i).isReg());
19502 MachineOperand PHISrcOp = AddendDef.getOperand(i);
19503 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
19504 if (&PHISrcInst == MI) {
19505 // Found a matching instruction.
19506 unsigned NewFMAOpc = 0;
19507 switch (MI->getOpcode()) {
19508 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
19509 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
19510 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
19511 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
19512 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
19513 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
19514 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
19515 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
19516 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
19517 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
19518 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
19519 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
19520 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
19521 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
19522 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
19523 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
19524 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
19525 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
19526 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
19527 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
19529 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
19530 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
19531 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
19532 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
19533 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
19534 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
19535 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
19536 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
19537 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
19538 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
19539 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
19540 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
19541 default: llvm_unreachable("Unrecognized FMA variant.");
19544 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
19545 MachineInstrBuilder MIB =
19546 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
19547 .addOperand(MI->getOperand(0))
19548 .addOperand(MI->getOperand(3))
19549 .addOperand(MI->getOperand(2))
19550 .addOperand(MI->getOperand(1));
19551 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
19552 MI->eraseFromParent();
19559 MachineBasicBlock *
19560 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
19561 MachineBasicBlock *BB) const {
19562 switch (MI->getOpcode()) {
19563 default: llvm_unreachable("Unexpected instr type to insert");
19564 case X86::TAILJMPd64:
19565 case X86::TAILJMPr64:
19566 case X86::TAILJMPm64:
19567 case X86::TAILJMPd64_REX:
19568 case X86::TAILJMPr64_REX:
19569 case X86::TAILJMPm64_REX:
19570 llvm_unreachable("TAILJMP64 would not be touched here.");
19571 case X86::TCRETURNdi64:
19572 case X86::TCRETURNri64:
19573 case X86::TCRETURNmi64:
19575 case X86::WIN_ALLOCA:
19576 return EmitLoweredWinAlloca(MI, BB);
19577 case X86::SEG_ALLOCA_32:
19578 case X86::SEG_ALLOCA_64:
19579 return EmitLoweredSegAlloca(MI, BB);
19580 case X86::TLSCall_32:
19581 case X86::TLSCall_64:
19582 return EmitLoweredTLSCall(MI, BB);
19583 case X86::CMOV_GR8:
19584 case X86::CMOV_FR32:
19585 case X86::CMOV_FR64:
19586 case X86::CMOV_V4F32:
19587 case X86::CMOV_V2F64:
19588 case X86::CMOV_V2I64:
19589 case X86::CMOV_V8F32:
19590 case X86::CMOV_V4F64:
19591 case X86::CMOV_V4I64:
19592 case X86::CMOV_V16F32:
19593 case X86::CMOV_V8F64:
19594 case X86::CMOV_V8I64:
19595 case X86::CMOV_GR16:
19596 case X86::CMOV_GR32:
19597 case X86::CMOV_RFP32:
19598 case X86::CMOV_RFP64:
19599 case X86::CMOV_RFP80:
19600 case X86::CMOV_V8I1:
19601 case X86::CMOV_V16I1:
19602 case X86::CMOV_V32I1:
19603 case X86::CMOV_V64I1:
19604 return EmitLoweredSelect(MI, BB);
19606 case X86::FP32_TO_INT16_IN_MEM:
19607 case X86::FP32_TO_INT32_IN_MEM:
19608 case X86::FP32_TO_INT64_IN_MEM:
19609 case X86::FP64_TO_INT16_IN_MEM:
19610 case X86::FP64_TO_INT32_IN_MEM:
19611 case X86::FP64_TO_INT64_IN_MEM:
19612 case X86::FP80_TO_INT16_IN_MEM:
19613 case X86::FP80_TO_INT32_IN_MEM:
19614 case X86::FP80_TO_INT64_IN_MEM: {
19615 MachineFunction *F = BB->getParent();
19616 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19617 DebugLoc DL = MI->getDebugLoc();
19619 // Change the floating point control register to use "round towards zero"
19620 // mode when truncating to an integer value.
19621 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
19622 addFrameReference(BuildMI(*BB, MI, DL,
19623 TII->get(X86::FNSTCW16m)), CWFrameIdx);
19625 // Load the old value of the high byte of the control word...
19627 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
19628 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
19631 // Set the high part to be round to zero...
19632 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
19635 // Reload the modified control word now...
19636 addFrameReference(BuildMI(*BB, MI, DL,
19637 TII->get(X86::FLDCW16m)), CWFrameIdx);
19639 // Restore the memory image of control word to original value
19640 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
19643 // Get the X86 opcode to use.
19645 switch (MI->getOpcode()) {
19646 default: llvm_unreachable("illegal opcode!");
19647 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
19648 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
19649 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
19650 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
19651 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
19652 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
19653 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
19654 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
19655 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
19659 MachineOperand &Op = MI->getOperand(0);
19661 AM.BaseType = X86AddressMode::RegBase;
19662 AM.Base.Reg = Op.getReg();
19664 AM.BaseType = X86AddressMode::FrameIndexBase;
19665 AM.Base.FrameIndex = Op.getIndex();
19667 Op = MI->getOperand(1);
19669 AM.Scale = Op.getImm();
19670 Op = MI->getOperand(2);
19672 AM.IndexReg = Op.getImm();
19673 Op = MI->getOperand(3);
19674 if (Op.isGlobal()) {
19675 AM.GV = Op.getGlobal();
19677 AM.Disp = Op.getImm();
19679 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
19680 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
19682 // Reload the original control word now.
19683 addFrameReference(BuildMI(*BB, MI, DL,
19684 TII->get(X86::FLDCW16m)), CWFrameIdx);
19686 MI->eraseFromParent(); // The pseudo instruction is gone now.
19689 // String/text processing lowering.
19690 case X86::PCMPISTRM128REG:
19691 case X86::VPCMPISTRM128REG:
19692 case X86::PCMPISTRM128MEM:
19693 case X86::VPCMPISTRM128MEM:
19694 case X86::PCMPESTRM128REG:
19695 case X86::VPCMPESTRM128REG:
19696 case X86::PCMPESTRM128MEM:
19697 case X86::VPCMPESTRM128MEM:
19698 assert(Subtarget->hasSSE42() &&
19699 "Target must have SSE4.2 or AVX features enabled");
19700 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
19702 // String/text processing lowering.
19703 case X86::PCMPISTRIREG:
19704 case X86::VPCMPISTRIREG:
19705 case X86::PCMPISTRIMEM:
19706 case X86::VPCMPISTRIMEM:
19707 case X86::PCMPESTRIREG:
19708 case X86::VPCMPESTRIREG:
19709 case X86::PCMPESTRIMEM:
19710 case X86::VPCMPESTRIMEM:
19711 assert(Subtarget->hasSSE42() &&
19712 "Target must have SSE4.2 or AVX features enabled");
19713 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
19715 // Thread synchronization.
19717 return EmitMonitor(MI, BB, Subtarget);
19721 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
19723 case X86::VASTART_SAVE_XMM_REGS:
19724 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
19726 case X86::VAARG_64:
19727 return EmitVAARG64WithCustomInserter(MI, BB);
19729 case X86::EH_SjLj_SetJmp32:
19730 case X86::EH_SjLj_SetJmp64:
19731 return emitEHSjLjSetJmp(MI, BB);
19733 case X86::EH_SjLj_LongJmp32:
19734 case X86::EH_SjLj_LongJmp64:
19735 return emitEHSjLjLongJmp(MI, BB);
19737 case TargetOpcode::STATEPOINT:
19738 // As an implementation detail, STATEPOINT shares the STACKMAP format at
19739 // this point in the process. We diverge later.
19740 return emitPatchPoint(MI, BB);
19742 case TargetOpcode::STACKMAP:
19743 case TargetOpcode::PATCHPOINT:
19744 return emitPatchPoint(MI, BB);
19746 case X86::VFMADDPDr213r:
19747 case X86::VFMADDPSr213r:
19748 case X86::VFMADDSDr213r:
19749 case X86::VFMADDSSr213r:
19750 case X86::VFMSUBPDr213r:
19751 case X86::VFMSUBPSr213r:
19752 case X86::VFMSUBSDr213r:
19753 case X86::VFMSUBSSr213r:
19754 case X86::VFNMADDPDr213r:
19755 case X86::VFNMADDPSr213r:
19756 case X86::VFNMADDSDr213r:
19757 case X86::VFNMADDSSr213r:
19758 case X86::VFNMSUBPDr213r:
19759 case X86::VFNMSUBPSr213r:
19760 case X86::VFNMSUBSDr213r:
19761 case X86::VFNMSUBSSr213r:
19762 case X86::VFMADDSUBPDr213r:
19763 case X86::VFMADDSUBPSr213r:
19764 case X86::VFMSUBADDPDr213r:
19765 case X86::VFMSUBADDPSr213r:
19766 case X86::VFMADDPDr213rY:
19767 case X86::VFMADDPSr213rY:
19768 case X86::VFMSUBPDr213rY:
19769 case X86::VFMSUBPSr213rY:
19770 case X86::VFNMADDPDr213rY:
19771 case X86::VFNMADDPSr213rY:
19772 case X86::VFNMSUBPDr213rY:
19773 case X86::VFNMSUBPSr213rY:
19774 case X86::VFMADDSUBPDr213rY:
19775 case X86::VFMADDSUBPSr213rY:
19776 case X86::VFMSUBADDPDr213rY:
19777 case X86::VFMSUBADDPSr213rY:
19778 return emitFMA3Instr(MI, BB);
19782 //===----------------------------------------------------------------------===//
19783 // X86 Optimization Hooks
19784 //===----------------------------------------------------------------------===//
19786 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
19789 const SelectionDAG &DAG,
19790 unsigned Depth) const {
19791 unsigned BitWidth = KnownZero.getBitWidth();
19792 unsigned Opc = Op.getOpcode();
19793 assert((Opc >= ISD::BUILTIN_OP_END ||
19794 Opc == ISD::INTRINSIC_WO_CHAIN ||
19795 Opc == ISD::INTRINSIC_W_CHAIN ||
19796 Opc == ISD::INTRINSIC_VOID) &&
19797 "Should use MaskedValueIsZero if you don't know whether Op"
19798 " is a target node!");
19800 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
19814 // These nodes' second result is a boolean.
19815 if (Op.getResNo() == 0)
19818 case X86ISD::SETCC:
19819 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
19821 case ISD::INTRINSIC_WO_CHAIN: {
19822 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
19823 unsigned NumLoBits = 0;
19826 case Intrinsic::x86_sse_movmsk_ps:
19827 case Intrinsic::x86_avx_movmsk_ps_256:
19828 case Intrinsic::x86_sse2_movmsk_pd:
19829 case Intrinsic::x86_avx_movmsk_pd_256:
19830 case Intrinsic::x86_mmx_pmovmskb:
19831 case Intrinsic::x86_sse2_pmovmskb_128:
19832 case Intrinsic::x86_avx2_pmovmskb: {
19833 // High bits of movmskp{s|d}, pmovmskb are known zero.
19835 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
19836 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
19837 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
19838 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
19839 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
19840 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
19841 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
19842 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
19844 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
19853 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
19855 const SelectionDAG &,
19856 unsigned Depth) const {
19857 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
19858 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
19859 return Op.getValueType().getScalarType().getSizeInBits();
19865 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
19866 /// node is a GlobalAddress + offset.
19867 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
19868 const GlobalValue* &GA,
19869 int64_t &Offset) const {
19870 if (N->getOpcode() == X86ISD::Wrapper) {
19871 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
19872 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
19873 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
19877 return TargetLowering::isGAPlusOffset(N, GA, Offset);
19880 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
19881 /// same as extracting the high 128-bit part of 256-bit vector and then
19882 /// inserting the result into the low part of a new 256-bit vector
19883 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
19884 EVT VT = SVOp->getValueType(0);
19885 unsigned NumElems = VT.getVectorNumElements();
19887 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19888 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
19889 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19890 SVOp->getMaskElt(j) >= 0)
19896 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
19897 /// same as extracting the low 128-bit part of 256-bit vector and then
19898 /// inserting the result into the high part of a new 256-bit vector
19899 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
19900 EVT VT = SVOp->getValueType(0);
19901 unsigned NumElems = VT.getVectorNumElements();
19903 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19904 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
19905 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19906 SVOp->getMaskElt(j) >= 0)
19912 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
19913 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
19914 TargetLowering::DAGCombinerInfo &DCI,
19915 const X86Subtarget* Subtarget) {
19917 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19918 SDValue V1 = SVOp->getOperand(0);
19919 SDValue V2 = SVOp->getOperand(1);
19920 EVT VT = SVOp->getValueType(0);
19921 unsigned NumElems = VT.getVectorNumElements();
19923 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19924 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19928 // V UNDEF BUILD_VECTOR UNDEF
19930 // CONCAT_VECTOR CONCAT_VECTOR
19933 // RESULT: V + zero extended
19935 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19936 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19937 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19940 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19943 // To match the shuffle mask, the first half of the mask should
19944 // be exactly the first vector, and all the rest a splat with the
19945 // first element of the second one.
19946 for (unsigned i = 0; i != NumElems/2; ++i)
19947 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19948 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19951 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19952 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19953 if (Ld->hasNUsesOfValue(1, 0)) {
19954 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19955 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19957 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19959 Ld->getPointerInfo(),
19960 Ld->getAlignment(),
19961 false/*isVolatile*/, true/*ReadMem*/,
19962 false/*WriteMem*/);
19964 // Make sure the newly-created LOAD is in the same position as Ld in
19965 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19966 // and update uses of Ld's output chain to use the TokenFactor.
19967 if (Ld->hasAnyUseOfValue(1)) {
19968 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19969 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19970 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19971 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19972 SDValue(ResNode.getNode(), 1));
19975 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19979 // Emit a zeroed vector and insert the desired subvector on its
19981 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19982 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19983 return DCI.CombineTo(N, InsV);
19986 //===--------------------------------------------------------------------===//
19987 // Combine some shuffles into subvector extracts and inserts:
19990 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19991 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19992 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19993 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19994 return DCI.CombineTo(N, InsV);
19997 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19998 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19999 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
20000 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
20001 return DCI.CombineTo(N, InsV);
20007 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
20010 /// This is the leaf of the recursive combinine below. When we have found some
20011 /// chain of single-use x86 shuffle instructions and accumulated the combined
20012 /// shuffle mask represented by them, this will try to pattern match that mask
20013 /// into either a single instruction if there is a special purpose instruction
20014 /// for this operation, or into a PSHUFB instruction which is a fully general
20015 /// instruction but should only be used to replace chains over a certain depth.
20016 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
20017 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
20018 TargetLowering::DAGCombinerInfo &DCI,
20019 const X86Subtarget *Subtarget) {
20020 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
20022 // Find the operand that enters the chain. Note that multiple uses are OK
20023 // here, we're not going to remove the operand we find.
20024 SDValue Input = Op.getOperand(0);
20025 while (Input.getOpcode() == ISD::BITCAST)
20026 Input = Input.getOperand(0);
20028 MVT VT = Input.getSimpleValueType();
20029 MVT RootVT = Root.getSimpleValueType();
20032 // Just remove no-op shuffle masks.
20033 if (Mask.size() == 1) {
20034 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
20039 // Use the float domain if the operand type is a floating point type.
20040 bool FloatDomain = VT.isFloatingPoint();
20042 // For floating point shuffles, we don't have free copies in the shuffle
20043 // instructions or the ability to load as part of the instruction, so
20044 // canonicalize their shuffles to UNPCK or MOV variants.
20046 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
20047 // vectors because it can have a load folded into it that UNPCK cannot. This
20048 // doesn't preclude something switching to the shorter encoding post-RA.
20050 // FIXME: Should teach these routines about AVX vector widths.
20051 if (FloatDomain && VT.getSizeInBits() == 128) {
20052 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
20053 bool Lo = Mask.equals({0, 0});
20056 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
20057 // is no slower than UNPCKLPD but has the option to fold the input operand
20058 // into even an unaligned memory load.
20059 if (Lo && Subtarget->hasSSE3()) {
20060 Shuffle = X86ISD::MOVDDUP;
20061 ShuffleVT = MVT::v2f64;
20063 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
20064 // than the UNPCK variants.
20065 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
20066 ShuffleVT = MVT::v4f32;
20068 if (Depth == 1 && Root->getOpcode() == Shuffle)
20069 return false; // Nothing to do!
20070 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20071 DCI.AddToWorklist(Op.getNode());
20072 if (Shuffle == X86ISD::MOVDDUP)
20073 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20075 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20076 DCI.AddToWorklist(Op.getNode());
20077 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20081 if (Subtarget->hasSSE3() &&
20082 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
20083 bool Lo = Mask.equals({0, 0, 2, 2});
20084 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
20085 MVT ShuffleVT = MVT::v4f32;
20086 if (Depth == 1 && Root->getOpcode() == Shuffle)
20087 return false; // Nothing to do!
20088 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20089 DCI.AddToWorklist(Op.getNode());
20090 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
20091 DCI.AddToWorklist(Op.getNode());
20092 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20096 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
20097 bool Lo = Mask.equals({0, 0, 1, 1});
20098 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20099 MVT ShuffleVT = MVT::v4f32;
20100 if (Depth == 1 && Root->getOpcode() == Shuffle)
20101 return false; // Nothing to do!
20102 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20103 DCI.AddToWorklist(Op.getNode());
20104 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20105 DCI.AddToWorklist(Op.getNode());
20106 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20112 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
20113 // variants as none of these have single-instruction variants that are
20114 // superior to the UNPCK formulation.
20115 if (!FloatDomain && VT.getSizeInBits() == 128 &&
20116 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20117 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
20118 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
20120 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
20121 bool Lo = Mask[0] == 0;
20122 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
20123 if (Depth == 1 && Root->getOpcode() == Shuffle)
20124 return false; // Nothing to do!
20126 switch (Mask.size()) {
20128 ShuffleVT = MVT::v8i16;
20131 ShuffleVT = MVT::v16i8;
20134 llvm_unreachable("Impossible mask size!");
20136 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
20137 DCI.AddToWorklist(Op.getNode());
20138 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
20139 DCI.AddToWorklist(Op.getNode());
20140 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20145 // Don't try to re-form single instruction chains under any circumstances now
20146 // that we've done encoding canonicalization for them.
20150 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
20151 // can replace them with a single PSHUFB instruction profitably. Intel's
20152 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
20153 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
20154 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
20155 SmallVector<SDValue, 16> PSHUFBMask;
20156 int NumBytes = VT.getSizeInBits() / 8;
20157 int Ratio = NumBytes / Mask.size();
20158 for (int i = 0; i < NumBytes; ++i) {
20159 if (Mask[i / Ratio] == SM_SentinelUndef) {
20160 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
20163 int M = Mask[i / Ratio] != SM_SentinelZero
20164 ? Ratio * Mask[i / Ratio] + i % Ratio
20166 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
20168 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
20169 Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Input);
20170 DCI.AddToWorklist(Op.getNode());
20171 SDValue PSHUFBMaskOp =
20172 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
20173 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
20174 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
20175 DCI.AddToWorklist(Op.getNode());
20176 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
20181 // Failed to find any combines.
20185 /// \brief Fully generic combining of x86 shuffle instructions.
