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 #define DEBUG_TYPE "x86-isel"
16 #include "X86ISelLowering.h"
18 #include "X86InstrBuilder.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/IntrinsicLowering.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCExpr.h"
41 #include "llvm/MC/MCSymbol.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VariadicFunction.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Target/TargetOptions.h"
55 STATISTIC(NumTailCalls, "Number of tail calls");
57 // Forward declarations.
58 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
61 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
62 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
63 /// simple subregister reference. Idx is an index in the 128 bits we
64 /// want. It need not be aligned to a 128-bit bounday. That makes
65 /// lowering EXTRACT_VECTOR_ELT operations easier.
66 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
67 SelectionDAG &DAG, DebugLoc dl) {
68 EVT VT = Vec.getValueType();
69 assert(VT.is256BitVector() && "Unexpected vector size!");
70 EVT ElVT = VT.getVectorElementType();
71 unsigned Factor = VT.getSizeInBits()/128;
72 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
73 VT.getVectorNumElements()/Factor);
75 // Extract from UNDEF is UNDEF.
76 if (Vec.getOpcode() == ISD::UNDEF)
77 return DAG.getUNDEF(ResultVT);
79 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
80 // we can match to VEXTRACTF128.
81 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
83 // This is the index of the first element of the 128-bit chunk
85 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
88 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
89 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
95 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
96 /// sets things up to match to an AVX VINSERTF128 instruction or a
97 /// simple superregister reference. Idx is an index in the 128 bits
98 /// we want. It need not be aligned to a 128-bit bounday. That makes
99 /// lowering INSERT_VECTOR_ELT operations easier.
100 static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
101 unsigned IdxVal, SelectionDAG &DAG,
103 // Inserting UNDEF is Result
104 if (Vec.getOpcode() == ISD::UNDEF)
107 EVT VT = Vec.getValueType();
108 assert(VT.is128BitVector() && "Unexpected vector size!");
110 EVT ElVT = VT.getVectorElementType();
111 EVT ResultVT = Result.getValueType();
113 // Insert the relevant 128 bits.
114 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
116 // This is the index of the first element of the 128-bit chunk
118 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
121 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
122 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
126 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
127 /// instructions. This is used because creating CONCAT_VECTOR nodes of
128 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
129 /// large BUILD_VECTORS.
130 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
131 unsigned NumElems, SelectionDAG &DAG,
133 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
134 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
137 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
138 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
139 bool is64Bit = Subtarget->is64Bit();
141 if (Subtarget->isTargetEnvMacho()) {
143 return new X86_64MachoTargetObjectFile();
144 return new TargetLoweringObjectFileMachO();
147 if (Subtarget->isTargetLinux())
148 return new X86LinuxTargetObjectFile();
149 if (Subtarget->isTargetELF())
150 return new TargetLoweringObjectFileELF();
151 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
152 return new TargetLoweringObjectFileCOFF();
153 llvm_unreachable("unknown subtarget type");
156 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
157 : TargetLowering(TM, createTLOF(TM)) {
158 Subtarget = &TM.getSubtarget<X86Subtarget>();
159 X86ScalarSSEf64 = Subtarget->hasSSE2();
160 X86ScalarSSEf32 = Subtarget->hasSSE1();
162 RegInfo = TM.getRegisterInfo();
163 TD = getDataLayout();
165 // Set up the TargetLowering object.
166 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
168 // X86 is weird, it always uses i8 for shift amounts and setcc results.
169 setBooleanContents(ZeroOrOneBooleanContent);
170 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
171 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
173 // For 64-bit since we have so many registers use the ILP scheduler, for
174 // 32-bit code use the register pressure specific scheduling.
175 // For Atom, always use ILP scheduling.
176 if (Subtarget->isAtom())
177 setSchedulingPreference(Sched::ILP);
178 else if (Subtarget->is64Bit())
179 setSchedulingPreference(Sched::ILP);
181 setSchedulingPreference(Sched::RegPressure);
182 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
184 // Bypass i32 with i8 on Atom when compiling with O2
185 if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default)
186 addBypassSlowDiv(32, 8);
188 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
189 // Setup Windows compiler runtime calls.
190 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
191 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
192 setLibcallName(RTLIB::SREM_I64, "_allrem");
193 setLibcallName(RTLIB::UREM_I64, "_aullrem");
194 setLibcallName(RTLIB::MUL_I64, "_allmul");
195 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
196 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
197 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
198 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
199 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
201 // The _ftol2 runtime function has an unusual calling conv, which
202 // is modeled by a special pseudo-instruction.
203 setLibcallName(RTLIB::FPTOUINT_F64_I64, 0);
204 setLibcallName(RTLIB::FPTOUINT_F32_I64, 0);
205 setLibcallName(RTLIB::FPTOUINT_F64_I32, 0);
206 setLibcallName(RTLIB::FPTOUINT_F32_I32, 0);
209 if (Subtarget->isTargetDarwin()) {
210 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
211 setUseUnderscoreSetJmp(false);
212 setUseUnderscoreLongJmp(false);
213 } else if (Subtarget->isTargetMingw()) {
214 // MS runtime is weird: it exports _setjmp, but longjmp!
215 setUseUnderscoreSetJmp(true);
216 setUseUnderscoreLongJmp(false);
218 setUseUnderscoreSetJmp(true);
219 setUseUnderscoreLongJmp(true);
222 // Set up the register classes.
223 addRegisterClass(MVT::i8, &X86::GR8RegClass);
224 addRegisterClass(MVT::i16, &X86::GR16RegClass);
225 addRegisterClass(MVT::i32, &X86::GR32RegClass);
226 if (Subtarget->is64Bit())
227 addRegisterClass(MVT::i64, &X86::GR64RegClass);
229 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
231 // We don't accept any truncstore of integer registers.
232 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
233 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
234 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
235 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
236 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
237 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
239 // SETOEQ and SETUNE require checking two conditions.
240 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
241 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
242 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
243 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
244 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
245 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
247 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
249 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
250 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
251 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
253 if (Subtarget->is64Bit()) {
254 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
255 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
256 } else if (!TM.Options.UseSoftFloat) {
257 // We have an algorithm for SSE2->double, and we turn this into a
258 // 64-bit FILD followed by conditional FADD for other targets.
259 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
260 // We have an algorithm for SSE2, and we turn this into a 64-bit
261 // FILD for other targets.
262 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
265 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
267 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
268 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
270 if (!TM.Options.UseSoftFloat) {
271 // SSE has no i16 to fp conversion, only i32
272 if (X86ScalarSSEf32) {
273 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
274 // f32 and f64 cases are Legal, f80 case is not
275 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
277 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
278 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
281 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
282 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
285 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
286 // are Legal, f80 is custom lowered.
287 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
288 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
290 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
292 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
293 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
295 if (X86ScalarSSEf32) {
296 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
297 // f32 and f64 cases are Legal, f80 case is not
298 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
300 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
301 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
304 // Handle FP_TO_UINT by promoting the destination to a larger signed
306 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
307 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
308 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
310 if (Subtarget->is64Bit()) {
311 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
312 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
313 } else if (!TM.Options.UseSoftFloat) {
314 // Since AVX is a superset of SSE3, only check for SSE here.
315 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
316 // Expand FP_TO_UINT into a select.
317 // FIXME: We would like to use a Custom expander here eventually to do
318 // the optimal thing for SSE vs. the default expansion in the legalizer.
319 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
321 // With SSE3 we can use fisttpll to convert to a signed i64; without
322 // SSE, we're stuck with a fistpll.
323 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
326 if (isTargetFTOL()) {
327 // Use the _ftol2 runtime function, which has a pseudo-instruction
328 // to handle its weird calling convention.
329 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
332 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
333 if (!X86ScalarSSEf64) {
334 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
335 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
336 if (Subtarget->is64Bit()) {
337 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
338 // Without SSE, i64->f64 goes through memory.
339 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
343 // Scalar integer divide and remainder are lowered to use operations that
344 // produce two results, to match the available instructions. This exposes
345 // the two-result form to trivial CSE, which is able to combine x/y and x%y
346 // into a single instruction.
348 // Scalar integer multiply-high is also lowered to use two-result
349 // operations, to match the available instructions. However, plain multiply
350 // (low) operations are left as Legal, as there are single-result
351 // instructions for this in x86. Using the two-result multiply instructions
352 // when both high and low results are needed must be arranged by dagcombine.
353 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
355 setOperationAction(ISD::MULHS, VT, Expand);
356 setOperationAction(ISD::MULHU, VT, Expand);
357 setOperationAction(ISD::SDIV, VT, Expand);
358 setOperationAction(ISD::UDIV, VT, Expand);
359 setOperationAction(ISD::SREM, VT, Expand);
360 setOperationAction(ISD::UREM, VT, Expand);
362 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
363 setOperationAction(ISD::ADDC, VT, Custom);
364 setOperationAction(ISD::ADDE, VT, Custom);
365 setOperationAction(ISD::SUBC, VT, Custom);
366 setOperationAction(ISD::SUBE, VT, Custom);
369 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
370 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
371 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
372 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
373 if (Subtarget->is64Bit())
374 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
375 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
376 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
377 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
378 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
379 setOperationAction(ISD::FREM , MVT::f32 , Expand);
380 setOperationAction(ISD::FREM , MVT::f64 , Expand);
381 setOperationAction(ISD::FREM , MVT::f80 , Expand);
382 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
384 // Promote the i8 variants and force them on up to i32 which has a shorter
386 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
387 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
388 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
389 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
390 if (Subtarget->hasBMI()) {
391 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
392 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
396 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
397 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
398 if (Subtarget->is64Bit())
399 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
402 if (Subtarget->hasLZCNT()) {
403 // When promoting the i8 variants, force them to i32 for a shorter
405 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
406 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
407 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
408 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
409 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
410 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
411 if (Subtarget->is64Bit())
412 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
414 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
415 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
416 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
417 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
418 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
419 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
420 if (Subtarget->is64Bit()) {
421 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
422 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
426 if (Subtarget->hasPOPCNT()) {
427 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
429 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
430 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
431 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
432 if (Subtarget->is64Bit())
433 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
436 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
437 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
439 // These should be promoted to a larger select which is supported.
440 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
441 // X86 wants to expand cmov itself.
442 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
443 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
444 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
445 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
446 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
447 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
448 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
449 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
450 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
451 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
452 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
453 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
454 if (Subtarget->is64Bit()) {
455 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
456 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
458 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
459 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intened to support
460 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
461 // support continuation, user-level threading, and etc.. As a result, no
462 // other SjLj exception interfaces are implemented and please don't build
463 // your own exception handling based on them.
464 // LLVM/Clang supports zero-cost DWARF exception handling.
465 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
466 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
469 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
470 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
471 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
472 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
473 if (Subtarget->is64Bit())
474 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
475 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
476 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
477 if (Subtarget->is64Bit()) {
478 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
479 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
480 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
481 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
482 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
484 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
485 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
486 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
487 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
488 if (Subtarget->is64Bit()) {
489 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
490 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
491 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
494 if (Subtarget->hasSSE1())
495 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
497 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
498 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
500 // On X86 and X86-64, atomic operations are lowered to locked instructions.
501 // Locked instructions, in turn, have implicit fence semantics (all memory
502 // operations are flushed before issuing the locked instruction, and they
503 // are not buffered), so we can fold away the common pattern of
504 // fence-atomic-fence.
505 setShouldFoldAtomicFences(true);
507 // Expand certain atomics
508 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
510 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
511 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
512 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
515 if (!Subtarget->is64Bit()) {
516 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
517 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
518 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
519 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
520 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
521 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
522 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
523 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
524 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
525 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
526 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
527 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
530 if (Subtarget->hasCmpxchg16b()) {
531 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
534 // FIXME - use subtarget debug flags
535 if (!Subtarget->isTargetDarwin() &&
536 !Subtarget->isTargetELF() &&
537 !Subtarget->isTargetCygMing()) {
538 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
541 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
542 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
543 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
544 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
545 if (Subtarget->is64Bit()) {
546 setExceptionPointerRegister(X86::RAX);
547 setExceptionSelectorRegister(X86::RDX);
549 setExceptionPointerRegister(X86::EAX);
550 setExceptionSelectorRegister(X86::EDX);
552 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
553 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
555 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
556 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
558 setOperationAction(ISD::TRAP, MVT::Other, Legal);
559 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
561 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
562 setOperationAction(ISD::VASTART , MVT::Other, Custom);
563 setOperationAction(ISD::VAEND , MVT::Other, Expand);
564 if (Subtarget->is64Bit()) {
565 setOperationAction(ISD::VAARG , MVT::Other, Custom);
566 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
568 setOperationAction(ISD::VAARG , MVT::Other, Expand);
569 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
572 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
573 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
575 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
576 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
577 MVT::i64 : MVT::i32, Custom);
578 else if (TM.Options.EnableSegmentedStacks)
579 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
580 MVT::i64 : MVT::i32, Custom);
582 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
583 MVT::i64 : MVT::i32, Expand);
585 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
586 // f32 and f64 use SSE.
587 // Set up the FP register classes.
588 addRegisterClass(MVT::f32, &X86::FR32RegClass);
589 addRegisterClass(MVT::f64, &X86::FR64RegClass);
591 // Use ANDPD to simulate FABS.
592 setOperationAction(ISD::FABS , MVT::f64, Custom);
593 setOperationAction(ISD::FABS , MVT::f32, Custom);
595 // Use XORP to simulate FNEG.
596 setOperationAction(ISD::FNEG , MVT::f64, Custom);
597 setOperationAction(ISD::FNEG , MVT::f32, Custom);
599 // Use ANDPD and ORPD to simulate FCOPYSIGN.
600 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
601 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
603 // Lower this to FGETSIGNx86 plus an AND.
604 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
605 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
607 // We don't support sin/cos/fmod
608 setOperationAction(ISD::FSIN , MVT::f64, Expand);
609 setOperationAction(ISD::FCOS , MVT::f64, Expand);
610 setOperationAction(ISD::FSIN , MVT::f32, Expand);
611 setOperationAction(ISD::FCOS , MVT::f32, Expand);
613 // Expand FP immediates into loads from the stack, except for the special
615 addLegalFPImmediate(APFloat(+0.0)); // xorpd
616 addLegalFPImmediate(APFloat(+0.0f)); // xorps
617 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
618 // Use SSE for f32, x87 for f64.
619 // Set up the FP register classes.
620 addRegisterClass(MVT::f32, &X86::FR32RegClass);
621 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
623 // Use ANDPS to simulate FABS.
624 setOperationAction(ISD::FABS , MVT::f32, Custom);
626 // Use XORP to simulate FNEG.
627 setOperationAction(ISD::FNEG , MVT::f32, Custom);
629 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
631 // Use ANDPS and ORPS to simulate FCOPYSIGN.
632 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
633 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
635 // We don't support sin/cos/fmod
636 setOperationAction(ISD::FSIN , MVT::f32, Expand);
637 setOperationAction(ISD::FCOS , MVT::f32, Expand);
639 // Special cases we handle for FP constants.
640 addLegalFPImmediate(APFloat(+0.0f)); // xorps
641 addLegalFPImmediate(APFloat(+0.0)); // FLD0
642 addLegalFPImmediate(APFloat(+1.0)); // FLD1
643 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
644 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
646 if (!TM.Options.UnsafeFPMath) {
647 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
648 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
650 } else if (!TM.Options.UseSoftFloat) {
651 // f32 and f64 in x87.
652 // Set up the FP register classes.
653 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
654 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
656 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
657 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
658 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
659 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
661 if (!TM.Options.UnsafeFPMath) {
662 setOperationAction(ISD::FSIN , MVT::f32 , Expand);
663 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
664 setOperationAction(ISD::FCOS , MVT::f32 , Expand);
665 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
667 addLegalFPImmediate(APFloat(+0.0)); // FLD0
668 addLegalFPImmediate(APFloat(+1.0)); // FLD1
669 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
670 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
671 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
672 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
673 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
674 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
677 // We don't support FMA.
678 setOperationAction(ISD::FMA, MVT::f64, Expand);
679 setOperationAction(ISD::FMA, MVT::f32, Expand);
681 // Long double always uses X87.
682 if (!TM.Options.UseSoftFloat) {
683 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
684 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
685 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
687 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
688 addLegalFPImmediate(TmpFlt); // FLD0
690 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
693 APFloat TmpFlt2(+1.0);
694 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
696 addLegalFPImmediate(TmpFlt2); // FLD1
697 TmpFlt2.changeSign();
698 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
701 if (!TM.Options.UnsafeFPMath) {
702 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
703 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
706 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
707 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
708 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
709 setOperationAction(ISD::FRINT, MVT::f80, Expand);
710 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
711 setOperationAction(ISD::FMA, MVT::f80, Expand);
714 // Always use a library call for pow.
715 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
716 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
717 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
719 setOperationAction(ISD::FLOG, MVT::f80, Expand);
720 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
721 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
722 setOperationAction(ISD::FEXP, MVT::f80, Expand);
723 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
725 // First set operation action for all vector types to either promote
726 // (for widening) or expand (for scalarization). Then we will selectively
727 // turn on ones that can be effectively codegen'd.
728 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
729 VT <= MVT::LAST_VECTOR_VALUETYPE; ++VT) {
730 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
731 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
732 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
733 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
734 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
735 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
736 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
737 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
738 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
739 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
740 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
741 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
742 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
743 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
744 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
745 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
746 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
747 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
748 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
749 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
750 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
751 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
752 setOperationAction(ISD::FMA, (MVT::SimpleValueType)VT, Expand);
753 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
754 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
755 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
756 setOperationAction(ISD::FFLOOR, (MVT::SimpleValueType)VT, Expand);
757 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
758 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
759 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
760 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
761 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
762 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
763 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
764 setOperationAction(ISD::CTTZ_ZERO_UNDEF, (MVT::SimpleValueType)VT, Expand);
765 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
766 setOperationAction(ISD::CTLZ_ZERO_UNDEF, (MVT::SimpleValueType)VT, Expand);
767 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
768 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
769 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
770 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
771 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
772 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
773 setOperationAction(ISD::SETCC, (MVT::SimpleValueType)VT, Expand);
774 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
775 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
776 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
777 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
778 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
779 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
780 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
781 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
782 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
783 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
784 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
785 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
786 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
787 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
788 setOperationAction(ISD::VSELECT, (MVT::SimpleValueType)VT, Expand);
789 for (int InnerVT = MVT::FIRST_VECTOR_VALUETYPE;
790 InnerVT <= MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
791 setTruncStoreAction((MVT::SimpleValueType)VT,
792 (MVT::SimpleValueType)InnerVT, Expand);
793 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
794 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
795 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
798 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
799 // with -msoft-float, disable use of MMX as well.
800 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
801 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
802 // No operations on x86mmx supported, everything uses intrinsics.
805 // MMX-sized vectors (other than x86mmx) are expected to be expanded
806 // into smaller operations.
807 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
808 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
809 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
810 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
811 setOperationAction(ISD::AND, MVT::v8i8, Expand);
812 setOperationAction(ISD::AND, MVT::v4i16, Expand);
813 setOperationAction(ISD::AND, MVT::v2i32, Expand);
814 setOperationAction(ISD::AND, MVT::v1i64, Expand);
815 setOperationAction(ISD::OR, MVT::v8i8, Expand);
816 setOperationAction(ISD::OR, MVT::v4i16, Expand);
817 setOperationAction(ISD::OR, MVT::v2i32, Expand);
818 setOperationAction(ISD::OR, MVT::v1i64, Expand);
819 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
820 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
821 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
822 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
823 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
824 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
825 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
826 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
827 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
828 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
829 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
830 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
831 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
832 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
833 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
834 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
835 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
837 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
838 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
840 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
841 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
842 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
843 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
844 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
845 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
846 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
847 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
848 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
849 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
851 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
854 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
855 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
857 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
858 // registers cannot be used even for integer operations.
859 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
860 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
861 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
862 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
864 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
865 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
866 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
867 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
868 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
869 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
870 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
871 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
872 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
873 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
874 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
875 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
876 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
877 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
878 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
879 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
880 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
882 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
883 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
884 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
885 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
887 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
888 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
889 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
891 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
893 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
894 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
895 MVT VT = (MVT::SimpleValueType)i;
896 // Do not attempt to custom lower non-power-of-2 vectors
897 if (!isPowerOf2_32(VT.getVectorNumElements()))
899 // Do not attempt to custom lower non-128-bit vectors
900 if (!VT.is128BitVector())
902 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
903 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
904 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
907 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
909 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
914 if (Subtarget->is64Bit()) {
915 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
916 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
919 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
920 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
921 MVT VT = (MVT::SimpleValueType)i;
923 // Do not attempt to promote non-128-bit vectors
924 if (!VT.is128BitVector())
927 setOperationAction(ISD::AND, VT, Promote);
928 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
929 setOperationAction(ISD::OR, VT, Promote);
930 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
931 setOperationAction(ISD::XOR, VT, Promote);
932 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
933 setOperationAction(ISD::LOAD, VT, Promote);
934 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
935 setOperationAction(ISD::SELECT, VT, Promote);
936 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
939 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
941 // Custom lower v2i64 and v2f64 selects.
942 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
943 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
944 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
945 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
947 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
948 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
950 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
951 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
952 // As there is no 64-bit GPR available, we need build a special custom
953 // sequence to convert from v2i32 to v2f32.
954 if (!Subtarget->is64Bit())
955 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
957 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
958 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
960 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
963 if (Subtarget->hasSSE41()) {
964 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
965 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
966 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
967 setOperationAction(ISD::FRINT, MVT::f32, Legal);
968 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
969 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
970 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
971 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
972 setOperationAction(ISD::FRINT, MVT::f64, Legal);
973 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
975 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
976 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
978 // FIXME: Do we need to handle scalar-to-vector here?
979 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
981 setOperationAction(ISD::VSELECT, MVT::v2f64, Legal);
982 setOperationAction(ISD::VSELECT, MVT::v2i64, Legal);
983 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
984 setOperationAction(ISD::VSELECT, MVT::v4i32, Legal);
985 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
987 // i8 and i16 vectors are custom , because the source register and source
988 // source memory operand types are not the same width. f32 vectors are
989 // custom since the immediate controlling the insert encodes additional
991 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
992 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
993 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
994 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
997 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
998 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1001 // FIXME: these should be Legal but thats only for the case where
1002 // the index is constant. For now custom expand to deal with that.
1003 if (Subtarget->is64Bit()) {
1004 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1009 if (Subtarget->hasSSE2()) {
1010 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1011 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1013 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1014 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1016 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1017 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1019 if (Subtarget->hasAVX2()) {
1020 setOperationAction(ISD::SRL, MVT::v2i64, Legal);
1021 setOperationAction(ISD::SRL, MVT::v4i32, Legal);
1023 setOperationAction(ISD::SHL, MVT::v2i64, Legal);
1024 setOperationAction(ISD::SHL, MVT::v4i32, Legal);
1026 setOperationAction(ISD::SRA, MVT::v4i32, Legal);
1028 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1031 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1032 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1034 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1038 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX()) {
1039 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1040 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1041 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1042 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1043 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1044 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1046 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1047 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1048 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1050 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1051 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1052 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1053 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1054 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1055 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1056 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1057 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1059 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1060 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1063 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1064 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1065 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1066 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1068 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1070 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
1072 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1073 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1074 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1076 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1077 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1078 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1080 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
1082 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1083 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1085 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1086 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1088 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1089 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1091 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1092 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1093 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1094 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1096 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1097 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1098 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1100 setOperationAction(ISD::VSELECT, MVT::v4f64, Legal);
1101 setOperationAction(ISD::VSELECT, MVT::v4i64, Legal);
1102 setOperationAction(ISD::VSELECT, MVT::v8i32, Legal);
1103 setOperationAction(ISD::VSELECT, MVT::v8f32, Legal);
1105 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1106 setOperationAction(ISD::FMA, MVT::v8f32, Custom);
1107 setOperationAction(ISD::FMA, MVT::v4f64, Custom);
1108 setOperationAction(ISD::FMA, MVT::v4f32, Custom);
1109 setOperationAction(ISD::FMA, MVT::v2f64, Custom);
1110 setOperationAction(ISD::FMA, MVT::f32, Custom);
1111 setOperationAction(ISD::FMA, MVT::f64, Custom);
1114 if (Subtarget->hasAVX2()) {
1115 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1116 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1117 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1118 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1120 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1121 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1122 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1123 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1125 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1126 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1127 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1128 // Don't lower v32i8 because there is no 128-bit byte mul
1130 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1132 setOperationAction(ISD::SRL, MVT::v4i64, Legal);
1133 setOperationAction(ISD::SRL, MVT::v8i32, Legal);
1135 setOperationAction(ISD::SHL, MVT::v4i64, Legal);
1136 setOperationAction(ISD::SHL, MVT::v8i32, Legal);
1138 setOperationAction(ISD::SRA, MVT::v8i32, Legal);
1140 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1141 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1142 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1143 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1145 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1146 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1147 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1148 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1150 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1151 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1152 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1153 // Don't lower v32i8 because there is no 128-bit byte mul
1155 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1156 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1158 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1159 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1161 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1164 // Custom lower several nodes for 256-bit types.
1165 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1166 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1167 MVT VT = (MVT::SimpleValueType)i;
1169 // Extract subvector is special because the value type
1170 // (result) is 128-bit but the source is 256-bit wide.
1171 if (VT.is128BitVector())
1172 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1174 // Do not attempt to custom lower other non-256-bit vectors
1175 if (!VT.is256BitVector())
1178 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1179 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1180 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1181 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1182 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1183 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1184 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1187 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1188 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1189 MVT VT = (MVT::SimpleValueType)i;
1191 // Do not attempt to promote non-256-bit vectors
1192 if (!VT.is256BitVector())
1195 setOperationAction(ISD::AND, VT, Promote);
1196 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1197 setOperationAction(ISD::OR, VT, Promote);
1198 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1199 setOperationAction(ISD::XOR, VT, Promote);
1200 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1201 setOperationAction(ISD::LOAD, VT, Promote);
1202 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1203 setOperationAction(ISD::SELECT, VT, Promote);
1204 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1208 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1209 // of this type with custom code.
1210 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1211 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1212 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1216 // We want to custom lower some of our intrinsics.
1217 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1218 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1221 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1222 // handle type legalization for these operations here.
1224 // FIXME: We really should do custom legalization for addition and
1225 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1226 // than generic legalization for 64-bit multiplication-with-overflow, though.
1227 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1228 // Add/Sub/Mul with overflow operations are custom lowered.
1230 setOperationAction(ISD::SADDO, VT, Custom);
1231 setOperationAction(ISD::UADDO, VT, Custom);
1232 setOperationAction(ISD::SSUBO, VT, Custom);
1233 setOperationAction(ISD::USUBO, VT, Custom);
1234 setOperationAction(ISD::SMULO, VT, Custom);
1235 setOperationAction(ISD::UMULO, VT, Custom);
1238 // There are no 8-bit 3-address imul/mul instructions
1239 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1240 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1242 if (!Subtarget->is64Bit()) {
1243 // These libcalls are not available in 32-bit.
1244 setLibcallName(RTLIB::SHL_I128, 0);
1245 setLibcallName(RTLIB::SRL_I128, 0);
1246 setLibcallName(RTLIB::SRA_I128, 0);
1249 // We have target-specific dag combine patterns for the following nodes:
1250 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1251 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1252 setTargetDAGCombine(ISD::VSELECT);
1253 setTargetDAGCombine(ISD::SELECT);
1254 setTargetDAGCombine(ISD::SHL);
1255 setTargetDAGCombine(ISD::SRA);
1256 setTargetDAGCombine(ISD::SRL);
1257 setTargetDAGCombine(ISD::OR);
1258 setTargetDAGCombine(ISD::AND);
1259 setTargetDAGCombine(ISD::ADD);
1260 setTargetDAGCombine(ISD::FADD);
1261 setTargetDAGCombine(ISD::FSUB);
1262 setTargetDAGCombine(ISD::FMA);
1263 setTargetDAGCombine(ISD::SUB);
1264 setTargetDAGCombine(ISD::LOAD);
1265 setTargetDAGCombine(ISD::STORE);
1266 setTargetDAGCombine(ISD::ZERO_EXTEND);
1267 setTargetDAGCombine(ISD::ANY_EXTEND);
1268 setTargetDAGCombine(ISD::SIGN_EXTEND);
1269 setTargetDAGCombine(ISD::TRUNCATE);
1270 setTargetDAGCombine(ISD::SINT_TO_FP);
1271 setTargetDAGCombine(ISD::SETCC);
1272 if (Subtarget->is64Bit())
1273 setTargetDAGCombine(ISD::MUL);
1274 setTargetDAGCombine(ISD::XOR);
1276 computeRegisterProperties();
1278 // On Darwin, -Os means optimize for size without hurting performance,
1279 // do not reduce the limit.
1280 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1281 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1282 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1283 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1284 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1285 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1286 setPrefLoopAlignment(4); // 2^4 bytes.
1287 benefitFromCodePlacementOpt = true;
1289 // Predictable cmov don't hurt on atom because it's in-order.
1290 predictableSelectIsExpensive = !Subtarget->isAtom();
1292 setPrefFunctionAlignment(4); // 2^4 bytes.
1296 EVT X86TargetLowering::getSetCCResultType(EVT VT) const {
1297 if (!VT.isVector()) return MVT::i8;
1298 return VT.changeVectorElementTypeToInteger();
1302 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1303 /// the desired ByVal argument alignment.
1304 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1307 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1308 if (VTy->getBitWidth() == 128)
1310 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1311 unsigned EltAlign = 0;
1312 getMaxByValAlign(ATy->getElementType(), EltAlign);
1313 if (EltAlign > MaxAlign)
1314 MaxAlign = EltAlign;
1315 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1316 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1317 unsigned EltAlign = 0;
1318 getMaxByValAlign(STy->getElementType(i), EltAlign);
1319 if (EltAlign > MaxAlign)
1320 MaxAlign = EltAlign;
1327 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1328 /// function arguments in the caller parameter area. For X86, aggregates
1329 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1330 /// are at 4-byte boundaries.
1331 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1332 if (Subtarget->is64Bit()) {
1333 // Max of 8 and alignment of type.
1334 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1341 if (Subtarget->hasSSE1())
1342 getMaxByValAlign(Ty, Align);
1346 /// getOptimalMemOpType - Returns the target specific optimal type for load
1347 /// and store operations as a result of memset, memcpy, and memmove
1348 /// lowering. If DstAlign is zero that means it's safe to destination
1349 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1350 /// means there isn't a need to check it against alignment requirement,
1351 /// probably because the source does not need to be loaded. If
1352 /// 'IsZeroVal' is true, that means it's safe to return a
1353 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1354 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1355 /// constant so it does not need to be loaded.
1356 /// It returns EVT::Other if the type should be determined using generic
1357 /// target-independent logic.
1359 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1360 unsigned DstAlign, unsigned SrcAlign,
1363 MachineFunction &MF) const {
1364 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1365 // linux. This is because the stack realignment code can't handle certain
1366 // cases like PR2962. This should be removed when PR2962 is fixed.
1367 const Function *F = MF.getFunction();
1369 !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat)) {
1371 (Subtarget->isUnalignedMemAccessFast() ||
1372 ((DstAlign == 0 || DstAlign >= 16) &&
1373 (SrcAlign == 0 || SrcAlign >= 16))) &&
1374 Subtarget->getStackAlignment() >= 16) {
1375 if (Subtarget->getStackAlignment() >= 32) {
1376 if (Subtarget->hasAVX2())
1378 if (Subtarget->hasAVX())
1381 if (Subtarget->hasSSE2())
1383 if (Subtarget->hasSSE1())
1385 } else if (!MemcpyStrSrc && Size >= 8 &&
1386 !Subtarget->is64Bit() &&
1387 Subtarget->getStackAlignment() >= 8 &&
1388 Subtarget->hasSSE2()) {
1389 // Do not use f64 to lower memcpy if source is string constant. It's
1390 // better to use i32 to avoid the loads.