20187 /// This should be the last combine run over the x86 shuffle instructions. Once
20188 /// they have been fully optimized, this will recursively consider all chains
20189 /// of single-use shuffle instructions, build a generic model of the cumulative
20190 /// shuffle operation, and check for simpler instructions which implement this
20191 /// operation. We use this primarily for two purposes:
20193 /// 1) Collapse generic shuffles to specialized single instructions when
20194 /// equivalent. In most cases, this is just an encoding size win, but
20195 /// sometimes we will collapse multiple generic shuffles into a single
20196 /// special-purpose shuffle.
20197 /// 2) Look for sequences of shuffle instructions with 3 or more total
20198 /// instructions, and replace them with the slightly more expensive SSSE3
20199 /// PSHUFB instruction if available. We do this as the last combining step
20200 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
20201 /// a suitable short sequence of other instructions. The PHUFB will either
20202 /// use a register or have to read from memory and so is slightly (but only
20203 /// slightly) more expensive than the other shuffle instructions.
20205 /// Because this is inherently a quadratic operation (for each shuffle in
20206 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
20207 /// This should never be an issue in practice as the shuffle lowering doesn't
20208 /// produce sequences of more than 8 instructions.
20210 /// FIXME: We will currently miss some cases where the redundant shuffling
20211 /// would simplify under the threshold for PSHUFB formation because of
20212 /// combine-ordering. To fix this, we should do the redundant instruction
20213 /// combining in this recursive walk.
20214 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
20215 ArrayRef<int> RootMask,
20216 int Depth, bool HasPSHUFB,
20218 TargetLowering::DAGCombinerInfo &DCI,
20219 const X86Subtarget *Subtarget) {
20220 // Bound the depth of our recursive combine because this is ultimately
20221 // quadratic in nature.
20225 // Directly rip through bitcasts to find the underlying operand.
20226 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
20227 Op = Op.getOperand(0);
20229 MVT VT = Op.getSimpleValueType();
20230 if (!VT.isVector())
20231 return false; // Bail if we hit a non-vector.
20233 assert(Root.getSimpleValueType().isVector() &&
20234 "Shuffles operate on vector types!");
20235 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
20236 "Can only combine shuffles of the same vector register size.");
20238 if (!isTargetShuffle(Op.getOpcode()))
20240 SmallVector<int, 16> OpMask;
20242 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
20243 // We only can combine unary shuffles which we can decode the mask for.
20244 if (!HaveMask || !IsUnary)
20247 assert(VT.getVectorNumElements() == OpMask.size() &&
20248 "Different mask size from vector size!");
20249 assert(((RootMask.size() > OpMask.size() &&
20250 RootMask.size() % OpMask.size() == 0) ||
20251 (OpMask.size() > RootMask.size() &&
20252 OpMask.size() % RootMask.size() == 0) ||
20253 OpMask.size() == RootMask.size()) &&
20254 "The smaller number of elements must divide the larger.");
20255 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
20256 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
20257 assert(((RootRatio == 1 && OpRatio == 1) ||
20258 (RootRatio == 1) != (OpRatio == 1)) &&
20259 "Must not have a ratio for both incoming and op masks!");
20261 SmallVector<int, 16> Mask;
20262 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
20264 // Merge this shuffle operation's mask into our accumulated mask. Note that
20265 // this shuffle's mask will be the first applied to the input, followed by the
20266 // root mask to get us all the way to the root value arrangement. The reason
20267 // for this order is that we are recursing up the operation chain.
20268 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
20269 int RootIdx = i / RootRatio;
20270 if (RootMask[RootIdx] < 0) {
20271 // This is a zero or undef lane, we're done.
20272 Mask.push_back(RootMask[RootIdx]);
20276 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
20277 int OpIdx = RootMaskedIdx / OpRatio;
20278 if (OpMask[OpIdx] < 0) {
20279 // The incoming lanes are zero or undef, it doesn't matter which ones we
20281 Mask.push_back(OpMask[OpIdx]);
20285 // Ok, we have non-zero lanes, map them through.
20286 Mask.push_back(OpMask[OpIdx] * OpRatio +
20287 RootMaskedIdx % OpRatio);
20290 // See if we can recurse into the operand to combine more things.
20291 switch (Op.getOpcode()) {
20292 case X86ISD::PSHUFB:
20294 case X86ISD::PSHUFD:
20295 case X86ISD::PSHUFHW:
20296 case X86ISD::PSHUFLW:
20297 if (Op.getOperand(0).hasOneUse() &&
20298 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20299 HasPSHUFB, DAG, DCI, Subtarget))
20303 case X86ISD::UNPCKL:
20304 case X86ISD::UNPCKH:
20305 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
20306 // We can't check for single use, we have to check that this shuffle is the only user.
20307 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
20308 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
20309 HasPSHUFB, DAG, DCI, Subtarget))
20314 // Minor canonicalization of the accumulated shuffle mask to make it easier
20315 // to match below. All this does is detect masks with squential pairs of
20316 // elements, and shrink them to the half-width mask. It does this in a loop
20317 // so it will reduce the size of the mask to the minimal width mask which
20318 // performs an equivalent shuffle.
20319 SmallVector<int, 16> WidenedMask;
20320 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
20321 Mask = std::move(WidenedMask);
20322 WidenedMask.clear();
20325 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
20329 /// \brief Get the PSHUF-style mask from PSHUF node.
20331 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
20332 /// PSHUF-style masks that can be reused with such instructions.
20333 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
20334 MVT VT = N.getSimpleValueType();
20335 SmallVector<int, 4> Mask;
20337 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
20341 // If we have more than 128-bits, only the low 128-bits of shuffle mask
20342 // matter. Check that the upper masks are repeats and remove them.
20343 if (VT.getSizeInBits() > 128) {
20344 int LaneElts = 128 / VT.getScalarSizeInBits();
20346 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
20347 for (int j = 0; j < LaneElts; ++j)
20348 assert(Mask[j] == Mask[i * LaneElts + j] - LaneElts &&
20349 "Mask doesn't repeat in high 128-bit lanes!");
20351 Mask.resize(LaneElts);
20354 switch (N.getOpcode()) {
20355 case X86ISD::PSHUFD:
20357 case X86ISD::PSHUFLW:
20360 case X86ISD::PSHUFHW:
20361 Mask.erase(Mask.begin(), Mask.begin() + 4);
20362 for (int &M : Mask)
20366 llvm_unreachable("No valid shuffle instruction found!");
20370 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
20372 /// We walk up the chain and look for a combinable shuffle, skipping over
20373 /// shuffles that we could hoist this shuffle's transformation past without
20374 /// altering anything.
20376 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
20378 TargetLowering::DAGCombinerInfo &DCI) {
20379 assert(N.getOpcode() == X86ISD::PSHUFD &&
20380 "Called with something other than an x86 128-bit half shuffle!");
20383 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
20384 // of the shuffles in the chain so that we can form a fresh chain to replace
20386 SmallVector<SDValue, 8> Chain;
20387 SDValue V = N.getOperand(0);
20388 for (; V.hasOneUse(); V = V.getOperand(0)) {
20389 switch (V.getOpcode()) {
20391 return SDValue(); // Nothing combined!
20394 // Skip bitcasts as we always know the type for the target specific
20398 case X86ISD::PSHUFD:
20399 // Found another dword shuffle.
20402 case X86ISD::PSHUFLW:
20403 // Check that the low words (being shuffled) are the identity in the
20404 // dword shuffle, and the high words are self-contained.
20405 if (Mask[0] != 0 || Mask[1] != 1 ||
20406 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
20409 Chain.push_back(V);
20412 case X86ISD::PSHUFHW:
20413 // Check that the high words (being shuffled) are the identity in the
20414 // dword shuffle, and the low words are self-contained.
20415 if (Mask[2] != 2 || Mask[3] != 3 ||
20416 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
20419 Chain.push_back(V);
20422 case X86ISD::UNPCKL:
20423 case X86ISD::UNPCKH:
20424 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
20425 // shuffle into a preceding word shuffle.
20426 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
20427 V.getSimpleValueType().getScalarType() != MVT::i16)
20430 // Search for a half-shuffle which we can combine with.
20431 unsigned CombineOp =
20432 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
20433 if (V.getOperand(0) != V.getOperand(1) ||
20434 !V->isOnlyUserOf(V.getOperand(0).getNode()))
20436 Chain.push_back(V);
20437 V = V.getOperand(0);
20439 switch (V.getOpcode()) {
20441 return SDValue(); // Nothing to combine.
20443 case X86ISD::PSHUFLW:
20444 case X86ISD::PSHUFHW:
20445 if (V.getOpcode() == CombineOp)
20448 Chain.push_back(V);
20452 V = V.getOperand(0);
20456 } while (V.hasOneUse());
20459 // Break out of the loop if we break out of the switch.
20463 if (!V.hasOneUse())
20464 // We fell out of the loop without finding a viable combining instruction.
20467 // Merge this node's mask and our incoming mask.
20468 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20469 for (int &M : Mask)
20471 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
20472 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20474 // Rebuild the chain around this new shuffle.
20475 while (!Chain.empty()) {
20476 SDValue W = Chain.pop_back_val();
20478 if (V.getValueType() != W.getOperand(0).getValueType())
20479 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
20481 switch (W.getOpcode()) {
20483 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
20485 case X86ISD::UNPCKL:
20486 case X86ISD::UNPCKH:
20487 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
20490 case X86ISD::PSHUFD:
20491 case X86ISD::PSHUFLW:
20492 case X86ISD::PSHUFHW:
20493 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
20497 if (V.getValueType() != N.getValueType())
20498 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
20500 // Return the new chain to replace N.
20504 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
20506 /// We walk up the chain, skipping shuffles of the other half and looking
20507 /// through shuffles which switch halves trying to find a shuffle of the same
20508 /// pair of dwords.
20509 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
20511 TargetLowering::DAGCombinerInfo &DCI) {
20513 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
20514 "Called with something other than an x86 128-bit half shuffle!");
20516 unsigned CombineOpcode = N.getOpcode();
20518 // Walk up a single-use chain looking for a combinable shuffle.
20519 SDValue V = N.getOperand(0);
20520 for (; V.hasOneUse(); V = V.getOperand(0)) {
20521 switch (V.getOpcode()) {
20523 return false; // Nothing combined!
20526 // Skip bitcasts as we always know the type for the target specific
20530 case X86ISD::PSHUFLW:
20531 case X86ISD::PSHUFHW:
20532 if (V.getOpcode() == CombineOpcode)
20535 // Other-half shuffles are no-ops.
20538 // Break out of the loop if we break out of the switch.
20542 if (!V.hasOneUse())
20543 // We fell out of the loop without finding a viable combining instruction.
20546 // Combine away the bottom node as its shuffle will be accumulated into
20547 // a preceding shuffle.
20548 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20550 // Record the old value.
20553 // Merge this node's mask and our incoming mask (adjusted to account for all
20554 // the pshufd instructions encountered).
20555 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20556 for (int &M : Mask)
20558 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
20559 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
20561 // Check that the shuffles didn't cancel each other out. If not, we need to
20562 // combine to the new one.
20564 // Replace the combinable shuffle with the combined one, updating all users
20565 // so that we re-evaluate the chain here.
20566 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
20571 /// \brief Try to combine x86 target specific shuffles.
20572 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
20573 TargetLowering::DAGCombinerInfo &DCI,
20574 const X86Subtarget *Subtarget) {
20576 MVT VT = N.getSimpleValueType();
20577 SmallVector<int, 4> Mask;
20579 switch (N.getOpcode()) {
20580 case X86ISD::PSHUFD:
20581 case X86ISD::PSHUFLW:
20582 case X86ISD::PSHUFHW:
20583 Mask = getPSHUFShuffleMask(N);
20584 assert(Mask.size() == 4);
20590 // Nuke no-op shuffles that show up after combining.
20591 if (isNoopShuffleMask(Mask))
20592 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
20594 // Look for simplifications involving one or two shuffle instructions.
20595 SDValue V = N.getOperand(0);
20596 switch (N.getOpcode()) {
20599 case X86ISD::PSHUFLW:
20600 case X86ISD::PSHUFHW:
20601 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
20603 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
20604 return SDValue(); // We combined away this shuffle, so we're done.
20606 // See if this reduces to a PSHUFD which is no more expensive and can
20607 // combine with more operations. Note that it has to at least flip the
20608 // dwords as otherwise it would have been removed as a no-op.
20609 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
20610 int DMask[] = {0, 1, 2, 3};
20611 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
20612 DMask[DOffset + 0] = DOffset + 1;
20613 DMask[DOffset + 1] = DOffset + 0;
20614 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
20615 V = DAG.getNode(ISD::BITCAST, DL, DVT, V);
20616 DCI.AddToWorklist(V.getNode());
20617 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
20618 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
20619 DCI.AddToWorklist(V.getNode());
20620 return DAG.getNode(ISD::BITCAST, DL, VT, V);
20623 // Look for shuffle patterns which can be implemented as a single unpack.
20624 // FIXME: This doesn't handle the location of the PSHUFD generically, and
20625 // only works when we have a PSHUFD followed by two half-shuffles.
20626 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
20627 (V.getOpcode() == X86ISD::PSHUFLW ||
20628 V.getOpcode() == X86ISD::PSHUFHW) &&
20629 V.getOpcode() != N.getOpcode() &&
20631 SDValue D = V.getOperand(0);
20632 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
20633 D = D.getOperand(0);
20634 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
20635 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
20636 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
20637 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20638 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
20640 for (int i = 0; i < 4; ++i) {
20641 WordMask[i + NOffset] = Mask[i] + NOffset;
20642 WordMask[i + VOffset] = VMask[i] + VOffset;
20644 // Map the word mask through the DWord mask.
20646 for (int i = 0; i < 8; ++i)
20647 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
20648 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
20649 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
20650 // We can replace all three shuffles with an unpack.
20651 V = DAG.getNode(ISD::BITCAST, DL, VT, D.getOperand(0));
20652 DCI.AddToWorklist(V.getNode());
20653 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
20662 case X86ISD::PSHUFD:
20663 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
20672 /// \brief Try to combine a shuffle into a target-specific add-sub node.
20674 /// We combine this directly on the abstract vector shuffle nodes so it is
20675 /// easier to generically match. We also insert dummy vector shuffle nodes for
20676 /// the operands which explicitly discard the lanes which are unused by this
20677 /// operation to try to flow through the rest of the combiner the fact that
20678 /// they're unused.
20679 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
20681 EVT VT = N->getValueType(0);
20683 // We only handle target-independent shuffles.
20684 // FIXME: It would be easy and harmless to use the target shuffle mask
20685 // extraction tool to support more.
20686 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
20689 auto *SVN = cast<ShuffleVectorSDNode>(N);
20690 ArrayRef<int> Mask = SVN->getMask();
20691 SDValue V1 = N->getOperand(0);
20692 SDValue V2 = N->getOperand(1);
20694 // We require the first shuffle operand to be the SUB node, and the second to
20695 // be the ADD node.
20696 // FIXME: We should support the commuted patterns.
20697 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
20700 // If there are other uses of these operations we can't fold them.
20701 if (!V1->hasOneUse() || !V2->hasOneUse())
20704 // Ensure that both operations have the same operands. Note that we can
20705 // commute the FADD operands.
20706 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
20707 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
20708 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
20711 // We're looking for blends between FADD and FSUB nodes. We insist on these
20712 // nodes being lined up in a specific expected pattern.
20713 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
20714 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
20715 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
20718 // Only specific types are legal at this point, assert so we notice if and
20719 // when these change.
20720 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
20721 VT == MVT::v4f64) &&
20722 "Unknown vector type encountered!");
20724 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
20727 /// PerformShuffleCombine - Performs several different shuffle combines.
20728 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
20729 TargetLowering::DAGCombinerInfo &DCI,
20730 const X86Subtarget *Subtarget) {
20732 SDValue N0 = N->getOperand(0);
20733 SDValue N1 = N->getOperand(1);
20734 EVT VT = N->getValueType(0);
20736 // Don't create instructions with illegal types after legalize types has run.
20737 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20738 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
20741 // If we have legalized the vector types, look for blends of FADD and FSUB
20742 // nodes that we can fuse into an ADDSUB node.
20743 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
20744 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
20747 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
20748 if (Subtarget->hasFp256() && VT.is256BitVector() &&
20749 N->getOpcode() == ISD::VECTOR_SHUFFLE)
20750 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
20752 // During Type Legalization, when promoting illegal vector types,
20753 // the backend might introduce new shuffle dag nodes and bitcasts.
20755 // This code performs the following transformation:
20756 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
20757 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
20759 // We do this only if both the bitcast and the BINOP dag nodes have
20760 // one use. Also, perform this transformation only if the new binary
20761 // operation is legal. This is to avoid introducing dag nodes that
20762 // potentially need to be further expanded (or custom lowered) into a
20763 // less optimal sequence of dag nodes.
20764 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
20765 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
20766 N0.getOpcode() == ISD::BITCAST) {
20767 SDValue BC0 = N0.getOperand(0);
20768 EVT SVT = BC0.getValueType();
20769 unsigned Opcode = BC0.getOpcode();
20770 unsigned NumElts = VT.getVectorNumElements();
20772 if (BC0.hasOneUse() && SVT.isVector() &&
20773 SVT.getVectorNumElements() * 2 == NumElts &&
20774 TLI.isOperationLegal(Opcode, VT)) {
20775 bool CanFold = false;
20787 unsigned SVTNumElts = SVT.getVectorNumElements();
20788 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20789 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
20790 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
20791 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
20792 CanFold = SVOp->getMaskElt(i) < 0;
20795 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
20796 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
20797 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
20798 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
20803 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
20804 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
20805 // consecutive, non-overlapping, and in the right order.
20806 SmallVector<SDValue, 16> Elts;
20807 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
20808 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
20810 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
20814 if (isTargetShuffle(N->getOpcode())) {
20816 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
20817 if (Shuffle.getNode())
20820 // Try recursively combining arbitrary sequences of x86 shuffle
20821 // instructions into higher-order shuffles. We do this after combining
20822 // specific PSHUF instruction sequences into their minimal form so that we
20823 // can evaluate how many specialized shuffle instructions are involved in
20824 // a particular chain.
20825 SmallVector<int, 1> NonceMask; // Just a placeholder.
20826 NonceMask.push_back(0);
20827 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
20828 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
20830 return SDValue(); // This routine will use CombineTo to replace N.
20836 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
20837 /// specific shuffle of a load can be folded into a single element load.
20838 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
20839 /// shuffles have been custom lowered so we need to handle those here.
20840 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
20841 TargetLowering::DAGCombinerInfo &DCI) {
20842 if (DCI.isBeforeLegalizeOps())
20845 SDValue InVec = N->getOperand(0);
20846 SDValue EltNo = N->getOperand(1);
20848 if (!isa<ConstantSDNode>(EltNo))
20851 EVT OriginalVT = InVec.getValueType();
20853 if (InVec.getOpcode() == ISD::BITCAST) {
20854 // Don't duplicate a load with other uses.