1394 if (Subtarget->is64Bit() && Size >= 8)
1399 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1400 /// current function. The returned value is a member of the
1401 /// MachineJumpTableInfo::JTEntryKind enum.
1402 unsigned X86TargetLowering::getJumpTableEncoding() const {
1403 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1405 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1406 Subtarget->isPICStyleGOT())
1407 return MachineJumpTableInfo::EK_Custom32;
1409 // Otherwise, use the normal jump table encoding heuristics.
1410 return TargetLowering::getJumpTableEncoding();
1414 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1415 const MachineBasicBlock *MBB,
1416 unsigned uid,MCContext &Ctx) const{
1417 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1418 Subtarget->isPICStyleGOT());
1419 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1421 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1422 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1425 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1427 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1428 SelectionDAG &DAG) const {
1429 if (!Subtarget->is64Bit())
1430 // This doesn't have DebugLoc associated with it, but is not really the
1431 // same as a Register.
1432 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1436 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1437 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1439 const MCExpr *X86TargetLowering::
1440 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1441 MCContext &Ctx) const {
1442 // X86-64 uses RIP relative addressing based on the jump table label.
1443 if (Subtarget->isPICStyleRIPRel())
1444 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1446 // Otherwise, the reference is relative to the PIC base.
1447 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1450 // FIXME: Why this routine is here? Move to RegInfo!
1451 std::pair<const TargetRegisterClass*, uint8_t>
1452 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1453 const TargetRegisterClass *RRC = 0;
1455 switch (VT.getSimpleVT().SimpleTy) {
1457 return TargetLowering::findRepresentativeClass(VT);
1458 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1459 RRC = Subtarget->is64Bit() ?
1460 (const TargetRegisterClass*)&X86::GR64RegClass :
1461 (const TargetRegisterClass*)&X86::GR32RegClass;
1464 RRC = &X86::VR64RegClass;
1466 case MVT::f32: case MVT::f64:
1467 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1468 case MVT::v4f32: case MVT::v2f64:
1469 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1471 RRC = &X86::VR128RegClass;
1474 return std::make_pair(RRC, Cost);
1477 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1478 unsigned &Offset) const {
1479 if (!Subtarget->isTargetLinux())
1482 if (Subtarget->is64Bit()) {
1483 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1485 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1498 //===----------------------------------------------------------------------===//
1499 // Return Value Calling Convention Implementation
1500 //===----------------------------------------------------------------------===//
1502 #include "X86GenCallingConv.inc"
1505 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1506 MachineFunction &MF, bool isVarArg,
1507 const SmallVectorImpl<ISD::OutputArg> &Outs,
1508 LLVMContext &Context) const {
1509 SmallVector<CCValAssign, 16> RVLocs;
1510 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1512 return CCInfo.CheckReturn(Outs, RetCC_X86);
1516 X86TargetLowering::LowerReturn(SDValue Chain,
1517 CallingConv::ID CallConv, bool isVarArg,
1518 const SmallVectorImpl<ISD::OutputArg> &Outs,
1519 const SmallVectorImpl<SDValue> &OutVals,
1520 DebugLoc dl, SelectionDAG &DAG) const {
1521 MachineFunction &MF = DAG.getMachineFunction();
1522 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1524 SmallVector<CCValAssign, 16> RVLocs;
1525 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1526 RVLocs, *DAG.getContext());
1527 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1529 // Add the regs to the liveout set for the function.
1530 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1531 for (unsigned i = 0; i != RVLocs.size(); ++i)
1532 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1533 MRI.addLiveOut(RVLocs[i].getLocReg());
1537 SmallVector<SDValue, 6> RetOps;
1538 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1539 // Operand #1 = Bytes To Pop
1540 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1543 // Copy the result values into the output registers.
1544 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1545 CCValAssign &VA = RVLocs[i];
1546 assert(VA.isRegLoc() && "Can only return in registers!");
1547 SDValue ValToCopy = OutVals[i];
1548 EVT ValVT = ValToCopy.getValueType();
1550 // Promote values to the appropriate types
1551 if (VA.getLocInfo() == CCValAssign::SExt)
1552 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1553 else if (VA.getLocInfo() == CCValAssign::ZExt)
1554 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1555 else if (VA.getLocInfo() == CCValAssign::AExt)
1556 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1557 else if (VA.getLocInfo() == CCValAssign::BCvt)
1558 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1560 // If this is x86-64, and we disabled SSE, we can't return FP values,
1561 // or SSE or MMX vectors.
1562 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1563 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1564 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1565 report_fatal_error("SSE register return with SSE disabled");
1567 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1568 // llvm-gcc has never done it right and no one has noticed, so this
1569 // should be OK for now.
1570 if (ValVT == MVT::f64 &&
1571 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1572 report_fatal_error("SSE2 register return with SSE2 disabled");
1574 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1575 // the RET instruction and handled by the FP Stackifier.
1576 if (VA.getLocReg() == X86::ST0 ||
1577 VA.getLocReg() == X86::ST1) {
1578 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1579 // change the value to the FP stack register class.
1580 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1581 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1582 RetOps.push_back(ValToCopy);
1583 // Don't emit a copytoreg.
1587 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1588 // which is returned in RAX / RDX.
1589 if (Subtarget->is64Bit()) {
1590 if (ValVT == MVT::x86mmx) {
1591 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1592 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1593 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1595 // If we don't have SSE2 available, convert to v4f32 so the generated
1596 // register is legal.
1597 if (!Subtarget->hasSSE2())
1598 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1603 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1604 Flag = Chain.getValue(1);
1607 // The x86-64 ABI for returning structs by value requires that we copy
1608 // the sret argument into %rax for the return. We saved the argument into
1609 // a virtual register in the entry block, so now we copy the value out
1611 if (Subtarget->is64Bit() &&
1612 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1613 MachineFunction &MF = DAG.getMachineFunction();
1614 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1615 unsigned Reg = FuncInfo->getSRetReturnReg();
1617 "SRetReturnReg should have been set in LowerFormalArguments().");
1618 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1620 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1621 Flag = Chain.getValue(1);
1623 // RAX now acts like a return value.
1624 MRI.addLiveOut(X86::RAX);
1627 RetOps[0] = Chain; // Update chain.
1629 // Add the flag if we have it.
1631 RetOps.push_back(Flag);
1633 return DAG.getNode(X86ISD::RET_FLAG, dl,
1634 MVT::Other, &RetOps[0], RetOps.size());
1637 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1638 if (N->getNumValues() != 1)
1640 if (!N->hasNUsesOfValue(1, 0))
1643 SDValue TCChain = Chain;
1644 SDNode *Copy = *N->use_begin();
1645 if (Copy->getOpcode() == ISD::CopyToReg) {
1646 // If the copy has a glue operand, we conservatively assume it isn't safe to
1647 // perform a tail call.
1648 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1650 TCChain = Copy->getOperand(0);
1651 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
1654 bool HasRet = false;
1655 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1657 if (UI->getOpcode() != X86ISD::RET_FLAG)
1670 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1671 ISD::NodeType ExtendKind) const {
1673 // TODO: Is this also valid on 32-bit?
1674 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1675 ReturnMVT = MVT::i8;
1677 ReturnMVT = MVT::i32;
1679 EVT MinVT = getRegisterType(Context, ReturnMVT);
1680 return VT.bitsLT(MinVT) ? MinVT : VT;
1683 /// LowerCallResult - Lower the result values of a call into the
1684 /// appropriate copies out of appropriate physical registers.
1687 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1688 CallingConv::ID CallConv, bool isVarArg,
1689 const SmallVectorImpl<ISD::InputArg> &Ins,
1690 DebugLoc dl, SelectionDAG &DAG,
1691 SmallVectorImpl<SDValue> &InVals) const {
1693 // Assign locations to each value returned by this call.
1694 SmallVector<CCValAssign, 16> RVLocs;
1695 bool Is64Bit = Subtarget->is64Bit();
1696 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1697 getTargetMachine(), RVLocs, *DAG.getContext());
1698 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1700 // Copy all of the result registers out of their specified physreg.
1701 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1702 CCValAssign &VA = RVLocs[i];
1703 EVT CopyVT = VA.getValVT();
1705 // If this is x86-64, and we disabled SSE, we can't return FP values
1706 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1707 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1708 report_fatal_error("SSE register return with SSE disabled");
1713 // If this is a call to a function that returns an fp value on the floating
1714 // point stack, we must guarantee the value is popped from the stack, so
1715 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1716 // if the return value is not used. We use the FpPOP_RETVAL instruction
1718 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1719 // If we prefer to use the value in xmm registers, copy it out as f80 and
1720 // use a truncate to move it from fp stack reg to xmm reg.
1721 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1722 SDValue Ops[] = { Chain, InFlag };
1723 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1724 MVT::Other, MVT::Glue, Ops, 2), 1);
1725 Val = Chain.getValue(0);
1727 // Round the f80 to the right size, which also moves it to the appropriate
1729 if (CopyVT != VA.getValVT())
1730 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1731 // This truncation won't change the value.
1732 DAG.getIntPtrConstant(1));
1734 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1735 CopyVT, InFlag).getValue(1);
1736 Val = Chain.getValue(0);
1738 InFlag = Chain.getValue(2);
1739 InVals.push_back(Val);
1746 //===----------------------------------------------------------------------===//
1747 // C & StdCall & Fast Calling Convention implementation
1748 //===----------------------------------------------------------------------===//
1749 // StdCall calling convention seems to be standard for many Windows' API
1750 // routines and around. It differs from C calling convention just a little:
1751 // callee should clean up the stack, not caller. Symbols should be also
1752 // decorated in some fancy way :) It doesn't support any vector arguments.
1753 // For info on fast calling convention see Fast Calling Convention (tail call)
1754 // implementation LowerX86_32FastCCCallTo.
1756 /// CallIsStructReturn - Determines whether a call uses struct return
1758 enum StructReturnType {
1763 static StructReturnType
1764 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1766 return NotStructReturn;
1768 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
1769 if (!Flags.isSRet())
1770 return NotStructReturn;
1771 if (Flags.isInReg())
1772 return RegStructReturn;
1773 return StackStructReturn;
1776 /// ArgsAreStructReturn - Determines whether a function uses struct
1777 /// return semantics.
1778 static StructReturnType
1779 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1781 return NotStructReturn;
1783 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
1784 if (!Flags.isSRet())
1785 return NotStructReturn;
1786 if (Flags.isInReg())
1787 return RegStructReturn;
1788 return StackStructReturn;
1791 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1792 /// by "Src" to address "Dst" with size and alignment information specified by
1793 /// the specific parameter attribute. The copy will be passed as a byval
1794 /// function parameter.
1796 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1797 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1799 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1801 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1802 /*isVolatile*/false, /*AlwaysInline=*/true,
1803 MachinePointerInfo(), MachinePointerInfo());
1806 /// IsTailCallConvention - Return true if the calling convention is one that
1807 /// supports tail call optimization.
1808 static bool IsTailCallConvention(CallingConv::ID CC) {
1809 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1812 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1813 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
1817 CallingConv::ID CalleeCC = CS.getCallingConv();
1818 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1824 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1825 /// a tailcall target by changing its ABI.
1826 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
1827 bool GuaranteedTailCallOpt) {
1828 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1832 X86TargetLowering::LowerMemArgument(SDValue Chain,
1833 CallingConv::ID CallConv,
1834 const SmallVectorImpl<ISD::InputArg> &Ins,
1835 DebugLoc dl, SelectionDAG &DAG,
1836 const CCValAssign &VA,
1837 MachineFrameInfo *MFI,
1839 // Create the nodes corresponding to a load from this parameter slot.
1840 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1841 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv,
1842 getTargetMachine().Options.GuaranteedTailCallOpt);
1843 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1846 // If value is passed by pointer we have address passed instead of the value
1848 if (VA.getLocInfo() == CCValAssign::Indirect)
1849 ValVT = VA.getLocVT();
1851 ValVT = VA.getValVT();
1853 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1854 // changed with more analysis.
1855 // In case of tail call optimization mark all arguments mutable. Since they
1856 // could be overwritten by lowering of arguments in case of a tail call.
1857 if (Flags.isByVal()) {
1858 unsigned Bytes = Flags.getByValSize();
1859 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1860 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1861 return DAG.getFrameIndex(FI, getPointerTy());
1863 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1864 VA.getLocMemOffset(), isImmutable);
1865 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1866 return DAG.getLoad(ValVT, dl, Chain, FIN,
1867 MachinePointerInfo::getFixedStack(FI),
1868 false, false, false, 0);
1873 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1874 CallingConv::ID CallConv,
1876 const SmallVectorImpl<ISD::InputArg> &Ins,
1879 SmallVectorImpl<SDValue> &InVals)
1881 MachineFunction &MF = DAG.getMachineFunction();
1882 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1884 const Function* Fn = MF.getFunction();
1885 if (Fn->hasExternalLinkage() &&
1886 Subtarget->isTargetCygMing() &&
1887 Fn->getName() == "main")
1888 FuncInfo->setForceFramePointer(true);
1890 MachineFrameInfo *MFI = MF.getFrameInfo();
1891 bool Is64Bit = Subtarget->is64Bit();
1892 bool IsWindows = Subtarget->isTargetWindows();
1893 bool IsWin64 = Subtarget->isTargetWin64();
1895 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1896 "Var args not supported with calling convention fastcc or ghc");
1898 // Assign locations to all of the incoming arguments.
1899 SmallVector<CCValAssign, 16> ArgLocs;
1900 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1901 ArgLocs, *DAG.getContext());
1903 // Allocate shadow area for Win64
1905 CCInfo.AllocateStack(32, 8);
1908 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1910 unsigned LastVal = ~0U;
1912 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1913 CCValAssign &VA = ArgLocs[i];
1914 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1916 assert(VA.getValNo() != LastVal &&
1917 "Don't support value assigned to multiple locs yet");
1919 LastVal = VA.getValNo();
1921 if (VA.isRegLoc()) {
1922 EVT RegVT = VA.getLocVT();
1923 const TargetRegisterClass *RC;
1924 if (RegVT == MVT::i32)
1925 RC = &X86::GR32RegClass;
1926 else if (Is64Bit && RegVT == MVT::i64)
1927 RC = &X86::GR64RegClass;
1928 else if (RegVT == MVT::f32)
1929 RC = &X86::FR32RegClass;
1930 else if (RegVT == MVT::f64)
1931 RC = &X86::FR64RegClass;
1932 else if (RegVT.is256BitVector())
1933 RC = &X86::VR256RegClass;
1934 else if (RegVT.is128BitVector())
1935 RC = &X86::VR128RegClass;
1936 else if (RegVT == MVT::x86mmx)
1937 RC = &X86::VR64RegClass;
1939 llvm_unreachable("Unknown argument type!");
1941 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1942 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1944 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1945 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1947 if (VA.getLocInfo() == CCValAssign::SExt)
1948 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1949 DAG.getValueType(VA.getValVT()));
1950 else if (VA.getLocInfo() == CCValAssign::ZExt)
1951 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1952 DAG.getValueType(VA.getValVT()));
1953 else if (VA.getLocInfo() == CCValAssign::BCvt)
1954 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1956 if (VA.isExtInLoc()) {
1957 // Handle MMX values passed in XMM regs.
1958 if (RegVT.isVector()) {
1959 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1962 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1965 assert(VA.isMemLoc());
1966 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1969 // If value is passed via pointer - do a load.
1970 if (VA.getLocInfo() == CCValAssign::Indirect)
1971 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1972 MachinePointerInfo(), false, false, false, 0);
1974 InVals.push_back(ArgValue);
1977 // The x86-64 ABI for returning structs by value requires that we copy
1978 // the sret argument into %rax for the return. Save the argument into
1979 // a virtual register so that we can access it from the return points.
1980 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1981 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1982 unsigned Reg = FuncInfo->getSRetReturnReg();
1984 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1985 FuncInfo->setSRetReturnReg(Reg);
1987 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1988 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1991 unsigned StackSize = CCInfo.getNextStackOffset();
1992 // Align stack specially for tail calls.
1993 if (FuncIsMadeTailCallSafe(CallConv,
1994 MF.getTarget().Options.GuaranteedTailCallOpt))
1995 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1997 // If the function takes variable number of arguments, make a frame index for
1998 // the start of the first vararg value... for expansion of llvm.va_start.
2000 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2001 CallConv != CallingConv::X86_ThisCall)) {
2002 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
2005 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
2007 // FIXME: We should really autogenerate these arrays
2008 static const uint16_t GPR64ArgRegsWin64[] = {
2009 X86::RCX, X86::RDX, X86::R8, X86::R9
2011 static const uint16_t GPR64ArgRegs64Bit[] = {
2012 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2014 static const uint16_t XMMArgRegs64Bit[] = {
2015 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2016 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2018 const uint16_t *GPR64ArgRegs;
2019 unsigned NumXMMRegs = 0;
2022 // The XMM registers which might contain var arg parameters are shadowed
2023 // in their paired GPR. So we only need to save the GPR to their home
2025 TotalNumIntRegs = 4;
2026 GPR64ArgRegs = GPR64ArgRegsWin64;
2028 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
2029 GPR64ArgRegs = GPR64ArgRegs64Bit;
2031 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
2034 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
2037 bool NoImplicitFloatOps = Fn->getFnAttributes().
2038 hasAttribute(Attributes::NoImplicitFloat);
2039 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2040 "SSE register cannot be used when SSE is disabled!");
2041 assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
2042 NoImplicitFloatOps) &&
2043 "SSE register cannot be used when SSE is disabled!");
2044 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2045 !Subtarget->hasSSE1())
2046 // Kernel mode asks for SSE to be disabled, so don't push them
2048 TotalNumXMMRegs = 0;
2051 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
2052 // Get to the caller-allocated home save location. Add 8 to account
2053 // for the return address.
2054 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2055 FuncInfo->setRegSaveFrameIndex(
2056 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2057 // Fixup to set vararg frame on shadow area (4 x i64).
2059 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2061 // For X86-64, if there are vararg parameters that are passed via
2062 // registers, then we must store them to their spots on the stack so
2063 // they may be loaded by deferencing the result of va_next.
2064 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2065 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
2066 FuncInfo->setRegSaveFrameIndex(
2067 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
2071 // Store the integer parameter registers.
2072 SmallVector<SDValue, 8> MemOps;
2073 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2075 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2076 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
2077 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2078 DAG.getIntPtrConstant(Offset));
2079 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
2080 &X86::GR64RegClass);
2081 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2083 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2084 MachinePointerInfo::getFixedStack(
2085 FuncInfo->getRegSaveFrameIndex(), Offset),
2087 MemOps.push_back(Store);
2091 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
2092 // Now store the XMM (fp + vector) parameter registers.
2093 SmallVector<SDValue, 11> SaveXMMOps;
2094 SaveXMMOps.push_back(Chain);
2096 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2097 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
2098 SaveXMMOps.push_back(ALVal);
2100 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2101 FuncInfo->getRegSaveFrameIndex()));
2102 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2103 FuncInfo->getVarArgsFPOffset()));
2105 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
2106 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
2107 &X86::VR128RegClass);
2108 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
2109 SaveXMMOps.push_back(Val);
2111 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2113 &SaveXMMOps[0], SaveXMMOps.size()));
2116 if (!MemOps.empty())
2117 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2118 &MemOps[0], MemOps.size());
2122 // Some CCs need callee pop.
2123 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2124 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2125 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2127 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2128 // If this is an sret function, the return should pop the hidden pointer.
2129 if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2130 argsAreStructReturn(Ins) == StackStructReturn)
2131 FuncInfo->setBytesToPopOnReturn(4);
2135 // RegSaveFrameIndex is X86-64 only.
2136 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2137 if (CallConv == CallingConv::X86_FastCall ||
2138 CallConv == CallingConv::X86_ThisCall)
2139 // fastcc functions can't have varargs.
2140 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2143 FuncInfo->setArgumentStackSize(StackSize);
2149 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2150 SDValue StackPtr, SDValue Arg,
2151 DebugLoc dl, SelectionDAG &DAG,
2152 const CCValAssign &VA,
2153 ISD::ArgFlagsTy Flags) const {
2154 unsigned LocMemOffset = VA.getLocMemOffset();
2155 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2156 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2157 if (Flags.isByVal())
2158 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2160 return DAG.getStore(Chain, dl, Arg, PtrOff,
2161 MachinePointerInfo::getStack(LocMemOffset),
2165 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2166 /// optimization is performed and it is required.
2168 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2169 SDValue &OutRetAddr, SDValue Chain,
2170 bool IsTailCall, bool Is64Bit,
2171 int FPDiff, DebugLoc dl) const {
2172 // Adjust the Return address stack slot.
2173 EVT VT = getPointerTy();
2174 OutRetAddr = getReturnAddressFrameIndex(DAG);
2176 // Load the "old" Return address.
2177 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2178 false, false, false, 0);
2179 return SDValue(OutRetAddr.getNode(), 1);
2182 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2183 /// optimization is performed and it is required (FPDiff!=0).
2185 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
2186 SDValue Chain, SDValue RetAddrFrIdx, EVT PtrVT,
2187 unsigned SlotSize, int FPDiff, DebugLoc dl) {
2188 // Store the return address to the appropriate stack slot.
2189 if (!FPDiff) return Chain;
2190 // Calculate the new stack slot for the return address.
2191 int NewReturnAddrFI =
2192 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
2193 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2194 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2195 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2201 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2202 SmallVectorImpl<SDValue> &InVals) const {
2203 SelectionDAG &DAG = CLI.DAG;
2204 DebugLoc &dl = CLI.DL;
2205 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2206 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2207 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2208 SDValue Chain = CLI.Chain;
2209 SDValue Callee = CLI.Callee;
2210 CallingConv::ID CallConv = CLI.CallConv;
2211 bool &isTailCall = CLI.IsTailCall;
2212 bool isVarArg = CLI.IsVarArg;
2214 MachineFunction &MF = DAG.getMachineFunction();
2215 bool Is64Bit = Subtarget->is64Bit();
2216 bool IsWin64 = Subtarget->isTargetWin64();
2217 bool IsWindows = Subtarget->isTargetWindows();
2218 StructReturnType SR = callIsStructReturn(Outs);
2219 bool IsSibcall = false;
2221 if (MF.getTarget().Options.DisableTailCalls)
2225 // Check if it's really possible to do a tail call.
2226 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2227 isVarArg, SR != NotStructReturn,
2228 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2229 Outs, OutVals, Ins, DAG);
2231 // Sibcalls are automatically detected tailcalls which do not require
2233 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2240 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2241 "Var args not supported with calling convention fastcc or ghc");
2243 // Analyze operands of the call, assigning locations to each operand.
2244 SmallVector<CCValAssign, 16> ArgLocs;
2245 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2246 ArgLocs, *DAG.getContext());
2248 // Allocate shadow area for Win64
2250 CCInfo.AllocateStack(32, 8);
2253 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2255 // Get a count of how many bytes are to be pushed on the stack.
2256 unsigned NumBytes = CCInfo.getNextStackOffset();
2258 // This is a sibcall. The memory operands are available in caller's
2259 // own caller's stack.
2261 else if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2262 IsTailCallConvention(CallConv))
2263 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2266 if (isTailCall && !IsSibcall) {
2267 // Lower arguments at fp - stackoffset + fpdiff.
2268 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2269 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2271 FPDiff = NumBytesCallerPushed - NumBytes;
2273 // Set the delta of movement of the returnaddr stackslot.
2274 // But only set if delta is greater than previous delta.
2275 if (FPDiff < X86Info->getTCReturnAddrDelta())
2276 X86Info->setTCReturnAddrDelta(FPDiff);
2280 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2282 SDValue RetAddrFrIdx;
2283 // Load return address for tail calls.
2284 if (isTailCall && FPDiff)
2285 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2286 Is64Bit, FPDiff, dl);
2288 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2289 SmallVector<SDValue, 8> MemOpChains;
2292 // Walk the register/memloc assignments, inserting copies/loads. In the case
2293 // of tail call optimization arguments are handle later.
2294 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2295 CCValAssign &VA = ArgLocs[i];
2296 EVT RegVT = VA.getLocVT();
2297 SDValue Arg = OutVals[i];
2298 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2299 bool isByVal = Flags.isByVal();
2301 // Promote the value if needed.
2302 switch (VA.getLocInfo()) {
2303 default: llvm_unreachable("Unknown loc info!");
2304 case CCValAssign::Full: break;
2305 case CCValAssign::SExt:
2306 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2308 case CCValAssign::ZExt:
2309 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2311 case CCValAssign::AExt:
2312 if (RegVT.is128BitVector()) {
2313 // Special case: passing MMX values in XMM registers.
2314 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2315 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2316 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2318 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2320 case CCValAssign::BCvt:
2321 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2323 case CCValAssign::Indirect: {
2324 // Store the argument.
2325 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2326 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2327 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2328 MachinePointerInfo::getFixedStack(FI),
2335 if (VA.isRegLoc()) {
2336 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2337 if (isVarArg && IsWin64) {
2338 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2339 // shadow reg if callee is a varargs function.
2340 unsigned ShadowReg = 0;
2341 switch (VA.getLocReg()) {
2342 case X86::XMM0: ShadowReg = X86::RCX; break;
2343 case X86::XMM1: ShadowReg = X86::RDX; break;
2344 case X86::XMM2: ShadowReg = X86::R8; break;
2345 case X86::XMM3: ShadowReg = X86::R9; break;
2348 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2350 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2351 assert(VA.isMemLoc());
2352 if (StackPtr.getNode() == 0)
2353 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2355 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2356 dl, DAG, VA, Flags));
2360 if (!MemOpChains.empty())
2361 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2362 &MemOpChains[0], MemOpChains.size());
2364 if (Subtarget->isPICStyleGOT()) {
2365 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2368 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2369 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy())));
2371 // If we are tail calling and generating PIC/GOT style code load the
2372 // address of the callee into ECX. The value in ecx is used as target of
2373 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2374 // for tail calls on PIC/GOT architectures. Normally we would just put the
2375 // address of GOT into ebx and then call target@PLT. But for tail calls
2376 // ebx would be restored (since ebx is callee saved) before jumping to the
2379 // Note: The actual moving to ECX is done further down.
2380 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2381 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2382 !G->getGlobal()->hasProtectedVisibility())
2383 Callee = LowerGlobalAddress(Callee, DAG);
2384 else if (isa<ExternalSymbolSDNode>(Callee))
2385 Callee = LowerExternalSymbol(Callee, DAG);
2389 if (Is64Bit && isVarArg && !IsWin64) {
2390 // From AMD64 ABI document:
2391 // For calls that may call functions that use varargs or stdargs
2392 // (prototype-less calls or calls to functions containing ellipsis (...) in
2393 // the declaration) %al is used as hidden argument to specify the number
2394 // of SSE registers used. The contents of %al do not need to match exactly
2395 // the number of registers, but must be an ubound on the number of SSE
2396 // registers used and is in the range 0 - 8 inclusive.
2398 // Count the number of XMM registers allocated.
2399 static const uint16_t XMMArgRegs[] = {
2400 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2401 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2403 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2404 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2405 && "SSE registers cannot be used when SSE is disabled");
2407 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2408 DAG.getConstant(NumXMMRegs, MVT::i8)));
2411 // For tail calls lower the arguments to the 'real' stack slot.
2413 // Force all the incoming stack arguments to be loaded from the stack
2414 // before any new outgoing arguments are stored to the stack, because the
2415 // outgoing stack slots may alias the incoming argument stack slots, and
2416 // the alias isn't otherwise explicit. This is slightly more conservative
2417 // than necessary, because it means that each store effectively depends
2418 // on every argument instead of just those arguments it would clobber.
2419 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2421 SmallVector<SDValue, 8> MemOpChains2;
2424 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2425 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2426 CCValAssign &VA = ArgLocs[i];
2429 assert(VA.isMemLoc());
2430 SDValue Arg = OutVals[i];
2431 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2432 // Create frame index.
2433 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2434 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2435 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2436 FIN = DAG.getFrameIndex(FI, getPointerTy());
2438 if (Flags.isByVal()) {
2439 // Copy relative to framepointer.
2440 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2441 if (StackPtr.getNode() == 0)
2442 StackPtr = DAG.getCopyFromReg(Chain, dl,
2443 RegInfo->getStackRegister(),
2445 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2447 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2451 // Store relative to framepointer.
2452 MemOpChains2.push_back(
2453 DAG.getStore(ArgChain, dl, Arg, FIN,
2454 MachinePointerInfo::getFixedStack(FI),
2460 if (!MemOpChains2.empty())
2461 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2462 &MemOpChains2[0], MemOpChains2.size());
2464 // Store the return address to the appropriate stack slot.
2465 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2466 getPointerTy(), RegInfo->getSlotSize(),
2470 // Build a sequence of copy-to-reg nodes chained together with token chain
2471 // and flag operands which copy the outgoing args into registers.
2473 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2474 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2475 RegsToPass[i].second, InFlag);
2476 InFlag = Chain.getValue(1);
2479 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2480 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2481 // In the 64-bit large code model, we have to make all calls
2482 // through a register, since the call instruction's 32-bit
2483 // pc-relative offset may not be large enough to hold the whole
2485 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2486 // If the callee is a GlobalAddress node (quite common, every direct call
2487 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2490 // We should use extra load for direct calls to dllimported functions in
2492 const GlobalValue *GV = G->getGlobal();
2493 if (!GV->hasDLLImportLinkage()) {
2494 unsigned char OpFlags = 0;
2495 bool ExtraLoad = false;
2496 unsigned WrapperKind = ISD::DELETED_NODE;
2498 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2499 // external symbols most go through the PLT in PIC mode. If the symbol
2500 // has hidden or protected visibility, or if it is static or local, then
2501 // we don't need to use the PLT - we can directly call it.
2502 if (Subtarget->isTargetELF() &&
2503 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2504 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2505 OpFlags = X86II::MO_PLT;
2506 } else if (Subtarget->isPICStyleStubAny() &&
2507 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2508 (!Subtarget->getTargetTriple().isMacOSX() ||
2509 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2510 // PC-relative references to external symbols should go through $stub,
2511 // unless we're building with the leopard linker or later, which
2512 // automatically synthesizes these stubs.
2513 OpFlags = X86II::MO_DARWIN_STUB;
2514 } else if (Subtarget->isPICStyleRIPRel() &&
2515 isa<Function>(GV) &&
2516 cast<Function>(GV)->getFnAttributes().
2517 hasAttribute(Attributes::NonLazyBind)) {
2518 // If the function is marked as non-lazy, generate an indirect call
2519 // which loads from the GOT directly. This avoids runtime overhead
2520 // at the cost of eager binding (and one extra byte of encoding).
2521 OpFlags = X86II::MO_GOTPCREL;
2522 WrapperKind = X86ISD::WrapperRIP;
2526 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2527 G->getOffset(), OpFlags);
2529 // Add a wrapper if needed.
2530 if (WrapperKind != ISD::DELETED_NODE)
2531 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2532 // Add extra indirection if needed.
2534 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2535 MachinePointerInfo::getGOT(),
2536 false, false, false, 0);
2538 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2539 unsigned char OpFlags = 0;
2541 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2542 // external symbols should go through the PLT.
2543 if (Subtarget->isTargetELF() &&
2544 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2545 OpFlags = X86II::MO_PLT;
2546 } else if (Subtarget->isPICStyleStubAny() &&
2547 (!Subtarget->getTargetTriple().isMacOSX() ||
2548 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2549 // PC-relative references to external symbols should go through $stub,
2550 // unless we're building with the leopard linker or later, which
2551 // automatically synthesizes these stubs.
2552 OpFlags = X86II::MO_DARWIN_STUB;
2555 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2559 // Returns a chain & a flag for retval copy to use.
2560 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2561 SmallVector<SDValue, 8> Ops;
2563 if (!IsSibcall && isTailCall) {
2564 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2565 DAG.getIntPtrConstant(0, true), InFlag);
2566 InFlag = Chain.getValue(1);
2569 Ops.push_back(Chain);
2570 Ops.push_back(Callee);
2573 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2575 // Add argument registers to the end of the list so that they are known live
2577 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2578 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2579 RegsToPass[i].second.getValueType()));
2581 // Add a register mask operand representing the call-preserved registers.
2582 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2583 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2584 assert(Mask && "Missing call preserved mask for calling convention");
2585 Ops.push_back(DAG.getRegisterMask(Mask));
2587 if (InFlag.getNode())
2588 Ops.push_back(InFlag);
2592 //// If this is the first return lowered for this function, add the regs
2593 //// to the liveout set for the function.
2594 // This isn't right, although it's probably harmless on x86; liveouts
2595 // should be computed from returns not tail calls. Consider a void
2596 // function making a tail call to a function returning int.
2597 return DAG.getNode(X86ISD::TC_RETURN, dl,
2598 NodeTys, &Ops[0], Ops.size());
2601 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2602 InFlag = Chain.getValue(1);
2604 // Create the CALLSEQ_END node.
2605 unsigned NumBytesForCalleeToPush;
2606 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2607 getTargetMachine().Options.GuaranteedTailCallOpt))
2608 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2609 else if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2610 SR == StackStructReturn)
2611 // If this is a call to a struct-return function, the callee
2612 // pops the hidden struct pointer, so we have to push it back.
2613 // This is common for Darwin/X86, Linux & Mingw32 targets.
2614 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
2615 NumBytesForCalleeToPush = 4;
2617 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2619 // Returns a flag for retval copy to use.
2621 Chain = DAG.getCALLSEQ_END(Chain,
2622 DAG.getIntPtrConstant(NumBytes, true),
2623 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2626 InFlag = Chain.getValue(1);
2629 // Handle result values, copying them out of physregs into vregs that we
2631 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2632 Ins, dl, DAG, InVals);
2636 //===----------------------------------------------------------------------===//
2637 // Fast Calling Convention (tail call) implementation
2638 //===----------------------------------------------------------------------===//
2640 // Like std call, callee cleans arguments, convention except that ECX is
2641 // reserved for storing the tail called function address. Only 2 registers are
2642 // free for argument passing (inreg). Tail call optimization is performed
2644 // * tailcallopt is enabled
2645 // * caller/callee are fastcc
2646 // On X86_64 architecture with GOT-style position independent code only local
2647 // (within module) calls are supported at the moment.
2648 // To keep the stack aligned according to platform abi the function
2649 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2650 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2651 // If a tail called function callee has more arguments than the caller the
2652 // caller needs to make sure that there is room to move the RETADDR to. This is
2653 // achieved by reserving an area the size of the argument delta right after the
2654 // original REtADDR, but before the saved framepointer or the spilled registers
2655 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2667 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2668 /// for a 16 byte align requirement.
2670 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2671 SelectionDAG& DAG) const {
2672 MachineFunction &MF = DAG.getMachineFunction();
2673 const TargetMachine &TM = MF.getTarget();
2674 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2675 unsigned StackAlignment = TFI.getStackAlignment();
2676 uint64_t AlignMask = StackAlignment - 1;
2677 int64_t Offset = StackSize;
2678 unsigned SlotSize = RegInfo->getSlotSize();
2679 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2680 // Number smaller than 12 so just add the difference.
2681 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2683 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2684 Offset = ((~AlignMask) & Offset) + StackAlignment +
2685 (StackAlignment-SlotSize);
2690 /// MatchingStackOffset - Return true if the given stack call argument is
2691 /// already available in the same position (relatively) of the caller's
2692 /// incoming argument stack.
2694 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2695 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2696 const X86InstrInfo *TII) {
2697 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2699 if (Arg.getOpcode() == ISD::CopyFromReg) {
2700 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2701 if (!TargetRegisterInfo::isVirtualRegister(VR))
2703 MachineInstr *Def = MRI->getVRegDef(VR);
2706 if (!Flags.isByVal()) {
2707 if (!TII->isLoadFromStackSlot(Def, FI))
2710 unsigned Opcode = Def->getOpcode();
2711 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2712 Def->getOperand(1).isFI()) {
2713 FI = Def->getOperand(1).getIndex();
2714 Bytes = Flags.getByValSize();
2718 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2719 if (Flags.isByVal())
2720 // ByVal argument is passed in as a pointer but it's now being
2721 // dereferenced. e.g.
2722 // define @foo(%struct.X* %A) {
2723 // tail call @bar(%struct.X* byval %A)
2726 SDValue Ptr = Ld->getBasePtr();
2727 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2730 FI = FINode->getIndex();
2731 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2732 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2733 FI = FINode->getIndex();
2734 Bytes = Flags.getByValSize();
2738 assert(FI != INT_MAX);
2739 if (!MFI->isFixedObjectIndex(FI))
2741 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2744 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2745 /// for tail call optimization. Targets which want to do tail call
2746 /// optimization should implement this function.
2748 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2749 CallingConv::ID CalleeCC,
2751 bool isCalleeStructRet,
2752 bool isCallerStructRet,
2754 const SmallVectorImpl<ISD::OutputArg> &Outs,
2755 const SmallVectorImpl<SDValue> &OutVals,
2756 const SmallVectorImpl<ISD::InputArg> &Ins,
2757 SelectionDAG& DAG) const {
2758 if (!IsTailCallConvention(CalleeCC) &&
2759 CalleeCC != CallingConv::C)
2762 // If -tailcallopt is specified, make fastcc functions tail-callable.
2763 const MachineFunction &MF = DAG.getMachineFunction();
2764 const Function *CallerF = DAG.getMachineFunction().getFunction();
2766 // If the function return type is x86_fp80 and the callee return type is not,
2767 // then the FP_EXTEND of the call result is not a nop. It's not safe to
2768 // perform a tailcall optimization here.
2769 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
2772 CallingConv::ID CallerCC = CallerF->getCallingConv();
2773 bool CCMatch = CallerCC == CalleeCC;
2775 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2776 if (IsTailCallConvention(CalleeCC) && CCMatch)
2781 // Look for obvious safe cases to perform tail call optimization that do not
2782 // require ABI changes. This is what gcc calls sibcall.
2784 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2785 // emit a special epilogue.
2786 if (RegInfo->needsStackRealignment(MF))
2789 // Also avoid sibcall optimization if either caller or callee uses struct
2790 // return semantics.
2791 if (isCalleeStructRet || isCallerStructRet)
2794 // An stdcall caller is expected to clean up its arguments; the callee
2795 // isn't going to do that.
2796 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2799 // Do not sibcall optimize vararg calls unless all arguments are passed via
2801 if (isVarArg && !Outs.empty()) {
2803 // Optimizing for varargs on Win64 is unlikely to be safe without
2804 // additional testing.
2805 if (Subtarget->isTargetWin64())
2808 SmallVector<CCValAssign, 16> ArgLocs;
2809 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2810 getTargetMachine(), ArgLocs, *DAG.getContext());
2812 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2813 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2814 if (!ArgLocs[i].isRegLoc())
2818 // If the call result is in ST0 / ST1, it needs to be popped off the x87
2819 // stack. Therefore, if it's not used by the call it is not safe to optimize
2820 // this into a sibcall.
2821 bool Unused = false;
2822 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2829 SmallVector<CCValAssign, 16> RVLocs;
2830 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2831 getTargetMachine(), RVLocs, *DAG.getContext());
2832 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2833 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2834 CCValAssign &VA = RVLocs[i];
2835 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2840 // If the calling conventions do not match, then we'd better make sure the
2841 // results are returned in the same way as what the caller expects.
2843 SmallVector<CCValAssign, 16> RVLocs1;
2844 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2845 getTargetMachine(), RVLocs1, *DAG.getContext());
2846 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2848 SmallVector<CCValAssign, 16> RVLocs2;
2849 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2850 getTargetMachine(), RVLocs2, *DAG.getContext());
2851 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2853 if (RVLocs1.size() != RVLocs2.size())
2855 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2856 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2858 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2860 if (RVLocs1[i].isRegLoc()) {
2861 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2864 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2870 // If the callee takes no arguments then go on to check the results of the
2872 if (!Outs.empty()) {
2873 // Check if stack adjustment is needed. For now, do not do this if any
2874 // argument is passed on the stack.
2875 SmallVector<CCValAssign, 16> ArgLocs;
2876 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2877 getTargetMachine(), ArgLocs, *DAG.getContext());
2879 // Allocate shadow area for Win64
2880 if (Subtarget->isTargetWin64()) {
2881 CCInfo.AllocateStack(32, 8);
2884 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2885 if (CCInfo.getNextStackOffset()) {
2886 MachineFunction &MF = DAG.getMachineFunction();
2887 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2890 // Check if the arguments are already laid out in the right way as
2891 // the caller's fixed stack objects.
2892 MachineFrameInfo *MFI = MF.getFrameInfo();
2893 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2894 const X86InstrInfo *TII =
2895 ((const X86TargetMachine&)getTargetMachine()).getInstrInfo();
2896 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2897 CCValAssign &VA = ArgLocs[i];
2898 SDValue Arg = OutVals[i];
2899 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2900 if (VA.getLocInfo() == CCValAssign::Indirect)
2902 if (!VA.isRegLoc()) {
2903 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2910 // If the tailcall address may be in a register, then make sure it's
2911 // possible to register allocate for it. In 32-bit, the call address can
2912 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2913 // callee-saved registers are restored. These happen to be the same
2914 // registers used to pass 'inreg' arguments so watch out for those.
2915 if (!Subtarget->is64Bit() &&
2916 !isa<GlobalAddressSDNode>(Callee) &&
2917 !isa<ExternalSymbolSDNode>(Callee)) {
2918 unsigned NumInRegs = 0;
2919 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2920 CCValAssign &VA = ArgLocs[i];
2923 unsigned Reg = VA.getLocReg();
2926 case X86::EAX: case X86::EDX: case X86::ECX:
2927 if (++NumInRegs == 3)
2939 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
2940 const TargetLibraryInfo *libInfo) const {
2941 return X86::createFastISel(funcInfo, libInfo);
2945 //===----------------------------------------------------------------------===//
2946 // Other Lowering Hooks
2947 //===----------------------------------------------------------------------===//
2949 static bool MayFoldLoad(SDValue Op) {
2950 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2953 static bool MayFoldIntoStore(SDValue Op) {
2954 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2957 static bool isTargetShuffle(unsigned Opcode) {
2959 default: return false;
2960 case X86ISD::PSHUFD:
2961 case X86ISD::PSHUFHW:
2962 case X86ISD::PSHUFLW:
2964 case X86ISD::PALIGN:
2965 case X86ISD::MOVLHPS:
2966 case X86ISD::MOVLHPD:
2967 case X86ISD::MOVHLPS:
2968 case X86ISD::MOVLPS:
2969 case X86ISD::MOVLPD:
2970 case X86ISD::MOVSHDUP:
2971 case X86ISD::MOVSLDUP:
2972 case X86ISD::MOVDDUP:
2975 case X86ISD::UNPCKL:
2976 case X86ISD::UNPCKH:
2977 case X86ISD::VPERMILP:
2978 case X86ISD::VPERM2X128:
2979 case X86ISD::VPERMI:
2984 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2985 SDValue V1, SelectionDAG &DAG) {
2987 default: llvm_unreachable("Unknown x86 shuffle node");
2988 case X86ISD::MOVSHDUP:
2989 case X86ISD::MOVSLDUP:
2990 case X86ISD::MOVDDUP:
2991 return DAG.getNode(Opc, dl, VT, V1);
2995 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2996 SDValue V1, unsigned TargetMask,
2997 SelectionDAG &DAG) {
2999 default: llvm_unreachable("Unknown x86 shuffle node");
3000 case X86ISD::PSHUFD:
3001 case X86ISD::PSHUFHW:
3002 case X86ISD::PSHUFLW:
3003 case X86ISD::VPERMILP:
3004 case X86ISD::VPERMI:
3005 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3009 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3010 SDValue V1, SDValue V2, unsigned TargetMask,
3011 SelectionDAG &DAG) {
3013 default: llvm_unreachable("Unknown x86 shuffle node");
3014 case X86ISD::PALIGN:
3016 case X86ISD::VPERM2X128:
3017 return DAG.getNode(Opc, dl, VT, V1, V2,
3018 DAG.getConstant(TargetMask, MVT::i8));
3022 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3023 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3025 default: llvm_unreachable("Unknown x86 shuffle node");
3026 case X86ISD::MOVLHPS:
3027 case X86ISD::MOVLHPD:
3028 case X86ISD::MOVHLPS:
3029 case X86ISD::MOVLPS:
3030 case X86ISD::MOVLPD:
3033 case X86ISD::UNPCKL:
3034 case X86ISD::UNPCKH:
3035 return DAG.getNode(Opc, dl, VT, V1, V2);
3039 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3040 MachineFunction &MF = DAG.getMachineFunction();
3041 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3042 int ReturnAddrIndex = FuncInfo->getRAIndex();
3044 if (ReturnAddrIndex == 0) {
3045 // Set up a frame object for the return address.
3046 unsigned SlotSize = RegInfo->getSlotSize();
3047 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
3049 FuncInfo->setRAIndex(ReturnAddrIndex);
3052 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3056 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3057 bool hasSymbolicDisplacement) {
3058 // Offset should fit into 32 bit immediate field.
3059 if (!isInt<32>(Offset))
3062 // If we don't have a symbolic displacement - we don't have any extra
3064 if (!hasSymbolicDisplacement)
3067 // FIXME: Some tweaks might be needed for medium code model.
3068 if (M != CodeModel::Small && M != CodeModel::Kernel)
3071 // For small code model we assume that latest object is 16MB before end of 31
3072 // bits boundary. We may also accept pretty large negative constants knowing
3073 // that all objects are in the positive half of address space.
3074 if (M == CodeModel::Small && Offset < 16*1024*1024)
3077 // For kernel code model we know that all object resist in the negative half
3078 // of 32bits address space. We may not accept negative offsets, since they may
3079 // be just off and we may accept pretty large positive ones.
3080 if (M == CodeModel::Kernel && Offset > 0)
3086 /// isCalleePop - Determines whether the callee is required to pop its
3087 /// own arguments. Callee pop is necessary to support tail calls.
3088 bool X86::isCalleePop(CallingConv::ID CallingConv,
3089 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3093 switch (CallingConv) {
3096 case CallingConv::X86_StdCall:
3098 case CallingConv::X86_FastCall:
3100 case CallingConv::X86_ThisCall:
3102 case CallingConv::Fast:
3104 case CallingConv::GHC:
3109 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3110 /// specific condition code, returning the condition code and the LHS/RHS of the
3111 /// comparison to make.
3112 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3113 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3115 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3116 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3117 // X > -1 -> X == 0, jump !sign.
3118 RHS = DAG.getConstant(0, RHS.getValueType());
3119 return X86::COND_NS;
3121 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3122 // X < 0 -> X == 0, jump on sign.
3125 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3127 RHS = DAG.getConstant(0, RHS.getValueType());
3128 return X86::COND_LE;
3132 switch (SetCCOpcode) {
3133 default: llvm_unreachable("Invalid integer condition!");
3134 case ISD::SETEQ: return X86::COND_E;
3135 case ISD::SETGT: return X86::COND_G;
3136 case ISD::SETGE: return X86::COND_GE;
3137 case ISD::SETLT: return X86::COND_L;
3138 case ISD::SETLE: return X86::COND_LE;
3139 case ISD::SETNE: return X86::COND_NE;
3140 case ISD::SETULT: return X86::COND_B;
3141 case ISD::SETUGT: return X86::COND_A;
3142 case ISD::SETULE: return X86::COND_BE;
3143 case ISD::SETUGE: return X86::COND_AE;
3147 // First determine if it is required or is profitable to flip the operands.
3149 // If LHS is a foldable load, but RHS is not, flip the condition.
3150 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3151 !ISD::isNON_EXTLoad(RHS.getNode())) {
3152 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3153 std::swap(LHS, RHS);
3156 switch (SetCCOpcode) {
3162 std::swap(LHS, RHS);
3166 // On a floating point condition, the flags are set as follows:
3168 // 0 | 0 | 0 | X > Y
3169 // 0 | 0 | 1 | X < Y
3170 // 1 | 0 | 0 | X == Y
3171 // 1 | 1 | 1 | unordered
3172 switch (SetCCOpcode) {
3173 default: llvm_unreachable("Condcode should be pre-legalized away");
3175 case ISD::SETEQ: return X86::COND_E;
3176 case ISD::SETOLT: // flipped
3178 case ISD::SETGT: return X86::COND_A;
3179 case ISD::SETOLE: // flipped
3181 case ISD::SETGE: return X86::COND_AE;
3182 case ISD::SETUGT: // flipped
3184 case ISD::SETLT: return X86::COND_B;
3185 case ISD::SETUGE: // flipped
3187 case ISD::SETLE: return X86::COND_BE;
3189 case ISD::SETNE: return X86::COND_NE;
3190 case ISD::SETUO: return X86::COND_P;
3191 case ISD::SETO: return X86::COND_NP;
3193 case ISD::SETUNE: return X86::COND_INVALID;
3197 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3198 /// code. Current x86 isa includes the following FP cmov instructions:
3199 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3200 static bool hasFPCMov(unsigned X86CC) {
3216 /// isFPImmLegal - Returns true if the target can instruction select the
3217 /// specified FP immediate natively. If false, the legalizer will
3218 /// materialize the FP immediate as a load from a constant pool.
3219 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3220 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3221 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3227 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3228 /// the specified range (L, H].
3229 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3230 return (Val < 0) || (Val >= Low && Val < Hi);
3233 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3234 /// specified value.
3235 static bool isUndefOrEqual(int Val, int CmpVal) {
3236 if (Val < 0 || Val == CmpVal)
3241 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3242 /// from position Pos and ending in Pos+Size, falls within the specified
3243 /// sequential range (L, L+Pos]. or is undef.
3244 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3245 unsigned Pos, unsigned Size, int Low) {
3246 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3247 if (!isUndefOrEqual(Mask[i], Low))
3252 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3253 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3254 /// the second operand.
3255 static bool isPSHUFDMask(ArrayRef<int> Mask, EVT VT) {
3256 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3257 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3258 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3259 return (Mask[0] < 2 && Mask[1] < 2);
3263 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3264 /// is suitable for input to PSHUFHW.
3265 static bool isPSHUFHWMask(ArrayRef<int> Mask, EVT VT, bool HasAVX2) {
3266 if (VT != MVT::v8i16 && (!HasAVX2 || VT != MVT::v16i16))
3269 // Lower quadword copied in order or undef.
3270 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3273 // Upper quadword shuffled.
3274 for (unsigned i = 4; i != 8; ++i)
3275 if (!isUndefOrInRange(Mask[i], 4, 8))
3278 if (VT == MVT::v16i16) {
3279 // Lower quadword copied in order or undef.
3280 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3283 // Upper quadword shuffled.
3284 for (unsigned i = 12; i != 16; ++i)
3285 if (!isUndefOrInRange(Mask[i], 12, 16))
3292 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3293 /// is suitable for input to PSHUFLW.
3294 static bool isPSHUFLWMask(ArrayRef<int> Mask, EVT VT, bool HasAVX2) {
3295 if (VT != MVT::v8i16 && (!HasAVX2 || VT != MVT::v16i16))
3298 // Upper quadword copied in order.
3299 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3302 // Lower quadword shuffled.
3303 for (unsigned i = 0; i != 4; ++i)
3304 if (!isUndefOrInRange(Mask[i], 0, 4))
3307 if (VT == MVT::v16i16) {
3308 // Upper quadword copied in order.
3309 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3312 // Lower quadword shuffled.
3313 for (unsigned i = 8; i != 12; ++i)
3314 if (!isUndefOrInRange(Mask[i], 8, 12))
3321 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3322 /// is suitable for input to PALIGNR.
3323 static bool isPALIGNRMask(ArrayRef<int> Mask, EVT VT,
3324 const X86Subtarget *Subtarget) {
3325 if ((VT.getSizeInBits() == 128 && !Subtarget->hasSSSE3()) ||
3326 (VT.getSizeInBits() == 256 && !Subtarget->hasAVX2()))
3329 unsigned NumElts = VT.getVectorNumElements();
3330 unsigned NumLanes = VT.getSizeInBits()/128;
3331 unsigned NumLaneElts = NumElts/NumLanes;
3333 // Do not handle 64-bit element shuffles with palignr.
3334 if (NumLaneElts == 2)
3337 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3339 for (i = 0; i != NumLaneElts; ++i) {
3344 // Lane is all undef, go to next lane
3345 if (i == NumLaneElts)
3348 int Start = Mask[i+l];
3350 // Make sure its in this lane in one of the sources
3351 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3352 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3355 // If not lane 0, then we must match lane 0
3356 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3359 // Correct second source to be contiguous with first source
3360 if (Start >= (int)NumElts)
3361 Start -= NumElts - NumLaneElts;
3363 // Make sure we're shifting in the right direction.
3364 if (Start <= (int)(i+l))
3369 // Check the rest of the elements to see if they are consecutive.
3370 for (++i; i != NumLaneElts; ++i) {
3371 int Idx = Mask[i+l];
3373 // Make sure its in this lane
3374 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3375 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3378 // If not lane 0, then we must match lane 0
3379 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3382 if (Idx >= (int)NumElts)
3383 Idx -= NumElts - NumLaneElts;
3385 if (!isUndefOrEqual(Idx, Start+i))
3394 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3395 /// the two vector operands have swapped position.
3396 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3397 unsigned NumElems) {
3398 for (unsigned i = 0; i != NumElems; ++i) {
3402 else if (idx < (int)NumElems)
3403 Mask[i] = idx + NumElems;
3405 Mask[i] = idx - NumElems;
3409 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3410 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
3411 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
3412 /// reverse of what x86 shuffles want.
3413 static bool isSHUFPMask(ArrayRef<int> Mask, EVT VT, bool HasAVX,
3414 bool Commuted = false) {
3415 if (!HasAVX && VT.getSizeInBits() == 256)
3418 unsigned NumElems = VT.getVectorNumElements();
3419 unsigned NumLanes = VT.getSizeInBits()/128;
3420 unsigned NumLaneElems = NumElems/NumLanes;
3422 if (NumLaneElems != 2 && NumLaneElems != 4)
3425 // VSHUFPSY divides the resulting vector into 4 chunks.
3426 // The sources are also splitted into 4 chunks, and each destination
3427 // chunk must come from a different source chunk.
3429 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3430 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3432 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3433 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3435 // VSHUFPDY divides the resulting vector into 4 chunks.
3436 // The sources are also splitted into 4 chunks, and each destination
3437 // chunk must come from a different source chunk.
3439 // SRC1 => X3 X2 X1 X0
3440 // SRC2 => Y3 Y2 Y1 Y0
3442 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
3444 unsigned HalfLaneElems = NumLaneElems/2;
3445 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
3446 for (unsigned i = 0; i != NumLaneElems; ++i) {
3447 int Idx = Mask[i+l];
3448 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
3449 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
3451 // For VSHUFPSY, the mask of the second half must be the same as the
3452 // first but with the appropriate offsets. This works in the same way as
3453 // VPERMILPS works with masks.
3454 if (NumElems != 8 || l == 0 || Mask[i] < 0)
3456 if (!isUndefOrEqual(Idx, Mask[i]+l))
3464 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3465 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3466 static bool isMOVHLPSMask(ArrayRef<int> Mask, EVT VT) {
3467 if (!VT.is128BitVector())
3470 unsigned NumElems = VT.getVectorNumElements();
3475 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3476 return isUndefOrEqual(Mask[0], 6) &&
3477 isUndefOrEqual(Mask[1], 7) &&
3478 isUndefOrEqual(Mask[2], 2) &&
3479 isUndefOrEqual(Mask[3], 3);
3482 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3483 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3485 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, EVT VT) {
3486 if (!VT.is128BitVector())
3489 unsigned NumElems = VT.getVectorNumElements();
3494 return isUndefOrEqual(Mask[0], 2) &&
3495 isUndefOrEqual(Mask[1], 3) &&
3496 isUndefOrEqual(Mask[2], 2) &&
3497 isUndefOrEqual(Mask[3], 3);
3500 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3501 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3502 static bool isMOVLPMask(ArrayRef<int> Mask, EVT VT) {
3503 if (!VT.is128BitVector())
3506 unsigned NumElems = VT.getVectorNumElements();
3508 if (NumElems != 2 && NumElems != 4)
3511 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3512 if (!isUndefOrEqual(Mask[i], i + NumElems))
3515 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
3516 if (!isUndefOrEqual(Mask[i], i))
3522 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3523 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3524 static bool isMOVLHPSMask(ArrayRef<int> Mask, EVT VT) {
3525 if (!VT.is128BitVector())
3528 unsigned NumElems = VT.getVectorNumElements();
3530 if (NumElems != 2 && NumElems != 4)
3533 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3534 if (!isUndefOrEqual(Mask[i], i))
3537 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3538 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
3545 // Some special combinations that can be optimized.
3548 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
3549 SelectionDAG &DAG) {
3550 EVT VT = SVOp->getValueType(0);
3551 DebugLoc dl = SVOp->getDebugLoc();
3553 if (VT != MVT::v8i32 && VT != MVT::v8f32)
3556 ArrayRef<int> Mask = SVOp->getMask();
3558 // These are the special masks that may be optimized.
3559 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
3560 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
3561 bool MatchEvenMask = true;
3562 bool MatchOddMask = true;
3563 for (int i=0; i<8; ++i) {
3564 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
3565 MatchEvenMask = false;
3566 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
3567 MatchOddMask = false;
3570 if (!MatchEvenMask && !MatchOddMask)
3573 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
3575 SDValue Op0 = SVOp->getOperand(0);
3576 SDValue Op1 = SVOp->getOperand(1);
3578 if (MatchEvenMask) {
3579 // Shift the second operand right to 32 bits.
3580 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
3581 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
3583 // Shift the first operand left to 32 bits.
3584 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
3585 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
3587 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
3588 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
3591 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3592 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3593 static bool isUNPCKLMask(ArrayRef<int> Mask, EVT VT,
3594 bool HasAVX2, bool V2IsSplat = false) {
3595 unsigned NumElts = VT.getVectorNumElements();
3597 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3598 "Unsupported vector type for unpckh");
3600 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
3601 (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
3604 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3605 // independently on 128-bit lanes.
3606 unsigned NumLanes = VT.getSizeInBits()/128;
3607 unsigned NumLaneElts = NumElts/NumLanes;
3609 for (unsigned l = 0; l != NumLanes; ++l) {
3610 for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
3611 i != (l+1)*NumLaneElts;
3614 int BitI1 = Mask[i+1];
3615 if (!isUndefOrEqual(BitI, j))
3618 if (!isUndefOrEqual(BitI1, NumElts))
3621 if (!isUndefOrEqual(BitI1, j + NumElts))
3630 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3631 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3632 static bool isUNPCKHMask(ArrayRef<int> Mask, EVT VT,
3633 bool HasAVX2, bool V2IsSplat = false) {
3634 unsigned NumElts = VT.getVectorNumElements();
3636 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3637 "Unsupported vector type for unpckh");
3639 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
3640 (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
3643 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3644 // independently on 128-bit lanes.
3645 unsigned NumLanes = VT.getSizeInBits()/128;
3646 unsigned NumLaneElts = NumElts/NumLanes;
3648 for (unsigned l = 0; l != NumLanes; ++l) {
3649 for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
3650 i != (l+1)*NumLaneElts; i += 2, ++j) {
3652 int BitI1 = Mask[i+1];
3653 if (!isUndefOrEqual(BitI, j))
3656 if (isUndefOrEqual(BitI1, NumElts))
3659 if (!isUndefOrEqual(BitI1, j+NumElts))
3667 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3668 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3670 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, EVT VT,
3672 unsigned NumElts = VT.getVectorNumElements();
3674 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3675 "Unsupported vector type for unpckh");
3677 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
3678 (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
3681 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3682 // FIXME: Need a better way to get rid of this, there's no latency difference
3683 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3684 // the former later. We should also remove the "_undef" special mask.
3685 if (NumElts == 4 && VT.getSizeInBits() == 256)
3688 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3689 // independently on 128-bit lanes.
3690 unsigned NumLanes = VT.getSizeInBits()/128;
3691 unsigned NumLaneElts = NumElts/NumLanes;
3693 for (unsigned l = 0; l != NumLanes; ++l) {
3694 for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
3695 i != (l+1)*NumLaneElts;
3698 int BitI1 = Mask[i+1];
3700 if (!isUndefOrEqual(BitI, j))
3702 if (!isUndefOrEqual(BitI1, j))
3710 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3711 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3713 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, EVT VT, bool HasAVX2) {
3714 unsigned NumElts = VT.getVectorNumElements();
3716 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3717 "Unsupported vector type for unpckh");
3719 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
3720 (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
3723 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3724 // independently on 128-bit lanes.
3725 unsigned NumLanes = VT.getSizeInBits()/128;
3726 unsigned NumLaneElts = NumElts/NumLanes;
3728 for (unsigned l = 0; l != NumLanes; ++l) {
3729 for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
3730 i != (l+1)*NumLaneElts; i += 2, ++j) {
3732 int BitI1 = Mask[i+1];
3733 if (!isUndefOrEqual(BitI, j))
3735 if (!isUndefOrEqual(BitI1, j))
3742 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3743 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3744 /// MOVSD, and MOVD, i.e. setting the lowest element.
3745 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
3746 if (VT.getVectorElementType().getSizeInBits() < 32)
3748 if (!VT.is128BitVector())
3751 unsigned NumElts = VT.getVectorNumElements();
3753 if (!isUndefOrEqual(Mask[0], NumElts))
3756 for (unsigned i = 1; i != NumElts; ++i)
3757 if (!isUndefOrEqual(Mask[i], i))
3763 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
3764 /// as permutations between 128-bit chunks or halves. As an example: this
3766 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
3767 /// The first half comes from the second half of V1 and the second half from the
3768 /// the second half of V2.
3769 static bool isVPERM2X128Mask(ArrayRef<int> Mask, EVT VT, bool HasAVX) {
3770 if (!HasAVX || !VT.is256BitVector())
3773 // The shuffle result is divided into half A and half B. In total the two
3774 // sources have 4 halves, namely: C, D, E, F. The final values of A and
3775 // B must come from C, D, E or F.
3776 unsigned HalfSize = VT.getVectorNumElements()/2;
3777 bool MatchA = false, MatchB = false;
3779 // Check if A comes from one of C, D, E, F.
3780 for (unsigned Half = 0; Half != 4; ++Half) {
3781 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
3787 // Check if B comes from one of C, D, E, F.
3788 for (unsigned Half = 0; Half != 4; ++Half) {
3789 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
3795 return MatchA && MatchB;
3798 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
3799 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
3800 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
3801 EVT VT = SVOp->getValueType(0);
3803 unsigned HalfSize = VT.getVectorNumElements()/2;
3805 unsigned FstHalf = 0, SndHalf = 0;
3806 for (unsigned i = 0; i < HalfSize; ++i) {
3807 if (SVOp->getMaskElt(i) > 0) {
3808 FstHalf = SVOp->getMaskElt(i)/HalfSize;
3812 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
3813 if (SVOp->getMaskElt(i) > 0) {
3814 SndHalf = SVOp->getMaskElt(i)/HalfSize;
3819 return (FstHalf | (SndHalf << 4));
3822 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
3823 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3824 /// Note that VPERMIL mask matching is different depending whether theunderlying
3825 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3826 /// to the same elements of the low, but to the higher half of the source.