20855 if (!InVec.hasOneUse())
20857 EVT BCVT = InVec.getOperand(0).getValueType();
20858 if (!BCVT.isVector() ||
20859 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
20861 InVec = InVec.getOperand(0);
20864 EVT CurrentVT = InVec.getValueType();
20866 if (!isTargetShuffle(InVec.getOpcode()))
20869 // Don't duplicate a load with other uses.
20870 if (!InVec.hasOneUse())
20873 SmallVector<int, 16> ShuffleMask;
20875 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
20876 ShuffleMask, UnaryShuffle))
20879 // Select the input vector, guarding against out of range extract vector.
20880 unsigned NumElems = CurrentVT.getVectorNumElements();
20881 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
20882 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
20883 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
20884 : InVec.getOperand(1);
20886 // If inputs to shuffle are the same for both ops, then allow 2 uses
20887 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
20888 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
20890 if (LdNode.getOpcode() == ISD::BITCAST) {
20891 // Don't duplicate a load with other uses.
20892 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
20895 AllowedUses = 1; // only allow 1 load use if we have a bitcast
20896 LdNode = LdNode.getOperand(0);
20899 if (!ISD::isNormalLoad(LdNode.getNode()))
20902 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
20904 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
20907 EVT EltVT = N->getValueType(0);
20908 // If there's a bitcast before the shuffle, check if the load type and
20909 // alignment is valid.
20910 unsigned Align = LN0->getAlignment();
20911 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20912 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
20913 EltVT.getTypeForEVT(*DAG.getContext()));
20915 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
20918 // All checks match so transform back to vector_shuffle so that DAG combiner
20919 // can finish the job
20922 // Create shuffle node taking into account the case that its a unary shuffle
20923 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20924 : InVec.getOperand(1);
20925 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20926 InVec.getOperand(0), Shuffle,
20928 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20929 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20933 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20934 /// special and don't usually play with other vector types, it's better to
20935 /// handle them early to be sure we emit efficient code by avoiding
20936 /// store-load conversions.
20937 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20938 if (N->getValueType(0) != MVT::x86mmx ||
20939 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20940 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20943 SDValue V = N->getOperand(0);
20944 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20945 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20946 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20947 N->getValueType(0), V.getOperand(0));
20952 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20953 /// generation and convert it from being a bunch of shuffles and extracts
20954 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20955 /// storing the value and loading scalars back, while for x64 we should
20956 /// use 64-bit extracts and shifts.
20957 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20958 TargetLowering::DAGCombinerInfo &DCI) {
20959 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20960 if (NewOp.getNode())
20963 SDValue InputVector = N->getOperand(0);
20964 SDLoc dl(InputVector);
20965 // Detect mmx to i32 conversion through a v2i32 elt extract.
20966 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20967 N->getValueType(0) == MVT::i32 &&
20968 InputVector.getValueType() == MVT::v2i32) {
20970 // The bitcast source is a direct mmx result.
20971 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20972 if (MMXSrc.getValueType() == MVT::x86mmx)
20973 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20974 N->getValueType(0),
20975 InputVector.getNode()->getOperand(0));
20977 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20978 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20979 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20980 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20981 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20982 MMXSrcOp.getValueType() == MVT::v1i64 &&
20983 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20984 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20985 N->getValueType(0),
20986 MMXSrcOp.getOperand(0));
20989 EVT VT = N->getValueType(0);
20991 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
20992 InputVector.getOpcode() == ISD::BITCAST &&
20993 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
20994 uint64_t ExtractedElt =
20995 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20996 uint64_t InputValue =
20997 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
20998 uint64_t Res = (InputValue >> ExtractedElt) & 1;
20999 return DAG.getConstant(Res, dl, MVT::i1);
21001 // Only operate on vectors of 4 elements, where the alternative shuffling
21002 // gets to be more expensive.
21003 if (InputVector.getValueType() != MVT::v4i32)
21006 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
21007 // single use which is a sign-extend or zero-extend, and all elements are
21009 SmallVector<SDNode *, 4> Uses;
21010 unsigned ExtractedElements = 0;
21011 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
21012 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
21013 if (UI.getUse().getResNo() != InputVector.getResNo())
21016 SDNode *Extract = *UI;
21017 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
21020 if (Extract->getValueType(0) != MVT::i32)
21022 if (!Extract->hasOneUse())
21024 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
21025 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
21027 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
21030 // Record which element was extracted.
21031 ExtractedElements |=
21032 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
21034 Uses.push_back(Extract);
21037 // If not all the elements were used, this may not be worthwhile.
21038 if (ExtractedElements != 15)
21041 // Ok, we've now decided to do the transformation.
21042 // If 64-bit shifts are legal, use the extract-shift sequence,
21043 // otherwise bounce the vector off the cache.
21044 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21047 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
21048 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
21049 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
21050 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
21051 DAG.getConstant(0, dl, VecIdxTy));
21052 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
21053 DAG.getConstant(1, dl, VecIdxTy));
21055 SDValue ShAmt = DAG.getConstant(32, dl,
21056 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
21057 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
21058 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
21059 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
21060 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
21061 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
21062 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
21064 // Store the value to a temporary stack slot.
21065 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
21066 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
21067 MachinePointerInfo(), false, false, 0);
21069 EVT ElementType = InputVector.getValueType().getVectorElementType();
21070 unsigned EltSize = ElementType.getSizeInBits() / 8;
21072 // Replace each use (extract) with a load of the appropriate element.
21073 for (unsigned i = 0; i < 4; ++i) {
21074 uint64_t Offset = EltSize * i;
21075 SDValue OffsetVal = DAG.getConstant(Offset, dl, TLI.getPointerTy());
21077 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
21078 StackPtr, OffsetVal);
21080 // Load the scalar.
21081 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
21082 ScalarAddr, MachinePointerInfo(),
21083 false, false, false, 0);
21088 // Replace the extracts
21089 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
21090 UE = Uses.end(); UI != UE; ++UI) {
21091 SDNode *Extract = *UI;
21093 SDValue Idx = Extract->getOperand(1);
21094 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
21095 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
21098 // The replacement was made in place; don't return anything.
21102 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
21103 static std::pair<unsigned, bool>
21104 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
21105 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
21106 if (!VT.isVector())
21107 return std::make_pair(0, false);
21109 bool NeedSplit = false;
21110 switch (VT.getSimpleVT().SimpleTy) {
21111 default: return std::make_pair(0, false);
21114 if (!Subtarget->hasVLX())
21115 return std::make_pair(0, false);
21119 if (!Subtarget->hasBWI())
21120 return std::make_pair(0, false);
21124 if (!Subtarget->hasAVX512())
21125 return std::make_pair(0, false);
21130 if (!Subtarget->hasAVX2())
21132 if (!Subtarget->hasAVX())
21133 return std::make_pair(0, false);
21138 if (!Subtarget->hasSSE2())
21139 return std::make_pair(0, false);
21142 // SSE2 has only a small subset of the operations.
21143 bool hasUnsigned = Subtarget->hasSSE41() ||
21144 (Subtarget->hasSSE2() && VT == MVT::v16i8);
21145 bool hasSigned = Subtarget->hasSSE41() ||
21146 (Subtarget->hasSSE2() && VT == MVT::v8i16);
21148 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21151 // Check for x CC y ? x : y.
21152 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21153 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21158 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21161 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21164 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21167 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21169 // Check for x CC y ? y : x -- a min/max with reversed arms.
21170 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21171 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21176 Opc = hasUnsigned ? X86ISD::UMAX : 0u; break;
21179 Opc = hasUnsigned ? X86ISD::UMIN : 0u; break;
21182 Opc = hasSigned ? X86ISD::SMAX : 0u; break;
21185 Opc = hasSigned ? X86ISD::SMIN : 0u; break;
21189 return std::make_pair(Opc, NeedSplit);
21193 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
21194 const X86Subtarget *Subtarget) {
21196 SDValue Cond = N->getOperand(0);
21197 SDValue LHS = N->getOperand(1);
21198 SDValue RHS = N->getOperand(2);
21200 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
21201 SDValue CondSrc = Cond->getOperand(0);
21202 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
21203 Cond = CondSrc->getOperand(0);
21206 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
21209 // A vselect where all conditions and data are constants can be optimized into
21210 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
21211 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
21212 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
21215 unsigned MaskValue = 0;
21216 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
21219 MVT VT = N->getSimpleValueType(0);
21220 unsigned NumElems = VT.getVectorNumElements();
21221 SmallVector<int, 8> ShuffleMask(NumElems, -1);
21222 for (unsigned i = 0; i < NumElems; ++i) {
21223 // Be sure we emit undef where we can.
21224 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
21225 ShuffleMask[i] = -1;
21227 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
21230 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21231 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
21233 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
21236 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
21238 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
21239 TargetLowering::DAGCombinerInfo &DCI,
21240 const X86Subtarget *Subtarget) {
21242 SDValue Cond = N->getOperand(0);
21243 // Get the LHS/RHS of the select.
21244 SDValue LHS = N->getOperand(1);
21245 SDValue RHS = N->getOperand(2);
21246 EVT VT = LHS.getValueType();
21247 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21249 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
21250 // instructions match the semantics of the common C idiom x<y?x:y but not
21251 // x<=y?x:y, because of how they handle negative zero (which can be
21252 // ignored in unsafe-math mode).
21253 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
21254 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
21255 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
21256 (Subtarget->hasSSE2() ||
21257 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
21258 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21260 unsigned Opcode = 0;
21261 // Check for x CC y ? x : y.
21262 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21263 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21267 // Converting this to a min would handle NaNs incorrectly, and swapping
21268 // the operands would cause it to handle comparisons between positive
21269 // and negative zero incorrectly.
21270 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21271 if (!DAG.getTarget().Options.UnsafeFPMath &&
21272 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21274 std::swap(LHS, RHS);
21276 Opcode = X86ISD::FMIN;
21279 // Converting this to a min would handle comparisons between positive
21280 // and negative zero incorrectly.
21281 if (!DAG.getTarget().Options.UnsafeFPMath &&
21282 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21284 Opcode = X86ISD::FMIN;
21287 // Converting this to a min would handle both negative zeros and NaNs
21288 // incorrectly, but we can swap the operands to fix both.
21289 std::swap(LHS, RHS);
21293 Opcode = X86ISD::FMIN;
21297 // Converting this to a max would handle comparisons between positive
21298 // and negative zero incorrectly.
21299 if (!DAG.getTarget().Options.UnsafeFPMath &&
21300 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
21302 Opcode = X86ISD::FMAX;
21305 // Converting this to a max would handle NaNs incorrectly, and swapping
21306 // the operands would cause it to handle comparisons between positive
21307 // and negative zero incorrectly.
21308 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
21309 if (!DAG.getTarget().Options.UnsafeFPMath &&
21310 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
21312 std::swap(LHS, RHS);
21314 Opcode = X86ISD::FMAX;
21317 // Converting this to a max would handle both negative zeros and NaNs
21318 // incorrectly, but we can swap the operands to fix both.
21319 std::swap(LHS, RHS);
21323 Opcode = X86ISD::FMAX;
21326 // Check for x CC y ? y : x -- a min/max with reversed arms.
21327 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
21328 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
21332 // Converting this to a min would handle comparisons between positive
21333 // and negative zero incorrectly, and swapping the operands would
21334 // cause it to handle NaNs incorrectly.
21335 if (!DAG.getTarget().Options.UnsafeFPMath &&
21336 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
21337 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21339 std::swap(LHS, RHS);
21341 Opcode = X86ISD::FMIN;
21344 // Converting this to a min would handle NaNs incorrectly.
21345 if (!DAG.getTarget().Options.UnsafeFPMath &&
21346 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
21348 Opcode = X86ISD::FMIN;
21351 // Converting this to a min would handle both negative zeros and NaNs
21352 // incorrectly, but we can swap the operands to fix both.
21353 std::swap(LHS, RHS);
21357 Opcode = X86ISD::FMIN;
21361 // Converting this to a max would handle NaNs incorrectly.
21362 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21364 Opcode = X86ISD::FMAX;
21367 // Converting this to a max would handle comparisons between positive
21368 // and negative zero incorrectly, and swapping the operands would
21369 // cause it to handle NaNs incorrectly.
21370 if (!DAG.getTarget().Options.UnsafeFPMath &&
21371 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
21372 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
21374 std::swap(LHS, RHS);
21376 Opcode = X86ISD::FMAX;
21379 // Converting this to a max would handle both negative zeros and NaNs
21380 // incorrectly, but we can swap the operands to fix both.
21381 std::swap(LHS, RHS);
21385 Opcode = X86ISD::FMAX;
21391 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
21394 EVT CondVT = Cond.getValueType();
21395 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
21396 CondVT.getVectorElementType() == MVT::i1) {
21397 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
21398 // lowering on KNL. In this case we convert it to
21399 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
21400 // The same situation for all 128 and 256-bit vectors of i8 and i16.
21401 // Since SKX these selects have a proper lowering.
21402 EVT OpVT = LHS.getValueType();
21403 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
21404 (OpVT.getVectorElementType() == MVT::i8 ||
21405 OpVT.getVectorElementType() == MVT::i16) &&
21406 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
21407 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
21408 DCI.AddToWorklist(Cond.getNode());
21409 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
21412 // If this is a select between two integer constants, try to do some
21414 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
21415 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
21416 // Don't do this for crazy integer types.
21417 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
21418 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
21419 // so that TrueC (the true value) is larger than FalseC.
21420 bool NeedsCondInvert = false;
21422 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
21423 // Efficiently invertible.
21424 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
21425 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
21426 isa<ConstantSDNode>(Cond.getOperand(1))))) {
21427 NeedsCondInvert = true;
21428 std::swap(TrueC, FalseC);
21431 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
21432 if (FalseC->getAPIntValue() == 0 &&
21433 TrueC->getAPIntValue().isPowerOf2()) {
21434 if (NeedsCondInvert) // Invert the condition if needed.
21435 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21436 DAG.getConstant(1, DL, Cond.getValueType()));
21438 // Zero extend the condition if needed.
21439 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
21441 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21442 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
21443 DAG.getConstant(ShAmt, DL, MVT::i8));
21446 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
21447 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21448 if (NeedsCondInvert) // Invert the condition if needed.
21449 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21450 DAG.getConstant(1, DL, Cond.getValueType()));
21452 // Zero extend the condition if needed.
21453 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21454 FalseC->getValueType(0), Cond);
21455 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21456 SDValue(FalseC, 0));
21459 // Optimize cases that will turn into an LEA instruction. This requires
21460 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21461 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21462 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21463 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21465 bool isFastMultiplier = false;
21467 switch ((unsigned char)Diff) {
21469 case 1: // result = add base, cond
21470 case 2: // result = lea base( , cond*2)
21471 case 3: // result = lea base(cond, cond*2)
21472 case 4: // result = lea base( , cond*4)
21473 case 5: // result = lea base(cond, cond*4)
21474 case 8: // result = lea base( , cond*8)
21475 case 9: // result = lea base(cond, cond*8)
21476 isFastMultiplier = true;
21481 if (isFastMultiplier) {
21482 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21483 if (NeedsCondInvert) // Invert the condition if needed.
21484 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
21485 DAG.getConstant(1, DL, Cond.getValueType()));
21487 // Zero extend the condition if needed.
21488 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21490 // Scale the condition by the difference.
21492 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21493 DAG.getConstant(Diff, DL,
21494 Cond.getValueType()));
21496 // Add the base if non-zero.
21497 if (FalseC->getAPIntValue() != 0)
21498 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21499 SDValue(FalseC, 0));
21506 // Canonicalize max and min:
21507 // (x > y) ? x : y -> (x >= y) ? x : y
21508 // (x < y) ? x : y -> (x <= y) ? x : y
21509 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
21510 // the need for an extra compare
21511 // against zero. e.g.
21512 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
21514 // testl %edi, %edi
21516 // cmovgl %edi, %eax
21520 // cmovsl %eax, %edi
21521 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
21522 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
21523 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
21524 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21529 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
21530 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
21531 Cond.getOperand(0), Cond.getOperand(1), NewCC);
21532 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
21537 // Early exit check
21538 if (!TLI.isTypeLegal(VT))
21541 // Match VSELECTs into subs with unsigned saturation.
21542 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
21543 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
21544 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
21545 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
21546 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
21548 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
21549 // left side invert the predicate to simplify logic below.
21551 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
21553 CC = ISD::getSetCCInverse(CC, true);
21554 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
21558 if (Other.getNode() && Other->getNumOperands() == 2 &&
21559 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
21560 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
21561 SDValue CondRHS = Cond->getOperand(1);
21563 // Look for a general sub with unsigned saturation first.
21564 // x >= y ? x-y : 0 --> subus x, y
21565 // x > y ? x-y : 0 --> subus x, y
21566 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
21567 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
21568 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
21570 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
21571 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
21572 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
21573 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
21574 // If the RHS is a constant we have to reverse the const
21575 // canonicalization.
21576 // x > C-1 ? x+-C : 0 --> subus x, C
21577 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
21578 CondRHSConst->getAPIntValue() ==
21579 (-OpRHSConst->getAPIntValue() - 1))
21580 return DAG.getNode(
21581 X86ISD::SUBUS, DL, VT, OpLHS,
21582 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
21584 // Another special case: If C was a sign bit, the sub has been
21585 // canonicalized into a xor.
21586 // FIXME: Would it be better to use computeKnownBits to determine
21587 // whether it's safe to decanonicalize the xor?
21588 // x s< 0 ? x^C : 0 --> subus x, C
21589 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
21590 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
21591 OpRHSConst->getAPIntValue().isSignBit())
21592 // Note that we have to rebuild the RHS constant here to ensure we
21593 // don't rely on particular values of undef lanes.
21594 return DAG.getNode(
21595 X86ISD::SUBUS, DL, VT, OpLHS,
21596 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
21601 // Try to match a min/max vector operation.
21602 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
21603 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
21604 unsigned Opc = ret.first;
21605 bool NeedSplit = ret.second;
21607 if (Opc && NeedSplit) {
21608 unsigned NumElems = VT.getVectorNumElements();
21609 // Extract the LHS vectors
21610 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
21611 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
21613 // Extract the RHS vectors
21614 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
21615 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
21617 // Create min/max for each subvector
21618 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
21619 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
21621 // Merge the result
21622 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
21624 return DAG.getNode(Opc, DL, VT, LHS, RHS);
21627 // Simplify vector selection if condition value type matches vselect
21629 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
21630 assert(Cond.getValueType().isVector() &&
21631 "vector select expects a vector selector!");
21633 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
21634 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
21636 // Try invert the condition if true value is not all 1s and false value
21638 if (!TValIsAllOnes && !FValIsAllZeros &&
21639 // Check if the selector will be produced by CMPP*/PCMP*
21640 Cond.getOpcode() == ISD::SETCC &&
21641 // Check if SETCC has already been promoted
21642 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
21643 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
21644 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
21646 if (TValIsAllZeros || FValIsAllOnes) {
21647 SDValue CC = Cond.getOperand(2);
21648 ISD::CondCode NewCC =
21649 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
21650 Cond.getOperand(0).getValueType().isInteger());
21651 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
21652 std::swap(LHS, RHS);
21653 TValIsAllOnes = FValIsAllOnes;
21654 FValIsAllZeros = TValIsAllZeros;
21658 if (TValIsAllOnes || FValIsAllZeros) {
21661 if (TValIsAllOnes && FValIsAllZeros)
21663 else if (TValIsAllOnes)
21664 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
21665 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
21666 else if (FValIsAllZeros)
21667 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
21668 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
21670 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
21674 // We should generate an X86ISD::BLENDI from a vselect if its argument
21675 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
21676 // constants. This specific pattern gets generated when we split a
21677 // selector for a 512 bit vector in a machine without AVX512 (but with
21678 // 256-bit vectors), during legalization:
21680 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
21682 // Iff we find this pattern and the build_vectors are built from
21683 // constants, we translate the vselect into a shuffle_vector that we
21684 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
21685 if ((N->getOpcode() == ISD::VSELECT ||
21686 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
21687 !DCI.isBeforeLegalize()) {
21688 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
21689 if (Shuffle.getNode())
21693 // If this is a *dynamic* select (non-constant condition) and we can match
21694 // this node with one of the variable blend instructions, restructure the
21695 // condition so that the blends can use the high bit of each element and use
21696 // SimplifyDemandedBits to simplify the condition operand.
21697 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
21698 !DCI.isBeforeLegalize() &&
21699 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
21700 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
21702 // Don't optimize vector selects that map to mask-registers.
21706 // We can only handle the cases where VSELECT is directly legal on the
21707 // subtarget. We custom lower VSELECT nodes with constant conditions and
21708 // this makes it hard to see whether a dynamic VSELECT will correctly
21709 // lower, so we both check the operation's status and explicitly handle the
21710 // cases where a *dynamic* blend will fail even though a constant-condition
21711 // blend could be custom lowered.
21712 // FIXME: We should find a better way to handle this class of problems.
21713 // Potentially, we should combine constant-condition vselect nodes
21714 // pre-legalization into shuffles and not mark as many types as custom
21716 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
21718 // FIXME: We don't support i16-element blends currently. We could and
21719 // should support them by making *all* the bits in the condition be set
21720 // rather than just the high bit and using an i8-element blend.
21721 if (VT.getScalarType() == MVT::i16)
21723 // Dynamic blending was only available from SSE4.1 onward.
21724 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
21726 // Byte blends are only available in AVX2
21727 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
21728 !Subtarget->hasAVX2())
21731 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
21732 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
21734 APInt KnownZero, KnownOne;
21735 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
21736 DCI.isBeforeLegalizeOps());
21737 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
21738 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
21740 // If we changed the computation somewhere in the DAG, this change
21741 // will affect all users of Cond.
21742 // Make sure it is fine and update all the nodes so that we do not
21743 // use the generic VSELECT anymore. Otherwise, we may perform
21744 // wrong optimizations as we messed up with the actual expectation
21745 // for the vector boolean values.
21746 if (Cond != TLO.Old) {
21747 // Check all uses of that condition operand to check whether it will be
21748 // consumed by non-BLEND instructions, which may depend on all bits are
21750 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21752 if (I->getOpcode() != ISD::VSELECT)
21753 // TODO: Add other opcodes eventually lowered into BLEND.
21756 // Update all the users of the condition, before committing the change,
21757 // so that the VSELECT optimizations that expect the correct vector
21758 // boolean value will not be triggered.
21759 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
21761 DAG.ReplaceAllUsesOfValueWith(
21763 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
21764 Cond, I->getOperand(1), I->getOperand(2)));
21765 DCI.CommitTargetLoweringOpt(TLO);
21768 // At this point, only Cond is changed. Change the condition
21769 // just for N to keep the opportunity to optimize all other
21770 // users their own way.
21771 DAG.ReplaceAllUsesOfValueWith(
21773 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
21774 TLO.New, N->getOperand(1), N->getOperand(2)));
21782 // Check whether a boolean test is testing a boolean value generated by
21783 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
21786 // Simplify the following patterns:
21787 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
21788 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
21789 // to (Op EFLAGS Cond)
21791 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
21792 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
21793 // to (Op EFLAGS !Cond)
21795 // where Op could be BRCOND or CMOV.
21797 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
21798 // Quit if not CMP and SUB with its value result used.
21799 if (Cmp.getOpcode() != X86ISD::CMP &&
21800 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
21803 // Quit if not used as a boolean value.
21804 if (CC != X86::COND_E && CC != X86::COND_NE)
21807 // Check CMP operands. One of them should be 0 or 1 and the other should be
21808 // an SetCC or extended from it.
21809 SDValue Op1 = Cmp.getOperand(0);
21810 SDValue Op2 = Cmp.getOperand(1);
21813 const ConstantSDNode* C = nullptr;
21814 bool needOppositeCond = (CC == X86::COND_E);
21815 bool checkAgainstTrue = false; // Is it a comparison against 1?
21817 if ((C = dyn_cast<ConstantSDNode>(Op1)))
21819 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
21821 else // Quit if all operands are not constants.
21824 if (C->getZExtValue() == 1) {
21825 needOppositeCond = !needOppositeCond;
21826 checkAgainstTrue = true;
21827 } else if (C->getZExtValue() != 0)
21828 // Quit if the constant is neither 0 or 1.
21831 bool truncatedToBoolWithAnd = false;
21832 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
21833 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
21834 SetCC.getOpcode() == ISD::TRUNCATE ||
21835 SetCC.getOpcode() == ISD::AND) {
21836 if (SetCC.getOpcode() == ISD::AND) {
21838 ConstantSDNode *CS;
21839 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
21840 CS->getZExtValue() == 1)
21842 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
21843 CS->getZExtValue() == 1)
21847 SetCC = SetCC.getOperand(OpIdx);
21848 truncatedToBoolWithAnd = true;
21850 SetCC = SetCC.getOperand(0);
21853 switch (SetCC.getOpcode()) {
21854 case X86ISD::SETCC_CARRY:
21855 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
21856 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
21857 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
21858 // truncated to i1 using 'and'.
21859 if (checkAgainstTrue && !truncatedToBoolWithAnd)
21861 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
21862 "Invalid use of SETCC_CARRY!");
21864 case X86ISD::SETCC:
21865 // Set the condition code or opposite one if necessary.
21866 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
21867 if (needOppositeCond)
21868 CC = X86::GetOppositeBranchCondition(CC);
21869 return SetCC.getOperand(1);
21870 case X86ISD::CMOV: {
21871 // Check whether false/true value has canonical one, i.e. 0 or 1.
21872 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
21873 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
21874 // Quit if true value is not a constant.
21877 // Quit if false value is not a constant.
21879 SDValue Op = SetCC.getOperand(0);
21880 // Skip 'zext' or 'trunc' node.
21881 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
21882 Op.getOpcode() == ISD::TRUNCATE)
21883 Op = Op.getOperand(0);
21884 // A special case for rdrand/rdseed, where 0 is set if false cond is
21886 if ((Op.getOpcode() != X86ISD::RDRAND &&
21887 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
21890 // Quit if false value is not the constant 0 or 1.
21891 bool FValIsFalse = true;
21892 if (FVal && FVal->getZExtValue() != 0) {
21893 if (FVal->getZExtValue() != 1)
21895 // If FVal is 1, opposite cond is needed.
21896 needOppositeCond = !needOppositeCond;
21897 FValIsFalse = false;
21899 // Quit if TVal is not the constant opposite of FVal.
21900 if (FValIsFalse && TVal->getZExtValue() != 1)
21902 if (!FValIsFalse && TVal->getZExtValue() != 0)
21904 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
21905 if (needOppositeCond)
21906 CC = X86::GetOppositeBranchCondition(CC);
21907 return SetCC.getOperand(3);
21914 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
21916 /// (X86or (X86setcc) (X86setcc))
21917 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
21918 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
21919 X86::CondCode &CC1, SDValue &Flags,
21921 if (Cond->getOpcode() == X86ISD::CMP) {
21922 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
21923 if (!CondOp1C || !CondOp1C->isNullValue())
21926 Cond = Cond->getOperand(0);
21931 SDValue SetCC0, SetCC1;
21932 switch (Cond->getOpcode()) {
21933 default: return false;
21940 SetCC0 = Cond->getOperand(0);
21941 SetCC1 = Cond->getOperand(1);
21945 // Make sure we have SETCC nodes, using the same flags value.
21946 if (SetCC0.getOpcode() != X86ISD::SETCC ||
21947 SetCC1.getOpcode() != X86ISD::SETCC ||
21948 SetCC0->getOperand(1) != SetCC1->getOperand(1))
21951 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
21952 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
21953 Flags = SetCC0->getOperand(1);
21957 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
21958 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
21959 TargetLowering::DAGCombinerInfo &DCI,
21960 const X86Subtarget *Subtarget) {
21963 // If the flag operand isn't dead, don't touch this CMOV.
21964 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
21967 SDValue FalseOp = N->getOperand(0);
21968 SDValue TrueOp = N->getOperand(1);
21969 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
21970 SDValue Cond = N->getOperand(3);
21972 if (CC == X86::COND_E || CC == X86::COND_NE) {
21973 switch (Cond.getOpcode()) {
21977 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
21978 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
21979 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21985 Flags = checkBoolTestSetCCCombine(Cond, CC);
21986 if (Flags.getNode() &&
21987 // Extra check as FCMOV only supports a subset of X86 cond.
21988 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21989 SDValue Ops[] = { FalseOp, TrueOp,
21990 DAG.getConstant(CC, DL, MVT::i8), Flags };
21991 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21994 // If this is a select between two integer constants, try to do some
21995 // optimizations. Note that the operands are ordered the opposite of SELECT
21997 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21998 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21999 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
22000 // larger than FalseC (the false value).
22001 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
22002 CC = X86::GetOppositeBranchCondition(CC);
22003 std::swap(TrueC, FalseC);
22004 std::swap(TrueOp, FalseOp);
22007 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
22008 // This is efficient for any integer data type (including i8/i16) and
22010 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
22011 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22012 DAG.getConstant(CC, DL, MVT::i8), Cond);
22014 // Zero extend the condition if needed.
22015 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
22017 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22018 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
22019 DAG.getConstant(ShAmt, DL, MVT::i8));
22020 if (N->getNumValues() == 2) // Dead flag value?
22021 return DCI.CombineTo(N, Cond, SDValue());
22025 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
22026 // for any integer data type, including i8/i16.
22027 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22028 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22029 DAG.getConstant(CC, DL, MVT::i8), Cond);
22031 // Zero extend the condition if needed.
22032 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22033 FalseC->getValueType(0), Cond);
22034 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22035 SDValue(FalseC, 0));
22037 if (N->getNumValues() == 2) // Dead flag value?
22038 return DCI.CombineTo(N, Cond, SDValue());
22042 // Optimize cases that will turn into an LEA instruction. This requires
22043 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22044 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22045 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22046 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22048 bool isFastMultiplier = false;
22050 switch ((unsigned char)Diff) {
22052 case 1: // result = add base, cond
22053 case 2: // result = lea base( , cond*2)
22054 case 3: // result = lea base(cond, cond*2)
22055 case 4: // result = lea base( , cond*4)
22056 case 5: // result = lea base(cond, cond*4)
22057 case 8: // result = lea base( , cond*8)
22058 case 9: // result = lea base(cond, cond*8)
22059 isFastMultiplier = true;
22064 if (isFastMultiplier) {
22065 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22066 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
22067 DAG.getConstant(CC, DL, MVT::i8), Cond);
22068 // Zero extend the condition if needed.
22069 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22071 // Scale the condition by the difference.
22073 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22074 DAG.getConstant(Diff, DL, Cond.getValueType()));
22076 // Add the base if non-zero.
22077 if (FalseC->getAPIntValue() != 0)
22078 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22079 SDValue(FalseC, 0));
22080 if (N->getNumValues() == 2) // Dead flag value?
22081 return DCI.CombineTo(N, Cond, SDValue());
22088 // Handle these cases:
22089 // (select (x != c), e, c) -> select (x != c), e, x),
22090 // (select (x == c), c, e) -> select (x == c), x, e)
22091 // where the c is an integer constant, and the "select" is the combination
22092 // of CMOV and CMP.
22094 // The rationale for this change is that the conditional-move from a constant
22095 // needs two instructions, however, conditional-move from a register needs
22096 // only one instruction.
22098 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
22099 // some instruction-combining opportunities. This opt needs to be
22100 // postponed as late as possible.
22102 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
22103 // the DCI.xxxx conditions are provided to postpone the optimization as
22104 // late as possible.
22106 ConstantSDNode *CmpAgainst = nullptr;
22107 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
22108 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
22109 !isa<ConstantSDNode>(Cond.getOperand(0))) {
22111 if (CC == X86::COND_NE &&
22112 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
22113 CC = X86::GetOppositeBranchCondition(CC);
22114 std::swap(TrueOp, FalseOp);
22117 if (CC == X86::COND_E &&
22118 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
22119 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
22120 DAG.getConstant(CC, DL, MVT::i8), Cond };
22121 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
22126 // Fold and/or of setcc's to double CMOV:
22127 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
22128 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
22130 // This combine lets us generate:
22131 // cmovcc1 (jcc1 if we don't have CMOV)
22137 // cmovne (jne if we don't have CMOV)
22138 // When we can't use the CMOV instruction, it might increase branch
22140 // When we can use CMOV, or when there is no mispredict, this improves
22141 // throughput and reduces register pressure.
22143 if (CC == X86::COND_NE) {
22145 X86::CondCode CC0, CC1;
22147 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
22149 std::swap(FalseOp, TrueOp);
22150 CC0 = X86::GetOppositeBranchCondition(CC0);
22151 CC1 = X86::GetOppositeBranchCondition(CC1);
22154 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
22156 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
22157 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
22158 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
22159 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
22167 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
22168 const X86Subtarget *Subtarget) {
22169 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
22171 default: return SDValue();
22172 // SSE/AVX/AVX2 blend intrinsics.
22173 case Intrinsic::x86_avx2_pblendvb:
22174 // Don't try to simplify this intrinsic if we don't have AVX2.
22175 if (!Subtarget->hasAVX2())
22178 case Intrinsic::x86_avx_blendv_pd_256:
22179 case Intrinsic::x86_avx_blendv_ps_256:
22180 // Don't try to simplify this intrinsic if we don't have AVX.
22181 if (!Subtarget->hasAVX())
22184 case Intrinsic::x86_sse41_blendvps:
22185 case Intrinsic::x86_sse41_blendvpd:
22186 case Intrinsic::x86_sse41_pblendvb: {
22187 SDValue Op0 = N->getOperand(1);
22188 SDValue Op1 = N->getOperand(2);
22189 SDValue Mask = N->getOperand(3);
22191 // Don't try to simplify this intrinsic if we don't have SSE4.1.
22192 if (!Subtarget->hasSSE41())
22195 // fold (blend A, A, Mask) -> A
22198 // fold (blend A, B, allZeros) -> A
22199 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
22201 // fold (blend A, B, allOnes) -> B
22202 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
22205 // Simplify the case where the mask is a constant i32 value.
22206 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
22207 if (C->isNullValue())
22209 if (C->isAllOnesValue())
22216 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
22217 case Intrinsic::x86_sse2_psrai_w:
22218 case Intrinsic::x86_sse2_psrai_d:
22219 case Intrinsic::x86_avx2_psrai_w:
22220 case Intrinsic::x86_avx2_psrai_d:
22221 case Intrinsic::x86_sse2_psra_w:
22222 case Intrinsic::x86_sse2_psra_d:
22223 case Intrinsic::x86_avx2_psra_w:
22224 case Intrinsic::x86_avx2_psra_d: {
22225 SDValue Op0 = N->getOperand(1);
22226 SDValue Op1 = N->getOperand(2);
22227 EVT VT = Op0.getValueType();
22228 assert(VT.isVector() && "Expected a vector type!");
22230 if (isa<BuildVectorSDNode>(Op1))
22231 Op1 = Op1.getOperand(0);
22233 if (!isa<ConstantSDNode>(Op1))
22236 EVT SVT = VT.getVectorElementType();
22237 unsigned SVTBits = SVT.getSizeInBits();
22239 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
22240 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
22241 uint64_t ShAmt = C.getZExtValue();
22243 // Don't try to convert this shift into a ISD::SRA if the shift
22244 // count is bigger than or equal to the element size.
22245 if (ShAmt >= SVTBits)
22248 // Trivial case: if the shift count is zero, then fold this
22249 // into the first operand.
22253 // Replace this packed shift intrinsic with a target independent
22256 SDValue Splat = DAG.getConstant(C, DL, VT);
22257 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
22262 /// PerformMulCombine - Optimize a single multiply with constant into two
22263 /// in order to implement it with two cheaper instructions, e.g.
22264 /// LEA + SHL, LEA + LEA.
22265 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
22266 TargetLowering::DAGCombinerInfo &DCI) {
22267 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
22270 EVT VT = N->getValueType(0);
22271 if (VT != MVT::i64 && VT != MVT::i32)
22274 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
22277 uint64_t MulAmt = C->getZExtValue();
22278 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
22281 uint64_t MulAmt1 = 0;
22282 uint64_t MulAmt2 = 0;
22283 if ((MulAmt % 9) == 0) {
22285 MulAmt2 = MulAmt / 9;
22286 } else if ((MulAmt % 5) == 0) {
22288 MulAmt2 = MulAmt / 5;
22289 } else if ((MulAmt % 3) == 0) {
22291 MulAmt2 = MulAmt / 3;
22294 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
22297 if (isPowerOf2_64(MulAmt2) &&
22298 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
22299 // If second multiplifer is pow2, issue it first. We want the multiply by
22300 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
22302 std::swap(MulAmt1, MulAmt2);
22305 if (isPowerOf2_64(MulAmt1))
22306 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
22307 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
22309 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
22310 DAG.getConstant(MulAmt1, DL, VT));
22312 if (isPowerOf2_64(MulAmt2))
22313 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
22314 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
22316 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
22317 DAG.getConstant(MulAmt2, DL, VT));
22319 // Do not add new nodes to DAG combiner worklist.
22320 DCI.CombineTo(N, NewMul, false);
22325 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
22326 SDValue N0 = N->getOperand(0);
22327 SDValue N1 = N->getOperand(1);
22328 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
22329 EVT VT = N0.getValueType();
22331 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
22332 // since the result of setcc_c is all zero's or all ones.
22333 if (VT.isInteger() && !VT.isVector() &&
22334 N1C && N0.getOpcode() == ISD::AND &&
22335 N0.getOperand(1).getOpcode() == ISD::Constant) {
22336 SDValue N00 = N0.getOperand(0);
22337 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
22338 ((N00.getOpcode() == ISD::ANY_EXTEND ||
22339 N00.getOpcode() == ISD::ZERO_EXTEND) &&
22340 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
22341 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
22342 APInt ShAmt = N1C->getAPIntValue();
22343 Mask = Mask.shl(ShAmt);
22346 return DAG.getNode(ISD::AND, DL, VT,
22347 N00, DAG.getConstant(Mask, DL, VT));
22352 // Hardware support for vector shifts is sparse which makes us scalarize the
22353 // vector operations in many cases. Also, on sandybridge ADD is faster than
22355 // (shl V, 1) -> add V,V
22356 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
22357 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
22358 assert(N0.getValueType().isVector() && "Invalid vector shift type");
22359 // We shift all of the values by one. In many cases we do not have
22360 // hardware support for this operation. This is better expressed as an ADD
22362 if (N1SplatC->getZExtValue() == 1)
22363 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
22369 /// \brief Returns a vector of 0s if the node in input is a vector logical
22370 /// shift by a constant amount which is known to be bigger than or equal
22371 /// to the vector element size in bits.