3827 /// In VPERMILPD the two lanes could be shuffled independently of each other
3828 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
3829 static bool isVPERMILPMask(ArrayRef<int> Mask, EVT VT, bool HasAVX) {
3833 unsigned NumElts = VT.getVectorNumElements();
3834 // Only match 256-bit with 32/64-bit types
3835 if (VT.getSizeInBits() != 256 || (NumElts != 4 && NumElts != 8))
3838 unsigned NumLanes = VT.getSizeInBits()/128;
3839 unsigned LaneSize = NumElts/NumLanes;
3840 for (unsigned l = 0; l != NumElts; l += LaneSize) {
3841 for (unsigned i = 0; i != LaneSize; ++i) {
3842 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
3844 if (NumElts != 8 || l == 0)
3846 // VPERMILPS handling
3849 if (!isUndefOrEqual(Mask[i+l], Mask[i]+l))
3857 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
3858 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3859 /// element of vector 2 and the other elements to come from vector 1 in order.
3860 static bool isCommutedMOVLMask(ArrayRef<int> Mask, EVT VT,
3861 bool V2IsSplat = false, bool V2IsUndef = false) {
3862 if (!VT.is128BitVector())
3865 unsigned NumOps = VT.getVectorNumElements();
3866 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3869 if (!isUndefOrEqual(Mask[0], 0))
3872 for (unsigned i = 1; i != NumOps; ++i)
3873 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3874 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3875 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3881 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3882 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3883 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3884 static bool isMOVSHDUPMask(ArrayRef<int> Mask, EVT VT,
3885 const X86Subtarget *Subtarget) {
3886 if (!Subtarget->hasSSE3())
3889 unsigned NumElems = VT.getVectorNumElements();
3891 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3892 (VT.getSizeInBits() == 256 && NumElems != 8))
3895 // "i+1" is the value the indexed mask element must have
3896 for (unsigned i = 0; i != NumElems; i += 2)
3897 if (!isUndefOrEqual(Mask[i], i+1) ||
3898 !isUndefOrEqual(Mask[i+1], i+1))
3904 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3905 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3906 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3907 static bool isMOVSLDUPMask(ArrayRef<int> Mask, EVT VT,
3908 const X86Subtarget *Subtarget) {
3909 if (!Subtarget->hasSSE3())
3912 unsigned NumElems = VT.getVectorNumElements();
3914 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3915 (VT.getSizeInBits() == 256 && NumElems != 8))
3918 // "i" is the value the indexed mask element must have
3919 for (unsigned i = 0; i != NumElems; i += 2)
3920 if (!isUndefOrEqual(Mask[i], i) ||
3921 !isUndefOrEqual(Mask[i+1], i))
3927 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
3928 /// specifies a shuffle of elements that is suitable for input to 256-bit
3929 /// version of MOVDDUP.
3930 static bool isMOVDDUPYMask(ArrayRef<int> Mask, EVT VT, bool HasAVX) {
3931 if (!HasAVX || !VT.is256BitVector())
3934 unsigned NumElts = VT.getVectorNumElements();
3938 for (unsigned i = 0; i != NumElts/2; ++i)
3939 if (!isUndefOrEqual(Mask[i], 0))
3941 for (unsigned i = NumElts/2; i != NumElts; ++i)
3942 if (!isUndefOrEqual(Mask[i], NumElts/2))
3947 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3948 /// specifies a shuffle of elements that is suitable for input to 128-bit
3949 /// version of MOVDDUP.
3950 static bool isMOVDDUPMask(ArrayRef<int> Mask, EVT VT) {
3951 if (!VT.is128BitVector())
3954 unsigned e = VT.getVectorNumElements() / 2;
3955 for (unsigned i = 0; i != e; ++i)
3956 if (!isUndefOrEqual(Mask[i], i))
3958 for (unsigned i = 0; i != e; ++i)
3959 if (!isUndefOrEqual(Mask[e+i], i))
3964 /// isVEXTRACTF128Index - Return true if the specified
3965 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3966 /// suitable for input to VEXTRACTF128.
3967 bool X86::isVEXTRACTF128Index(SDNode *N) {
3968 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3971 // The index should be aligned on a 128-bit boundary.
3973 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3975 unsigned VL = N->getValueType(0).getVectorNumElements();
3976 unsigned VBits = N->getValueType(0).getSizeInBits();
3977 unsigned ElSize = VBits / VL;
3978 bool Result = (Index * ElSize) % 128 == 0;
3983 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3984 /// operand specifies a subvector insert that is suitable for input to
3986 bool X86::isVINSERTF128Index(SDNode *N) {
3987 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3990 // The index should be aligned on a 128-bit boundary.
3992 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3994 unsigned VL = N->getValueType(0).getVectorNumElements();
3995 unsigned VBits = N->getValueType(0).getSizeInBits();
3996 unsigned ElSize = VBits / VL;
3997 bool Result = (Index * ElSize) % 128 == 0;
4002 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4003 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4004 /// Handles 128-bit and 256-bit.
4005 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4006 EVT VT = N->getValueType(0);
4008 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4009 "Unsupported vector type for PSHUF/SHUFP");
4011 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4012 // independently on 128-bit lanes.
4013 unsigned NumElts = VT.getVectorNumElements();
4014 unsigned NumLanes = VT.getSizeInBits()/128;
4015 unsigned NumLaneElts = NumElts/NumLanes;
4017 assert((NumLaneElts == 2 || NumLaneElts == 4) &&
4018 "Only supports 2 or 4 elements per lane");
4020 unsigned Shift = (NumLaneElts == 4) ? 1 : 0;
4022 for (unsigned i = 0; i != NumElts; ++i) {
4023 int Elt = N->getMaskElt(i);
4024 if (Elt < 0) continue;
4025 Elt &= NumLaneElts - 1;
4026 unsigned ShAmt = (i << Shift) % 8;
4027 Mask |= Elt << ShAmt;
4033 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4034 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4035 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4036 EVT VT = N->getValueType(0);
4038 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4039 "Unsupported vector type for PSHUFHW");
4041 unsigned NumElts = VT.getVectorNumElements();
4044 for (unsigned l = 0; l != NumElts; l += 8) {
4045 // 8 nodes per lane, but we only care about the last 4.
4046 for (unsigned i = 0; i < 4; ++i) {
4047 int Elt = N->getMaskElt(l+i+4);
4048 if (Elt < 0) continue;
4049 Elt &= 0x3; // only 2-bits.
4050 Mask |= Elt << (i * 2);
4057 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4058 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4059 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4060 EVT VT = N->getValueType(0);
4062 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4063 "Unsupported vector type for PSHUFHW");
4065 unsigned NumElts = VT.getVectorNumElements();
4068 for (unsigned l = 0; l != NumElts; l += 8) {
4069 // 8 nodes per lane, but we only care about the first 4.
4070 for (unsigned i = 0; i < 4; ++i) {
4071 int Elt = N->getMaskElt(l+i);
4072 if (Elt < 0) continue;
4073 Elt &= 0x3; // only 2-bits
4074 Mask |= Elt << (i * 2);
4081 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4082 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4083 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4084 EVT VT = SVOp->getValueType(0);
4085 unsigned EltSize = VT.getVectorElementType().getSizeInBits() >> 3;
4087 unsigned NumElts = VT.getVectorNumElements();
4088 unsigned NumLanes = VT.getSizeInBits()/128;
4089 unsigned NumLaneElts = NumElts/NumLanes;
4093 for (i = 0; i != NumElts; ++i) {
4094 Val = SVOp->getMaskElt(i);
4098 if (Val >= (int)NumElts)
4099 Val -= NumElts - NumLaneElts;
4101 assert(Val - i > 0 && "PALIGNR imm should be positive");
4102 return (Val - i) * EltSize;
4105 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
4106 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4108 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
4109 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4110 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
4113 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4115 EVT VecVT = N->getOperand(0).getValueType();
4116 EVT ElVT = VecVT.getVectorElementType();
4118 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4119 return Index / NumElemsPerChunk;
4122 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
4123 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4125 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
4126 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4127 llvm_unreachable("Illegal insert subvector for VINSERTF128");
4130 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4132 EVT VecVT = N->getValueType(0);
4133 EVT ElVT = VecVT.getVectorElementType();
4135 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4136 return Index / NumElemsPerChunk;
4139 /// getShuffleCLImmediate - Return the appropriate immediate to shuffle
4140 /// the specified VECTOR_SHUFFLE mask with VPERMQ and VPERMPD instructions.
4141 /// Handles 256-bit.
4142 static unsigned getShuffleCLImmediate(ShuffleVectorSDNode *N) {
4143 EVT VT = N->getValueType(0);
4145 unsigned NumElts = VT.getVectorNumElements();
4147 assert((VT.is256BitVector() && NumElts == 4) &&
4148 "Unsupported vector type for VPERMQ/VPERMPD");
4151 for (unsigned i = 0; i != NumElts; ++i) {
4152 int Elt = N->getMaskElt(i);
4155 Mask |= Elt << (i*2);
4160 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4162 bool X86::isZeroNode(SDValue Elt) {
4163 return ((isa<ConstantSDNode>(Elt) &&
4164 cast<ConstantSDNode>(Elt)->isNullValue()) ||
4165 (isa<ConstantFPSDNode>(Elt) &&
4166 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
4169 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4170 /// their permute mask.
4171 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4172 SelectionDAG &DAG) {
4173 EVT VT = SVOp->getValueType(0);
4174 unsigned NumElems = VT.getVectorNumElements();
4175 SmallVector<int, 8> MaskVec;
4177 for (unsigned i = 0; i != NumElems; ++i) {
4178 int Idx = SVOp->getMaskElt(i);
4180 if (Idx < (int)NumElems)
4185 MaskVec.push_back(Idx);
4187 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
4188 SVOp->getOperand(0), &MaskVec[0]);
4191 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4192 /// match movhlps. The lower half elements should come from upper half of
4193 /// V1 (and in order), and the upper half elements should come from the upper
4194 /// half of V2 (and in order).
4195 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, EVT VT) {
4196 if (!VT.is128BitVector())
4198 if (VT.getVectorNumElements() != 4)
4200 for (unsigned i = 0, e = 2; i != e; ++i)
4201 if (!isUndefOrEqual(Mask[i], i+2))
4203 for (unsigned i = 2; i != 4; ++i)
4204 if (!isUndefOrEqual(Mask[i], i+4))
4209 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4210 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4212 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4213 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4215 N = N->getOperand(0).getNode();
4216 if (!ISD::isNON_EXTLoad(N))
4219 *LD = cast<LoadSDNode>(N);
4223 // Test whether the given value is a vector value which will be legalized
4225 static bool WillBeConstantPoolLoad(SDNode *N) {
4226 if (N->getOpcode() != ISD::BUILD_VECTOR)
4229 // Check for any non-constant elements.
4230 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4231 switch (N->getOperand(i).getNode()->getOpcode()) {
4233 case ISD::ConstantFP:
4240 // Vectors of all-zeros and all-ones are materialized with special
4241 // instructions rather than being loaded.
4242 return !ISD::isBuildVectorAllZeros(N) &&
4243 !ISD::isBuildVectorAllOnes(N);
4246 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4247 /// match movlp{s|d}. The lower half elements should come from lower half of
4248 /// V1 (and in order), and the upper half elements should come from the upper
4249 /// half of V2 (and in order). And since V1 will become the source of the
4250 /// MOVLP, it must be either a vector load or a scalar load to vector.
4251 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4252 ArrayRef<int> Mask, EVT VT) {
4253 if (!VT.is128BitVector())
4256 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4258 // Is V2 is a vector load, don't do this transformation. We will try to use
4259 // load folding shufps op.
4260 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
4263 unsigned NumElems = VT.getVectorNumElements();
4265 if (NumElems != 2 && NumElems != 4)
4267 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4268 if (!isUndefOrEqual(Mask[i], i))
4270 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4271 if (!isUndefOrEqual(Mask[i], i+NumElems))
4276 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4278 static bool isSplatVector(SDNode *N) {
4279 if (N->getOpcode() != ISD::BUILD_VECTOR)
4282 SDValue SplatValue = N->getOperand(0);
4283 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4284 if (N->getOperand(i) != SplatValue)
4289 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4290 /// to an zero vector.
4291 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4292 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4293 SDValue V1 = N->getOperand(0);
4294 SDValue V2 = N->getOperand(1);
4295 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4296 for (unsigned i = 0; i != NumElems; ++i) {
4297 int Idx = N->getMaskElt(i);
4298 if (Idx >= (int)NumElems) {
4299 unsigned Opc = V2.getOpcode();
4300 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4302 if (Opc != ISD::BUILD_VECTOR ||
4303 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4305 } else if (Idx >= 0) {
4306 unsigned Opc = V1.getOpcode();
4307 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4309 if (Opc != ISD::BUILD_VECTOR ||
4310 !X86::isZeroNode(V1.getOperand(Idx)))
4317 /// getZeroVector - Returns a vector of specified type with all zero elements.
4319 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4320 SelectionDAG &DAG, DebugLoc dl) {
4321 assert(VT.isVector() && "Expected a vector type");
4322 unsigned Size = VT.getSizeInBits();
4324 // Always build SSE zero vectors as <4 x i32> bitcasted
4325 // to their dest type. This ensures they get CSE'd.
4327 if (Size == 128) { // SSE
4328 if (Subtarget->hasSSE2()) { // SSE2
4329 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4330 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4332 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4333 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4335 } else if (Size == 256) { // AVX
4336 if (Subtarget->hasAVX2()) { // AVX2
4337 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4338 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4339 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
4341 // 256-bit logic and arithmetic instructions in AVX are all
4342 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4343 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4344 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4345 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
4348 llvm_unreachable("Unexpected vector type");
4350 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4353 /// getOnesVector - Returns a vector of specified type with all bits set.
4354 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4355 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4356 /// Then bitcast to their original type, ensuring they get CSE'd.
4357 static SDValue getOnesVector(EVT VT, bool HasAVX2, SelectionDAG &DAG,
4359 assert(VT.isVector() && "Expected a vector type");
4360 unsigned Size = VT.getSizeInBits();
4362 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4365 if (HasAVX2) { // AVX2
4366 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4367 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
4369 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4370 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4372 } else if (Size == 128) {
4373 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4375 llvm_unreachable("Unexpected vector type");
4377 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4380 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4381 /// that point to V2 points to its first element.
4382 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
4383 for (unsigned i = 0; i != NumElems; ++i) {
4384 if (Mask[i] > (int)NumElems) {
4390 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4391 /// operation of specified width.
4392 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4394 unsigned NumElems = VT.getVectorNumElements();
4395 SmallVector<int, 8> Mask;
4396 Mask.push_back(NumElems);
4397 for (unsigned i = 1; i != NumElems; ++i)
4399 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4402 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4403 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4405 unsigned NumElems = VT.getVectorNumElements();
4406 SmallVector<int, 8> Mask;
4407 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4409 Mask.push_back(i + NumElems);
4411 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4414 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4415 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4417 unsigned NumElems = VT.getVectorNumElements();
4418 SmallVector<int, 8> Mask;
4419 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4420 Mask.push_back(i + Half);
4421 Mask.push_back(i + NumElems + Half);
4423 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4426 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4427 // a generic shuffle instruction because the target has no such instructions.
4428 // Generate shuffles which repeat i16 and i8 several times until they can be
4429 // represented by v4f32 and then be manipulated by target suported shuffles.
4430 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4431 EVT VT = V.getValueType();
4432 int NumElems = VT.getVectorNumElements();
4433 DebugLoc dl = V.getDebugLoc();
4435 while (NumElems > 4) {
4436 if (EltNo < NumElems/2) {
4437 V = getUnpackl(DAG, dl, VT, V, V);
4439 V = getUnpackh(DAG, dl, VT, V, V);
4440 EltNo -= NumElems/2;
4447 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4448 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4449 EVT VT = V.getValueType();
4450 DebugLoc dl = V.getDebugLoc();
4451 unsigned Size = VT.getSizeInBits();
4454 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4455 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4456 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4458 } else if (Size == 256) {
4459 // To use VPERMILPS to splat scalars, the second half of indicies must
4460 // refer to the higher part, which is a duplication of the lower one,
4461 // because VPERMILPS can only handle in-lane permutations.
4462 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4463 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4465 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4466 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4469 llvm_unreachable("Vector size not supported");
4471 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4474 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4475 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4476 EVT SrcVT = SV->getValueType(0);
4477 SDValue V1 = SV->getOperand(0);
4478 DebugLoc dl = SV->getDebugLoc();
4480 int EltNo = SV->getSplatIndex();
4481 int NumElems = SrcVT.getVectorNumElements();
4482 unsigned Size = SrcVT.getSizeInBits();
4484 assert(((Size == 128 && NumElems > 4) || Size == 256) &&
4485 "Unknown how to promote splat for type");
4487 // Extract the 128-bit part containing the splat element and update
4488 // the splat element index when it refers to the higher register.
4490 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
4491 if (EltNo >= NumElems/2)
4492 EltNo -= NumElems/2;
4495 // All i16 and i8 vector types can't be used directly by a generic shuffle
4496 // instruction because the target has no such instruction. Generate shuffles
4497 // which repeat i16 and i8 several times until they fit in i32, and then can
4498 // be manipulated by target suported shuffles.
4499 EVT EltVT = SrcVT.getVectorElementType();
4500 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4501 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4503 // Recreate the 256-bit vector and place the same 128-bit vector
4504 // into the low and high part. This is necessary because we want
4505 // to use VPERM* to shuffle the vectors
4507 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
4510 return getLegalSplat(DAG, V1, EltNo);
4513 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4514 /// vector of zero or undef vector. This produces a shuffle where the low
4515 /// element of V2 is swizzled into the zero/undef vector, landing at element
4516 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4517 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4519 const X86Subtarget *Subtarget,
4520 SelectionDAG &DAG) {
4521 EVT VT = V2.getValueType();
4523 ? getZeroVector(VT, Subtarget, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4524 unsigned NumElems = VT.getVectorNumElements();
4525 SmallVector<int, 16> MaskVec;
4526 for (unsigned i = 0; i != NumElems; ++i)
4527 // If this is the insertion idx, put the low elt of V2 here.
4528 MaskVec.push_back(i == Idx ? NumElems : i);
4529 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4532 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4533 /// target specific opcode. Returns true if the Mask could be calculated.
4534 /// Sets IsUnary to true if only uses one source.
4535 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4536 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4537 unsigned NumElems = VT.getVectorNumElements();
4541 switch(N->getOpcode()) {
4543 ImmN = N->getOperand(N->getNumOperands()-1);
4544 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4546 case X86ISD::UNPCKH:
4547 DecodeUNPCKHMask(VT, Mask);
4549 case X86ISD::UNPCKL:
4550 DecodeUNPCKLMask(VT, Mask);
4552 case X86ISD::MOVHLPS:
4553 DecodeMOVHLPSMask(NumElems, Mask);
4555 case X86ISD::MOVLHPS:
4556 DecodeMOVLHPSMask(NumElems, Mask);
4558 case X86ISD::PSHUFD:
4559 case X86ISD::VPERMILP:
4560 ImmN = N->getOperand(N->getNumOperands()-1);
4561 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4564 case X86ISD::PSHUFHW:
4565 ImmN = N->getOperand(N->getNumOperands()-1);
4566 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4569 case X86ISD::PSHUFLW:
4570 ImmN = N->getOperand(N->getNumOperands()-1);
4571 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4574 case X86ISD::VPERMI:
4575 ImmN = N->getOperand(N->getNumOperands()-1);
4576 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4580 case X86ISD::MOVSD: {
4581 // The index 0 always comes from the first element of the second source,
4582 // this is why MOVSS and MOVSD are used in the first place. The other
4583 // elements come from the other positions of the first source vector
4584 Mask.push_back(NumElems);
4585 for (unsigned i = 1; i != NumElems; ++i) {
4590 case X86ISD::VPERM2X128:
4591 ImmN = N->getOperand(N->getNumOperands()-1);
4592 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4593 if (Mask.empty()) return false;
4595 case X86ISD::MOVDDUP:
4596 case X86ISD::MOVLHPD:
4597 case X86ISD::MOVLPD:
4598 case X86ISD::MOVLPS:
4599 case X86ISD::MOVSHDUP:
4600 case X86ISD::MOVSLDUP:
4601 case X86ISD::PALIGN:
4602 // Not yet implemented
4604 default: llvm_unreachable("unknown target shuffle node");
4610 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4611 /// element of the result of the vector shuffle.
4612 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4615 return SDValue(); // Limit search depth.
4617 SDValue V = SDValue(N, 0);
4618 EVT VT = V.getValueType();
4619 unsigned Opcode = V.getOpcode();
4621 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4622 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4623 int Elt = SV->getMaskElt(Index);
4626 return DAG.getUNDEF(VT.getVectorElementType());
4628 unsigned NumElems = VT.getVectorNumElements();
4629 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4630 : SV->getOperand(1);
4631 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4634 // Recurse into target specific vector shuffles to find scalars.
4635 if (isTargetShuffle(Opcode)) {
4636 MVT ShufVT = V.getValueType().getSimpleVT();
4637 unsigned NumElems = ShufVT.getVectorNumElements();
4638 SmallVector<int, 16> ShuffleMask;
4641 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4644 int Elt = ShuffleMask[Index];
4646 return DAG.getUNDEF(ShufVT.getVectorElementType());
4648 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4650 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4654 // Actual nodes that may contain scalar elements
4655 if (Opcode == ISD::BITCAST) {
4656 V = V.getOperand(0);
4657 EVT SrcVT = V.getValueType();
4658 unsigned NumElems = VT.getVectorNumElements();
4660 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4664 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4665 return (Index == 0) ? V.getOperand(0)
4666 : DAG.getUNDEF(VT.getVectorElementType());
4668 if (V.getOpcode() == ISD::BUILD_VECTOR)
4669 return V.getOperand(Index);
4674 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4675 /// shuffle operation which come from a consecutively from a zero. The
4676 /// search can start in two different directions, from left or right.
4678 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, unsigned NumElems,
4679 bool ZerosFromLeft, SelectionDAG &DAG) {
4681 for (i = 0; i != NumElems; ++i) {
4682 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4683 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
4684 if (!(Elt.getNode() &&
4685 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4692 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
4693 /// correspond consecutively to elements from one of the vector operands,
4694 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4696 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
4697 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
4698 unsigned NumElems, unsigned &OpNum) {
4699 bool SeenV1 = false;
4700 bool SeenV2 = false;
4702 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
4703 int Idx = SVOp->getMaskElt(i);
4704 // Ignore undef indicies
4708 if (Idx < (int)NumElems)
4713 // Only accept consecutive elements from the same vector
4714 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4718 OpNum = SeenV1 ? 0 : 1;
4722 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4723 /// logical left shift of a vector.
4724 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4725 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4726 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4727 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4728 false /* check zeros from right */, DAG);
4734 // Considering the elements in the mask that are not consecutive zeros,
4735 // check if they consecutively come from only one of the source vectors.
4737 // V1 = {X, A, B, C} 0
4739 // vector_shuffle V1, V2 <1, 2, 3, X>
4741 if (!isShuffleMaskConsecutive(SVOp,
4742 0, // Mask Start Index
4743 NumElems-NumZeros, // Mask End Index(exclusive)
4744 NumZeros, // Where to start looking in the src vector
4745 NumElems, // Number of elements in vector
4746 OpSrc)) // Which source operand ?
4751 ShVal = SVOp->getOperand(OpSrc);
4755 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4756 /// logical left shift of a vector.
4757 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4758 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4759 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4760 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4761 true /* check zeros from left */, DAG);
4767 // Considering the elements in the mask that are not consecutive zeros,
4768 // check if they consecutively come from only one of the source vectors.
4770 // 0 { A, B, X, X } = V2
4772 // vector_shuffle V1, V2 <X, X, 4, 5>
4774 if (!isShuffleMaskConsecutive(SVOp,
4775 NumZeros, // Mask Start Index
4776 NumElems, // Mask End Index(exclusive)
4777 0, // Where to start looking in the src vector
4778 NumElems, // Number of elements in vector
4779 OpSrc)) // Which source operand ?
4784 ShVal = SVOp->getOperand(OpSrc);
4788 /// isVectorShift - Returns true if the shuffle can be implemented as a
4789 /// logical left or right shift of a vector.
4790 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4791 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4792 // Although the logic below support any bitwidth size, there are no
4793 // shift instructions which handle more than 128-bit vectors.
4794 if (!SVOp->getValueType(0).is128BitVector())
4797 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4798 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4804 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4806 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4807 unsigned NumNonZero, unsigned NumZero,
4809 const X86Subtarget* Subtarget,
4810 const TargetLowering &TLI) {
4814 DebugLoc dl = Op.getDebugLoc();
4817 for (unsigned i = 0; i < 16; ++i) {
4818 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4819 if (ThisIsNonZero && First) {
4821 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4823 V = DAG.getUNDEF(MVT::v8i16);
4828 SDValue ThisElt(0, 0), LastElt(0, 0);
4829 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4830 if (LastIsNonZero) {
4831 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4832 MVT::i16, Op.getOperand(i-1));
4834 if (ThisIsNonZero) {
4835 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4836 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4837 ThisElt, DAG.getConstant(8, MVT::i8));
4839 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4843 if (ThisElt.getNode())
4844 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4845 DAG.getIntPtrConstant(i/2));
4849 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4852 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4854 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4855 unsigned NumNonZero, unsigned NumZero,
4857 const X86Subtarget* Subtarget,
4858 const TargetLowering &TLI) {
4862 DebugLoc dl = Op.getDebugLoc();
4865 for (unsigned i = 0; i < 8; ++i) {
4866 bool isNonZero = (NonZeros & (1 << i)) != 0;
4870 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4872 V = DAG.getUNDEF(MVT::v8i16);
4875 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4876 MVT::v8i16, V, Op.getOperand(i),
4877 DAG.getIntPtrConstant(i));
4884 /// getVShift - Return a vector logical shift node.
4886 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4887 unsigned NumBits, SelectionDAG &DAG,
4888 const TargetLowering &TLI, DebugLoc dl) {
4889 assert(VT.is128BitVector() && "Unknown type for VShift");
4890 EVT ShVT = MVT::v2i64;
4891 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4892 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4893 return DAG.getNode(ISD::BITCAST, dl, VT,
4894 DAG.getNode(Opc, dl, ShVT, SrcOp,
4895 DAG.getConstant(NumBits,
4896 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4900 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4901 SelectionDAG &DAG) const {
4903 // Check if the scalar load can be widened into a vector load. And if
4904 // the address is "base + cst" see if the cst can be "absorbed" into
4905 // the shuffle mask.
4906 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4907 SDValue Ptr = LD->getBasePtr();
4908 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4910 EVT PVT = LD->getValueType(0);
4911 if (PVT != MVT::i32 && PVT != MVT::f32)
4916 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4917 FI = FINode->getIndex();
4919 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4920 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4921 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4922 Offset = Ptr.getConstantOperandVal(1);
4923 Ptr = Ptr.getOperand(0);
4928 // FIXME: 256-bit vector instructions don't require a strict alignment,
4929 // improve this code to support it better.
4930 unsigned RequiredAlign = VT.getSizeInBits()/8;
4931 SDValue Chain = LD->getChain();
4932 // Make sure the stack object alignment is at least 16 or 32.
4933 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4934 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4935 if (MFI->isFixedObjectIndex(FI)) {
4936 // Can't change the alignment. FIXME: It's possible to compute
4937 // the exact stack offset and reference FI + adjust offset instead.
4938 // If someone *really* cares about this. That's the way to implement it.
4941 MFI->setObjectAlignment(FI, RequiredAlign);
4945 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4946 // Ptr + (Offset & ~15).
4949 if ((Offset % RequiredAlign) & 3)
4951 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4953 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4954 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4956 int EltNo = (Offset - StartOffset) >> 2;
4957 unsigned NumElems = VT.getVectorNumElements();
4959 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4960 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4961 LD->getPointerInfo().getWithOffset(StartOffset),
4962 false, false, false, 0);
4964 SmallVector<int, 8> Mask;
4965 for (unsigned i = 0; i != NumElems; ++i)
4966 Mask.push_back(EltNo);
4968 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4974 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4975 /// vector of type 'VT', see if the elements can be replaced by a single large
4976 /// load which has the same value as a build_vector whose operands are 'elts'.
4978 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4980 /// FIXME: we'd also like to handle the case where the last elements are zero
4981 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4982 /// There's even a handy isZeroNode for that purpose.
4983 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4984 DebugLoc &DL, SelectionDAG &DAG) {
4985 EVT EltVT = VT.getVectorElementType();
4986 unsigned NumElems = Elts.size();
4988 LoadSDNode *LDBase = NULL;
4989 unsigned LastLoadedElt = -1U;
4991 // For each element in the initializer, see if we've found a load or an undef.
4992 // If we don't find an initial load element, or later load elements are
4993 // non-consecutive, bail out.
4994 for (unsigned i = 0; i < NumElems; ++i) {
4995 SDValue Elt = Elts[i];
4997 if (!Elt.getNode() ||
4998 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5001 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5003 LDBase = cast<LoadSDNode>(Elt.getNode());
5007 if (Elt.getOpcode() == ISD::UNDEF)
5010 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5011 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5016 // If we have found an entire vector of loads and undefs, then return a large
5017 // load of the entire vector width starting at the base pointer. If we found
5018 // consecutive loads for the low half, generate a vzext_load node.
5019 if (LastLoadedElt == NumElems - 1) {
5020 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5021 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5022 LDBase->getPointerInfo(),
5023 LDBase->isVolatile(), LDBase->isNonTemporal(),
5024 LDBase->isInvariant(), 0);
5025 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5026 LDBase->getPointerInfo(),
5027 LDBase->isVolatile(), LDBase->isNonTemporal(),
5028 LDBase->isInvariant(), LDBase->getAlignment());
5030 if (NumElems == 4 && LastLoadedElt == 1 &&
5031 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5032 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5033 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5035 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, 2, MVT::i64,
5036 LDBase->getPointerInfo(),
5037 LDBase->getAlignment(),
5038 false/*isVolatile*/, true/*ReadMem*/,
5041 // Make sure the newly-created LOAD is in the same position as LDBase in
5042 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5043 // update uses of LDBase's output chain to use the TokenFactor.
5044 if (LDBase->hasAnyUseOfValue(1)) {
5045 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5046 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5047 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5048 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5049 SDValue(ResNode.getNode(), 1));
5052 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5057 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5058 /// to generate a splat value for the following cases:
5059 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5060 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5061 /// a scalar load, or a constant.
5062 /// The VBROADCAST node is returned when a pattern is found,
5063 /// or SDValue() otherwise.
5065 X86TargetLowering::LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const {
5066 if (!Subtarget->hasAVX())
5069 EVT VT = Op.getValueType();
5070 DebugLoc dl = Op.getDebugLoc();
5072 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5073 "Unsupported vector type for broadcast.");
5078 switch (Op.getOpcode()) {
5080 // Unknown pattern found.
5083 case ISD::BUILD_VECTOR: {
5084 // The BUILD_VECTOR node must be a splat.
5085 if (!isSplatVector(Op.getNode()))
5088 Ld = Op.getOperand(0);
5089 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5090 Ld.getOpcode() == ISD::ConstantFP);
5092 // The suspected load node has several users. Make sure that all
5093 // of its users are from the BUILD_VECTOR node.
5094 // Constants may have multiple users.
5095 if (!ConstSplatVal && !Ld->hasNUsesOfValue(VT.getVectorNumElements(), 0))
5100 case ISD::VECTOR_SHUFFLE: {
5101 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5103 // Shuffles must have a splat mask where the first element is
5105 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5108 SDValue Sc = Op.getOperand(0);
5109 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5110 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5112 if (!Subtarget->hasAVX2())
5115 // Use the register form of the broadcast instruction available on AVX2.
5116 if (VT.is256BitVector())
5117 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5118 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5121 Ld = Sc.getOperand(0);
5122 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5123 Ld.getOpcode() == ISD::ConstantFP);
5125 // The scalar_to_vector node and the suspected
5126 // load node must have exactly one user.
5127 // Constants may have multiple users.
5128 if (!ConstSplatVal && (!Sc.hasOneUse() || !Ld.hasOneUse()))
5134 bool Is256 = VT.is256BitVector();
5136 // Handle the broadcasting a single constant scalar from the constant pool
5137 // into a vector. On Sandybridge it is still better to load a constant vector
5138 // from the constant pool and not to broadcast it from a scalar.
5139 if (ConstSplatVal && Subtarget->hasAVX2()) {
5140 EVT CVT = Ld.getValueType();
5141 assert(!CVT.isVector() && "Must not broadcast a vector type");
5142 unsigned ScalarSize = CVT.getSizeInBits();
5144 if (ScalarSize == 32 || (Is256 && ScalarSize == 64)) {
5145 const Constant *C = 0;
5146 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5147 C = CI->getConstantIntValue();
5148 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5149 C = CF->getConstantFPValue();
5151 assert(C && "Invalid constant type");
5153 SDValue CP = DAG.getConstantPool(C, getPointerTy());
5154 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5155 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5156 MachinePointerInfo::getConstantPool(),
5157 false, false, false, Alignment);
5159 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5163 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5164 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5166 // Handle AVX2 in-register broadcasts.
5167 if (!IsLoad && Subtarget->hasAVX2() &&
5168 (ScalarSize == 32 || (Is256 && ScalarSize == 64)))
5169 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5171 // The scalar source must be a normal load.
5175 if (ScalarSize == 32 || (Is256 && ScalarSize == 64))
5176 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5178 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5179 // double since there is no vbroadcastsd xmm
5180 if (Subtarget->hasAVX2() && Ld.getValueType().isInteger()) {
5181 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5182 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5185 // Unsupported broadcast.
5190 X86TargetLowering::buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const {
5191 EVT VT = Op.getValueType();
5193 // Skip if insert_vec_elt is not supported.
5194 if (!isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5197 DebugLoc DL = Op.getDebugLoc();
5198 unsigned NumElems = Op.getNumOperands();
5202 SmallVector<unsigned, 4> InsertIndices;
5203 SmallVector<int, 8> Mask(NumElems, -1);
5205 for (unsigned i = 0; i != NumElems; ++i) {
5206 unsigned Opc = Op.getOperand(i).getOpcode();
5208 if (Opc == ISD::UNDEF)
5211 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5212 // Quit if more than 1 elements need inserting.
5213 if (InsertIndices.size() > 1)
5216 InsertIndices.push_back(i);
5220 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5221 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5223 // Quit if extracted from vector of different type.
5224 if (ExtractedFromVec.getValueType() != VT)
5227 // Quit if non-constant index.
5228 if (!isa<ConstantSDNode>(ExtIdx))
5231 if (VecIn1.getNode() == 0)
5232 VecIn1 = ExtractedFromVec;
5233 else if (VecIn1 != ExtractedFromVec) {
5234 if (VecIn2.getNode() == 0)
5235 VecIn2 = ExtractedFromVec;
5236 else if (VecIn2 != ExtractedFromVec)
5237 // Quit if more than 2 vectors to shuffle
5241 unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5243 if (ExtractedFromVec == VecIn1)
5245 else if (ExtractedFromVec == VecIn2)
5246 Mask[i] = Idx + NumElems;
5249 if (VecIn1.getNode() == 0)
5252 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5253 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5254 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5255 unsigned Idx = InsertIndices[i];
5256 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5257 DAG.getIntPtrConstant(Idx));
5264 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5265 DebugLoc dl = Op.getDebugLoc();
5267 EVT VT = Op.getValueType();
5268 EVT ExtVT = VT.getVectorElementType();
5269 unsigned NumElems = Op.getNumOperands();
5271 // Vectors containing all zeros can be matched by pxor and xorps later
5272 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5273 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5274 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5275 if (VT == MVT::v4i32 || VT == MVT::v8i32)
5278 return getZeroVector(VT, Subtarget, DAG, dl);
5281 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5282 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5283 // vpcmpeqd on 256-bit vectors.
5284 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5285 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasAVX2()))
5288 return getOnesVector(VT, Subtarget->hasAVX2(), DAG, dl);
5291 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
5292 if (Broadcast.getNode())
5295 unsigned EVTBits = ExtVT.getSizeInBits();
5297 unsigned NumZero = 0;
5298 unsigned NumNonZero = 0;
5299 unsigned NonZeros = 0;
5300 bool IsAllConstants = true;
5301 SmallSet<SDValue, 8> Values;
5302 for (unsigned i = 0; i < NumElems; ++i) {
5303 SDValue Elt = Op.getOperand(i);
5304 if (Elt.getOpcode() == ISD::UNDEF)
5307 if (Elt.getOpcode() != ISD::Constant &&
5308 Elt.getOpcode() != ISD::ConstantFP)
5309 IsAllConstants = false;
5310 if (X86::isZeroNode(Elt))
5313 NonZeros |= (1 << i);
5318 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5319 if (NumNonZero == 0)
5320 return DAG.getUNDEF(VT);
5322 // Special case for single non-zero, non-undef, element.
5323 if (NumNonZero == 1) {
5324 unsigned Idx = CountTrailingZeros_32(NonZeros);
5325 SDValue Item = Op.getOperand(Idx);
5327 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5328 // the value are obviously zero, truncate the value to i32 and do the
5329 // insertion that way. Only do this if the value is non-constant or if the
5330 // value is a constant being inserted into element 0. It is cheaper to do
5331 // a constant pool load than it is to do a movd + shuffle.
5332 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5333 (!IsAllConstants || Idx == 0)) {
5334 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5336 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5337 EVT VecVT = MVT::v4i32;
5338 unsigned VecElts = 4;
5340 // Truncate the value (which may itself be a constant) to i32, and
5341 // convert it to a vector with movd (S2V+shuffle to zero extend).
5342 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5343 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5344 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5346 // Now we have our 32-bit value zero extended in the low element of
5347 // a vector. If Idx != 0, swizzle it into place.
5349 SmallVector<int, 4> Mask;
5350 Mask.push_back(Idx);
5351 for (unsigned i = 1; i != VecElts; ++i)
5353 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
5356 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5360 // If we have a constant or non-constant insertion into the low element of
5361 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5362 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5363 // depending on what the source datatype is.
5366 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5368 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5369 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5370 if (VT.is256BitVector()) {
5371 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5372 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5373 Item, DAG.getIntPtrConstant(0));
5375 assert(VT.is128BitVector() && "Expected an SSE value type!");
5376 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5377 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5378 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5381 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5382 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5383 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5384 if (VT.is256BitVector()) {
5385 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5386 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5388 assert(VT.is128BitVector() && "Expected an SSE value type!");
5389 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5391 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5395 // Is it a vector logical left shift?
5396 if (NumElems == 2 && Idx == 1 &&
5397 X86::isZeroNode(Op.getOperand(0)) &&
5398 !X86::isZeroNode(Op.getOperand(1))) {
5399 unsigned NumBits = VT.getSizeInBits();
5400 return getVShift(true, VT,
5401 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5402 VT, Op.getOperand(1)),
5403 NumBits/2, DAG, *this, dl);
5406 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5409 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5410 // is a non-constant being inserted into an element other than the low one,
5411 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5412 // movd/movss) to move this into the low element, then shuffle it into
5414 if (EVTBits == 32) {
5415 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5417 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5418 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
5419 SmallVector<int, 8> MaskVec;
5420 for (unsigned i = 0; i != NumElems; ++i)
5421 MaskVec.push_back(i == Idx ? 0 : 1);
5422 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5426 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5427 if (Values.size() == 1) {
5428 if (EVTBits == 32) {
5429 // Instead of a shuffle like this:
5430 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5431 // Check if it's possible to issue this instead.
5432 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5433 unsigned Idx = CountTrailingZeros_32(NonZeros);
5434 SDValue Item = Op.getOperand(Idx);
5435 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5436 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5441 // A vector full of immediates; various special cases are already
5442 // handled, so this is best done with a single constant-pool load.
5446 // For AVX-length vectors, build the individual 128-bit pieces and use
5447 // shuffles to put them in place.
5448 if (VT.is256BitVector()) {
5449 SmallVector<SDValue, 32> V;
5450 for (unsigned i = 0; i != NumElems; ++i)
5451 V.push_back(Op.getOperand(i));
5453 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5455 // Build both the lower and upper subvector.
5456 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5457 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5460 // Recreate the wider vector with the lower and upper part.
5461 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5464 // Let legalizer expand 2-wide build_vectors.
5465 if (EVTBits == 64) {
5466 if (NumNonZero == 1) {
5467 // One half is zero or undef.
5468 unsigned Idx = CountTrailingZeros_32(NonZeros);
5469 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5470 Op.getOperand(Idx));
5471 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5476 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5477 if (EVTBits == 8 && NumElems == 16) {
5478 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5480 if (V.getNode()) return V;
5483 if (EVTBits == 16 && NumElems == 8) {
5484 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5486 if (V.getNode()) return V;
5489 // If element VT is == 32 bits, turn it into a number of shuffles.
5490 SmallVector<SDValue, 8> V(NumElems);
5491 if (NumElems == 4 && NumZero > 0) {
5492 for (unsigned i = 0; i < 4; ++i) {
5493 bool isZero = !(NonZeros & (1 << i));
5495 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5497 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5500 for (unsigned i = 0; i < 2; ++i) {
5501 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5504 V[i] = V[i*2]; // Must be a zero vector.
5507 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5510 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5513 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5518 bool Reverse1 = (NonZeros & 0x3) == 2;
5519 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5523 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5524 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5526 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5529 if (Values.size() > 1 && VT.is128BitVector()) {
5530 // Check for a build vector of consecutive loads.
5531 for (unsigned i = 0; i < NumElems; ++i)
5532 V[i] = Op.getOperand(i);
5534 // Check for elements which are consecutive loads.
5535 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5539 // Check for a build vector from mostly shuffle plus few inserting.
5540 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5544 // For SSE 4.1, use insertps to put the high elements into the low element.
5545 if (getSubtarget()->hasSSE41()) {
5547 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5548 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5550 Result = DAG.getUNDEF(VT);
5552 for (unsigned i = 1; i < NumElems; ++i) {
5553 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5554 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5555 Op.getOperand(i), DAG.getIntPtrConstant(i));
5560 // Otherwise, expand into a number of unpckl*, start by extending each of
5561 // our (non-undef) elements to the full vector width with the element in the
5562 // bottom slot of the vector (which generates no code for SSE).
5563 for (unsigned i = 0; i < NumElems; ++i) {
5564 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5565 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5567 V[i] = DAG.getUNDEF(VT);
5570 // Next, we iteratively mix elements, e.g. for v4f32:
5571 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5572 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5573 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5574 unsigned EltStride = NumElems >> 1;
5575 while (EltStride != 0) {
5576 for (unsigned i = 0; i < EltStride; ++i) {
5577 // If V[i+EltStride] is undef and this is the first round of mixing,
5578 // then it is safe to just drop this shuffle: V[i] is already in the
5579 // right place, the one element (since it's the first round) being
5580 // inserted as undef can be dropped. This isn't safe for successive
5581 // rounds because they will permute elements within both vectors.
5582 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5583 EltStride == NumElems/2)
5586 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5595 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5596 // to create 256-bit vectors from two other 128-bit ones.
5597 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5598 DebugLoc dl = Op.getDebugLoc();
5599 EVT ResVT = Op.getValueType();
5601 assert(ResVT.is256BitVector() && "Value type must be 256-bit wide");
5603 SDValue V1 = Op.getOperand(0);
5604 SDValue V2 = Op.getOperand(1);
5605 unsigned NumElems = ResVT.getVectorNumElements();
5607 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5610 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5611 assert(Op.getNumOperands() == 2);
5613 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5614 // from two other 128-bit ones.
5615 return LowerAVXCONCAT_VECTORS(Op, DAG);
5618 // Try to lower a shuffle node into a simple blend instruction.
5620 LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
5621 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5622 SDValue V1 = SVOp->getOperand(0);
5623 SDValue V2 = SVOp->getOperand(1);
5624 DebugLoc dl = SVOp->getDebugLoc();
5625 MVT VT = SVOp->getValueType(0).getSimpleVT();
5626 unsigned NumElems = VT.getVectorNumElements();
5628 if (!Subtarget->hasSSE41())
5634 switch (VT.SimpleTy) {
5635 default: return SDValue();
5637 ISDNo = X86ISD::BLENDPW;
5642 ISDNo = X86ISD::BLENDPS;
5647 ISDNo = X86ISD::BLENDPD;
5652 if (!Subtarget->hasAVX())
5654 ISDNo = X86ISD::BLENDPS;
5659 if (!Subtarget->hasAVX())
5661 ISDNo = X86ISD::BLENDPD;
5665 assert(ISDNo && "Invalid Op Number");
5667 unsigned MaskVals = 0;
5669 for (unsigned i = 0; i != NumElems; ++i) {
5670 int EltIdx = SVOp->getMaskElt(i);
5671 if (EltIdx == (int)i || EltIdx < 0)
5673 else if (EltIdx == (int)(i + NumElems))
5674 continue; // Bit is set to zero;
5679 V1 = DAG.getNode(ISD::BITCAST, dl, OpTy, V1);
5680 V2 = DAG.getNode(ISD::BITCAST, dl, OpTy, V2);
5681 SDValue Ret = DAG.getNode(ISDNo, dl, OpTy, V1, V2,
5682 DAG.getConstant(MaskVals, MVT::i32));
5683 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
5686 // v8i16 shuffles - Prefer shuffles in the following order:
5687 // 1. [all] pshuflw, pshufhw, optional move
5688 // 2. [ssse3] 1 x pshufb
5689 // 3. [ssse3] 2 x pshufb + 1 x por
5690 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5692 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
5693 SelectionDAG &DAG) {
5694 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5695 SDValue V1 = SVOp->getOperand(0);
5696 SDValue V2 = SVOp->getOperand(1);
5697 DebugLoc dl = SVOp->getDebugLoc();
5698 SmallVector<int, 8> MaskVals;
5700 // Determine if more than 1 of the words in each of the low and high quadwords
5701 // of the result come from the same quadword of one of the two inputs. Undef
5702 // mask values count as coming from any quadword, for better codegen.
5703 unsigned LoQuad[] = { 0, 0, 0, 0 };
5704 unsigned HiQuad[] = { 0, 0, 0, 0 };
5705 std::bitset<4> InputQuads;
5706 for (unsigned i = 0; i < 8; ++i) {
5707 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
5708 int EltIdx = SVOp->getMaskElt(i);
5709 MaskVals.push_back(EltIdx);
5718 InputQuads.set(EltIdx / 4);
5721 int BestLoQuad = -1;
5722 unsigned MaxQuad = 1;
5723 for (unsigned i = 0; i < 4; ++i) {
5724 if (LoQuad[i] > MaxQuad) {
5726 MaxQuad = LoQuad[i];
5730 int BestHiQuad = -1;
5732 for (unsigned i = 0; i < 4; ++i) {
5733 if (HiQuad[i] > MaxQuad) {
5735 MaxQuad = HiQuad[i];
5739 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5740 // of the two input vectors, shuffle them into one input vector so only a
5741 // single pshufb instruction is necessary. If There are more than 2 input
5742 // quads, disable the next transformation since it does not help SSSE3.
5743 bool V1Used = InputQuads[0] || InputQuads[1];
5744 bool V2Used = InputQuads[2] || InputQuads[3];
5745 if (Subtarget->hasSSSE3()) {
5746 if (InputQuads.count() == 2 && V1Used && V2Used) {
5747 BestLoQuad = InputQuads[0] ? 0 : 1;
5748 BestHiQuad = InputQuads[2] ? 2 : 3;
5750 if (InputQuads.count() > 2) {
5756 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5757 // the shuffle mask. If a quad is scored as -1, that means that it contains
5758 // words from all 4 input quadwords.
5760 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5762 BestLoQuad < 0 ? 0 : BestLoQuad,
5763 BestHiQuad < 0 ? 1 : BestHiQuad
5765 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5766 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5767 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5768 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5770 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5771 // source words for the shuffle, to aid later transformations.
5772 bool AllWordsInNewV = true;
5773 bool InOrder[2] = { true, true };
5774 for (unsigned i = 0; i != 8; ++i) {
5775 int idx = MaskVals[i];
5777 InOrder[i/4] = false;
5778 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5780 AllWordsInNewV = false;
5784 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5785 if (AllWordsInNewV) {
5786 for (int i = 0; i != 8; ++i) {
5787 int idx = MaskVals[i];
5790 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5791 if ((idx != i) && idx < 4)
5793 if ((idx != i) && idx > 3)
5802 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5803 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5804 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5805 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5806 unsigned TargetMask = 0;
5807 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5808 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5809 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5810 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
5811 getShufflePSHUFLWImmediate(SVOp);
5812 V1 = NewV.getOperand(0);
5813 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5817 // If we have SSSE3, and all words of the result are from 1 input vector,
5818 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5819 // is present, fall back to case 4.
5820 if (Subtarget->hasSSSE3()) {
5821 SmallVector<SDValue,16> pshufbMask;
5823 // If we have elements from both input vectors, set the high bit of the
5824 // shuffle mask element to zero out elements that come from V2 in the V1
5825 // mask, and elements that come from V1 in the V2 mask, so that the two
5826 // results can be OR'd together.
5827 bool TwoInputs = V1Used && V2Used;
5828 for (unsigned i = 0; i != 8; ++i) {
5829 int EltIdx = MaskVals[i] * 2;
5830 int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
5831 int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
5832 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
5833 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
5835 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5836 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5837 DAG.getNode(ISD::BUILD_VECTOR, dl,
5838 MVT::v16i8, &pshufbMask[0], 16));
5840 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5842 // Calculate the shuffle mask for the second input, shuffle it, and
5843 // OR it with the first shuffled input.
5845 for (unsigned i = 0; i != 8; ++i) {
5846 int EltIdx = MaskVals[i] * 2;
5847 int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
5848 int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
5849 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
5850 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
5852 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5853 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5854 DAG.getNode(ISD::BUILD_VECTOR, dl,
5855 MVT::v16i8, &pshufbMask[0], 16));
5856 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5857 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5860 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5861 // and update MaskVals with new element order.
5862 std::bitset<8> InOrder;
5863 if (BestLoQuad >= 0) {
5864 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
5865 for (int i = 0; i != 4; ++i) {
5866 int idx = MaskVals[i];
5869 } else if ((idx / 4) == BestLoQuad) {
5874 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5877 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
5878 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5879 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5881 getShufflePSHUFLWImmediate(SVOp), DAG);
5885 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5886 // and update MaskVals with the new element order.
5887 if (BestHiQuad >= 0) {
5888 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
5889 for (unsigned i = 4; i != 8; ++i) {
5890 int idx = MaskVals[i];
5893 } else if ((idx / 4) == BestHiQuad) {
5894 MaskV[i] = (idx & 3) + 4;
5898 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5901 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
5902 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5903 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5905 getShufflePSHUFHWImmediate(SVOp), DAG);
5909 // In case BestHi & BestLo were both -1, which means each quadword has a word
5910 // from each of the four input quadwords, calculate the InOrder bitvector now
5911 // before falling through to the insert/extract cleanup.
5912 if (BestLoQuad == -1 && BestHiQuad == -1) {
5914 for (int i = 0; i != 8; ++i)
5915 if (MaskVals[i] < 0 || MaskVals[i] == i)
5919 // The other elements are put in the right place using pextrw and pinsrw.
5920 for (unsigned i = 0; i != 8; ++i) {
5923 int EltIdx = MaskVals[i];
5926 SDValue ExtOp = (EltIdx < 8) ?
5927 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5928 DAG.getIntPtrConstant(EltIdx)) :
5929 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5930 DAG.getIntPtrConstant(EltIdx - 8));
5931 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5932 DAG.getIntPtrConstant(i));
5937 // v16i8 shuffles - Prefer shuffles in the following order:
5938 // 1. [ssse3] 1 x pshufb
5939 // 2. [ssse3] 2 x pshufb + 1 x por
5940 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5942 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5944 const X86TargetLowering &TLI) {
5945 SDValue V1 = SVOp->getOperand(0);
5946 SDValue V2 = SVOp->getOperand(1);
5947 DebugLoc dl = SVOp->getDebugLoc();
5948 ArrayRef<int> MaskVals = SVOp->getMask();
5950 // If we have SSSE3, case 1 is generated when all result bytes come from
5951 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5952 // present, fall back to case 3.
5954 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5955 if (TLI.getSubtarget()->hasSSSE3()) {
5956 SmallVector<SDValue,16> pshufbMask;
5958 // If all result elements are from one input vector, then only translate
5959 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5961 // Otherwise, we have elements from both input vectors, and must zero out
5962 // elements that come from V2 in the first mask, and V1 in the second mask
5963 // so that we can OR them together.
5964 for (unsigned i = 0; i != 16; ++i) {
5965 int EltIdx = MaskVals[i];
5966 if (EltIdx < 0 || EltIdx >= 16)
5968 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5970 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5971 DAG.getNode(ISD::BUILD_VECTOR, dl,
5972 MVT::v16i8, &pshufbMask[0], 16));
5974 // As PSHUFB will zero elements with negative indices, it's safe to ignore
5975 // the 2nd operand if it's undefined or zero.
5976 if (V2.getOpcode() == ISD::UNDEF ||
5977 ISD::isBuildVectorAllZeros(V2.getNode()))
5980 // Calculate the shuffle mask for the second input, shuffle it, and
5981 // OR it with the first shuffled input.
5983 for (unsigned i = 0; i != 16; ++i) {
5984 int EltIdx = MaskVals[i];
5985 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
5986 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5988 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5989 DAG.getNode(ISD::BUILD_VECTOR, dl,
5990 MVT::v16i8, &pshufbMask[0], 16));
5991 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5994 // No SSSE3 - Calculate in place words and then fix all out of place words
5995 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5996 // the 16 different words that comprise the two doublequadword input vectors.
5997 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5998 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
6000 for (int i = 0; i != 8; ++i) {
6001 int Elt0 = MaskVals[i*2];
6002 int Elt1 = MaskVals[i*2+1];
6004 // This word of the result is all undef, skip it.
6005 if (Elt0 < 0 && Elt1 < 0)
6008 // This word of the result is already in the correct place, skip it.
6009 if ((Elt0 == i*2) && (Elt1 == i*2+1))
6012 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
6013 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
6016 // If Elt0 and Elt1 are defined, are consecutive, and can be load
6017 // using a single extract together, load it and store it.
6018 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
6019 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6020 DAG.getIntPtrConstant(Elt1 / 2));
6021 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6022 DAG.getIntPtrConstant(i));
6026 // If Elt1 is defined, extract it from the appropriate source. If the
6027 // source byte is not also odd, shift the extracted word left 8 bits
6028 // otherwise clear the bottom 8 bits if we need to do an or.
6030 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6031 DAG.getIntPtrConstant(Elt1 / 2));
6032 if ((Elt1 & 1) == 0)
6033 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
6035 TLI.getShiftAmountTy(InsElt.getValueType())));
6037 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
6038 DAG.getConstant(0xFF00, MVT::i16));
6040 // If Elt0 is defined, extract it from the appropriate source. If the
6041 // source byte is not also even, shift the extracted word right 8 bits. If
6042 // Elt1 was also defined, OR the extracted values together before
6043 // inserting them in the result.
6045 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
6046 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
6047 if ((Elt0 & 1) != 0)
6048 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
6050 TLI.getShiftAmountTy(InsElt0.getValueType())));
6052 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
6053 DAG.getConstant(0x00FF, MVT::i16));
6054 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
6057 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6058 DAG.getIntPtrConstant(i));
6060 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
6063 // v32i8 shuffles - Translate to VPSHUFB if possible.
6065 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
6066 const X86Subtarget *Subtarget,
6067 SelectionDAG &DAG) {
6068 EVT VT = SVOp->getValueType(0);
6069 SDValue V1 = SVOp->getOperand(0);
6070 SDValue V2 = SVOp->getOperand(1);
6071 DebugLoc dl = SVOp->getDebugLoc();
6072 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
6074 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6075 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
6076 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
6078 // VPSHUFB may be generated if
6079 // (1) one of input vector is undefined or zeroinitializer.
6080 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
6081 // And (2) the mask indexes don't cross the 128-bit lane.
6082 if (VT != MVT::v32i8 || !Subtarget->hasAVX2() ||
6083 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
6086 if (V1IsAllZero && !V2IsAllZero) {
6087 CommuteVectorShuffleMask(MaskVals, 32);
6090 SmallVector<SDValue, 32> pshufbMask;
6091 for (unsigned i = 0; i != 32; i++) {
6092 int EltIdx = MaskVals[i];
6093 if (EltIdx < 0 || EltIdx >= 32)
6096 if ((EltIdx >= 16 && i < 16) || (EltIdx < 16 && i >= 16))
6097 // Cross lane is not allowed.
6101 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6103 return DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, V1,
6104 DAG.getNode(ISD::BUILD_VECTOR, dl,
6105 MVT::v32i8, &pshufbMask[0], 32));
6108 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
6109 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
6110 /// done when every pair / quad of shuffle mask elements point to elements in
6111 /// the right sequence. e.g.
6112 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
6114 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
6115 SelectionDAG &DAG, DebugLoc dl) {
6116 MVT VT = SVOp->getValueType(0).getSimpleVT();
6117 unsigned NumElems = VT.getVectorNumElements();
6120 switch (VT.SimpleTy) {
6121 default: llvm_unreachable("Unexpected!");
6122 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
6123 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
6124 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
6125 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
6126 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
6127 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
6130 SmallVector<int, 8> MaskVec;
6131 for (unsigned i = 0; i != NumElems; i += Scale) {
6133 for (unsigned j = 0; j != Scale; ++j) {
6134 int EltIdx = SVOp->getMaskElt(i+j);
6138 StartIdx = (EltIdx / Scale);
6139 if (EltIdx != (int)(StartIdx*Scale + j))
6142 MaskVec.push_back(StartIdx);
6145 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
6146 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
6147 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
6150 /// getVZextMovL - Return a zero-extending vector move low node.
6152 static SDValue getVZextMovL(EVT VT, EVT OpVT,
6153 SDValue SrcOp, SelectionDAG &DAG,
6154 const X86Subtarget *Subtarget, DebugLoc dl) {
6155 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
6156 LoadSDNode *LD = NULL;
6157 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
6158 LD = dyn_cast<LoadSDNode>(SrcOp);
6160 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
6162 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
6163 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
6164 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6165 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
6166 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
6168 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
6169 return DAG.getNode(ISD::BITCAST, dl, VT,
6170 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6171 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6179 return DAG.getNode(ISD::BITCAST, dl, VT,
6180 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6181 DAG.getNode(ISD::BITCAST, dl,
6185 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
6186 /// which could not be matched by any known target speficic shuffle
6188 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6190 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
6191 if (NewOp.getNode())
6194 EVT VT = SVOp->getValueType(0);
6196 unsigned NumElems = VT.getVectorNumElements();
6197 unsigned NumLaneElems = NumElems / 2;
6199 DebugLoc dl = SVOp->getDebugLoc();
6200 MVT EltVT = VT.getVectorElementType().getSimpleVT();
6201 EVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
6204 SmallVector<int, 16> Mask;
6205 for (unsigned l = 0; l < 2; ++l) {
6206 // Build a shuffle mask for the output, discovering on the fly which
6207 // input vectors to use as shuffle operands (recorded in InputUsed).
6208 // If building a suitable shuffle vector proves too hard, then bail
6209 // out with UseBuildVector set.
6210 bool UseBuildVector = false;
6211 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
6212 unsigned LaneStart = l * NumLaneElems;
6213 for (unsigned i = 0; i != NumLaneElems; ++i) {
6214 // The mask element. This indexes into the input.
6215 int Idx = SVOp->getMaskElt(i+LaneStart);
6217 // the mask element does not index into any input vector.
6222 // The input vector this mask element indexes into.
6223 int Input = Idx / NumLaneElems;
6225 // Turn the index into an offset from the start of the input vector.
6226 Idx -= Input * NumLaneElems;
6228 // Find or create a shuffle vector operand to hold this input.
6230 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
6231 if (InputUsed[OpNo] == Input)
6232 // This input vector is already an operand.
6234 if (InputUsed[OpNo] < 0) {
6235 // Create a new operand for this input vector.
6236 InputUsed[OpNo] = Input;
6241 if (OpNo >= array_lengthof(InputUsed)) {
6242 // More than two input vectors used! Give up on trying to create a
6243 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
6244 UseBuildVector = true;
6248 // Add the mask index for the new shuffle vector.
6249 Mask.push_back(Idx + OpNo * NumLaneElems);
6252 if (UseBuildVector) {
6253 SmallVector<SDValue, 16> SVOps;
6254 for (unsigned i = 0; i != NumLaneElems; ++i) {
6255 // The mask element. This indexes into the input.
6256 int Idx = SVOp->getMaskElt(i+LaneStart);
6258 SVOps.push_back(DAG.getUNDEF(EltVT));
6262 // The input vector this mask element indexes into.
6263 int Input = Idx / NumElems;
6265 // Turn the index into an offset from the start of the input vector.
6266 Idx -= Input * NumElems;
6268 // Extract the vector element by hand.
6269 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6270 SVOp->getOperand(Input),
6271 DAG.getIntPtrConstant(Idx)));
6274 // Construct the output using a BUILD_VECTOR.
6275 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &SVOps[0],
6277 } else if (InputUsed[0] < 0) {
6278 // No input vectors were used! The result is undefined.
6279 Output[l] = DAG.getUNDEF(NVT);
6281 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
6282 (InputUsed[0] % 2) * NumLaneElems,
6284 // If only one input was used, use an undefined vector for the other.
6285 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
6286 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
6287 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
6288 // At least one input vector was used. Create a new shuffle vector.
6289 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
6295 // Concatenate the result back
6296 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
6299 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
6300 /// 4 elements, and match them with several different shuffle types.
6302 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6303 SDValue V1 = SVOp->getOperand(0);
6304 SDValue V2 = SVOp->getOperand(1);
6305 DebugLoc dl = SVOp->getDebugLoc();
6306 EVT VT = SVOp->getValueType(0);
6308 assert(VT.is128BitVector() && "Unsupported vector size");
6310 std::pair<int, int> Locs[4];
6311 int Mask1[] = { -1, -1, -1, -1 };
6312 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
6316 for (unsigned i = 0; i != 4; ++i) {
6317 int Idx = PermMask[i];
6319 Locs[i] = std::make_pair(-1, -1);
6321 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
6323 Locs[i] = std::make_pair(0, NumLo);
6327 Locs[i] = std::make_pair(1, NumHi);
6329 Mask1[2+NumHi] = Idx;
6335 if (NumLo <= 2 && NumHi <= 2) {
6336 // If no more than two elements come from either vector. This can be
6337 // implemented with two shuffles. First shuffle gather the elements.
6338 // The second shuffle, which takes the first shuffle as both of its
6339 // vector operands, put the elements into the right order.