22372 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
22373 const X86Subtarget *Subtarget) {
22374 EVT VT = N->getValueType(0);
22376 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
22377 (!Subtarget->hasInt256() ||
22378 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
22381 SDValue Amt = N->getOperand(1);
22383 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
22384 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
22385 APInt ShiftAmt = AmtSplat->getAPIntValue();
22386 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
22388 // SSE2/AVX2 logical shifts always return a vector of 0s
22389 // if the shift amount is bigger than or equal to
22390 // the element size. The constant shift amount will be
22391 // encoded as a 8-bit immediate.
22392 if (ShiftAmt.trunc(8).uge(MaxAmount))
22393 return getZeroVector(VT, Subtarget, DAG, DL);
22399 /// PerformShiftCombine - Combine shifts.
22400 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
22401 TargetLowering::DAGCombinerInfo &DCI,
22402 const X86Subtarget *Subtarget) {
22403 if (N->getOpcode() == ISD::SHL) {
22404 SDValue V = PerformSHLCombine(N, DAG);
22405 if (V.getNode()) return V;
22408 if (N->getOpcode() != ISD::SRA) {
22409 // Try to fold this logical shift into a zero vector.
22410 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
22411 if (V.getNode()) return V;
22417 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
22418 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
22419 // and friends. Likewise for OR -> CMPNEQSS.
22420 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
22421 TargetLowering::DAGCombinerInfo &DCI,
22422 const X86Subtarget *Subtarget) {
22425 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
22426 // we're requiring SSE2 for both.
22427 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
22428 SDValue N0 = N->getOperand(0);
22429 SDValue N1 = N->getOperand(1);
22430 SDValue CMP0 = N0->getOperand(1);
22431 SDValue CMP1 = N1->getOperand(1);
22434 // The SETCCs should both refer to the same CMP.
22435 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
22438 SDValue CMP00 = CMP0->getOperand(0);
22439 SDValue CMP01 = CMP0->getOperand(1);
22440 EVT VT = CMP00.getValueType();
22442 if (VT == MVT::f32 || VT == MVT::f64) {
22443 bool ExpectingFlags = false;
22444 // Check for any users that want flags:
22445 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
22446 !ExpectingFlags && UI != UE; ++UI)
22447 switch (UI->getOpcode()) {
22452 ExpectingFlags = true;
22454 case ISD::CopyToReg:
22455 case ISD::SIGN_EXTEND:
22456 case ISD::ZERO_EXTEND:
22457 case ISD::ANY_EXTEND:
22461 if (!ExpectingFlags) {
22462 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
22463 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
22465 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
22466 X86::CondCode tmp = cc0;
22471 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
22472 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
22473 // FIXME: need symbolic constants for these magic numbers.
22474 // See X86ATTInstPrinter.cpp:printSSECC().
22475 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
22476 if (Subtarget->hasAVX512()) {
22477 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
22479 DAG.getConstant(x86cc, DL, MVT::i8));
22480 if (N->getValueType(0) != MVT::i1)
22481 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
22485 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
22486 CMP00.getValueType(), CMP00, CMP01,
22487 DAG.getConstant(x86cc, DL,
22490 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
22491 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
22493 if (is64BitFP && !Subtarget->is64Bit()) {
22494 // On a 32-bit target, we cannot bitcast the 64-bit float to a
22495 // 64-bit integer, since that's not a legal type. Since
22496 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
22497 // bits, but can do this little dance to extract the lowest 32 bits
22498 // and work with those going forward.
22499 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
22501 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
22503 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
22504 Vector32, DAG.getIntPtrConstant(0, DL));
22508 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
22510 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
22511 DAG.getConstant(1, DL, IntVT));
22512 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
22514 return OneBitOfTruth;
22522 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
22523 /// so it can be folded inside ANDNP.
22524 static bool CanFoldXORWithAllOnes(const SDNode *N) {
22525 EVT VT = N->getValueType(0);
22527 // Match direct AllOnes for 128 and 256-bit vectors
22528 if (ISD::isBuildVectorAllOnes(N))
22531 // Look through a bit convert.
22532 if (N->getOpcode() == ISD::BITCAST)
22533 N = N->getOperand(0).getNode();
22535 // Sometimes the operand may come from a insert_subvector building a 256-bit
22537 if (VT.is256BitVector() &&
22538 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
22539 SDValue V1 = N->getOperand(0);
22540 SDValue V2 = N->getOperand(1);
22542 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
22543 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
22544 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
22545 ISD::isBuildVectorAllOnes(V2.getNode()))
22552 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
22553 // register. In most cases we actually compare or select YMM-sized registers
22554 // and mixing the two types creates horrible code. This method optimizes
22555 // some of the transition sequences.
22556 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
22557 TargetLowering::DAGCombinerInfo &DCI,
22558 const X86Subtarget *Subtarget) {
22559 EVT VT = N->getValueType(0);
22560 if (!VT.is256BitVector())
22563 assert((N->getOpcode() == ISD::ANY_EXTEND ||
22564 N->getOpcode() == ISD::ZERO_EXTEND ||
22565 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
22567 SDValue Narrow = N->getOperand(0);
22568 EVT NarrowVT = Narrow->getValueType(0);
22569 if (!NarrowVT.is128BitVector())
22572 if (Narrow->getOpcode() != ISD::XOR &&
22573 Narrow->getOpcode() != ISD::AND &&
22574 Narrow->getOpcode() != ISD::OR)
22577 SDValue N0 = Narrow->getOperand(0);
22578 SDValue N1 = Narrow->getOperand(1);
22581 // The Left side has to be a trunc.
22582 if (N0.getOpcode() != ISD::TRUNCATE)
22585 // The type of the truncated inputs.
22586 EVT WideVT = N0->getOperand(0)->getValueType(0);
22590 // The right side has to be a 'trunc' or a constant vector.
22591 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
22592 ConstantSDNode *RHSConstSplat = nullptr;
22593 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
22594 RHSConstSplat = RHSBV->getConstantSplatNode();
22595 if (!RHSTrunc && !RHSConstSplat)
22598 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22600 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
22603 // Set N0 and N1 to hold the inputs to the new wide operation.
22604 N0 = N0->getOperand(0);
22605 if (RHSConstSplat) {
22606 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
22607 SDValue(RHSConstSplat, 0));
22608 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
22609 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
22610 } else if (RHSTrunc) {
22611 N1 = N1->getOperand(0);
22614 // Generate the wide operation.
22615 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
22616 unsigned Opcode = N->getOpcode();
22618 case ISD::ANY_EXTEND:
22620 case ISD::ZERO_EXTEND: {
22621 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
22622 APInt Mask = APInt::getAllOnesValue(InBits);
22623 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
22624 return DAG.getNode(ISD::AND, DL, VT,
22625 Op, DAG.getConstant(Mask, DL, VT));
22627 case ISD::SIGN_EXTEND:
22628 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
22629 Op, DAG.getValueType(NarrowVT));
22631 llvm_unreachable("Unexpected opcode");
22635 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
22636 TargetLowering::DAGCombinerInfo &DCI,
22637 const X86Subtarget *Subtarget) {
22638 SDValue N0 = N->getOperand(0);
22639 SDValue N1 = N->getOperand(1);
22642 // A vector zext_in_reg may be represented as a shuffle,
22643 // feeding into a bitcast (this represents anyext) feeding into
22644 // an and with a mask.
22645 // We'd like to try to combine that into a shuffle with zero
22646 // plus a bitcast, removing the and.
22647 if (N0.getOpcode() != ISD::BITCAST ||
22648 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
22651 // The other side of the AND should be a splat of 2^C, where C
22652 // is the number of bits in the source type.
22653 if (N1.getOpcode() == ISD::BITCAST)
22654 N1 = N1.getOperand(0);
22655 if (N1.getOpcode() != ISD::BUILD_VECTOR)
22657 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
22659 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
22660 EVT SrcType = Shuffle->getValueType(0);
22662 // We expect a single-source shuffle
22663 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
22666 unsigned SrcSize = SrcType.getScalarSizeInBits();
22668 APInt SplatValue, SplatUndef;
22669 unsigned SplatBitSize;
22671 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
22672 SplatBitSize, HasAnyUndefs))
22675 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
22676 // Make sure the splat matches the mask we expect
22677 if (SplatBitSize > ResSize ||
22678 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
22681 // Make sure the input and output size make sense
22682 if (SrcSize >= ResSize || ResSize % SrcSize)
22685 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
22686 // The number of u's between each two values depends on the ratio between
22687 // the source and dest type.
22688 unsigned ZextRatio = ResSize / SrcSize;
22689 bool IsZext = true;
22690 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
22691 if (i % ZextRatio) {
22692 if (Shuffle->getMaskElt(i) > 0) {
22698 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
22699 // Expected element number
22709 // Ok, perform the transformation - replace the shuffle with
22710 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
22711 // (instead of undef) where the k elements come from the zero vector.
22712 SmallVector<int, 8> Mask;
22713 unsigned NumElems = SrcType.getVectorNumElements();
22714 for (unsigned i = 0; i < NumElems; ++i)
22716 Mask.push_back(NumElems);
22718 Mask.push_back(i / ZextRatio);
22720 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
22721 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
22722 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
22725 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
22726 TargetLowering::DAGCombinerInfo &DCI,
22727 const X86Subtarget *Subtarget) {
22728 if (DCI.isBeforeLegalizeOps())
22731 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
22734 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
22737 EVT VT = N->getValueType(0);
22738 SDValue N0 = N->getOperand(0);
22739 SDValue N1 = N->getOperand(1);
22742 // Create BEXTR instructions
22743 // BEXTR is ((X >> imm) & (2**size-1))
22744 if (VT == MVT::i32 || VT == MVT::i64) {
22745 // Check for BEXTR.
22746 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
22747 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
22748 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
22749 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22750 if (MaskNode && ShiftNode) {
22751 uint64_t Mask = MaskNode->getZExtValue();
22752 uint64_t Shift = ShiftNode->getZExtValue();
22753 if (isMask_64(Mask)) {
22754 uint64_t MaskSize = countPopulation(Mask);
22755 if (Shift + MaskSize <= VT.getSizeInBits())
22756 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
22757 DAG.getConstant(Shift | (MaskSize << 8), DL,
22766 // Want to form ANDNP nodes:
22767 // 1) In the hopes of then easily combining them with OR and AND nodes
22768 // to form PBLEND/PSIGN.
22769 // 2) To match ANDN packed intrinsics
22770 if (VT != MVT::v2i64 && VT != MVT::v4i64)
22773 // Check LHS for vnot
22774 if (N0.getOpcode() == ISD::XOR &&
22775 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
22776 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
22777 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
22779 // Check RHS for vnot
22780 if (N1.getOpcode() == ISD::XOR &&
22781 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
22782 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
22783 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
22788 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
22789 TargetLowering::DAGCombinerInfo &DCI,
22790 const X86Subtarget *Subtarget) {
22791 if (DCI.isBeforeLegalizeOps())
22794 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
22798 SDValue N0 = N->getOperand(0);
22799 SDValue N1 = N->getOperand(1);
22800 EVT VT = N->getValueType(0);
22802 // look for psign/blend
22803 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
22804 if (!Subtarget->hasSSSE3() ||
22805 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
22808 // Canonicalize pandn to RHS
22809 if (N0.getOpcode() == X86ISD::ANDNP)
22811 // or (and (m, y), (pandn m, x))
22812 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
22813 SDValue Mask = N1.getOperand(0);
22814 SDValue X = N1.getOperand(1);
22816 if (N0.getOperand(0) == Mask)
22817 Y = N0.getOperand(1);
22818 if (N0.getOperand(1) == Mask)
22819 Y = N0.getOperand(0);
22821 // Check to see if the mask appeared in both the AND and ANDNP and
22825 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
22826 // Look through mask bitcast.
22827 if (Mask.getOpcode() == ISD::BITCAST)
22828 Mask = Mask.getOperand(0);
22829 if (X.getOpcode() == ISD::BITCAST)
22830 X = X.getOperand(0);
22831 if (Y.getOpcode() == ISD::BITCAST)
22832 Y = Y.getOperand(0);
22834 EVT MaskVT = Mask.getValueType();
22836 // Validate that the Mask operand is a vector sra node.
22837 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
22838 // there is no psrai.b
22839 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
22840 unsigned SraAmt = ~0;
22841 if (Mask.getOpcode() == ISD::SRA) {
22842 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
22843 if (auto *AmtConst = AmtBV->getConstantSplatNode())
22844 SraAmt = AmtConst->getZExtValue();
22845 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
22846 SDValue SraC = Mask.getOperand(1);
22847 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
22849 if ((SraAmt + 1) != EltBits)
22854 // Now we know we at least have a plendvb with the mask val. See if
22855 // we can form a psignb/w/d.
22856 // psign = x.type == y.type == mask.type && y = sub(0, x);
22857 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
22858 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
22859 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
22860 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
22861 "Unsupported VT for PSIGN");
22862 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
22863 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22865 // PBLENDVB only available on SSE 4.1
22866 if (!Subtarget->hasSSE41())
22869 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
22871 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
22872 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
22873 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
22874 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
22875 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
22879 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
22882 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
22883 MachineFunction &MF = DAG.getMachineFunction();
22885 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
22887 // SHLD/SHRD instructions have lower register pressure, but on some
22888 // platforms they have higher latency than the equivalent
22889 // series of shifts/or that would otherwise be generated.
22890 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
22891 // have higher latencies and we are not optimizing for size.
22892 if (!OptForSize && Subtarget->isSHLDSlow())
22895 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
22897 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
22899 if (!N0.hasOneUse() || !N1.hasOneUse())
22902 SDValue ShAmt0 = N0.getOperand(1);
22903 if (ShAmt0.getValueType() != MVT::i8)
22905 SDValue ShAmt1 = N1.getOperand(1);
22906 if (ShAmt1.getValueType() != MVT::i8)
22908 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
22909 ShAmt0 = ShAmt0.getOperand(0);
22910 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
22911 ShAmt1 = ShAmt1.getOperand(0);
22914 unsigned Opc = X86ISD::SHLD;
22915 SDValue Op0 = N0.getOperand(0);
22916 SDValue Op1 = N1.getOperand(0);
22917 if (ShAmt0.getOpcode() == ISD::SUB) {
22918 Opc = X86ISD::SHRD;
22919 std::swap(Op0, Op1);
22920 std::swap(ShAmt0, ShAmt1);
22923 unsigned Bits = VT.getSizeInBits();
22924 if (ShAmt1.getOpcode() == ISD::SUB) {
22925 SDValue Sum = ShAmt1.getOperand(0);
22926 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
22927 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
22928 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
22929 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
22930 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
22931 return DAG.getNode(Opc, DL, VT,
22933 DAG.getNode(ISD::TRUNCATE, DL,
22936 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
22937 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
22939 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
22940 return DAG.getNode(Opc, DL, VT,
22941 N0.getOperand(0), N1.getOperand(0),
22942 DAG.getNode(ISD::TRUNCATE, DL,
22949 // Generate NEG and CMOV for integer abs.
22950 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
22951 EVT VT = N->getValueType(0);
22953 // Since X86 does not have CMOV for 8-bit integer, we don't convert
22954 // 8-bit integer abs to NEG and CMOV.
22955 if (VT.isInteger() && VT.getSizeInBits() == 8)
22958 SDValue N0 = N->getOperand(0);
22959 SDValue N1 = N->getOperand(1);
22962 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
22963 // and change it to SUB and CMOV.
22964 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
22965 N0.getOpcode() == ISD::ADD &&
22966 N0.getOperand(1) == N1 &&
22967 N1.getOpcode() == ISD::SRA &&
22968 N1.getOperand(0) == N0.getOperand(0))
22969 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
22970 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
22971 // Generate SUB & CMOV.
22972 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
22973 DAG.getConstant(0, DL, VT), N0.getOperand(0));
22975 SDValue Ops[] = { N0.getOperand(0), Neg,
22976 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
22977 SDValue(Neg.getNode(), 1) };
22978 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
22983 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
22984 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
22985 TargetLowering::DAGCombinerInfo &DCI,
22986 const X86Subtarget *Subtarget) {
22987 if (DCI.isBeforeLegalizeOps())
22990 if (Subtarget->hasCMov()) {
22991 SDValue RV = performIntegerAbsCombine(N, DAG);
22999 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
23000 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
23001 TargetLowering::DAGCombinerInfo &DCI,
23002 const X86Subtarget *Subtarget) {
23003 LoadSDNode *Ld = cast<LoadSDNode>(N);
23004 EVT RegVT = Ld->getValueType(0);
23005 EVT MemVT = Ld->getMemoryVT();
23007 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23009 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
23010 // into two 16-byte operations.
23011 ISD::LoadExtType Ext = Ld->getExtensionType();
23012 unsigned Alignment = Ld->getAlignment();
23013 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
23014 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
23015 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
23016 unsigned NumElems = RegVT.getVectorNumElements();
23020 SDValue Ptr = Ld->getBasePtr();
23021 SDValue Increment = DAG.getConstant(16, dl, TLI.getPointerTy());
23023 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
23025 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23026 Ld->getPointerInfo(), Ld->isVolatile(),
23027 Ld->isNonTemporal(), Ld->isInvariant(),
23029 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23030 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
23031 Ld->getPointerInfo(), Ld->isVolatile(),
23032 Ld->isNonTemporal(), Ld->isInvariant(),
23033 std::min(16U, Alignment));
23034 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
23036 Load2.getValue(1));
23038 SDValue NewVec = DAG.getUNDEF(RegVT);
23039 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
23040 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
23041 return DCI.CombineTo(N, NewVec, TF, true);
23047 /// PerformMLOADCombine - Resolve extending loads
23048 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
23049 TargetLowering::DAGCombinerInfo &DCI,
23050 const X86Subtarget *Subtarget) {
23051 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
23052 if (Mld->getExtensionType() != ISD::SEXTLOAD)
23055 EVT VT = Mld->getValueType(0);
23056 unsigned NumElems = VT.getVectorNumElements();
23057 EVT LdVT = Mld->getMemoryVT();
23060 assert(LdVT != VT && "Cannot extend to the same type");
23061 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
23062 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
23063 // From, To sizes and ElemCount must be pow of two
23064 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23065 "Unexpected size for extending masked load");
23067 unsigned SizeRatio = ToSz / FromSz;
23068 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
23070 // Create a type on which we perform the shuffle
23071 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23072 LdVT.getScalarType(), NumElems*SizeRatio);
23073 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23075 // Convert Src0 value
23076 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
23077 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
23078 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23079 for (unsigned i = 0; i != NumElems; ++i)
23080 ShuffleVec[i] = i * SizeRatio;
23082 // Can't shuffle using an illegal type.