6340 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6342 int Mask2[] = { -1, -1, -1, -1 };
6344 for (unsigned i = 0; i != 4; ++i)
6345 if (Locs[i].first != -1) {
6346 unsigned Idx = (i < 2) ? 0 : 4;
6347 Idx += Locs[i].first * 2 + Locs[i].second;
6351 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6354 if (NumLo == 3 || NumHi == 3) {
6355 // Otherwise, we must have three elements from one vector, call it X, and
6356 // one element from the other, call it Y. First, use a shufps to build an
6357 // intermediate vector with the one element from Y and the element from X
6358 // that will be in the same half in the final destination (the indexes don't
6359 // matter). Then, use a shufps to build the final vector, taking the half
6360 // containing the element from Y from the intermediate, and the other half
6363 // Normalize it so the 3 elements come from V1.
6364 CommuteVectorShuffleMask(PermMask, 4);
6368 // Find the element from V2.
6370 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6371 int Val = PermMask[HiIndex];
6378 Mask1[0] = PermMask[HiIndex];
6380 Mask1[2] = PermMask[HiIndex^1];
6382 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6385 Mask1[0] = PermMask[0];
6386 Mask1[1] = PermMask[1];
6387 Mask1[2] = HiIndex & 1 ? 6 : 4;
6388 Mask1[3] = HiIndex & 1 ? 4 : 6;
6389 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6392 Mask1[0] = HiIndex & 1 ? 2 : 0;
6393 Mask1[1] = HiIndex & 1 ? 0 : 2;
6394 Mask1[2] = PermMask[2];
6395 Mask1[3] = PermMask[3];
6400 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6403 // Break it into (shuffle shuffle_hi, shuffle_lo).
6404 int LoMask[] = { -1, -1, -1, -1 };
6405 int HiMask[] = { -1, -1, -1, -1 };
6407 int *MaskPtr = LoMask;
6408 unsigned MaskIdx = 0;
6411 for (unsigned i = 0; i != 4; ++i) {
6418 int Idx = PermMask[i];
6420 Locs[i] = std::make_pair(-1, -1);
6421 } else if (Idx < 4) {
6422 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6423 MaskPtr[LoIdx] = Idx;
6426 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6427 MaskPtr[HiIdx] = Idx;
6432 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6433 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6434 int MaskOps[] = { -1, -1, -1, -1 };
6435 for (unsigned i = 0; i != 4; ++i)
6436 if (Locs[i].first != -1)
6437 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
6438 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6441 static bool MayFoldVectorLoad(SDValue V) {
6442 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6443 V = V.getOperand(0);
6445 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6446 V = V.getOperand(0);
6447 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
6448 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
6449 // BUILD_VECTOR (load), undef
6450 V = V.getOperand(0);
6452 return MayFoldLoad(V);
6455 // FIXME: the version above should always be used. Since there's
6456 // a bug where several vector shuffles can't be folded because the
6457 // DAG is not updated during lowering and a node claims to have two
6458 // uses while it only has one, use this version, and let isel match
6459 // another instruction if the load really happens to have more than
6460 // one use. Remove this version after this bug get fixed.
6461 // rdar://8434668, PR8156
6462 static bool RelaxedMayFoldVectorLoad(SDValue V) {
6463 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6464 V = V.getOperand(0);
6465 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6466 V = V.getOperand(0);
6467 if (ISD::isNormalLoad(V.getNode()))
6473 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
6474 EVT VT = Op.getValueType();
6476 // Canonizalize to v2f64.
6477 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6478 return DAG.getNode(ISD::BITCAST, dl, VT,
6479 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6484 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
6486 SDValue V1 = Op.getOperand(0);
6487 SDValue V2 = Op.getOperand(1);
6488 EVT VT = Op.getValueType();
6490 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6492 if (HasSSE2 && VT == MVT::v2f64)
6493 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6495 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
6496 return DAG.getNode(ISD::BITCAST, dl, VT,
6497 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
6498 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
6499 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
6503 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
6504 SDValue V1 = Op.getOperand(0);
6505 SDValue V2 = Op.getOperand(1);
6506 EVT VT = Op.getValueType();
6508 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6509 "unsupported shuffle type");
6511 if (V2.getOpcode() == ISD::UNDEF)
6515 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6519 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6520 SDValue V1 = Op.getOperand(0);
6521 SDValue V2 = Op.getOperand(1);
6522 EVT VT = Op.getValueType();
6523 unsigned NumElems = VT.getVectorNumElements();
6525 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6526 // operand of these instructions is only memory, so check if there's a
6527 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6529 bool CanFoldLoad = false;
6531 // Trivial case, when V2 comes from a load.
6532 if (MayFoldVectorLoad(V2))
6535 // When V1 is a load, it can be folded later into a store in isel, example:
6536 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6538 // (MOVLPSmr addr:$src1, VR128:$src2)
6539 // So, recognize this potential and also use MOVLPS or MOVLPD
6540 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6543 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6545 if (HasSSE2 && NumElems == 2)
6546 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6549 // If we don't care about the second element, proceed to use movss.
6550 if (SVOp->getMaskElt(1) != -1)
6551 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6554 // movl and movlp will both match v2i64, but v2i64 is never matched by
6555 // movl earlier because we make it strict to avoid messing with the movlp load
6556 // folding logic (see the code above getMOVLP call). Match it here then,
6557 // this is horrible, but will stay like this until we move all shuffle
6558 // matching to x86 specific nodes. Note that for the 1st condition all
6559 // types are matched with movsd.
6561 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
6562 // as to remove this logic from here, as much as possible
6563 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
6564 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6565 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6568 assert(VT != MVT::v4i32 && "unsupported shuffle type");
6570 // Invert the operand order and use SHUFPS to match it.
6571 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
6572 getShuffleSHUFImmediate(SVOp), DAG);
6575 // Reduce a vector shuffle to zext.
6577 X86TargetLowering::lowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
6578 // PMOVZX is only available from SSE41.
6579 if (!Subtarget->hasSSE41())
6582 EVT VT = Op.getValueType();
6584 // Only AVX2 support 256-bit vector integer extending.
6585 if (!Subtarget->hasAVX2() && VT.is256BitVector())
6588 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6589 DebugLoc DL = Op.getDebugLoc();
6590 SDValue V1 = Op.getOperand(0);
6591 SDValue V2 = Op.getOperand(1);
6592 unsigned NumElems = VT.getVectorNumElements();
6594 // Extending is an unary operation and the element type of the source vector
6595 // won't be equal to or larger than i64.
6596 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
6597 VT.getVectorElementType() == MVT::i64)
6600 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
6601 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
6602 while ((1U << Shift) < NumElems) {
6603 if (SVOp->getMaskElt(1U << Shift) == 1)
6606 // The maximal ratio is 8, i.e. from i8 to i64.
6611 // Check the shuffle mask.
6612 unsigned Mask = (1U << Shift) - 1;
6613 for (unsigned i = 0; i != NumElems; ++i) {
6614 int EltIdx = SVOp->getMaskElt(i);
6615 if ((i & Mask) != 0 && EltIdx != -1)
6617 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
6621 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
6622 EVT NeVT = EVT::getIntegerVT(*DAG.getContext(), NBits);
6623 EVT NVT = EVT::getVectorVT(*DAG.getContext(), NeVT, NumElems >> Shift);
6625 if (!isTypeLegal(NVT))
6628 // Simplify the operand as it's prepared to be fed into shuffle.
6629 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
6630 if (V1.getOpcode() == ISD::BITCAST &&
6631 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
6632 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6634 .getOperand(0).getValueType().getSizeInBits() == SignificantBits) {
6635 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
6636 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
6637 ConstantSDNode *CIdx =
6638 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
6639 // If it's foldable, i.e. normal load with single use, we will let code
6640 // selection to fold it. Otherwise, we will short the conversion sequence.
6641 if (CIdx && CIdx->getZExtValue() == 0 &&
6642 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse()))
6643 V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V);
6646 return DAG.getNode(ISD::BITCAST, DL, VT,
6647 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
6651 X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
6652 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6653 EVT VT = Op.getValueType();
6654 DebugLoc dl = Op.getDebugLoc();
6655 SDValue V1 = Op.getOperand(0);
6656 SDValue V2 = Op.getOperand(1);
6658 if (isZeroShuffle(SVOp))
6659 return getZeroVector(VT, Subtarget, DAG, dl);
6661 // Handle splat operations
6662 if (SVOp->isSplat()) {
6663 unsigned NumElem = VT.getVectorNumElements();
6664 int Size = VT.getSizeInBits();
6666 // Use vbroadcast whenever the splat comes from a foldable load
6667 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
6668 if (Broadcast.getNode())
6671 // Handle splats by matching through known shuffle masks
6672 if ((Size == 128 && NumElem <= 4) ||
6673 (Size == 256 && NumElem < 8))
6676 // All remaning splats are promoted to target supported vector shuffles.
6677 return PromoteSplat(SVOp, DAG);
6680 // Check integer expanding shuffles.
6681 SDValue NewOp = lowerVectorIntExtend(Op, DAG);
6682 if (NewOp.getNode())
6685 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6687 if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
6688 VT == MVT::v16i16 || VT == MVT::v32i8) {
6689 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6690 if (NewOp.getNode())
6691 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6692 } else if ((VT == MVT::v4i32 ||
6693 (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6694 // FIXME: Figure out a cleaner way to do this.
6695 // Try to make use of movq to zero out the top part.
6696 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6697 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6698 if (NewOp.getNode()) {
6699 EVT NewVT = NewOp.getValueType();
6700 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
6701 NewVT, true, false))
6702 return getVZextMovL(VT, NewVT, NewOp.getOperand(0),
6703 DAG, Subtarget, dl);
6705 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6706 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6707 if (NewOp.getNode()) {
6708 EVT NewVT = NewOp.getValueType();
6709 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
6710 return getVZextMovL(VT, NewVT, NewOp.getOperand(1),
6711 DAG, Subtarget, dl);
6719 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6720 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6721 SDValue V1 = Op.getOperand(0);
6722 SDValue V2 = Op.getOperand(1);
6723 EVT VT = Op.getValueType();
6724 DebugLoc dl = Op.getDebugLoc();
6725 unsigned NumElems = VT.getVectorNumElements();
6726 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6727 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6728 bool V1IsSplat = false;
6729 bool V2IsSplat = false;
6730 bool HasSSE2 = Subtarget->hasSSE2();
6731 bool HasAVX = Subtarget->hasAVX();
6732 bool HasAVX2 = Subtarget->hasAVX2();
6733 MachineFunction &MF = DAG.getMachineFunction();
6734 bool OptForSize = MF.getFunction()->getFnAttributes().
6735 hasAttribute(Attributes::OptimizeForSize);
6737 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
6739 if (V1IsUndef && V2IsUndef)
6740 return DAG.getUNDEF(VT);
6742 assert(!V1IsUndef && "Op 1 of shuffle should not be undef");
6744 // Vector shuffle lowering takes 3 steps:
6746 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6747 // narrowing and commutation of operands should be handled.
6748 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6750 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6751 // so the shuffle can be broken into other shuffles and the legalizer can
6752 // try the lowering again.
6754 // The general idea is that no vector_shuffle operation should be left to
6755 // be matched during isel, all of them must be converted to a target specific
6758 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6759 // narrowing and commutation of operands should be handled. The actual code
6760 // doesn't include all of those, work in progress...
6761 SDValue NewOp = NormalizeVectorShuffle(Op, DAG);
6762 if (NewOp.getNode())
6765 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
6767 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6768 // unpckh_undef). Only use pshufd if speed is more important than size.
6769 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasAVX2))
6770 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
6771 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasAVX2))
6772 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
6774 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
6775 V2IsUndef && RelaxedMayFoldVectorLoad(V1))
6776 return getMOVDDup(Op, dl, V1, DAG);
6778 if (isMOVHLPS_v_undef_Mask(M, VT))
6779 return getMOVHighToLow(Op, dl, DAG);
6781 // Use to match splats
6782 if (HasSSE2 && isUNPCKHMask(M, VT, HasAVX2) && V2IsUndef &&
6783 (VT == MVT::v2f64 || VT == MVT::v2i64))
6784 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
6786 if (isPSHUFDMask(M, VT)) {
6787 // The actual implementation will match the mask in the if above and then
6788 // during isel it can match several different instructions, not only pshufd
6789 // as its name says, sad but true, emulate the behavior for now...
6790 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6791 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6793 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
6795 if (HasAVX && (VT == MVT::v4f32 || VT == MVT::v2f64))
6796 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask, DAG);
6798 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6799 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6801 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
6805 // Check if this can be converted into a logical shift.
6806 bool isLeft = false;
6809 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6810 if (isShift && ShVal.hasOneUse()) {
6811 // If the shifted value has multiple uses, it may be cheaper to use
6812 // v_set0 + movlhps or movhlps, etc.
6813 EVT EltVT = VT.getVectorElementType();
6814 ShAmt *= EltVT.getSizeInBits();
6815 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6818 if (isMOVLMask(M, VT)) {
6819 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6820 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6821 if (!isMOVLPMask(M, VT)) {
6822 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6823 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6825 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6826 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6830 // FIXME: fold these into legal mask.
6831 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasAVX2))
6832 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6834 if (isMOVHLPSMask(M, VT))
6835 return getMOVHighToLow(Op, dl, DAG);
6837 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
6838 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6840 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
6841 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6843 if (isMOVLPMask(M, VT))
6844 return getMOVLP(Op, dl, DAG, HasSSE2);
6846 if (ShouldXformToMOVHLPS(M, VT) ||
6847 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
6848 return CommuteVectorShuffle(SVOp, DAG);
6851 // No better options. Use a vshldq / vsrldq.
6852 EVT EltVT = VT.getVectorElementType();
6853 ShAmt *= EltVT.getSizeInBits();
6854 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6857 bool Commuted = false;
6858 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6859 // 1,1,1,1 -> v8i16 though.
6860 V1IsSplat = isSplatVector(V1.getNode());
6861 V2IsSplat = isSplatVector(V2.getNode());
6863 // Canonicalize the splat or undef, if present, to be on the RHS.
6864 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
6865 CommuteVectorShuffleMask(M, NumElems);
6867 std::swap(V1IsSplat, V2IsSplat);
6871 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
6872 // Shuffling low element of v1 into undef, just return v1.
6875 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6876 // the instruction selector will not match, so get a canonical MOVL with
6877 // swapped operands to undo the commute.
6878 return getMOVL(DAG, dl, VT, V2, V1);
6881 if (isUNPCKLMask(M, VT, HasAVX2))
6882 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6884 if (isUNPCKHMask(M, VT, HasAVX2))
6885 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6888 // Normalize mask so all entries that point to V2 points to its first
6889 // element then try to match unpck{h|l} again. If match, return a
6890 // new vector_shuffle with the corrected mask.p
6891 SmallVector<int, 8> NewMask(M.begin(), M.end());
6892 NormalizeMask(NewMask, NumElems);
6893 if (isUNPCKLMask(NewMask, VT, HasAVX2, true))
6894 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6895 if (isUNPCKHMask(NewMask, VT, HasAVX2, true))
6896 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6900 // Commute is back and try unpck* again.
6901 // FIXME: this seems wrong.
6902 CommuteVectorShuffleMask(M, NumElems);
6904 std::swap(V1IsSplat, V2IsSplat);
6907 if (isUNPCKLMask(M, VT, HasAVX2))
6908 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6910 if (isUNPCKHMask(M, VT, HasAVX2))
6911 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6914 // Normalize the node to match x86 shuffle ops if needed
6915 if (!V2IsUndef && (isSHUFPMask(M, VT, HasAVX, /* Commuted */ true)))
6916 return CommuteVectorShuffle(SVOp, DAG);
6918 // The checks below are all present in isShuffleMaskLegal, but they are
6919 // inlined here right now to enable us to directly emit target specific
6920 // nodes, and remove one by one until they don't return Op anymore.
6922 if (isPALIGNRMask(M, VT, Subtarget))
6923 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6924 getShufflePALIGNRImmediate(SVOp),
6927 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6928 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6929 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6930 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
6933 if (isPSHUFHWMask(M, VT, HasAVX2))
6934 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6935 getShufflePSHUFHWImmediate(SVOp),
6938 if (isPSHUFLWMask(M, VT, HasAVX2))
6939 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6940 getShufflePSHUFLWImmediate(SVOp),
6943 if (isSHUFPMask(M, VT, HasAVX))
6944 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
6945 getShuffleSHUFImmediate(SVOp), DAG);
6947 if (isUNPCKL_v_undef_Mask(M, VT, HasAVX2))
6948 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
6949 if (isUNPCKH_v_undef_Mask(M, VT, HasAVX2))
6950 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
6952 //===--------------------------------------------------------------------===//
6953 // Generate target specific nodes for 128 or 256-bit shuffles only
6954 // supported in the AVX instruction set.
6957 // Handle VMOVDDUPY permutations
6958 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasAVX))
6959 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
6961 // Handle VPERMILPS/D* permutations
6962 if (isVPERMILPMask(M, VT, HasAVX)) {
6963 if (HasAVX2 && VT == MVT::v8i32)
6964 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
6965 getShuffleSHUFImmediate(SVOp), DAG);
6966 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
6967 getShuffleSHUFImmediate(SVOp), DAG);
6970 // Handle VPERM2F128/VPERM2I128 permutations
6971 if (isVPERM2X128Mask(M, VT, HasAVX))
6972 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
6973 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
6975 SDValue BlendOp = LowerVECTOR_SHUFFLEtoBlend(SVOp, Subtarget, DAG);
6976 if (BlendOp.getNode())
6979 if (V2IsUndef && HasAVX2 && (VT == MVT::v8i32 || VT == MVT::v8f32)) {
6980 SmallVector<SDValue, 8> permclMask;
6981 for (unsigned i = 0; i != 8; ++i) {
6982 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MVT::i32));
6984 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32,
6986 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
6987 return DAG.getNode(X86ISD::VPERMV, dl, VT,
6988 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
6991 if (V2IsUndef && HasAVX2 && (VT == MVT::v4i64 || VT == MVT::v4f64))
6992 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1,
6993 getShuffleCLImmediate(SVOp), DAG);
6996 //===--------------------------------------------------------------------===//
6997 // Since no target specific shuffle was selected for this generic one,
6998 // lower it into other known shuffles. FIXME: this isn't true yet, but
6999 // this is the plan.
7002 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
7003 if (VT == MVT::v8i16) {
7004 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
7005 if (NewOp.getNode())
7009 if (VT == MVT::v16i8) {
7010 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
7011 if (NewOp.getNode())
7015 if (VT == MVT::v32i8) {
7016 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
7017 if (NewOp.getNode())
7021 // Handle all 128-bit wide vectors with 4 elements, and match them with
7022 // several different shuffle types.
7023 if (NumElems == 4 && VT.is128BitVector())
7024 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
7026 // Handle general 256-bit shuffles
7027 if (VT.is256BitVector())
7028 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
7034 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
7035 SelectionDAG &DAG) const {
7036 EVT VT = Op.getValueType();
7037 DebugLoc dl = Op.getDebugLoc();
7039 if (!Op.getOperand(0).getValueType().is128BitVector())
7042 if (VT.getSizeInBits() == 8) {
7043 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
7044 Op.getOperand(0), Op.getOperand(1));
7045 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7046 DAG.getValueType(VT));
7047 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7050 if (VT.getSizeInBits() == 16) {
7051 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7052 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
7054 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7055 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7056 DAG.getNode(ISD::BITCAST, dl,
7060 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
7061 Op.getOperand(0), Op.getOperand(1));
7062 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7063 DAG.getValueType(VT));
7064 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7067 if (VT == MVT::f32) {
7068 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
7069 // the result back to FR32 register. It's only worth matching if the
7070 // result has a single use which is a store or a bitcast to i32. And in
7071 // the case of a store, it's not worth it if the index is a constant 0,
7072 // because a MOVSSmr can be used instead, which is smaller and faster.
7073 if (!Op.hasOneUse())
7075 SDNode *User = *Op.getNode()->use_begin();
7076 if ((User->getOpcode() != ISD::STORE ||
7077 (isa<ConstantSDNode>(Op.getOperand(1)) &&
7078 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
7079 (User->getOpcode() != ISD::BITCAST ||
7080 User->getValueType(0) != MVT::i32))
7082 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7083 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
7086 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
7089 if (VT == MVT::i32 || VT == MVT::i64) {
7090 // ExtractPS/pextrq works with constant index.
7091 if (isa<ConstantSDNode>(Op.getOperand(1)))
7099 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
7100 SelectionDAG &DAG) const {
7101 if (!isa<ConstantSDNode>(Op.getOperand(1)))
7104 SDValue Vec = Op.getOperand(0);
7105 EVT VecVT = Vec.getValueType();
7107 // If this is a 256-bit vector result, first extract the 128-bit vector and
7108 // then extract the element from the 128-bit vector.
7109 if (VecVT.is256BitVector()) {
7110 DebugLoc dl = Op.getNode()->getDebugLoc();
7111 unsigned NumElems = VecVT.getVectorNumElements();
7112 SDValue Idx = Op.getOperand(1);
7113 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7115 // Get the 128-bit vector.
7116 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
7118 if (IdxVal >= NumElems/2)
7119 IdxVal -= NumElems/2;
7120 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
7121 DAG.getConstant(IdxVal, MVT::i32));
7124 assert(VecVT.is128BitVector() && "Unexpected vector length");
7126 if (Subtarget->hasSSE41()) {
7127 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
7132 EVT VT = Op.getValueType();
7133 DebugLoc dl = Op.getDebugLoc();
7134 // TODO: handle v16i8.
7135 if (VT.getSizeInBits() == 16) {
7136 SDValue Vec = Op.getOperand(0);
7137 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7139 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7140 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7141 DAG.getNode(ISD::BITCAST, dl,
7144 // Transform it so it match pextrw which produces a 32-bit result.
7145 EVT EltVT = MVT::i32;
7146 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
7147 Op.getOperand(0), Op.getOperand(1));
7148 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
7149 DAG.getValueType(VT));
7150 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7153 if (VT.getSizeInBits() == 32) {
7154 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7158 // SHUFPS the element to the lowest double word, then movss.
7159 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
7160 EVT VVT = Op.getOperand(0).getValueType();
7161 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7162 DAG.getUNDEF(VVT), Mask);
7163 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7164 DAG.getIntPtrConstant(0));
7167 if (VT.getSizeInBits() == 64) {
7168 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
7169 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
7170 // to match extract_elt for f64.
7171 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7175 // UNPCKHPD the element to the lowest double word, then movsd.
7176 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
7177 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
7178 int Mask[2] = { 1, -1 };
7179 EVT VVT = Op.getOperand(0).getValueType();
7180 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7181 DAG.getUNDEF(VVT), Mask);
7182 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7183 DAG.getIntPtrConstant(0));
7190 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
7191 SelectionDAG &DAG) const {
7192 EVT VT = Op.getValueType();
7193 EVT EltVT = VT.getVectorElementType();
7194 DebugLoc dl = Op.getDebugLoc();
7196 SDValue N0 = Op.getOperand(0);
7197 SDValue N1 = Op.getOperand(1);
7198 SDValue N2 = Op.getOperand(2);
7200 if (!VT.is128BitVector())
7203 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
7204 isa<ConstantSDNode>(N2)) {
7206 if (VT == MVT::v8i16)
7207 Opc = X86ISD::PINSRW;
7208 else if (VT == MVT::v16i8)
7209 Opc = X86ISD::PINSRB;
7211 Opc = X86ISD::PINSRB;
7213 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
7215 if (N1.getValueType() != MVT::i32)
7216 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7217 if (N2.getValueType() != MVT::i32)
7218 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7219 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
7222 if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
7223 // Bits [7:6] of the constant are the source select. This will always be
7224 // zero here. The DAG Combiner may combine an extract_elt index into these
7225 // bits. For example (insert (extract, 3), 2) could be matched by putting
7226 // the '3' into bits [7:6] of X86ISD::INSERTPS.
7227 // Bits [5:4] of the constant are the destination select. This is the
7228 // value of the incoming immediate.
7229 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
7230 // combine either bitwise AND or insert of float 0.0 to set these bits.
7231 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
7232 // Create this as a scalar to vector..
7233 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
7234 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
7237 if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
7238 // PINSR* works with constant index.
7245 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
7246 EVT VT = Op.getValueType();
7247 EVT EltVT = VT.getVectorElementType();
7249 DebugLoc dl = Op.getDebugLoc();
7250 SDValue N0 = Op.getOperand(0);
7251 SDValue N1 = Op.getOperand(1);
7252 SDValue N2 = Op.getOperand(2);
7254 // If this is a 256-bit vector result, first extract the 128-bit vector,
7255 // insert the element into the extracted half and then place it back.
7256 if (VT.is256BitVector()) {
7257 if (!isa<ConstantSDNode>(N2))
7260 // Get the desired 128-bit vector half.
7261 unsigned NumElems = VT.getVectorNumElements();
7262 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7263 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
7265 // Insert the element into the desired half.
7266 bool Upper = IdxVal >= NumElems/2;
7267 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
7268 DAG.getConstant(Upper ? IdxVal-NumElems/2 : IdxVal, MVT::i32));
7270 // Insert the changed part back to the 256-bit vector
7271 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
7274 if (Subtarget->hasSSE41())
7275 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7277 if (EltVT == MVT::i8)
7280 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7281 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7282 // as its second argument.
7283 if (N1.getValueType() != MVT::i32)
7284 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7285 if (N2.getValueType() != MVT::i32)
7286 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7287 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7292 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
7293 LLVMContext *Context = DAG.getContext();
7294 DebugLoc dl = Op.getDebugLoc();
7295 EVT OpVT = Op.getValueType();
7297 // If this is a 256-bit vector result, first insert into a 128-bit
7298 // vector and then insert into the 256-bit vector.
7299 if (!OpVT.is128BitVector()) {
7300 // Insert into a 128-bit vector.
7301 EVT VT128 = EVT::getVectorVT(*Context,
7302 OpVT.getVectorElementType(),
7303 OpVT.getVectorNumElements() / 2);
7305 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7307 // Insert the 128-bit vector.
7308 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
7311 if (OpVT == MVT::v1i64 &&
7312 Op.getOperand(0).getValueType() == MVT::i64)
7313 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7315 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7316 assert(OpVT.is128BitVector() && "Expected an SSE type!");
7317 return DAG.getNode(ISD::BITCAST, dl, OpVT,
7318 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7321 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7322 // a simple subregister reference or explicit instructions to grab
7323 // upper bits of a vector.
7324 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7325 SelectionDAG &DAG) {
7326 if (Subtarget->hasAVX()) {
7327 DebugLoc dl = Op.getNode()->getDebugLoc();
7328 SDValue Vec = Op.getNode()->getOperand(0);
7329 SDValue Idx = Op.getNode()->getOperand(1);
7331 if (Op.getNode()->getValueType(0).is128BitVector() &&
7332 Vec.getNode()->getValueType(0).is256BitVector() &&
7333 isa<ConstantSDNode>(Idx)) {
7334 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7335 return Extract128BitVector(Vec, IdxVal, DAG, dl);
7341 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7342 // simple superregister reference or explicit instructions to insert
7343 // the upper bits of a vector.
7344 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7345 SelectionDAG &DAG) {
7346 if (Subtarget->hasAVX()) {
7347 DebugLoc dl = Op.getNode()->getDebugLoc();
7348 SDValue Vec = Op.getNode()->getOperand(0);
7349 SDValue SubVec = Op.getNode()->getOperand(1);
7350 SDValue Idx = Op.getNode()->getOperand(2);
7352 if (Op.getNode()->getValueType(0).is256BitVector() &&
7353 SubVec.getNode()->getValueType(0).is128BitVector() &&
7354 isa<ConstantSDNode>(Idx)) {
7355 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7356 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
7362 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7363 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7364 // one of the above mentioned nodes. It has to be wrapped because otherwise
7365 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7366 // be used to form addressing mode. These wrapped nodes will be selected
7369 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7370 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7372 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7374 unsigned char OpFlag = 0;
7375 unsigned WrapperKind = X86ISD::Wrapper;
7376 CodeModel::Model M = getTargetMachine().getCodeModel();
7378 if (Subtarget->isPICStyleRIPRel() &&
7379 (M == CodeModel::Small || M == CodeModel::Kernel))
7380 WrapperKind = X86ISD::WrapperRIP;
7381 else if (Subtarget->isPICStyleGOT())
7382 OpFlag = X86II::MO_GOTOFF;
7383 else if (Subtarget->isPICStyleStubPIC())
7384 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7386 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7388 CP->getOffset(), OpFlag);
7389 DebugLoc DL = CP->getDebugLoc();
7390 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7391 // With PIC, the address is actually $g + Offset.
7393 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7394 DAG.getNode(X86ISD::GlobalBaseReg,
7395 DebugLoc(), getPointerTy()),
7402 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7403 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7405 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7407 unsigned char OpFlag = 0;
7408 unsigned WrapperKind = X86ISD::Wrapper;
7409 CodeModel::Model M = getTargetMachine().getCodeModel();
7411 if (Subtarget->isPICStyleRIPRel() &&
7412 (M == CodeModel::Small || M == CodeModel::Kernel))
7413 WrapperKind = X86ISD::WrapperRIP;
7414 else if (Subtarget->isPICStyleGOT())
7415 OpFlag = X86II::MO_GOTOFF;
7416 else if (Subtarget->isPICStyleStubPIC())
7417 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7419 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7421 DebugLoc DL = JT->getDebugLoc();
7422 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7424 // With PIC, the address is actually $g + Offset.
7426 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7427 DAG.getNode(X86ISD::GlobalBaseReg,
7428 DebugLoc(), getPointerTy()),
7435 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7436 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7438 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7440 unsigned char OpFlag = 0;
7441 unsigned WrapperKind = X86ISD::Wrapper;
7442 CodeModel::Model M = getTargetMachine().getCodeModel();
7444 if (Subtarget->isPICStyleRIPRel() &&
7445 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7446 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7447 OpFlag = X86II::MO_GOTPCREL;
7448 WrapperKind = X86ISD::WrapperRIP;
7449 } else if (Subtarget->isPICStyleGOT()) {
7450 OpFlag = X86II::MO_GOT;
7451 } else if (Subtarget->isPICStyleStubPIC()) {
7452 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7453 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7454 OpFlag = X86II::MO_DARWIN_NONLAZY;
7457 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7459 DebugLoc DL = Op.getDebugLoc();
7460 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7463 // With PIC, the address is actually $g + Offset.
7464 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7465 !Subtarget->is64Bit()) {
7466 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7467 DAG.getNode(X86ISD::GlobalBaseReg,
7468 DebugLoc(), getPointerTy()),
7472 // For symbols that require a load from a stub to get the address, emit the
7474 if (isGlobalStubReference(OpFlag))
7475 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7476 MachinePointerInfo::getGOT(), false, false, false, 0);
7482 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7483 // Create the TargetBlockAddressAddress node.
7484 unsigned char OpFlags =
7485 Subtarget->ClassifyBlockAddressReference();
7486 CodeModel::Model M = getTargetMachine().getCodeModel();
7487 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
7488 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
7489 DebugLoc dl = Op.getDebugLoc();
7490 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
7493 if (Subtarget->isPICStyleRIPRel() &&
7494 (M == CodeModel::Small || M == CodeModel::Kernel))
7495 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7497 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7499 // With PIC, the address is actually $g + Offset.
7500 if (isGlobalRelativeToPICBase(OpFlags)) {
7501 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7502 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7510 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
7512 SelectionDAG &DAG) const {
7513 // Create the TargetGlobalAddress node, folding in the constant
7514 // offset if it is legal.
7515 unsigned char OpFlags =
7516 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
7517 CodeModel::Model M = getTargetMachine().getCodeModel();
7519 if (OpFlags == X86II::MO_NO_FLAG &&
7520 X86::isOffsetSuitableForCodeModel(Offset, M)) {
7521 // A direct static reference to a global.
7522 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
7525 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
7528 if (Subtarget->isPICStyleRIPRel() &&
7529 (M == CodeModel::Small || M == CodeModel::Kernel))
7530 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7532 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7534 // With PIC, the address is actually $g + Offset.
7535 if (isGlobalRelativeToPICBase(OpFlags)) {
7536 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7537 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7541 // For globals that require a load from a stub to get the address, emit the
7543 if (isGlobalStubReference(OpFlags))
7544 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7545 MachinePointerInfo::getGOT(), false, false, false, 0);
7547 // If there was a non-zero offset that we didn't fold, create an explicit
7550 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7551 DAG.getConstant(Offset, getPointerTy()));
7557 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
7558 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
7559 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
7560 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
7564 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
7565 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
7566 unsigned char OperandFlags, bool LocalDynamic = false) {
7567 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7568 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7569 DebugLoc dl = GA->getDebugLoc();
7570 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7571 GA->getValueType(0),
7575 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
7579 SDValue Ops[] = { Chain, TGA, *InFlag };
7580 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 3);
7582 SDValue Ops[] = { Chain, TGA };
7583 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 2);
7586 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
7587 MFI->setAdjustsStack(true);
7589 SDValue Flag = Chain.getValue(1);
7590 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
7593 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
7595 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7598 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
7599 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7600 DAG.getNode(X86ISD::GlobalBaseReg,
7601 DebugLoc(), PtrVT), InFlag);
7602 InFlag = Chain.getValue(1);
7604 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
7607 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
7609 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7611 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
7612 X86::RAX, X86II::MO_TLSGD);
7615 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
7619 DebugLoc dl = GA->getDebugLoc();
7621 // Get the start address of the TLS block for this module.
7622 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
7623 .getInfo<X86MachineFunctionInfo>();
7624 MFI->incNumLocalDynamicTLSAccesses();
7628 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
7629 X86II::MO_TLSLD, /*LocalDynamic=*/true);
7632 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7633 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT), InFlag);
7634 InFlag = Chain.getValue(1);
7635 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
7636 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
7639 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
7643 unsigned char OperandFlags = X86II::MO_DTPOFF;
7644 unsigned WrapperKind = X86ISD::Wrapper;
7645 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7646 GA->getValueType(0),
7647 GA->getOffset(), OperandFlags);
7648 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7650 // Add x@dtpoff with the base.
7651 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
7654 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
7655 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7656 const EVT PtrVT, TLSModel::Model model,
7657 bool is64Bit, bool isPIC) {
7658 DebugLoc dl = GA->getDebugLoc();
7660 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
7661 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
7662 is64Bit ? 257 : 256));
7664 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
7665 DAG.getIntPtrConstant(0),
7666 MachinePointerInfo(Ptr),
7667 false, false, false, 0);
7669 unsigned char OperandFlags = 0;
7670 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
7672 unsigned WrapperKind = X86ISD::Wrapper;
7673 if (model == TLSModel::LocalExec) {
7674 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
7675 } else if (model == TLSModel::InitialExec) {
7677 OperandFlags = X86II::MO_GOTTPOFF;
7678 WrapperKind = X86ISD::WrapperRIP;
7680 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
7683 llvm_unreachable("Unexpected model");
7686 // emit "addl x@ntpoff,%eax" (local exec)
7687 // or "addl x@indntpoff,%eax" (initial exec)
7688 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
7689 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7690 GA->getValueType(0),
7691 GA->getOffset(), OperandFlags);
7692 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7694 if (model == TLSModel::InitialExec) {
7695 if (isPIC && !is64Bit) {
7696 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
7697 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT),
7701 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
7702 MachinePointerInfo::getGOT(), false, false, false,
7706 // The address of the thread local variable is the add of the thread
7707 // pointer with the offset of the variable.
7708 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
7712 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
7714 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
7715 const GlobalValue *GV = GA->getGlobal();
7717 if (Subtarget->isTargetELF()) {
7718 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
7721 case TLSModel::GeneralDynamic:
7722 if (Subtarget->is64Bit())
7723 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
7724 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
7725 case TLSModel::LocalDynamic:
7726 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
7727 Subtarget->is64Bit());
7728 case TLSModel::InitialExec:
7729 case TLSModel::LocalExec:
7730 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
7731 Subtarget->is64Bit(),
7732 getTargetMachine().getRelocationModel() == Reloc::PIC_);
7734 llvm_unreachable("Unknown TLS model.");
7737 if (Subtarget->isTargetDarwin()) {
7738 // Darwin only has one model of TLS. Lower to that.
7739 unsigned char OpFlag = 0;
7740 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
7741 X86ISD::WrapperRIP : X86ISD::Wrapper;
7743 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7745 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
7746 !Subtarget->is64Bit();
7748 OpFlag = X86II::MO_TLVP_PIC_BASE;
7750 OpFlag = X86II::MO_TLVP;
7751 DebugLoc DL = Op.getDebugLoc();
7752 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7753 GA->getValueType(0),
7754 GA->getOffset(), OpFlag);
7755 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7757 // With PIC32, the address is actually $g + Offset.
7759 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7760 DAG.getNode(X86ISD::GlobalBaseReg,
7761 DebugLoc(), getPointerTy()),
7764 // Lowering the machine isd will make sure everything is in the right
7766 SDValue Chain = DAG.getEntryNode();
7767 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7768 SDValue Args[] = { Chain, Offset };
7769 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7771 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7772 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7773 MFI->setAdjustsStack(true);
7775 // And our return value (tls address) is in the standard call return value
7777 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7778 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
7782 if (Subtarget->isTargetWindows()) {
7783 // Just use the implicit TLS architecture
7784 // Need to generate someting similar to:
7785 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
7787 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
7788 // mov rcx, qword [rdx+rcx*8]
7789 // mov eax, .tls$:tlsvar
7790 // [rax+rcx] contains the address
7791 // Windows 64bit: gs:0x58
7792 // Windows 32bit: fs:__tls_array
7794 // If GV is an alias then use the aliasee for determining
7795 // thread-localness.
7796 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
7797 GV = GA->resolveAliasedGlobal(false);
7798 DebugLoc dl = GA->getDebugLoc();
7799 SDValue Chain = DAG.getEntryNode();
7801 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
7802 // %gs:0x58 (64-bit).
7803 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
7804 ? Type::getInt8PtrTy(*DAG.getContext(),
7806 : Type::getInt32PtrTy(*DAG.getContext(),
7809 SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain,
7810 Subtarget->is64Bit()
7811 ? DAG.getIntPtrConstant(0x58)
7812 : DAG.getExternalSymbol("_tls_array",
7814 MachinePointerInfo(Ptr),
7815 false, false, false, 0);
7817 // Load the _tls_index variable
7818 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
7819 if (Subtarget->is64Bit())
7820 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
7821 IDX, MachinePointerInfo(), MVT::i32,
7824 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
7825 false, false, false, 0);
7827 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
7829 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
7831 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
7832 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
7833 false, false, false, 0);
7835 // Get the offset of start of .tls section
7836 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7837 GA->getValueType(0),
7838 GA->getOffset(), X86II::MO_SECREL);
7839 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
7841 // The address of the thread local variable is the add of the thread
7842 // pointer with the offset of the variable.
7843 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
7846 llvm_unreachable("TLS not implemented for this target.");
7850 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
7851 /// and take a 2 x i32 value to shift plus a shift amount.
7852 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
7853 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7854 EVT VT = Op.getValueType();
7855 unsigned VTBits = VT.getSizeInBits();
7856 DebugLoc dl = Op.getDebugLoc();
7857 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7858 SDValue ShOpLo = Op.getOperand(0);
7859 SDValue ShOpHi = Op.getOperand(1);
7860 SDValue ShAmt = Op.getOperand(2);
7861 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7862 DAG.getConstant(VTBits - 1, MVT::i8))
7863 : DAG.getConstant(0, VT);
7866 if (Op.getOpcode() == ISD::SHL_PARTS) {
7867 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7868 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7870 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7871 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7874 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7875 DAG.getConstant(VTBits, MVT::i8));
7876 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7877 AndNode, DAG.getConstant(0, MVT::i8));
7880 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7881 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7882 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7884 if (Op.getOpcode() == ISD::SHL_PARTS) {
7885 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7886 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7888 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7889 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7892 SDValue Ops[2] = { Lo, Hi };
7893 return DAG.getMergeValues(Ops, 2, dl);
7896 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7897 SelectionDAG &DAG) const {
7898 EVT SrcVT = Op.getOperand(0).getValueType();
7900 if (SrcVT.isVector())
7903 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7904 "Unknown SINT_TO_FP to lower!");
7906 // These are really Legal; return the operand so the caller accepts it as
7908 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7910 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7911 Subtarget->is64Bit()) {
7915 DebugLoc dl = Op.getDebugLoc();
7916 unsigned Size = SrcVT.getSizeInBits()/8;
7917 MachineFunction &MF = DAG.getMachineFunction();
7918 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7919 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7920 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7922 MachinePointerInfo::getFixedStack(SSFI),
7924 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7927 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7929 SelectionDAG &DAG) const {
7931 DebugLoc DL = Op.getDebugLoc();
7933 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7935 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7937 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7939 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7941 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7942 MachineMemOperand *MMO;
7944 int SSFI = FI->getIndex();
7946 DAG.getMachineFunction()
7947 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7948 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7950 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7951 StackSlot = StackSlot.getOperand(1);
7953 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7954 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7956 Tys, Ops, array_lengthof(Ops),
7960 Chain = Result.getValue(1);
7961 SDValue InFlag = Result.getValue(2);
7963 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7964 // shouldn't be necessary except that RFP cannot be live across
7965 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7966 MachineFunction &MF = DAG.getMachineFunction();
7967 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7968 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7969 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7970 Tys = DAG.getVTList(MVT::Other);
7972 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7974 MachineMemOperand *MMO =
7975 DAG.getMachineFunction()
7976 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7977 MachineMemOperand::MOStore, SSFISize, SSFISize);
7979 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7980 Ops, array_lengthof(Ops),
7981 Op.getValueType(), MMO);
7982 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7983 MachinePointerInfo::getFixedStack(SSFI),
7984 false, false, false, 0);
7990 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7991 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7992 SelectionDAG &DAG) const {
7993 // This algorithm is not obvious. Here it is what we're trying to output:
7996 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
7997 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
8001 pshufd $0x4e, %xmm0, %xmm1
8006 DebugLoc dl = Op.getDebugLoc();
8007 LLVMContext *Context = DAG.getContext();
8009 // Build some magic constants.
8010 const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
8011 Constant *C0 = ConstantDataVector::get(*Context, CV0);
8012 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
8014 SmallVector<Constant*,2> CV1;
8016 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
8018 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
8019 Constant *C1 = ConstantVector::get(CV1);
8020 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
8022 // Load the 64-bit value into an XMM register.
8023 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
8025 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
8026 MachinePointerInfo::getConstantPool(),
8027 false, false, false, 16);
8028 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
8029 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
8032 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
8033 MachinePointerInfo::getConstantPool(),
8034 false, false, false, 16);
8035 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
8036 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
8039 if (Subtarget->hasSSE3()) {
8040 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
8041 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
8043 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
8044 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
8046 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
8047 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
8051 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
8052 DAG.getIntPtrConstant(0));
8055 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
8056 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
8057 SelectionDAG &DAG) const {
8058 DebugLoc dl = Op.getDebugLoc();
8059 // FP constant to bias correct the final result.
8060 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
8063 // Load the 32-bit value into an XMM register.
8064 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
8067 // Zero out the upper parts of the register.
8068 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
8070 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8071 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
8072 DAG.getIntPtrConstant(0));
8074 // Or the load with the bias.
8075 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
8076 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8077 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8079 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8080 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8081 MVT::v2f64, Bias)));
8082 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8083 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
8084 DAG.getIntPtrConstant(0));
8086 // Subtract the bias.
8087 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
8089 // Handle final rounding.
8090 EVT DestVT = Op.getValueType();
8092 if (DestVT.bitsLT(MVT::f64))
8093 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
8094 DAG.getIntPtrConstant(0));
8095 if (DestVT.bitsGT(MVT::f64))
8096 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
8098 // Handle final rounding.
8102 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
8103 SelectionDAG &DAG) const {
8104 SDValue N0 = Op.getOperand(0);
8105 EVT SVT = N0.getValueType();
8106 DebugLoc dl = Op.getDebugLoc();
8108 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
8109 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
8110 "Custom UINT_TO_FP is not supported!");
8112 EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, SVT.getVectorNumElements());
8113 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
8114 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
8117 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
8118 SelectionDAG &DAG) const {
8119 SDValue N0 = Op.getOperand(0);
8120 DebugLoc dl = Op.getDebugLoc();
8122 if (Op.getValueType().isVector())
8123 return lowerUINT_TO_FP_vec(Op, DAG);
8125 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
8126 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
8127 // the optimization here.
8128 if (DAG.SignBitIsZero(N0))
8129 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
8131 EVT SrcVT = N0.getValueType();
8132 EVT DstVT = Op.getValueType();
8133 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
8134 return LowerUINT_TO_FP_i64(Op, DAG);
8135 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
8136 return LowerUINT_TO_FP_i32(Op, DAG);
8137 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
8140 // Make a 64-bit buffer, and use it to build an FILD.
8141 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
8142 if (SrcVT == MVT::i32) {
8143 SDValue WordOff = DAG.getConstant(4, getPointerTy());
8144 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
8145 getPointerTy(), StackSlot, WordOff);
8146 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8147 StackSlot, MachinePointerInfo(),
8149 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
8150 OffsetSlot, MachinePointerInfo(),
8152 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
8156 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
8157 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8158 StackSlot, MachinePointerInfo(),
8160 // For i64 source, we need to add the appropriate power of 2 if the input
8161 // was negative. This is the same as the optimization in
8162 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
8163 // we must be careful to do the computation in x87 extended precision, not
8164 // in SSE. (The generic code can't know it's OK to do this, or how to.)
8165 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
8166 MachineMemOperand *MMO =
8167 DAG.getMachineFunction()
8168 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8169 MachineMemOperand::MOLoad, 8, 8);
8171 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
8172 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
8173 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
8176 APInt FF(32, 0x5F800000ULL);
8178 // Check whether the sign bit is set.
8179 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
8180 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
8183 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
8184 SDValue FudgePtr = DAG.getConstantPool(
8185 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
8188 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
8189 SDValue Zero = DAG.getIntPtrConstant(0);
8190 SDValue Four = DAG.getIntPtrConstant(4);
8191 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
8193 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
8195 // Load the value out, extending it from f32 to f80.
8196 // FIXME: Avoid the extend by constructing the right constant pool?
8197 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
8198 FudgePtr, MachinePointerInfo::getConstantPool(),
8199 MVT::f32, false, false, 4);
8200 // Extend everything to 80 bits to force it to be done on x87.
8201 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
8202 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
8205 std::pair<SDValue,SDValue> X86TargetLowering::
8206 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned, bool IsReplace) const {
8207 DebugLoc DL = Op.getDebugLoc();
8209 EVT DstTy = Op.getValueType();
8211 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
8212 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
8216 assert(DstTy.getSimpleVT() <= MVT::i64 &&
8217 DstTy.getSimpleVT() >= MVT::i16 &&
8218 "Unknown FP_TO_INT to lower!");
8220 // These are really Legal.
8221 if (DstTy == MVT::i32 &&
8222 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8223 return std::make_pair(SDValue(), SDValue());
8224 if (Subtarget->is64Bit() &&
8225 DstTy == MVT::i64 &&
8226 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8227 return std::make_pair(SDValue(), SDValue());
8229 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
8230 // stack slot, or into the FTOL runtime function.
8231 MachineFunction &MF = DAG.getMachineFunction();
8232 unsigned MemSize = DstTy.getSizeInBits()/8;
8233 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8234 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8237 if (!IsSigned && isIntegerTypeFTOL(DstTy))
8238 Opc = X86ISD::WIN_FTOL;
8240 switch (DstTy.getSimpleVT().SimpleTy) {
8241 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
8242 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
8243 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
8244 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
8247 SDValue Chain = DAG.getEntryNode();
8248 SDValue Value = Op.getOperand(0);
8249 EVT TheVT = Op.getOperand(0).getValueType();
8250 // FIXME This causes a redundant load/store if the SSE-class value is already
8251 // in memory, such as if it is on the callstack.
8252 if (isScalarFPTypeInSSEReg(TheVT)) {
8253 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
8254 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
8255 MachinePointerInfo::getFixedStack(SSFI),
8257 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
8259 Chain, StackSlot, DAG.getValueType(TheVT)
8262 MachineMemOperand *MMO =
8263 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8264 MachineMemOperand::MOLoad, MemSize, MemSize);
8265 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
8267 Chain = Value.getValue(1);
8268 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8269 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8272 MachineMemOperand *MMO =
8273 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8274 MachineMemOperand::MOStore, MemSize, MemSize);
8276 if (Opc != X86ISD::WIN_FTOL) {
8277 // Build the FP_TO_INT*_IN_MEM
8278 SDValue Ops[] = { Chain, Value, StackSlot };
8279 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
8280 Ops, 3, DstTy, MMO);
8281 return std::make_pair(FIST, StackSlot);
8283 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
8284 DAG.getVTList(MVT::Other, MVT::Glue),
8286 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
8287 MVT::i32, ftol.getValue(1));
8288 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
8289 MVT::i32, eax.getValue(2));
8290 SDValue Ops[] = { eax, edx };
8291 SDValue pair = IsReplace
8292 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops, 2)
8293 : DAG.getMergeValues(Ops, 2, DL);
8294 return std::make_pair(pair, SDValue());
8298 SDValue X86TargetLowering::lowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const {
8299 DebugLoc DL = Op.getDebugLoc();
8300 EVT VT = Op.getValueType();
8301 SDValue In = Op.getOperand(0);
8302 EVT SVT = In.getValueType();
8304 if (!VT.is256BitVector() || !SVT.is128BitVector() ||
8305 VT.getVectorNumElements() != SVT.getVectorNumElements())
8308 assert(Subtarget->hasAVX() && "256-bit vector is observed without AVX!");
8310 // AVX2 has better support of integer extending.
8311 if (Subtarget->hasAVX2())
8312 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
8314 SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
8315 static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
8316 SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
8317 DAG.getVectorShuffle(MVT::v8i16, DL, In, DAG.getUNDEF(MVT::v8i16), &Mask[0]));
8319 return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
8322 SDValue X86TargetLowering::lowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
8323 DebugLoc DL = Op.getDebugLoc();
8324 EVT VT = Op.getValueType();
8325 EVT SVT = Op.getOperand(0).getValueType();
8327 if (!VT.is128BitVector() || !SVT.is256BitVector() ||
8328 VT.getVectorNumElements() != SVT.getVectorNumElements())
8331 assert(Subtarget->hasAVX() && "256-bit vector is observed without AVX!");
8333 unsigned NumElems = VT.getVectorNumElements();
8334 EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
8337 SDValue In = Op.getOperand(0);
8338 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
8339 // Prepare truncation shuffle mask
8340 for (unsigned i = 0; i != NumElems; ++i)
8342 SDValue V = DAG.getVectorShuffle(NVT, DL,
8343 DAG.getNode(ISD::BITCAST, DL, NVT, In),
8344 DAG.getUNDEF(NVT), &MaskVec[0]);
8345 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
8346 DAG.getIntPtrConstant(0));
8349 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
8350 SelectionDAG &DAG) const {
8351 if (Op.getValueType().isVector()) {
8352 if (Op.getValueType() == MVT::v8i16)
8353 return DAG.getNode(ISD::TRUNCATE, Op.getDebugLoc(), Op.getValueType(),
8354 DAG.getNode(ISD::FP_TO_SINT, Op.getDebugLoc(),
8355 MVT::v8i32, Op.getOperand(0)));
8359 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8360 /*IsSigned=*/ true, /*IsReplace=*/ false);
8361 SDValue FIST = Vals.first, StackSlot = Vals.second;
8362 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
8363 if (FIST.getNode() == 0) return Op;
8365 if (StackSlot.getNode())
8367 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
8368 FIST, StackSlot, MachinePointerInfo(),
8369 false, false, false, 0);
8371 // The node is the result.
8375 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
8376 SelectionDAG &DAG) const {
8377 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8378 /*IsSigned=*/ false, /*IsReplace=*/ false);
8379 SDValue FIST = Vals.first, StackSlot = Vals.second;
8380 assert(FIST.getNode() && "Unexpected failure");
8382 if (StackSlot.getNode())
8384 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
8385 FIST, StackSlot, MachinePointerInfo(),
8386 false, false, false, 0);
8388 // The node is the result.
8392 SDValue X86TargetLowering::lowerFP_EXTEND(SDValue Op,
8393 SelectionDAG &DAG) const {
8394 DebugLoc DL = Op.getDebugLoc();
8395 EVT VT = Op.getValueType();
8396 SDValue In = Op.getOperand(0);
8397 EVT SVT = In.getValueType();
8399 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
8401 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
8402 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
8403 In, DAG.getUNDEF(SVT)));
8406 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
8407 LLVMContext *Context = DAG.getContext();
8408 DebugLoc dl = Op.getDebugLoc();
8409 EVT VT = Op.getValueType();
8411 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
8412 if (VT.isVector()) {
8413 EltVT = VT.getVectorElementType();
8414 NumElts = VT.getVectorNumElements();
8417 if (EltVT == MVT::f64)
8418 C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
8420 C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
8421 C = ConstantVector::getSplat(NumElts, C);
8422 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
8423 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8424 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8425 MachinePointerInfo::getConstantPool(),
8426 false, false, false, Alignment);
8427 if (VT.isVector()) {
8428 MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8429 return DAG.getNode(ISD::BITCAST, dl, VT,
8430 DAG.getNode(ISD::AND, dl, ANDVT,
8431 DAG.getNode(ISD::BITCAST, dl, ANDVT,
8433 DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
8435 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
8438 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
8439 LLVMContext *Context = DAG.getContext();
8440 DebugLoc dl = Op.getDebugLoc();
8441 EVT VT = Op.getValueType();
8443 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
8444 if (VT.isVector()) {
8445 EltVT = VT.getVectorElementType();
8446 NumElts = VT.getVectorNumElements();
8449 if (EltVT == MVT::f64)
8450 C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
8452 C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
8453 C = ConstantVector::getSplat(NumElts, C);
8454 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
8455 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8456 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8457 MachinePointerInfo::getConstantPool(),
8458 false, false, false, Alignment);
8459 if (VT.isVector()) {
8460 MVT XORVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8461 return DAG.getNode(ISD::BITCAST, dl, VT,
8462 DAG.getNode(ISD::XOR, dl, XORVT,
8463 DAG.getNode(ISD::BITCAST, dl, XORVT,
8465 DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
8468 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
8471 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
8472 LLVMContext *Context = DAG.getContext();
8473 SDValue Op0 = Op.getOperand(0);
8474 SDValue Op1 = Op.getOperand(1);
8475 DebugLoc dl = Op.getDebugLoc();
8476 EVT VT = Op.getValueType();
8477 EVT SrcVT = Op1.getValueType();
8479 // If second operand is smaller, extend it first.
8480 if (SrcVT.bitsLT(VT)) {
8481 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
8484 // And if it is bigger, shrink it first.
8485 if (SrcVT.bitsGT(VT)) {
8486 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
8490 // At this point the operands and the result should have the same
8491 // type, and that won't be f80 since that is not custom lowered.
8493 // First get the sign bit of second operand.
8494 SmallVector<Constant*,4> CV;
8495 if (SrcVT == MVT::f64) {
8496 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
8497 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8499 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
8500 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8501 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8502 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8504 Constant *C = ConstantVector::get(CV);
8505 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8506 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
8507 MachinePointerInfo::getConstantPool(),
8508 false, false, false, 16);
8509 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
8511 // Shift sign bit right or left if the two operands have different types.
8512 if (SrcVT.bitsGT(VT)) {
8513 // Op0 is MVT::f32, Op1 is MVT::f64.
8514 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
8515 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
8516 DAG.getConstant(32, MVT::i32));
8517 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
8518 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
8519 DAG.getIntPtrConstant(0));
8522 // Clear first operand sign bit.
8524 if (VT == MVT::f64) {
8525 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
8526 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8528 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
8529 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8530 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8531 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8533 C = ConstantVector::get(CV);
8534 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8535 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8536 MachinePointerInfo::getConstantPool(),
8537 false, false, false, 16);
8538 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
8540 // Or the value with the sign bit.
8541 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
8544 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
8545 SDValue N0 = Op.getOperand(0);
8546 DebugLoc dl = Op.getDebugLoc();
8547 EVT VT = Op.getValueType();
8549 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
8550 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
8551 DAG.getConstant(1, VT));
8552 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
8555 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
8557 SDValue X86TargetLowering::LowerVectorAllZeroTest(SDValue Op, SelectionDAG &DAG) const {
8558 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
8560 if (!Subtarget->hasSSE41())
8563 if (!Op->hasOneUse())
8566 SDNode *N = Op.getNode();
8567 DebugLoc DL = N->getDebugLoc();
8569 SmallVector<SDValue, 8> Opnds;
8570 DenseMap<SDValue, unsigned> VecInMap;
8571 EVT VT = MVT::Other;
8573 // Recognize a special case where a vector is casted into wide integer to
8575 Opnds.push_back(N->getOperand(0));
8576 Opnds.push_back(N->getOperand(1));
8578 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
8579 SmallVector<SDValue, 8>::const_iterator I = Opnds.begin() + Slot;
8580 // BFS traverse all OR'd operands.
8581 if (I->getOpcode() == ISD::OR) {
8582 Opnds.push_back(I->getOperand(0));
8583 Opnds.push_back(I->getOperand(1));
8584 // Re-evaluate the number of nodes to be traversed.
8585 e += 2; // 2 more nodes (LHS and RHS) are pushed.
8589 // Quit if a non-EXTRACT_VECTOR_ELT
8590 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8593 // Quit if without a constant index.
8594 SDValue Idx = I->getOperand(1);
8595 if (!isa<ConstantSDNode>(Idx))
8598 SDValue ExtractedFromVec = I->getOperand(0);
8599 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
8600 if (M == VecInMap.end()) {
8601 VT = ExtractedFromVec.getValueType();
8602 // Quit if not 128/256-bit vector.
8603 if (!VT.is128BitVector() && !VT.is256BitVector())
8605 // Quit if not the same type.
8606 if (VecInMap.begin() != VecInMap.end() &&
8607 VT != VecInMap.begin()->first.getValueType())
8609 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
8611 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
8614 assert((VT.is128BitVector() || VT.is256BitVector()) &&
8615 "Not extracted from 128-/256-bit vector.");
8617 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
8618 SmallVector<SDValue, 8> VecIns;
8620 for (DenseMap<SDValue, unsigned>::const_iterator
8621 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
8622 // Quit if not all elements are used.
8623 if (I->second != FullMask)
8625 VecIns.push_back(I->first);
8628 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8630 // Cast all vectors into TestVT for PTEST.
8631 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
8632 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
8634 // If more than one full vectors are evaluated, OR them first before PTEST.
8635 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
8636 // Each iteration will OR 2 nodes and append the result until there is only
8637 // 1 node left, i.e. the final OR'd value of all vectors.
8638 SDValue LHS = VecIns[Slot];
8639 SDValue RHS = VecIns[Slot + 1];
8640 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
8643 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
8644 VecIns.back(), VecIns.back());
8647 /// Emit nodes that will be selected as "test Op0,Op0", or something
8649 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
8650 SelectionDAG &DAG) const {
8651 DebugLoc dl = Op.getDebugLoc();
8653 // CF and OF aren't always set the way we want. Determine which
8654 // of these we need.
8655 bool NeedCF = false;
8656 bool NeedOF = false;
8659 case X86::COND_A: case X86::COND_AE:
8660 case X86::COND_B: case X86::COND_BE:
8663 case X86::COND_G: case X86::COND_GE:
8664 case X86::COND_L: case X86::COND_LE:
8665 case X86::COND_O: case X86::COND_NO:
8670 // See if we can use the EFLAGS value from the operand instead of
8671 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
8672 // we prove that the arithmetic won't overflow, we can't use OF or CF.
8673 if (Op.getResNo() != 0 || NeedOF || NeedCF)
8674 // Emit a CMP with 0, which is the TEST pattern.
8675 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8676 DAG.getConstant(0, Op.getValueType()));
8678 unsigned Opcode = 0;
8679 unsigned NumOperands = 0;
8681 // Truncate operations may prevent the merge of the SETCC instruction
8682 // and the arithmetic intruction before it. Attempt to truncate the operands
8683 // of the arithmetic instruction and use a reduced bit-width instruction.
8684 bool NeedTruncation = false;
8685 SDValue ArithOp = Op;
8686 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
8687 SDValue Arith = Op->getOperand(0);
8688 // Both the trunc and the arithmetic op need to have one user each.
8689 if (Arith->hasOneUse())
8690 switch (Arith.getOpcode()) {
8697 NeedTruncation = true;
8703 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
8704 // which may be the result of a CAST. We use the variable 'Op', which is the
8705 // non-casted variable when we check for possible users.
8706 switch (ArithOp.getOpcode()) {
8708 // Due to an isel shortcoming, be conservative if this add is likely to be
8709 // selected as part of a load-modify-store instruction. When the root node
8710 // in a match is a store, isel doesn't know how to remap non-chain non-flag
8711 // uses of other nodes in the match, such as the ADD in this case. This
8712 // leads to the ADD being left around and reselected, with the result being
8713 // two adds in the output. Alas, even if none our users are stores, that
8714 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
8715 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
8716 // climbing the DAG back to the root, and it doesn't seem to be worth the
8718 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8719 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8720 if (UI->getOpcode() != ISD::CopyToReg &&
8721 UI->getOpcode() != ISD::SETCC &&
8722 UI->getOpcode() != ISD::STORE)
8725 if (ConstantSDNode *C =
8726 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
8727 // An add of one will be selected as an INC.
8728 if (C->getAPIntValue() == 1) {
8729 Opcode = X86ISD::INC;
8734 // An add of negative one (subtract of one) will be selected as a DEC.
8735 if (C->getAPIntValue().isAllOnesValue()) {
8736 Opcode = X86ISD::DEC;
8742 // Otherwise use a regular EFLAGS-setting add.
8743 Opcode = X86ISD::ADD;
8747 // If the primary and result isn't used, don't bother using X86ISD::AND,
8748 // because a TEST instruction will be better.
8749 bool NonFlagUse = false;
8750 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8751 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
8753 unsigned UOpNo = UI.getOperandNo();
8754 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
8755 // Look pass truncate.
8756 UOpNo = User->use_begin().getOperandNo();
8757 User = *User->use_begin();
8760 if (User->getOpcode() != ISD::BRCOND &&
8761 User->getOpcode() != ISD::SETCC &&
8762 !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) {
8775 // Due to the ISEL shortcoming noted above, be conservative if this op is
8776 // likely to be selected as part of a load-modify-store instruction.
8777 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8778 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8779 if (UI->getOpcode() == ISD::STORE)
8782 // Otherwise use a regular EFLAGS-setting instruction.
8783 switch (ArithOp.getOpcode()) {
8784 default: llvm_unreachable("unexpected operator!");
8785 case ISD::SUB: Opcode = X86ISD::SUB; break;
8786 case ISD::XOR: Opcode = X86ISD::XOR; break;
8787 case ISD::AND: Opcode = X86ISD::AND; break;
8789 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
8790 SDValue EFLAGS = LowerVectorAllZeroTest(Op, DAG);
8791 if (EFLAGS.getNode())
8794 Opcode = X86ISD::OR;
8808 return SDValue(Op.getNode(), 1);
8814 // If we found that truncation is beneficial, perform the truncation and
8816 if (NeedTruncation) {
8817 EVT VT = Op.getValueType();
8818 SDValue WideVal = Op->getOperand(0);
8819 EVT WideVT = WideVal.getValueType();
8820 unsigned ConvertedOp = 0;
8821 // Use a target machine opcode to prevent further DAGCombine
8822 // optimizations that may separate the arithmetic operations
8823 // from the setcc node.