23083 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23084 && "WideVecVT should be legal");
23085 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
23086 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
23088 // Prepare the new mask
23090 SDValue Mask = Mld->getMask();
23091 if (Mask.getValueType() == VT) {
23092 // Mask and original value have the same type
23093 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
23094 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23095 for (unsigned i = 0; i != NumElems; ++i)
23096 ShuffleVec[i] = i * SizeRatio;
23097 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23098 ShuffleVec[i] = NumElems*SizeRatio;
23099 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23100 DAG.getConstant(0, dl, WideVecVT),
23104 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23105 unsigned WidenNumElts = NumElems*SizeRatio;
23106 unsigned MaskNumElts = VT.getVectorNumElements();
23107 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23110 unsigned NumConcat = WidenNumElts / MaskNumElts;
23111 SmallVector<SDValue, 16> Ops(NumConcat);
23112 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23114 for (unsigned i = 1; i != NumConcat; ++i)
23117 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23120 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
23121 Mld->getBasePtr(), NewMask, WideSrc0,
23122 Mld->getMemoryVT(), Mld->getMemOperand(),
23124 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
23125 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
23128 /// PerformMSTORECombine - Resolve truncating stores
23129 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
23130 const X86Subtarget *Subtarget) {
23131 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
23132 if (!Mst->isTruncatingStore())
23135 EVT VT = Mst->getValue().getValueType();
23136 unsigned NumElems = VT.getVectorNumElements();
23137 EVT StVT = Mst->getMemoryVT();
23140 assert(StVT != VT && "Cannot truncate to the same type");
23141 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23142 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23144 // From, To sizes and ElemCount must be pow of two
23145 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
23146 "Unexpected size for truncating masked store");
23147 // We are going to use the original vector elt for storing.
23148 // Accumulated smaller vector elements must be a multiple of the store size.
23149 assert (((NumElems * FromSz) % ToSz) == 0 &&
23150 "Unexpected ratio for truncating masked store");
23152 unsigned SizeRatio = FromSz / ToSz;
23153 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23155 // Create a type on which we perform the shuffle
23156 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23157 StVT.getScalarType(), NumElems*SizeRatio);
23159 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23161 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
23162 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
23163 for (unsigned i = 0; i != NumElems; ++i)
23164 ShuffleVec[i] = i * SizeRatio;
23166 // Can't shuffle using an illegal type.
23167 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
23168 && "WideVecVT should be legal");
23170 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23171 DAG.getUNDEF(WideVecVT),
23175 SDValue Mask = Mst->getMask();
23176 if (Mask.getValueType() == VT) {
23177 // Mask and original value have the same type
23178 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
23179 for (unsigned i = 0; i != NumElems; ++i)
23180 ShuffleVec[i] = i * SizeRatio;
23181 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
23182 ShuffleVec[i] = NumElems*SizeRatio;
23183 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
23184 DAG.getConstant(0, dl, WideVecVT),
23188 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
23189 unsigned WidenNumElts = NumElems*SizeRatio;
23190 unsigned MaskNumElts = VT.getVectorNumElements();
23191 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
23194 unsigned NumConcat = WidenNumElts / MaskNumElts;
23195 SmallVector<SDValue, 16> Ops(NumConcat);
23196 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
23198 for (unsigned i = 1; i != NumConcat; ++i)
23201 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
23204 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
23205 NewMask, StVT, Mst->getMemOperand(), false);
23207 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
23208 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
23209 const X86Subtarget *Subtarget) {
23210 StoreSDNode *St = cast<StoreSDNode>(N);
23211 EVT VT = St->getValue().getValueType();
23212 EVT StVT = St->getMemoryVT();
23214 SDValue StoredVal = St->getOperand(1);
23215 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23217 // If we are saving a concatenation of two XMM registers and 32-byte stores
23218 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
23219 unsigned Alignment = St->getAlignment();
23220 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
23221 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
23222 StVT == VT && !IsAligned) {
23223 unsigned NumElems = VT.getVectorNumElements();
23227 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
23228 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
23230 SDValue Stride = DAG.getConstant(16, dl, TLI.getPointerTy());
23231 SDValue Ptr0 = St->getBasePtr();
23232 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
23234 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
23235 St->getPointerInfo(), St->isVolatile(),
23236 St->isNonTemporal(), Alignment);
23237 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
23238 St->getPointerInfo(), St->isVolatile(),
23239 St->isNonTemporal(),
23240 std::min(16U, Alignment));
23241 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
23244 // Optimize trunc store (of multiple scalars) to shuffle and store.
23245 // First, pack all of the elements in one place. Next, store to memory
23246 // in fewer chunks.
23247 if (St->isTruncatingStore() && VT.isVector()) {
23248 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23249 unsigned NumElems = VT.getVectorNumElements();
23250 assert(StVT != VT && "Cannot truncate to the same type");
23251 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
23252 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
23254 // From, To sizes and ElemCount must be pow of two
23255 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
23256 // We are going to use the original vector elt for storing.
23257 // Accumulated smaller vector elements must be a multiple of the store size.
23258 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
23260 unsigned SizeRatio = FromSz / ToSz;
23262 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
23264 // Create a type on which we perform the shuffle
23265 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
23266 StVT.getScalarType(), NumElems*SizeRatio);
23268 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
23270 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
23271 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
23272 for (unsigned i = 0; i != NumElems; ++i)
23273 ShuffleVec[i] = i * SizeRatio;
23275 // Can't shuffle using an illegal type.
23276 if (!TLI.isTypeLegal(WideVecVT))
23279 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
23280 DAG.getUNDEF(WideVecVT),
23282 // At this point all of the data is stored at the bottom of the
23283 // register. We now need to save it to mem.
23285 // Find the largest store unit
23286 MVT StoreType = MVT::i8;
23287 for (MVT Tp : MVT::integer_valuetypes()) {
23288 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
23292 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
23293 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
23294 (64 <= NumElems * ToSz))
23295 StoreType = MVT::f64;
23297 // Bitcast the original vector into a vector of store-size units
23298 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
23299 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
23300 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
23301 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
23302 SmallVector<SDValue, 8> Chains;
23303 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8, dl,
23304 TLI.getPointerTy());
23305 SDValue Ptr = St->getBasePtr();
23307 // Perform one or more big stores into memory.
23308 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
23309 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
23310 StoreType, ShuffWide,
23311 DAG.getIntPtrConstant(i, dl));
23312 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
23313 St->getPointerInfo(), St->isVolatile(),
23314 St->isNonTemporal(), St->getAlignment());
23315 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
23316 Chains.push_back(Ch);
23319 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
23322 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
23323 // the FP state in cases where an emms may be missing.
23324 // A preferable solution to the general problem is to figure out the right
23325 // places to insert EMMS. This qualifies as a quick hack.
23327 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
23328 if (VT.getSizeInBits() != 64)
23331 const Function *F = DAG.getMachineFunction().getFunction();
23332 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
23334 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
23335 if ((VT.isVector() ||
23336 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
23337 isa<LoadSDNode>(St->getValue()) &&
23338 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
23339 St->getChain().hasOneUse() && !St->isVolatile()) {
23340 SDNode* LdVal = St->getValue().getNode();
23341 LoadSDNode *Ld = nullptr;
23342 int TokenFactorIndex = -1;
23343 SmallVector<SDValue, 8> Ops;
23344 SDNode* ChainVal = St->getChain().getNode();
23345 // Must be a store of a load. We currently handle two cases: the load
23346 // is a direct child, and it's under an intervening TokenFactor. It is
23347 // possible to dig deeper under nested TokenFactors.
23348 if (ChainVal == LdVal)
23349 Ld = cast<LoadSDNode>(St->getChain());
23350 else if (St->getValue().hasOneUse() &&
23351 ChainVal->getOpcode() == ISD::TokenFactor) {
23352 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
23353 if (ChainVal->getOperand(i).getNode() == LdVal) {
23354 TokenFactorIndex = i;
23355 Ld = cast<LoadSDNode>(St->getValue());
23357 Ops.push_back(ChainVal->getOperand(i));
23361 if (!Ld || !ISD::isNormalLoad(Ld))
23364 // If this is not the MMX case, i.e. we are just turning i64 load/store
23365 // into f64 load/store, avoid the transformation if there are multiple
23366 // uses of the loaded value.
23367 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
23372 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
23373 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
23375 if (Subtarget->is64Bit() || F64IsLegal) {
23376 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
23377 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
23378 Ld->getPointerInfo(), Ld->isVolatile(),
23379 Ld->isNonTemporal(), Ld->isInvariant(),
23380 Ld->getAlignment());
23381 SDValue NewChain = NewLd.getValue(1);
23382 if (TokenFactorIndex != -1) {
23383 Ops.push_back(NewChain);
23384 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23386 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
23387 St->getPointerInfo(),
23388 St->isVolatile(), St->isNonTemporal(),
23389 St->getAlignment());
23392 // Otherwise, lower to two pairs of 32-bit loads / stores.
23393 SDValue LoAddr = Ld->getBasePtr();
23394 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
23395 DAG.getConstant(4, LdDL, MVT::i32));
23397 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
23398 Ld->getPointerInfo(),
23399 Ld->isVolatile(), Ld->isNonTemporal(),
23400 Ld->isInvariant(), Ld->getAlignment());
23401 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
23402 Ld->getPointerInfo().getWithOffset(4),
23403 Ld->isVolatile(), Ld->isNonTemporal(),
23405 MinAlign(Ld->getAlignment(), 4));
23407 SDValue NewChain = LoLd.getValue(1);
23408 if (TokenFactorIndex != -1) {
23409 Ops.push_back(LoLd);
23410 Ops.push_back(HiLd);
23411 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
23414 LoAddr = St->getBasePtr();
23415 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
23416 DAG.getConstant(4, StDL, MVT::i32));
23418 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
23419 St->getPointerInfo(),
23420 St->isVolatile(), St->isNonTemporal(),
23421 St->getAlignment());
23422 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
23423 St->getPointerInfo().getWithOffset(4),
23425 St->isNonTemporal(),
23426 MinAlign(St->getAlignment(), 4));
23427 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
23430 // This is similar to the above case, but here we handle a scalar 64-bit
23431 // integer store that is extracted from a vector on a 32-bit target.
23432 // If we have SSE2, then we can treat it like a floating-point double
23433 // to get past legalization. The execution dependencies fixup pass will
23434 // choose the optimal machine instruction for the store if this really is
23435 // an integer or v2f32 rather than an f64.
23436 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
23437 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
23438 SDValue OldExtract = St->getOperand(1);
23439 SDValue ExtOp0 = OldExtract.getOperand(0);
23440 unsigned VecSize = ExtOp0.getValueSizeInBits();
23441 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
23442 SDValue BitCast = DAG.getNode(ISD::BITCAST, dl, VecVT, ExtOp0);
23443 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
23444 BitCast, OldExtract.getOperand(1));
23445 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
23446 St->getPointerInfo(), St->isVolatile(),
23447 St->isNonTemporal(), St->getAlignment());
23453 /// Return 'true' if this vector operation is "horizontal"
23454 /// and return the operands for the horizontal operation in LHS and RHS. A
23455 /// horizontal operation performs the binary operation on successive elements
23456 /// of its first operand, then on successive elements of its second operand,
23457 /// returning the resulting values in a vector. For example, if
23458 /// A = < float a0, float a1, float a2, float a3 >
23460 /// B = < float b0, float b1, float b2, float b3 >
23461 /// then the result of doing a horizontal operation on A and B is
23462 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
23463 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
23464 /// A horizontal-op B, for some already available A and B, and if so then LHS is
23465 /// set to A, RHS to B, and the routine returns 'true'.
23466 /// Note that the binary operation should have the property that if one of the
23467 /// operands is UNDEF then the result is UNDEF.
23468 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
23469 // Look for the following pattern: if
23470 // A = < float a0, float a1, float a2, float a3 >
23471 // B = < float b0, float b1, float b2, float b3 >
23473 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
23474 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
23475 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
23476 // which is A horizontal-op B.
23478 // At least one of the operands should be a vector shuffle.
23479 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
23480 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
23483 MVT VT = LHS.getSimpleValueType();
23485 assert((VT.is128BitVector() || VT.is256BitVector()) &&
23486 "Unsupported vector type for horizontal add/sub");
23488 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
23489 // operate independently on 128-bit lanes.
23490 unsigned NumElts = VT.getVectorNumElements();
23491 unsigned NumLanes = VT.getSizeInBits()/128;
23492 unsigned NumLaneElts = NumElts / NumLanes;
23493 assert((NumLaneElts % 2 == 0) &&
23494 "Vector type should have an even number of elements in each lane");
23495 unsigned HalfLaneElts = NumLaneElts/2;
23497 // View LHS in the form
23498 // LHS = VECTOR_SHUFFLE A, B, LMask
23499 // If LHS is not a shuffle then pretend it is the shuffle
23500 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
23501 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
23504 SmallVector<int, 16> LMask(NumElts);
23505 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23506 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
23507 A = LHS.getOperand(0);
23508 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
23509 B = LHS.getOperand(1);
23510 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
23511 std::copy(Mask.begin(), Mask.end(), LMask.begin());
23513 if (LHS.getOpcode() != ISD::UNDEF)
23515 for (unsigned i = 0; i != NumElts; ++i)
23519 // Likewise, view RHS in the form
23520 // RHS = VECTOR_SHUFFLE C, D, RMask
23522 SmallVector<int, 16> RMask(NumElts);
23523 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
23524 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
23525 C = RHS.getOperand(0);
23526 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
23527 D = RHS.getOperand(1);
23528 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
23529 std::copy(Mask.begin(), Mask.end(), RMask.begin());
23531 if (RHS.getOpcode() != ISD::UNDEF)
23533 for (unsigned i = 0; i != NumElts; ++i)
23537 // Check that the shuffles are both shuffling the same vectors.
23538 if (!(A == C && B == D) && !(A == D && B == C))
23541 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
23542 if (!A.getNode() && !B.getNode())
23545 // If A and B occur in reverse order in RHS, then "swap" them (which means
23546 // rewriting the mask).
23548 ShuffleVectorSDNode::commuteMask(RMask);
23550 // At this point LHS and RHS are equivalent to
23551 // LHS = VECTOR_SHUFFLE A, B, LMask
23552 // RHS = VECTOR_SHUFFLE A, B, RMask
23553 // Check that the masks correspond to performing a horizontal operation.
23554 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
23555 for (unsigned i = 0; i != NumLaneElts; ++i) {
23556 int LIdx = LMask[i+l], RIdx = RMask[i+l];
23558 // Ignore any UNDEF components.
23559 if (LIdx < 0 || RIdx < 0 ||
23560 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
23561 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
23564 // Check that successive elements are being operated on. If not, this is
23565 // not a horizontal operation.
23566 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
23567 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
23568 if (!(LIdx == Index && RIdx == Index + 1) &&
23569 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
23574 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
23575 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
23579 /// Do target-specific dag combines on floating point adds.
23580 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
23581 const X86Subtarget *Subtarget) {
23582 EVT VT = N->getValueType(0);
23583 SDValue LHS = N->getOperand(0);
23584 SDValue RHS = N->getOperand(1);
23586 // Try to synthesize horizontal adds from adds of shuffles.
23587 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23588 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23589 isHorizontalBinOp(LHS, RHS, true))
23590 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
23594 /// Do target-specific dag combines on floating point subs.
23595 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
23596 const X86Subtarget *Subtarget) {
23597 EVT VT = N->getValueType(0);
23598 SDValue LHS = N->getOperand(0);
23599 SDValue RHS = N->getOperand(1);
23601 // Try to synthesize horizontal subs from subs of shuffles.
23602 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
23603 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
23604 isHorizontalBinOp(LHS, RHS, false))
23605 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
23609 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
23610 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
23611 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
23613 // F[X]OR(0.0, x) -> x
23614 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23615 if (C->getValueAPF().isPosZero())
23616 return N->getOperand(1);
23618 // F[X]OR(x, 0.0) -> x
23619 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23620 if (C->getValueAPF().isPosZero())
23621 return N->getOperand(0);
23625 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
23626 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
23627 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
23629 // Only perform optimizations if UnsafeMath is used.
23630 if (!DAG.getTarget().Options.UnsafeFPMath)
23633 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
23634 // into FMINC and FMAXC, which are Commutative operations.
23635 unsigned NewOp = 0;
23636 switch (N->getOpcode()) {
23637 default: llvm_unreachable("unknown opcode");
23638 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
23639 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
23642 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
23643 N->getOperand(0), N->getOperand(1));
23646 /// Do target-specific dag combines on X86ISD::FAND nodes.
23647 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
23648 // FAND(0.0, x) -> 0.0
23649 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23650 if (C->getValueAPF().isPosZero())
23651 return N->getOperand(0);
23653 // FAND(x, 0.0) -> 0.0
23654 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23655 if (C->getValueAPF().isPosZero())
23656 return N->getOperand(1);
23661 /// Do target-specific dag combines on X86ISD::FANDN nodes
23662 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
23663 // FANDN(0.0, x) -> x
23664 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
23665 if (C->getValueAPF().isPosZero())
23666 return N->getOperand(1);
23668 // FANDN(x, 0.0) -> 0.0
23669 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
23670 if (C->getValueAPF().isPosZero())
23671 return N->getOperand(1);
23676 static SDValue PerformBTCombine(SDNode *N,
23678 TargetLowering::DAGCombinerInfo &DCI) {
23679 // BT ignores high bits in the bit index operand.
23680 SDValue Op1 = N->getOperand(1);
23681 if (Op1.hasOneUse()) {
23682 unsigned BitWidth = Op1.getValueSizeInBits();
23683 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
23684 APInt KnownZero, KnownOne;
23685 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
23686 !DCI.isBeforeLegalizeOps());
23687 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23688 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
23689 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
23690 DCI.CommitTargetLoweringOpt(TLO);
23695 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
23696 SDValue Op = N->getOperand(0);
23697 if (Op.getOpcode() == ISD::BITCAST)
23698 Op = Op.getOperand(0);
23699 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
23700 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
23701 VT.getVectorElementType().getSizeInBits() ==
23702 OpVT.getVectorElementType().getSizeInBits()) {
23703 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
23708 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
23709 const X86Subtarget *Subtarget) {
23710 EVT VT = N->getValueType(0);
23711 if (!VT.isVector())
23714 SDValue N0 = N->getOperand(0);
23715 SDValue N1 = N->getOperand(1);
23716 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
23719 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
23720 // both SSE and AVX2 since there is no sign-extended shift right
23721 // operation on a vector with 64-bit elements.
23722 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
23723 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
23724 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
23725 N0.getOpcode() == ISD::SIGN_EXTEND)) {
23726 SDValue N00 = N0.getOperand(0);
23728 // EXTLOAD has a better solution on AVX2,
23729 // it may be replaced with X86ISD::VSEXT node.
23730 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
23731 if (!ISD::isNormalLoad(N00.getNode()))
23734 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
23735 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
23737 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
23743 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
23744 TargetLowering::DAGCombinerInfo &DCI,
23745 const X86Subtarget *Subtarget) {
23746 SDValue N0 = N->getOperand(0);
23747 EVT VT = N->getValueType(0);
23748 EVT SVT = VT.getScalarType();
23749 EVT InVT = N0->getValueType(0);
23750 EVT InSVT = InVT.getScalarType();
23753 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
23754 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
23755 // This exposes the sext to the sdivrem lowering, so that it directly extends
23756 // from AH (which we otherwise need to do contortions to access).
23757 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
23758 InVT == MVT::i8 && VT == MVT::i32) {
23759 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23760 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
23761 N0.getOperand(0), N0.getOperand(1));
23762 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23763 return R.getValue(1);
23766 if (!DCI.isBeforeLegalizeOps()) {
23767 if (N0.getValueType() == MVT::i1) {
23768 SDValue Zero = DAG.getConstant(0, DL, VT);
23770 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
23771 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
23776 if (VT.isVector()) {
23777 auto ExtendToVec128 = [&DAG](SDLoc DL, SDValue N) {
23778 EVT InVT = N->getValueType(0);
23779 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
23780 128 / InVT.getScalarSizeInBits());
23781 SmallVector<SDValue, 8> Opnds(128 / InVT.getSizeInBits(),
23782 DAG.getUNDEF(InVT));
23784 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
23787 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
23788 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
23789 if (VT.getSizeInBits() == 128 &&
23790 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
23791 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
23792 SDValue ExOp = ExtendToVec128(DL, N0);
23793 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
23796 // On pre-AVX2 targets, split into 128-bit nodes of
23797 // ISD::SIGN_EXTEND_VECTOR_INREG.
23798 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
23799 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
23800 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
23801 unsigned NumVecs = VT.getSizeInBits() / 128;
23802 unsigned NumSubElts = 128 / SVT.getSizeInBits();
23803 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
23804 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
23806 SmallVector<SDValue, 8> Opnds;
23807 for (unsigned i = 0, Offset = 0; i != NumVecs;
23808 ++i, Offset += NumSubElts) {
23809 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
23810 DAG.getIntPtrConstant(Offset, DL));
23811 SrcVec = ExtendToVec128(DL, SrcVec);
23812 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
23813 Opnds.push_back(SrcVec);
23815 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
23819 if (!Subtarget->hasFp256())
23822 if (VT.isVector() && VT.getSizeInBits() == 256) {
23823 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23831 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
23832 const X86Subtarget* Subtarget) {
23834 EVT VT = N->getValueType(0);
23836 // Let legalize expand this if it isn't a legal type yet.
23837 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
23840 EVT ScalarVT = VT.getScalarType();
23841 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
23842 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
23845 SDValue A = N->getOperand(0);
23846 SDValue B = N->getOperand(1);
23847 SDValue C = N->getOperand(2);
23849 bool NegA = (A.getOpcode() == ISD::FNEG);
23850 bool NegB = (B.getOpcode() == ISD::FNEG);
23851 bool NegC = (C.getOpcode() == ISD::FNEG);
23853 // Negative multiplication when NegA xor NegB
23854 bool NegMul = (NegA != NegB);
23856 A = A.getOperand(0);
23858 B = B.getOperand(0);
23860 C = C.getOperand(0);
23864 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
23866 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
23868 return DAG.getNode(Opcode, dl, VT, A, B, C);
23871 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
23872 TargetLowering::DAGCombinerInfo &DCI,
23873 const X86Subtarget *Subtarget) {
23874 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
23875 // (and (i32 x86isd::setcc_carry), 1)
23876 // This eliminates the zext. This transformation is necessary because
23877 // ISD::SETCC is always legalized to i8.
23879 SDValue N0 = N->getOperand(0);
23880 EVT VT = N->getValueType(0);
23882 if (N0.getOpcode() == ISD::AND &&
23884 N0.getOperand(0).hasOneUse()) {
23885 SDValue N00 = N0.getOperand(0);
23886 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23887 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23888 if (!C || C->getZExtValue() != 1)
23890 return DAG.getNode(ISD::AND, dl, VT,
23891 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23892 N00.getOperand(0), N00.getOperand(1)),
23893 DAG.getConstant(1, dl, VT));
23897 if (N0.getOpcode() == ISD::TRUNCATE &&
23899 N0.getOperand(0).hasOneUse()) {
23900 SDValue N00 = N0.getOperand(0);
23901 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23902 return DAG.getNode(ISD::AND, dl, VT,
23903 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
23904 N00.getOperand(0), N00.getOperand(1)),
23905 DAG.getConstant(1, dl, VT));
23908 if (VT.is256BitVector()) {
23909 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
23914 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
23915 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
23916 // This exposes the zext to the udivrem lowering, so that it directly extends
23917 // from AH (which we otherwise need to do contortions to access).
23918 if (N0.getOpcode() == ISD::UDIVREM &&
23919 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
23920 (VT == MVT::i32 || VT == MVT::i64)) {
23921 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
23922 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
23923 N0.getOperand(0), N0.getOperand(1));
23924 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
23925 return R.getValue(1);
23931 // Optimize x == -y --> x+y == 0
23932 // x != -y --> x+y != 0
23933 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
23934 const X86Subtarget* Subtarget) {
23935 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
23936 SDValue LHS = N->getOperand(0);
23937 SDValue RHS = N->getOperand(1);
23938 EVT VT = N->getValueType(0);
23941 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
23942 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
23943 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
23944 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
23945 LHS.getOperand(1));
23946 return DAG.getSetCC(DL, N->getValueType(0), addV,
23947 DAG.getConstant(0, DL, addV.getValueType()), CC);
23949 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
23950 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
23951 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
23952 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
23953 RHS.getOperand(1));
23954 return DAG.getSetCC(DL, N->getValueType(0), addV,
23955 DAG.getConstant(0, DL, addV.getValueType()), CC);
23958 if (VT.getScalarType() == MVT::i1 &&
23959 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
23961 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23962 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23963 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23965 if (!IsSEXT0 || !IsVZero1) {
23966 // Swap the operands and update the condition code.
23967 std::swap(LHS, RHS);
23968 CC = ISD::getSetCCSwappedOperands(CC);
23970 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
23971 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
23972 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
23975 if (IsSEXT0 && IsVZero1) {
23976 assert(VT == LHS.getOperand(0).getValueType() &&
23977 "Uexpected operand type");
23978 if (CC == ISD::SETGT)
23979 return DAG.getConstant(0, DL, VT);
23980 if (CC == ISD::SETLE)
23981 return DAG.getConstant(1, DL, VT);
23982 if (CC == ISD::SETEQ || CC == ISD::SETGE)
23983 return DAG.getNOT(DL, LHS.getOperand(0), VT);
23985 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
23986 "Unexpected condition code!");
23987 return LHS.getOperand(0);
23994 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
23995 SelectionDAG &DAG) {
23997 MVT VT = Load->getSimpleValueType(0);
23998 MVT EVT = VT.getVectorElementType();
23999 SDValue Addr = Load->getOperand(1);
24000 SDValue NewAddr = DAG.getNode(
24001 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
24002 DAG.getConstant(Index * EVT.getStoreSize(), dl,
24003 Addr.getSimpleValueType()));
24006 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
24007 DAG.getMachineFunction().getMachineMemOperand(
24008 Load->getMemOperand(), 0, EVT.getStoreSize()));
24012 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
24013 const X86Subtarget *Subtarget) {
24015 MVT VT = N->getOperand(1)->getSimpleValueType(0);
24016 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
24017 "X86insertps is only defined for v4x32");
24019 SDValue Ld = N->getOperand(1);
24020 if (MayFoldLoad(Ld)) {
24021 // Extract the countS bits from the immediate so we can get the proper
24022 // address when narrowing the vector load to a specific element.
24023 // When the second source op is a memory address, insertps doesn't use
24024 // countS and just gets an f32 from that address.
24025 unsigned DestIndex =
24026 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
24028 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
24030 // Create this as a scalar to vector to match the instruction pattern.
24031 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
24032 // countS bits are ignored when loading from memory on insertps, which
24033 // means we don't need to explicitly set them to 0.
24034 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
24035 LoadScalarToVector, N->getOperand(2));
24040 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
24041 SDValue V0 = N->getOperand(0);
24042 SDValue V1 = N->getOperand(1);
24044 EVT VT = N->getValueType(0);
24046 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
24047 // operands and changing the mask to 1. This saves us a bunch of
24048 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
24049 // x86InstrInfo knows how to commute this back after instruction selection
24050 // if it would help register allocation.
24052 // TODO: If optimizing for size or a processor that doesn't suffer from
24053 // partial register update stalls, this should be transformed into a MOVSD
24054 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
24056 if (VT == MVT::v2f64)
24057 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
24058 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
24059 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
24060 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
24066 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
24067 // as "sbb reg,reg", since it can be extended without zext and produces
24068 // an all-ones bit which is more useful than 0/1 in some cases.
24069 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
24072 return DAG.getNode(ISD::AND, DL, VT,
24073 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
24074 DAG.getConstant(X86::COND_B, DL, MVT::i8),
24076 DAG.getConstant(1, DL, VT));
24077 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
24078 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
24079 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
24080 DAG.getConstant(X86::COND_B, DL, MVT::i8),
24084 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
24085 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
24086 TargetLowering::DAGCombinerInfo &DCI,
24087 const X86Subtarget *Subtarget) {
24089 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
24090 SDValue EFLAGS = N->getOperand(1);
24092 if (CC == X86::COND_A) {
24093 // Try to convert COND_A into COND_B in an attempt to facilitate
24094 // materializing "setb reg".
24096 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
24097 // cannot take an immediate as its first operand.
24099 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
24100 EFLAGS.getValueType().isInteger() &&
24101 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
24102 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
24103 EFLAGS.getNode()->getVTList(),
24104 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
24105 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
24106 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
24110 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
24111 // a zext and produces an all-ones bit which is more useful than 0/1 in some
24113 if (CC == X86::COND_B)
24114 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
24118 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24119 if (Flags.getNode()) {
24120 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24121 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
24127 // Optimize branch condition evaluation.
24129 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
24130 TargetLowering::DAGCombinerInfo &DCI,
24131 const X86Subtarget *Subtarget) {
24133 SDValue Chain = N->getOperand(0);
24134 SDValue Dest = N->getOperand(1);
24135 SDValue EFLAGS = N->getOperand(3);
24136 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
24140 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
24141 if (Flags.getNode()) {
24142 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
24143 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
24150 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
24151 SelectionDAG &DAG) {
24152 // Take advantage of vector comparisons producing 0 or -1 in each lane to
24153 // optimize away operation when it's from a constant.
24155 // The general transformation is:
24156 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
24157 // AND(VECTOR_CMP(x,y), constant2)
24158 // constant2 = UNARYOP(constant)
24160 // Early exit if this isn't a vector operation, the operand of the
24161 // unary operation isn't a bitwise AND, or if the sizes of the operations
24162 // aren't the same.
24163 EVT VT = N->getValueType(0);
24164 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
24165 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
24166 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
24169 // Now check that the other operand of the AND is a constant. We could
24170 // make the transformation for non-constant splats as well, but it's unclear
24171 // that would be a benefit as it would not eliminate any operations, just
24172 // perform one more step in scalar code before moving to the vector unit.
24173 if (BuildVectorSDNode *BV =
24174 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
24175 // Bail out if the vector isn't a constant.
24176 if (!BV->isConstant())
24179 // Everything checks out. Build up the new and improved node.
24181 EVT IntVT = BV->getValueType(0);
24182 // Create a new constant of the appropriate type for the transformed
24184 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
24185 // The AND node needs bitcasts to/from an integer vector type around it.
24186 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
24187 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
24188 N->getOperand(0)->getOperand(0), MaskConst);
24189 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
24196 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
24197 const X86Subtarget *Subtarget) {
24198 // First try to optimize away the conversion entirely when it's
24199 // conditionally from a constant. Vectors only.
24200 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
24201 if (Res != SDValue())
24204 // Now move on to more general possibilities.
24205 SDValue Op0 = N->getOperand(0);
24206 EVT InVT = Op0->getValueType(0);
24208 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
24209 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
24211 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
24212 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
24213 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
24216 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
24217 // a 32-bit target where SSE doesn't support i64->FP operations.
24218 if (Op0.getOpcode() == ISD::LOAD) {
24219 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
24220 EVT VT = Ld->getValueType(0);
24222 // This transformation is not supported if the result type is f16
24223 if (N->getValueType(0) == MVT::f16)
24226 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
24227 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
24228 !Subtarget->is64Bit() && VT == MVT::i64) {
24229 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
24230 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
24231 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
24238 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
24239 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
24240 X86TargetLowering::DAGCombinerInfo &DCI) {
24241 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
24242 // the result is either zero or one (depending on the input carry bit).
24243 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
24244 if (X86::isZeroNode(N->getOperand(0)) &&
24245 X86::isZeroNode(N->getOperand(1)) &&
24246 // We don't have a good way to replace an EFLAGS use, so only do this when
24248 SDValue(N, 1).use_empty()) {
24250 EVT VT = N->getValueType(0);
24251 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
24252 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
24253 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
24254 DAG.getConstant(X86::COND_B, DL,
24257 DAG.getConstant(1, DL, VT));
24258 return DCI.CombineTo(N, Res1, CarryOut);
24264 // fold (add Y, (sete X, 0)) -> adc 0, Y
24265 // (add Y, (setne X, 0)) -> sbb -1, Y
24266 // (sub (sete X, 0), Y) -> sbb 0, Y
24267 // (sub (setne X, 0), Y) -> adc -1, Y
24268 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
24271 // Look through ZExts.
24272 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
24273 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
24276 SDValue SetCC = Ext.getOperand(0);
24277 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
24280 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
24281 if (CC != X86::COND_E && CC != X86::COND_NE)
24284 SDValue Cmp = SetCC.getOperand(1);
24285 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
24286 !X86::isZeroNode(Cmp.getOperand(1)) ||
24287 !Cmp.getOperand(0).getValueType().isInteger())
24290 SDValue CmpOp0 = Cmp.getOperand(0);
24291 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
24292 DAG.getConstant(1, DL, CmpOp0.getValueType()));
24294 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
24295 if (CC == X86::COND_NE)
24296 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
24297 DL, OtherVal.getValueType(), OtherVal,
24298 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
24300 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
24301 DL, OtherVal.getValueType(), OtherVal,
24302 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
24305 /// PerformADDCombine - Do target-specific dag combines on integer adds.
24306 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
24307 const X86Subtarget *Subtarget) {
24308 EVT VT = N->getValueType(0);
24309 SDValue Op0 = N->getOperand(0);
24310 SDValue Op1 = N->getOperand(1);
24312 // Try to synthesize horizontal adds from adds of shuffles.
24313 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24314 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24315 isHorizontalBinOp(Op0, Op1, true))
24316 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
24318 return OptimizeConditionalInDecrement(N, DAG);
24321 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
24322 const X86Subtarget *Subtarget) {
24323 SDValue Op0 = N->getOperand(0);
24324 SDValue Op1 = N->getOperand(1);
24326 // X86 can't encode an immediate LHS of a sub. See if we can push the
24327 // negation into a preceding instruction.
24328 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
24329 // If the RHS of the sub is a XOR with one use and a constant, invert the
24330 // immediate. Then add one to the LHS of the sub so we can turn
24331 // X-Y -> X+~Y+1, saving one register.
24332 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
24333 isa<ConstantSDNode>(Op1.getOperand(1))) {
24334 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
24335 EVT VT = Op0.getValueType();
24336 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
24338 DAG.getConstant(~XorC, SDLoc(Op1), VT));
24339 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
24340 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
24344 // Try to synthesize horizontal adds from adds of shuffles.
24345 EVT VT = N->getValueType(0);
24346 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
24347 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
24348 isHorizontalBinOp(Op0, Op1, true))
24349 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
24351 return OptimizeConditionalInDecrement(N, DAG);
24354 /// performVZEXTCombine - Performs build vector combines
24355 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
24356 TargetLowering::DAGCombinerInfo &DCI,
24357 const X86Subtarget *Subtarget) {
24359 MVT VT = N->getSimpleValueType(0);
24360 SDValue Op = N->getOperand(0);
24361 MVT OpVT = Op.getSimpleValueType();
24362 MVT OpEltVT = OpVT.getVectorElementType();
24363 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
24365 // (vzext (bitcast (vzext (x)) -> (vzext x)
24367 while (V.getOpcode() == ISD::BITCAST)
24368 V = V.getOperand(0);
24370 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
24371 MVT InnerVT = V.getSimpleValueType();
24372 MVT InnerEltVT = InnerVT.getVectorElementType();
24374 // If the element sizes match exactly, we can just do one larger vzext. This
24375 // is always an exact type match as vzext operates on integer types.
24376 if (OpEltVT == InnerEltVT) {
24377 assert(OpVT == InnerVT && "Types must match for vzext!");
24378 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
24381 // The only other way we can combine them is if only a single element of the
24382 // inner vzext is used in the input to the outer vzext.
24383 if (InnerEltVT.getSizeInBits() < InputBits)
24386 // In this case, the inner vzext is completely dead because we're going to
24387 // only look at bits inside of the low element. Just do the outer vzext on
24388 // a bitcast of the input to the inner.
24389 return DAG.getNode(X86ISD::VZEXT, DL, VT,
24390 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
24393 // Check if we can bypass extracting and re-inserting an element of an input
24394 // vector. Essentialy:
24395 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
24396 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
24397 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
24398 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
24399 SDValue ExtractedV = V.getOperand(0);
24400 SDValue OrigV = ExtractedV.getOperand(0);
24401 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
24402 if (ExtractIdx->getZExtValue() == 0) {
24403 MVT OrigVT = OrigV.getSimpleValueType();
24404 // Extract a subvector if necessary...
24405 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
24406 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
24407 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
24408 OrigVT.getVectorNumElements() / Ratio);
24409 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
24410 DAG.getIntPtrConstant(0, DL));
24412 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
24413 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
24420 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
24421 DAGCombinerInfo &DCI) const {
24422 SelectionDAG &DAG = DCI.DAG;
24423 switch (N->getOpcode()) {
24425 case ISD::EXTRACT_VECTOR_ELT:
24426 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
24429 case X86ISD::SHRUNKBLEND:
24430 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
24431 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
24432 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
24433 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
24434 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
24435 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
24436 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
24439 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
24440 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
24441 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
24442 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
24443 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
24444 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
24445 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
24446 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
24447 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
24448 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
24449 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
24451 case X86ISD::FOR: return PerformFORCombine(N, DAG);
24453 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
24454 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
24455 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
24456 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
24457 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
24458 case ISD::ANY_EXTEND:
24459 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
24460 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
24461 case ISD::SIGN_EXTEND_INREG:
24462 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
24463 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
24464 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
24465 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
24466 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
24467 case X86ISD::SHUFP: // Handle all target specific shuffles
24468 case X86ISD::PALIGNR:
24469 case X86ISD::UNPCKH:
24470 case X86ISD::UNPCKL:
24471 case X86ISD::MOVHLPS:
24472 case X86ISD::MOVLHPS:
24473 case X86ISD::PSHUFB:
24474 case X86ISD::PSHUFD:
24475 case X86ISD::PSHUFHW:
24476 case X86ISD::PSHUFLW:
24477 case X86ISD::MOVSS:
24478 case X86ISD::MOVSD:
24479 case X86ISD::VPERMILPI:
24480 case X86ISD::VPERM2X128:
24481 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
24482 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
24483 case ISD::INTRINSIC_WO_CHAIN:
24484 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
24485 case X86ISD::INSERTPS: {
24486 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
24487 return PerformINSERTPSCombine(N, DAG, Subtarget);
24490 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
24496 /// isTypeDesirableForOp - Return true if the target has native support for
24497 /// the specified value type and it is 'desirable' to use the type for the
24498 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
24499 /// instruction encodings are longer and some i16 instructions are slow.
24500 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
24501 if (!isTypeLegal(VT))
24503 if (VT != MVT::i16)
24510 case ISD::SIGN_EXTEND:
24511 case ISD::ZERO_EXTEND:
24512 case ISD::ANY_EXTEND:
24525 /// IsDesirableToPromoteOp - This method query the target whether it is
24526 /// beneficial for dag combiner to promote the specified node. If true, it
24527 /// should return the desired promotion type by reference.
24528 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
24529 EVT VT = Op.getValueType();
24530 if (VT != MVT::i16)
24533 bool Promote = false;
24534 bool Commute = false;
24535 switch (Op.getOpcode()) {
24538 LoadSDNode *LD = cast<LoadSDNode>(Op);
24539 // If the non-extending load has a single use and it's not live out, then it
24540 // might be folded.
24541 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
24542 Op.hasOneUse()*/) {
24543 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
24544 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
24545 // The only case where we'd want to promote LOAD (rather then it being
24546 // promoted as an operand is when it's only use is liveout.
24547 if (UI->getOpcode() != ISD::CopyToReg)
24554 case ISD::SIGN_EXTEND:
24555 case ISD::ZERO_EXTEND:
24556 case ISD::ANY_EXTEND:
24561 SDValue N0 = Op.getOperand(0);
24562 // Look out for (store (shl (load), x)).
24563 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
24576 SDValue N0 = Op.getOperand(0);
24577 SDValue N1 = Op.getOperand(1);
24578 if (!Commute && MayFoldLoad(N1))
24580 // Avoid disabling potential load folding opportunities.
24581 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
24583 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
24593 //===----------------------------------------------------------------------===//
24594 // X86 Inline Assembly Support
24595 //===----------------------------------------------------------------------===//
24597 // Helper to match a string separated by whitespace.
24598 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
24599 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
24601 for (StringRef Piece : Pieces) {
24602 if (!S.startswith(Piece)) // Check if the piece matches.
24605 S = S.substr(Piece.size());
24606 StringRef::size_type Pos = S.find_first_not_of(" \t");
24607 if (Pos == 0) // We matched a prefix.
24616 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
24618 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
24619 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
24620 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
24621 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
24623 if (AsmPieces.size() == 3)
24625 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
24632 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
24633 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
24635 std::string AsmStr = IA->getAsmString();
24637 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
24638 if (!Ty || Ty->getBitWidth() % 16 != 0)
24641 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
24642 SmallVector<StringRef, 4> AsmPieces;
24643 SplitString(AsmStr, AsmPieces, ";\n");
24645 switch (AsmPieces.size()) {
24646 default: return false;
24648 // FIXME: this should verify that we are targeting a 486 or better. If not,
24649 // we will turn this bswap into something that will be lowered to logical
24650 // ops instead of emitting the bswap asm. For now, we don't support 486 or
24651 // lower so don't worry about this.
24653 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
24654 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
24655 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
24656 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
24657 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
24658 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
24659 // No need to check constraints, nothing other than the equivalent of
24660 // "=r,0" would be valid here.
24661 return IntrinsicLowering::LowerToByteSwap(CI);
24664 // rorw $$8, ${0:w} --> llvm.bswap.i16
24665 if (CI->getType()->isIntegerTy(16) &&
24666 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24667 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
24668 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
24670 const std::string &ConstraintsStr = IA->getConstraintString();
24671 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24672 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24673 if (clobbersFlagRegisters(AsmPieces))
24674 return IntrinsicLowering::LowerToByteSwap(CI);
24678 if (CI->getType()->isIntegerTy(32) &&
24679 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
24680 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
24681 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
24682 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
24684 const std::string &ConstraintsStr = IA->getConstraintString();
24685 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
24686 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
24687 if (clobbersFlagRegisters(AsmPieces))
24688 return IntrinsicLowering::LowerToByteSwap(CI);
24691 if (CI->getType()->isIntegerTy(64)) {
24692 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
24693 if (Constraints.size() >= 2 &&
24694 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
24695 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
24696 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
24697 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
24698 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
24699 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
24700 return IntrinsicLowering::LowerToByteSwap(CI);
24708 /// getConstraintType - Given a constraint letter, return the type of
24709 /// constraint it is for this target.
24710 X86TargetLowering::ConstraintType
24711 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
24712 if (Constraint.size() == 1) {
24713 switch (Constraint[0]) {
24724 return C_RegisterClass;
24748 return TargetLowering::getConstraintType(Constraint);
24751 /// Examine constraint type and operand type and determine a weight value.
24752 /// This object must already have been set up with the operand type
24753 /// and the current alternative constraint selected.
24754 TargetLowering::ConstraintWeight
24755 X86TargetLowering::getSingleConstraintMatchWeight(
24756 AsmOperandInfo &info, const char *constraint) const {
24757 ConstraintWeight weight = CW_Invalid;
24758 Value *CallOperandVal = info.CallOperandVal;
24759 // If we don't have a value, we can't do a match,
24760 // but allow it at the lowest weight.
24761 if (!CallOperandVal)
24763 Type *type = CallOperandVal->getType();
24764 // Look at the constraint type.
24765 switch (*constraint) {
24767 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
24778 if (CallOperandVal->getType()->isIntegerTy())
24779 weight = CW_SpecificReg;
24784 if (type->isFloatingPointTy())
24785 weight = CW_SpecificReg;
24788 if (type->isX86_MMXTy() && Subtarget->hasMMX())
24789 weight = CW_SpecificReg;
24793 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
24794 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
24795 weight = CW_Register;
24798 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
24799 if (C->getZExtValue() <= 31)
24800 weight = CW_Constant;
24804 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24805 if (C->getZExtValue() <= 63)
24806 weight = CW_Constant;
24810 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24811 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
24812 weight = CW_Constant;
24816 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24817 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
24818 weight = CW_Constant;
24822 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24823 if (C->getZExtValue() <= 3)
24824 weight = CW_Constant;
24828 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24829 if (C->getZExtValue() <= 0xff)
24830 weight = CW_Constant;
24835 if (isa<ConstantFP>(CallOperandVal)) {
24836 weight = CW_Constant;
24840 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24841 if ((C->getSExtValue() >= -0x80000000LL) &&
24842 (C->getSExtValue() <= 0x7fffffffLL))
24843 weight = CW_Constant;
24847 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
24848 if (C->getZExtValue() <= 0xffffffff)
24849 weight = CW_Constant;
24856 /// LowerXConstraint - try to replace an X constraint, which matches anything,
24857 /// with another that has more specific requirements based on the type of the
24858 /// corresponding operand.
24859 const char *X86TargetLowering::
24860 LowerXConstraint(EVT ConstraintVT) const {
24861 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
24862 // 'f' like normal targets.
24863 if (ConstraintVT.isFloatingPoint()) {
24864 if (Subtarget->hasSSE2())
24866 if (Subtarget->hasSSE1())
24870 return TargetLowering::LowerXConstraint(ConstraintVT);
24873 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
24874 /// vector. If it is invalid, don't add anything to Ops.
24875 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
24876 std::string &Constraint,
24877 std::vector<SDValue>&Ops,
24878 SelectionDAG &DAG) const {
24881 // Only support length 1 constraints for now.
24882 if (Constraint.length() > 1) return;
24884 char ConstraintLetter = Constraint[0];
24885 switch (ConstraintLetter) {
24888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24889 if (C->getZExtValue() <= 31) {
24890 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24891 Op.getValueType());
24897 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24898 if (C->getZExtValue() <= 63) {
24899 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24900 Op.getValueType());
24906 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24907 if (isInt<8>(C->getSExtValue())) {
24908 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24909 Op.getValueType());
24915 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24916 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
24917 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
24918 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
24919 Op.getValueType());
24925 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24926 if (C->getZExtValue() <= 3) {
24927 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24928 Op.getValueType());
24934 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24935 if (C->getZExtValue() <= 255) {
24936 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24937 Op.getValueType());
24943 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24944 if (C->getZExtValue() <= 127) {
24945 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24946 Op.getValueType());
24952 // 32-bit signed value
24953 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24954 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24955 C->getSExtValue())) {
24956 // Widen to 64 bits here to get it sign extended.
24957 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
24960 // FIXME gcc accepts some relocatable values here too, but only in certain
24961 // memory models; it's complicated.
24966 // 32-bit unsigned value
24967 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
24968 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
24969 C->getZExtValue())) {
24970 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
24971 Op.getValueType());
24975 // FIXME gcc accepts some relocatable values here too, but only in certain
24976 // memory models; it's complicated.
24980 // Literal immediates are always ok.
24981 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
24982 // Widen to 64 bits here to get it sign extended.
24983 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
24987 // In any sort of PIC mode addresses need to be computed at runtime by
24988 // adding in a register or some sort of table lookup. These can't
24989 // be used as immediates.
24990 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
24993 // If we are in non-pic codegen mode, we allow the address of a global (with
24994 // an optional displacement) to be used with 'i'.
24995 GlobalAddressSDNode *GA = nullptr;
24996 int64_t Offset = 0;
24998 // Match either (GA), (GA+C), (GA+C1+C2), etc.
25000 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
25001 Offset += GA->getOffset();
25003 } else if (Op.getOpcode() == ISD::ADD) {
25004 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
25005 Offset += C->getZExtValue();
25006 Op = Op.getOperand(0);
25009 } else if (Op.getOpcode() == ISD::SUB) {
25010 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
25011 Offset += -C->getZExtValue();
25012 Op = Op.getOperand(0);
25017 // Otherwise, this isn't something we can handle, reject it.
25021 const GlobalValue *GV = GA->getGlobal();
25022 // If we require an extra load to get this address, as in PIC mode, we
25023 // can't accept it.
25024 if (isGlobalStubReference(
25025 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
25028 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
25029 GA->getValueType(0), Offset);
25034 if (Result.getNode()) {
25035 Ops.push_back(Result);
25038 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
25041 std::pair<unsigned, const TargetRegisterClass *>
25042 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
25043 const std::string &Constraint,
25045 // First, see if this is a constraint that directly corresponds to an LLVM
25047 if (Constraint.size() == 1) {
25048 // GCC Constraint Letters
25049 switch (Constraint[0]) {
25051 // TODO: Slight differences here in allocation order and leaving
25052 // RIP in the class. Do they matter any more here than they do
25053 // in the normal allocation?
25054 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
25055 if (Subtarget->is64Bit()) {
25056 if (VT == MVT::i32 || VT == MVT::f32)
25057 return std::make_pair(0U, &X86::GR32RegClass);
25058 if (VT == MVT::i16)
25059 return std::make_pair(0U, &X86::GR16RegClass);
25060 if (VT == MVT::i8 || VT == MVT::i1)
25061 return std::make_pair(0U, &X86::GR8RegClass);
25062 if (VT == MVT::i64 || VT == MVT::f64)
25063 return std::make_pair(0U, &X86::GR64RegClass);
25066 // 32-bit fallthrough
25067 case 'Q': // Q_REGS
25068 if (VT == MVT::i32 || VT == MVT::f32)
25069 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
25070 if (VT == MVT::i16)
25071 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
25072 if (VT == MVT::i8 || VT == MVT::i1)
25073 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
25074 if (VT == MVT::i64)
25075 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
25077 case 'r': // GENERAL_REGS
25078 case 'l': // INDEX_REGS
25079 if (VT == MVT::i8 || VT == MVT::i1)
25080 return std::make_pair(0U, &X86::GR8RegClass);
25081 if (VT == MVT::i16)
25082 return std::make_pair(0U, &X86::GR16RegClass);
25083 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
25084 return std::make_pair(0U, &X86::GR32RegClass);
25085 return std::make_pair(0U, &X86::GR64RegClass);
25086 case 'R': // LEGACY_REGS
25087 if (VT == MVT::i8 || VT == MVT::i1)
25088 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
25089 if (VT == MVT::i16)
25090 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
25091 if (VT == MVT::i32 || !Subtarget->is64Bit())
25092 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
25093 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
25094 case 'f': // FP Stack registers.
25095 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
25096 // value to the correct fpstack register class.
25097 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
25098 return std::make_pair(0U, &X86::RFP32RegClass);
25099 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
25100 return std::make_pair(0U, &X86::RFP64RegClass);
25101 return std::make_pair(0U, &X86::RFP80RegClass);
25102 case 'y': // MMX_REGS if MMX allowed.
25103 if (!Subtarget->hasMMX()) break;
25104 return std::make_pair(0U, &X86::VR64RegClass);
25105 case 'Y': // SSE_REGS if SSE2 allowed
25106 if (!Subtarget->hasSSE2()) break;
25108 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
25109 if (!Subtarget->hasSSE1()) break;
25111 switch (VT.SimpleTy) {
25113 // Scalar SSE types.
25116 return std::make_pair(0U, &X86::FR32RegClass);
25119 return std::make_pair(0U, &X86::FR64RegClass);
25127 return std::make_pair(0U, &X86::VR128RegClass);
25135 return std::make_pair(0U, &X86::VR256RegClass);
25140 return std::make_pair(0U, &X86::VR512RegClass);
25146 // Use the default implementation in TargetLowering to convert the register
25147 // constraint into a member of a register class.
25148 std::pair<unsigned, const TargetRegisterClass*> Res;
25149 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
25151 // Not found as a standard register?
25153 // Map st(0) -> st(7) -> ST0
25154 if (Constraint.size() == 7 && Constraint[0] == '{' &&
25155 tolower(Constraint[1]) == 's' &&
25156 tolower(Constraint[2]) == 't' &&
25157 Constraint[3] == '(' &&
25158 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
25159 Constraint[5] == ')' &&
25160 Constraint[6] == '}') {
25162 Res.first = X86::FP0+Constraint[4]-'0';
25163 Res.second = &X86::RFP80RegClass;
25167 // GCC allows "st(0)" to be called just plain "st".
25168 if (StringRef("{st}").equals_lower(Constraint)) {
25169 Res.first = X86::FP0;
25170 Res.second = &X86::RFP80RegClass;
25175 if (StringRef("{flags}").equals_lower(Constraint)) {
25176 Res.first = X86::EFLAGS;
25177 Res.second = &X86::CCRRegClass;
25181 // 'A' means EAX + EDX.
25182 if (Constraint == "A") {
25183 Res.first = X86::EAX;
25184 Res.second = &X86::GR32_ADRegClass;
25190 // Otherwise, check to see if this is a register class of the wrong value
25191 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
25192 // turn into {ax},{dx}.
25193 if (Res.second->hasType(VT))
25194 return Res; // Correct type already, nothing to do.
25196 // All of the single-register GCC register classes map their values onto
25197 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
25198 // really want an 8-bit or 32-bit register, map to the appropriate register
25199 // class and return the appropriate register.
25200 if (Res.second == &X86::GR16RegClass) {
25201 if (VT == MVT::i8 || VT == MVT::i1) {
25202 unsigned DestReg = 0;
25203 switch (Res.first) {
25205 case X86::AX: DestReg = X86::AL; break;
25206 case X86::DX: DestReg = X86::DL; break;
25207 case X86::CX: DestReg = X86::CL; break;
25208 case X86::BX: DestReg = X86::BL; break;
25211 Res.first = DestReg;
25212 Res.second = &X86::GR8RegClass;
25214 } else if (VT == MVT::i32 || VT == MVT::f32) {
25215 unsigned DestReg = 0;
25216 switch (Res.first) {
25218 case X86::AX: DestReg = X86::EAX; break;
25219 case X86::DX: DestReg = X86::EDX; break;
25220 case X86::CX: DestReg = X86::ECX; break;
25221 case X86::BX: DestReg = X86::EBX; break;
25222 case X86::SI: DestReg = X86::ESI; break;
25223 case X86::DI: DestReg = X86::EDI; break;
25224 case X86::BP: DestReg = X86::EBP; break;
25225 case X86::SP: DestReg = X86::ESP; break;
25228 Res.first = DestReg;
25229 Res.second = &X86::GR32RegClass;
25231 } else if (VT == MVT::i64 || VT == MVT::f64) {
25232 unsigned DestReg = 0;
25233 switch (Res.first) {
25235 case X86::AX: DestReg = X86::RAX; break;
25236 case X86::DX: DestReg = X86::RDX; break;
25237 case X86::CX: DestReg = X86::RCX; break;
25238 case X86::BX: DestReg = X86::RBX; break;
25239 case X86::SI: DestReg = X86::RSI; break;
25240 case X86::DI: DestReg = X86::RDI; break;
25241 case X86::BP: DestReg = X86::RBP; break;
25242 case X86::SP: DestReg = X86::RSP; break;
25245 Res.first = DestReg;
25246 Res.second = &X86::GR64RegClass;
25249 } else if (Res.second == &X86::FR32RegClass ||
25250 Res.second == &X86::FR64RegClass ||
25251 Res.second == &X86::VR128RegClass ||
25252 Res.second == &X86::VR256RegClass ||
25253 Res.second == &X86::FR32XRegClass ||
25254 Res.second == &X86::FR64XRegClass ||
25255 Res.second == &X86::VR128XRegClass ||
25256 Res.second == &X86::VR256XRegClass ||
25257 Res.second == &X86::VR512RegClass) {
25258 // Handle references to XMM physical registers that got mapped into the
25259 // wrong class. This can happen with constraints like {xmm0} where the
25260 // target independent register mapper will just pick the first match it can
25261 // find, ignoring the required type.
25263 if (VT == MVT::f32 || VT == MVT::i32)
25264 Res.second = &X86::FR32RegClass;
25265 else if (VT == MVT::f64 || VT == MVT::i64)
25266 Res.second = &X86::FR64RegClass;
25267 else if (X86::VR128RegClass.hasType(VT))
25268 Res.second = &X86::VR128RegClass;
25269 else if (X86::VR256RegClass.hasType(VT))
25270 Res.second = &X86::VR256RegClass;
25271 else if (X86::VR512RegClass.hasType(VT))
25272 Res.second = &X86::VR512RegClass;
25278 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
25280 // Scaling factors are not free at all.
25281 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
25282 // will take 2 allocations in the out of order engine instead of 1
25283 // for plain addressing mode, i.e. inst (reg1).
25285 // vaddps (%rsi,%drx), %ymm0, %ymm1
25286 // Requires two allocations (one for the load, one for the computation)
25288 // vaddps (%rsi), %ymm0, %ymm1
25289 // Requires just 1 allocation, i.e., freeing allocations for other operations
25290 // and having less micro operations to execute.
25292 // For some X86 architectures, this is even worse because for instance for
25293 // stores, the complex addressing mode forces the instruction to use the
25294 // "load" ports instead of the dedicated "store" port.
25295 // E.g., on Haswell:
25296 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
25297 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
25298 if (isLegalAddressingMode(AM, Ty))
25299 // Scale represents reg2 * scale, thus account for 1
25300 // as soon as we use a second register.
25301 return AM.Scale != 0;
25305 bool X86TargetLowering::isTargetFTOL() const {
25306 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();