8824 switch (WideVal.getOpcode()) {
8826 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
8827 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
8828 case ISD::AND: ConvertedOp = X86ISD::AND; break;
8829 case ISD::OR: ConvertedOp = X86ISD::OR; break;
8830 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
8834 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8835 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
8836 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
8837 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
8838 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
8844 // Emit a CMP with 0, which is the TEST pattern.
8845 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8846 DAG.getConstant(0, Op.getValueType()));
8848 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
8849 SmallVector<SDValue, 4> Ops;
8850 for (unsigned i = 0; i != NumOperands; ++i)
8851 Ops.push_back(Op.getOperand(i));
8853 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
8854 DAG.ReplaceAllUsesWith(Op, New);
8855 return SDValue(New.getNode(), 1);
8858 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
8860 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
8861 SelectionDAG &DAG) const {
8862 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
8863 if (C->getAPIntValue() == 0)
8864 return EmitTest(Op0, X86CC, DAG);
8866 DebugLoc dl = Op0.getDebugLoc();
8867 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
8868 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
8869 // Use SUB instead of CMP to enable CSE between SUB and CMP.
8870 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
8871 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
8873 return SDValue(Sub.getNode(), 1);
8875 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
8878 /// Convert a comparison if required by the subtarget.
8879 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
8880 SelectionDAG &DAG) const {
8881 // If the subtarget does not support the FUCOMI instruction, floating-point
8882 // comparisons have to be converted.
8883 if (Subtarget->hasCMov() ||
8884 Cmp.getOpcode() != X86ISD::CMP ||
8885 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
8886 !Cmp.getOperand(1).getValueType().isFloatingPoint())
8889 // The instruction selector will select an FUCOM instruction instead of
8890 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
8891 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
8892 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
8893 DebugLoc dl = Cmp.getDebugLoc();
8894 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
8895 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
8896 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
8897 DAG.getConstant(8, MVT::i8));
8898 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
8899 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
8902 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
8903 /// if it's possible.
8904 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
8905 DebugLoc dl, SelectionDAG &DAG) const {
8906 SDValue Op0 = And.getOperand(0);
8907 SDValue Op1 = And.getOperand(1);
8908 if (Op0.getOpcode() == ISD::TRUNCATE)
8909 Op0 = Op0.getOperand(0);
8910 if (Op1.getOpcode() == ISD::TRUNCATE)
8911 Op1 = Op1.getOperand(0);
8914 if (Op1.getOpcode() == ISD::SHL)
8915 std::swap(Op0, Op1);
8916 if (Op0.getOpcode() == ISD::SHL) {
8917 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
8918 if (And00C->getZExtValue() == 1) {
8919 // If we looked past a truncate, check that it's only truncating away
8921 unsigned BitWidth = Op0.getValueSizeInBits();
8922 unsigned AndBitWidth = And.getValueSizeInBits();
8923 if (BitWidth > AndBitWidth) {
8925 DAG.ComputeMaskedBits(Op0, Zeros, Ones);
8926 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
8930 RHS = Op0.getOperand(1);
8932 } else if (Op1.getOpcode() == ISD::Constant) {
8933 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
8934 uint64_t AndRHSVal = AndRHS->getZExtValue();
8935 SDValue AndLHS = Op0;
8937 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
8938 LHS = AndLHS.getOperand(0);
8939 RHS = AndLHS.getOperand(1);
8942 // Use BT if the immediate can't be encoded in a TEST instruction.
8943 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
8945 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
8949 if (LHS.getNode()) {
8950 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
8951 // instruction. Since the shift amount is in-range-or-undefined, we know
8952 // that doing a bittest on the i32 value is ok. We extend to i32 because
8953 // the encoding for the i16 version is larger than the i32 version.
8954 // Also promote i16 to i32 for performance / code size reason.
8955 if (LHS.getValueType() == MVT::i8 ||
8956 LHS.getValueType() == MVT::i16)
8957 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
8959 // If the operand types disagree, extend the shift amount to match. Since
8960 // BT ignores high bits (like shifts) we can use anyextend.
8961 if (LHS.getValueType() != RHS.getValueType())
8962 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
8964 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
8965 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
8966 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8967 DAG.getConstant(Cond, MVT::i8), BT);
8973 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
8975 if (Op.getValueType().isVector()) return LowerVSETCC(Op, DAG);
8977 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
8978 SDValue Op0 = Op.getOperand(0);
8979 SDValue Op1 = Op.getOperand(1);
8980 DebugLoc dl = Op.getDebugLoc();
8981 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8983 // Optimize to BT if possible.
8984 // Lower (X & (1 << N)) == 0 to BT(X, N).
8985 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
8986 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
8987 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
8988 Op1.getOpcode() == ISD::Constant &&
8989 cast<ConstantSDNode>(Op1)->isNullValue() &&
8990 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8991 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
8992 if (NewSetCC.getNode())
8996 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
8998 if (Op1.getOpcode() == ISD::Constant &&
8999 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
9000 cast<ConstantSDNode>(Op1)->isNullValue()) &&
9001 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
9003 // If the input is a setcc, then reuse the input setcc or use a new one with
9004 // the inverted condition.
9005 if (Op0.getOpcode() == X86ISD::SETCC) {
9006 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
9007 bool Invert = (CC == ISD::SETNE) ^
9008 cast<ConstantSDNode>(Op1)->isNullValue();
9009 if (!Invert) return Op0;
9011 CCode = X86::GetOppositeBranchCondition(CCode);
9012 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9013 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
9017 bool isFP = Op1.getValueType().isFloatingPoint();
9018 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
9019 if (X86CC == X86::COND_INVALID)
9022 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
9023 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
9024 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9025 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
9028 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
9029 // ones, and then concatenate the result back.
9030 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
9031 EVT VT = Op.getValueType();
9033 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
9034 "Unsupported value type for operation");
9036 unsigned NumElems = VT.getVectorNumElements();
9037 DebugLoc dl = Op.getDebugLoc();
9038 SDValue CC = Op.getOperand(2);
9040 // Extract the LHS vectors
9041 SDValue LHS = Op.getOperand(0);
9042 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
9043 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
9045 // Extract the RHS vectors
9046 SDValue RHS = Op.getOperand(1);
9047 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
9048 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
9050 // Issue the operation on the smaller types and concatenate the result back
9051 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9052 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9053 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9054 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
9055 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
9059 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
9061 SDValue Op0 = Op.getOperand(0);
9062 SDValue Op1 = Op.getOperand(1);
9063 SDValue CC = Op.getOperand(2);
9064 EVT VT = Op.getValueType();
9065 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
9066 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
9067 DebugLoc dl = Op.getDebugLoc();
9071 EVT EltVT = Op0.getValueType().getVectorElementType();
9072 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
9078 // SSE Condition code mapping:
9087 switch (SetCCOpcode) {
9088 default: llvm_unreachable("Unexpected SETCC condition");
9090 case ISD::SETEQ: SSECC = 0; break;
9092 case ISD::SETGT: Swap = true; // Fallthrough
9094 case ISD::SETOLT: SSECC = 1; break;
9096 case ISD::SETGE: Swap = true; // Fallthrough
9098 case ISD::SETOLE: SSECC = 2; break;
9099 case ISD::SETUO: SSECC = 3; break;
9101 case ISD::SETNE: SSECC = 4; break;
9102 case ISD::SETULE: Swap = true; // Fallthrough
9103 case ISD::SETUGE: SSECC = 5; break;
9104 case ISD::SETULT: Swap = true; // Fallthrough
9105 case ISD::SETUGT: SSECC = 6; break;
9106 case ISD::SETO: SSECC = 7; break;
9108 case ISD::SETONE: SSECC = 8; break;
9111 std::swap(Op0, Op1);
9113 // In the two special cases we can't handle, emit two comparisons.
9116 unsigned CombineOpc;
9117 if (SetCCOpcode == ISD::SETUEQ) {
9118 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
9120 assert(SetCCOpcode == ISD::SETONE);
9121 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
9124 SDValue Cmp0 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9125 DAG.getConstant(CC0, MVT::i8));
9126 SDValue Cmp1 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9127 DAG.getConstant(CC1, MVT::i8));
9128 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
9130 // Handle all other FP comparisons here.
9131 return DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9132 DAG.getConstant(SSECC, MVT::i8));
9135 // Break 256-bit integer vector compare into smaller ones.
9136 if (VT.is256BitVector() && !Subtarget->hasAVX2())
9137 return Lower256IntVSETCC(Op, DAG);
9139 // We are handling one of the integer comparisons here. Since SSE only has
9140 // GT and EQ comparisons for integer, swapping operands and multiple
9141 // operations may be required for some comparisons.
9143 bool Swap = false, Invert = false, FlipSigns = false;
9145 switch (SetCCOpcode) {
9146 default: llvm_unreachable("Unexpected SETCC condition");
9147 case ISD::SETNE: Invert = true;
9148 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
9149 case ISD::SETLT: Swap = true;
9150 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
9151 case ISD::SETGE: Swap = true;
9152 case ISD::SETLE: Opc = X86ISD::PCMPGT; Invert = true; break;
9153 case ISD::SETULT: Swap = true;
9154 case ISD::SETUGT: Opc = X86ISD::PCMPGT; FlipSigns = true; break;
9155 case ISD::SETUGE: Swap = true;
9156 case ISD::SETULE: Opc = X86ISD::PCMPGT; FlipSigns = true; Invert = true; break;
9159 std::swap(Op0, Op1);
9161 // Check that the operation in question is available (most are plain SSE2,
9162 // but PCMPGTQ and PCMPEQQ have different requirements).
9163 if (VT == MVT::v2i64) {
9164 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42())
9166 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41())
9170 // Since SSE has no unsigned integer comparisons, we need to flip the sign
9171 // bits of the inputs before performing those operations.
9173 EVT EltVT = VT.getVectorElementType();
9174 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
9176 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
9177 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
9179 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
9180 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
9183 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
9185 // If the logical-not of the result is required, perform that now.
9187 Result = DAG.getNOT(dl, Result, VT);
9192 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
9193 static bool isX86LogicalCmp(SDValue Op) {
9194 unsigned Opc = Op.getNode()->getOpcode();
9195 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
9196 Opc == X86ISD::SAHF)
9198 if (Op.getResNo() == 1 &&
9199 (Opc == X86ISD::ADD ||
9200 Opc == X86ISD::SUB ||
9201 Opc == X86ISD::ADC ||
9202 Opc == X86ISD::SBB ||
9203 Opc == X86ISD::SMUL ||
9204 Opc == X86ISD::UMUL ||
9205 Opc == X86ISD::INC ||
9206 Opc == X86ISD::DEC ||
9207 Opc == X86ISD::OR ||
9208 Opc == X86ISD::XOR ||
9209 Opc == X86ISD::AND))
9212 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
9218 static bool isZero(SDValue V) {
9219 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9220 return C && C->isNullValue();
9223 static bool isAllOnes(SDValue V) {
9224 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9225 return C && C->isAllOnesValue();
9228 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
9229 if (V.getOpcode() != ISD::TRUNCATE)
9232 SDValue VOp0 = V.getOperand(0);
9233 unsigned InBits = VOp0.getValueSizeInBits();
9234 unsigned Bits = V.getValueSizeInBits();
9235 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
9238 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
9239 bool addTest = true;
9240 SDValue Cond = Op.getOperand(0);
9241 SDValue Op1 = Op.getOperand(1);
9242 SDValue Op2 = Op.getOperand(2);
9243 DebugLoc DL = Op.getDebugLoc();
9246 if (Cond.getOpcode() == ISD::SETCC) {
9247 SDValue NewCond = LowerSETCC(Cond, DAG);
9248 if (NewCond.getNode())
9252 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
9253 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
9254 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
9255 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
9256 if (Cond.getOpcode() == X86ISD::SETCC &&
9257 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
9258 isZero(Cond.getOperand(1).getOperand(1))) {
9259 SDValue Cmp = Cond.getOperand(1);
9261 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
9263 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
9264 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
9265 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
9267 SDValue CmpOp0 = Cmp.getOperand(0);
9268 // Apply further optimizations for special cases
9269 // (select (x != 0), -1, 0) -> neg & sbb
9270 // (select (x == 0), 0, -1) -> neg & sbb
9271 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
9272 if (YC->isNullValue() &&
9273 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
9274 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
9275 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
9276 DAG.getConstant(0, CmpOp0.getValueType()),
9278 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9279 DAG.getConstant(X86::COND_B, MVT::i8),
9280 SDValue(Neg.getNode(), 1));
9284 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
9285 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
9286 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9288 SDValue Res = // Res = 0 or -1.
9289 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9290 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
9292 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
9293 Res = DAG.getNOT(DL, Res, Res.getValueType());
9295 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
9296 if (N2C == 0 || !N2C->isNullValue())
9297 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
9302 // Look past (and (setcc_carry (cmp ...)), 1).
9303 if (Cond.getOpcode() == ISD::AND &&
9304 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
9305 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
9306 if (C && C->getAPIntValue() == 1)
9307 Cond = Cond.getOperand(0);
9310 // If condition flag is set by a X86ISD::CMP, then use it as the condition
9311 // setting operand in place of the X86ISD::SETCC.
9312 unsigned CondOpcode = Cond.getOpcode();
9313 if (CondOpcode == X86ISD::SETCC ||
9314 CondOpcode == X86ISD::SETCC_CARRY) {
9315 CC = Cond.getOperand(0);
9317 SDValue Cmp = Cond.getOperand(1);
9318 unsigned Opc = Cmp.getOpcode();
9319 EVT VT = Op.getValueType();
9321 bool IllegalFPCMov = false;
9322 if (VT.isFloatingPoint() && !VT.isVector() &&
9323 !isScalarFPTypeInSSEReg(VT)) // FPStack?
9324 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
9326 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
9327 Opc == X86ISD::BT) { // FIXME
9331 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
9332 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
9333 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
9334 Cond.getOperand(0).getValueType() != MVT::i8)) {
9335 SDValue LHS = Cond.getOperand(0);
9336 SDValue RHS = Cond.getOperand(1);
9340 switch (CondOpcode) {
9341 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
9342 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
9343 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
9344 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
9345 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
9346 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
9347 default: llvm_unreachable("unexpected overflowing operator");
9349 if (CondOpcode == ISD::UMULO)
9350 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
9353 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
9355 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
9357 if (CondOpcode == ISD::UMULO)
9358 Cond = X86Op.getValue(2);
9360 Cond = X86Op.getValue(1);
9362 CC = DAG.getConstant(X86Cond, MVT::i8);
9367 // Look pass the truncate if the high bits are known zero.
9368 if (isTruncWithZeroHighBitsInput(Cond, DAG))
9369 Cond = Cond.getOperand(0);
9371 // We know the result of AND is compared against zero. Try to match
9373 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
9374 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
9375 if (NewSetCC.getNode()) {
9376 CC = NewSetCC.getOperand(0);
9377 Cond = NewSetCC.getOperand(1);
9384 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9385 Cond = EmitTest(Cond, X86::COND_NE, DAG);
9388 // a < b ? -1 : 0 -> RES = ~setcc_carry
9389 // a < b ? 0 : -1 -> RES = setcc_carry
9390 // a >= b ? -1 : 0 -> RES = setcc_carry
9391 // a >= b ? 0 : -1 -> RES = ~setcc_carry
9392 if (Cond.getOpcode() == X86ISD::SUB) {
9393 Cond = ConvertCmpIfNecessary(Cond, DAG);
9394 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
9396 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
9397 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
9398 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9399 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
9400 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
9401 return DAG.getNOT(DL, Res, Res.getValueType());
9406 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
9407 // widen the cmov and push the truncate through. This avoids introducing a new
9408 // branch during isel and doesn't add any extensions.
9409 if (Op.getValueType() == MVT::i8 &&
9410 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
9411 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
9412 if (T1.getValueType() == T2.getValueType() &&
9413 // Blacklist CopyFromReg to avoid partial register stalls.
9414 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
9415 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
9416 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
9417 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
9421 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
9422 // condition is true.
9423 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
9424 SDValue Ops[] = { Op2, Op1, CC, Cond };
9425 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
9428 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
9429 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
9430 // from the AND / OR.
9431 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
9432 Opc = Op.getOpcode();
9433 if (Opc != ISD::OR && Opc != ISD::AND)
9435 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
9436 Op.getOperand(0).hasOneUse() &&
9437 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
9438 Op.getOperand(1).hasOneUse());
9441 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
9442 // 1 and that the SETCC node has a single use.
9443 static bool isXor1OfSetCC(SDValue Op) {
9444 if (Op.getOpcode() != ISD::XOR)
9446 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
9447 if (N1C && N1C->getAPIntValue() == 1) {
9448 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
9449 Op.getOperand(0).hasOneUse();
9454 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
9455 bool addTest = true;
9456 SDValue Chain = Op.getOperand(0);
9457 SDValue Cond = Op.getOperand(1);
9458 SDValue Dest = Op.getOperand(2);
9459 DebugLoc dl = Op.getDebugLoc();
9461 bool Inverted = false;
9463 if (Cond.getOpcode() == ISD::SETCC) {
9464 // Check for setcc([su]{add,sub,mul}o == 0).
9465 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
9466 isa<ConstantSDNode>(Cond.getOperand(1)) &&
9467 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
9468 Cond.getOperand(0).getResNo() == 1 &&
9469 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
9470 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
9471 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
9472 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
9473 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
9474 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
9476 Cond = Cond.getOperand(0);
9478 SDValue NewCond = LowerSETCC(Cond, DAG);
9479 if (NewCond.getNode())
9484 // FIXME: LowerXALUO doesn't handle these!!
9485 else if (Cond.getOpcode() == X86ISD::ADD ||
9486 Cond.getOpcode() == X86ISD::SUB ||
9487 Cond.getOpcode() == X86ISD::SMUL ||
9488 Cond.getOpcode() == X86ISD::UMUL)
9489 Cond = LowerXALUO(Cond, DAG);
9492 // Look pass (and (setcc_carry (cmp ...)), 1).
9493 if (Cond.getOpcode() == ISD::AND &&
9494 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
9495 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
9496 if (C && C->getAPIntValue() == 1)
9497 Cond = Cond.getOperand(0);
9500 // If condition flag is set by a X86ISD::CMP, then use it as the condition
9501 // setting operand in place of the X86ISD::SETCC.
9502 unsigned CondOpcode = Cond.getOpcode();
9503 if (CondOpcode == X86ISD::SETCC ||
9504 CondOpcode == X86ISD::SETCC_CARRY) {
9505 CC = Cond.getOperand(0);
9507 SDValue Cmp = Cond.getOperand(1);
9508 unsigned Opc = Cmp.getOpcode();
9509 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
9510 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
9514 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
9518 // These can only come from an arithmetic instruction with overflow,
9519 // e.g. SADDO, UADDO.
9520 Cond = Cond.getNode()->getOperand(1);
9526 CondOpcode = Cond.getOpcode();
9527 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
9528 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
9529 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
9530 Cond.getOperand(0).getValueType() != MVT::i8)) {
9531 SDValue LHS = Cond.getOperand(0);
9532 SDValue RHS = Cond.getOperand(1);
9536 switch (CondOpcode) {
9537 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
9538 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
9539 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
9540 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
9541 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
9542 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
9543 default: llvm_unreachable("unexpected overflowing operator");
9546 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
9547 if (CondOpcode == ISD::UMULO)
9548 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
9551 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
9553 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
9555 if (CondOpcode == ISD::UMULO)
9556 Cond = X86Op.getValue(2);
9558 Cond = X86Op.getValue(1);
9560 CC = DAG.getConstant(X86Cond, MVT::i8);
9564 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
9565 SDValue Cmp = Cond.getOperand(0).getOperand(1);
9566 if (CondOpc == ISD::OR) {
9567 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
9568 // two branches instead of an explicit OR instruction with a
9570 if (Cmp == Cond.getOperand(1).getOperand(1) &&
9571 isX86LogicalCmp(Cmp)) {
9572 CC = Cond.getOperand(0).getOperand(0);
9573 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9574 Chain, Dest, CC, Cmp);
9575 CC = Cond.getOperand(1).getOperand(0);
9579 } else { // ISD::AND
9580 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
9581 // two branches instead of an explicit AND instruction with a
9582 // separate test. However, we only do this if this block doesn't
9583 // have a fall-through edge, because this requires an explicit
9584 // jmp when the condition is false.
9585 if (Cmp == Cond.getOperand(1).getOperand(1) &&
9586 isX86LogicalCmp(Cmp) &&
9587 Op.getNode()->hasOneUse()) {
9588 X86::CondCode CCode =
9589 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
9590 CCode = X86::GetOppositeBranchCondition(CCode);
9591 CC = DAG.getConstant(CCode, MVT::i8);
9592 SDNode *User = *Op.getNode()->use_begin();
9593 // Look for an unconditional branch following this conditional branch.
9594 // We need this because we need to reverse the successors in order
9595 // to implement FCMP_OEQ.
9596 if (User->getOpcode() == ISD::BR) {
9597 SDValue FalseBB = User->getOperand(1);
9599 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9600 assert(NewBR == User);
9604 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9605 Chain, Dest, CC, Cmp);
9606 X86::CondCode CCode =
9607 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
9608 CCode = X86::GetOppositeBranchCondition(CCode);
9609 CC = DAG.getConstant(CCode, MVT::i8);
9615 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
9616 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
9617 // It should be transformed during dag combiner except when the condition
9618 // is set by a arithmetics with overflow node.
9619 X86::CondCode CCode =
9620 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
9621 CCode = X86::GetOppositeBranchCondition(CCode);
9622 CC = DAG.getConstant(CCode, MVT::i8);
9623 Cond = Cond.getOperand(0).getOperand(1);
9625 } else if (Cond.getOpcode() == ISD::SETCC &&
9626 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
9627 // For FCMP_OEQ, we can emit
9628 // two branches instead of an explicit AND instruction with a
9629 // separate test. However, we only do this if this block doesn't
9630 // have a fall-through edge, because this requires an explicit
9631 // jmp when the condition is false.
9632 if (Op.getNode()->hasOneUse()) {
9633 SDNode *User = *Op.getNode()->use_begin();
9634 // Look for an unconditional branch following this conditional branch.
9635 // We need this because we need to reverse the successors in order
9636 // to implement FCMP_OEQ.
9637 if (User->getOpcode() == ISD::BR) {
9638 SDValue FalseBB = User->getOperand(1);
9640 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9641 assert(NewBR == User);
9645 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
9646 Cond.getOperand(0), Cond.getOperand(1));
9647 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9648 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9649 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9650 Chain, Dest, CC, Cmp);
9651 CC = DAG.getConstant(X86::COND_P, MVT::i8);
9656 } else if (Cond.getOpcode() == ISD::SETCC &&
9657 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
9658 // For FCMP_UNE, we can emit
9659 // two branches instead of an explicit AND instruction with a
9660 // separate test. However, we only do this if this block doesn't
9661 // have a fall-through edge, because this requires an explicit
9662 // jmp when the condition is false.
9663 if (Op.getNode()->hasOneUse()) {
9664 SDNode *User = *Op.getNode()->use_begin();
9665 // Look for an unconditional branch following this conditional branch.
9666 // We need this because we need to reverse the successors in order
9667 // to implement FCMP_UNE.
9668 if (User->getOpcode() == ISD::BR) {
9669 SDValue FalseBB = User->getOperand(1);
9671 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9672 assert(NewBR == User);
9675 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
9676 Cond.getOperand(0), Cond.getOperand(1));
9677 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9678 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9679 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9680 Chain, Dest, CC, Cmp);
9681 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
9691 // Look pass the truncate if the high bits are known zero.
9692 if (isTruncWithZeroHighBitsInput(Cond, DAG))
9693 Cond = Cond.getOperand(0);
9695 // We know the result of AND is compared against zero. Try to match
9697 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
9698 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
9699 if (NewSetCC.getNode()) {
9700 CC = NewSetCC.getOperand(0);
9701 Cond = NewSetCC.getOperand(1);
9708 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9709 Cond = EmitTest(Cond, X86::COND_NE, DAG);
9711 Cond = ConvertCmpIfNecessary(Cond, DAG);
9712 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9713 Chain, Dest, CC, Cond);
9717 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
9718 // Calls to _alloca is needed to probe the stack when allocating more than 4k
9719 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
9720 // that the guard pages used by the OS virtual memory manager are allocated in
9721 // correct sequence.
9723 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
9724 SelectionDAG &DAG) const {
9725 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
9726 getTargetMachine().Options.EnableSegmentedStacks) &&
9727 "This should be used only on Windows targets or when segmented stacks "
9729 assert(!Subtarget->isTargetEnvMacho() && "Not implemented");
9730 DebugLoc dl = Op.getDebugLoc();
9733 SDValue Chain = Op.getOperand(0);
9734 SDValue Size = Op.getOperand(1);
9735 // FIXME: Ensure alignment here
9737 bool Is64Bit = Subtarget->is64Bit();
9738 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
9740 if (getTargetMachine().Options.EnableSegmentedStacks) {
9741 MachineFunction &MF = DAG.getMachineFunction();
9742 MachineRegisterInfo &MRI = MF.getRegInfo();
9745 // The 64 bit implementation of segmented stacks needs to clobber both r10
9746 // r11. This makes it impossible to use it along with nested parameters.
9747 const Function *F = MF.getFunction();
9749 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
9751 if (I->hasNestAttr())
9752 report_fatal_error("Cannot use segmented stacks with functions that "
9753 "have nested arguments.");
9756 const TargetRegisterClass *AddrRegClass =
9757 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
9758 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
9759 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
9760 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
9761 DAG.getRegister(Vreg, SPTy));
9762 SDValue Ops1[2] = { Value, Chain };
9763 return DAG.getMergeValues(Ops1, 2, dl);
9766 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
9768 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
9769 Flag = Chain.getValue(1);
9770 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
9772 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
9773 Flag = Chain.getValue(1);
9775 Chain = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
9778 SDValue Ops1[2] = { Chain.getValue(0), Chain };
9779 return DAG.getMergeValues(Ops1, 2, dl);
9783 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
9784 MachineFunction &MF = DAG.getMachineFunction();
9785 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
9787 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
9788 DebugLoc DL = Op.getDebugLoc();
9790 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
9791 // vastart just stores the address of the VarArgsFrameIndex slot into the
9792 // memory location argument.
9793 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
9795 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
9796 MachinePointerInfo(SV), false, false, 0);
9800 // gp_offset (0 - 6 * 8)
9801 // fp_offset (48 - 48 + 8 * 16)
9802 // overflow_arg_area (point to parameters coming in memory).
9804 SmallVector<SDValue, 8> MemOps;
9805 SDValue FIN = Op.getOperand(1);
9807 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
9808 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
9810 FIN, MachinePointerInfo(SV), false, false, 0);
9811 MemOps.push_back(Store);
9814 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
9815 FIN, DAG.getIntPtrConstant(4));
9816 Store = DAG.getStore(Op.getOperand(0), DL,
9817 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
9819 FIN, MachinePointerInfo(SV, 4), false, false, 0);
9820 MemOps.push_back(Store);
9822 // Store ptr to overflow_arg_area
9823 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
9824 FIN, DAG.getIntPtrConstant(4));
9825 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
9827 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
9828 MachinePointerInfo(SV, 8),
9830 MemOps.push_back(Store);
9832 // Store ptr to reg_save_area.
9833 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
9834 FIN, DAG.getIntPtrConstant(8));
9835 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
9837 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
9838 MachinePointerInfo(SV, 16), false, false, 0);
9839 MemOps.push_back(Store);
9840 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
9841 &MemOps[0], MemOps.size());
9844 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
9845 assert(Subtarget->is64Bit() &&
9846 "LowerVAARG only handles 64-bit va_arg!");
9847 assert((Subtarget->isTargetLinux() ||
9848 Subtarget->isTargetDarwin()) &&
9849 "Unhandled target in LowerVAARG");
9850 assert(Op.getNode()->getNumOperands() == 4);
9851 SDValue Chain = Op.getOperand(0);
9852 SDValue SrcPtr = Op.getOperand(1);
9853 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
9854 unsigned Align = Op.getConstantOperandVal(3);
9855 DebugLoc dl = Op.getDebugLoc();
9857 EVT ArgVT = Op.getNode()->getValueType(0);
9858 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
9859 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
9862 // Decide which area this value should be read from.
9863 // TODO: Implement the AMD64 ABI in its entirety. This simple
9864 // selection mechanism works only for the basic types.
9865 if (ArgVT == MVT::f80) {
9866 llvm_unreachable("va_arg for f80 not yet implemented");
9867 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
9868 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
9869 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
9870 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
9872 llvm_unreachable("Unhandled argument type in LowerVAARG");
9876 // Sanity Check: Make sure using fp_offset makes sense.
9877 assert(!getTargetMachine().Options.UseSoftFloat &&
9878 !(DAG.getMachineFunction()
9879 .getFunction()->getFnAttributes()
9880 .hasAttribute(Attributes::NoImplicitFloat)) &&
9881 Subtarget->hasSSE1());
9884 // Insert VAARG_64 node into the DAG
9885 // VAARG_64 returns two values: Variable Argument Address, Chain
9886 SmallVector<SDValue, 11> InstOps;
9887 InstOps.push_back(Chain);
9888 InstOps.push_back(SrcPtr);
9889 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
9890 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
9891 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
9892 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
9893 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
9894 VTs, &InstOps[0], InstOps.size(),
9896 MachinePointerInfo(SV),
9901 Chain = VAARG.getValue(1);
9903 // Load the next argument and return it
9904 return DAG.getLoad(ArgVT, dl,
9907 MachinePointerInfo(),
9908 false, false, false, 0);
9911 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
9912 SelectionDAG &DAG) {
9913 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
9914 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
9915 SDValue Chain = Op.getOperand(0);
9916 SDValue DstPtr = Op.getOperand(1);
9917 SDValue SrcPtr = Op.getOperand(2);
9918 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
9919 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9920 DebugLoc DL = Op.getDebugLoc();
9922 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
9923 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
9925 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
9928 // getTargetVShiftNOde - Handle vector element shifts where the shift amount
9929 // may or may not be a constant. Takes immediate version of shift as input.
9930 static SDValue getTargetVShiftNode(unsigned Opc, DebugLoc dl, EVT VT,
9931 SDValue SrcOp, SDValue ShAmt,
9932 SelectionDAG &DAG) {
9933 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
9935 if (isa<ConstantSDNode>(ShAmt)) {
9936 // Constant may be a TargetConstant. Use a regular constant.
9937 uint32_t ShiftAmt = cast<ConstantSDNode>(ShAmt)->getZExtValue();
9939 default: llvm_unreachable("Unknown target vector shift node");
9943 return DAG.getNode(Opc, dl, VT, SrcOp,
9944 DAG.getConstant(ShiftAmt, MVT::i32));
9948 // Change opcode to non-immediate version
9950 default: llvm_unreachable("Unknown target vector shift node");
9951 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
9952 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
9953 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
9956 // Need to build a vector containing shift amount
9957 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
9960 ShOps[1] = DAG.getConstant(0, MVT::i32);
9961 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
9962 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &ShOps[0], 4);
9964 // The return type has to be a 128-bit type with the same element
9965 // type as the input type.
9966 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9967 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
9969 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
9970 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
9973 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
9974 DebugLoc dl = Op.getDebugLoc();
9975 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9977 default: return SDValue(); // Don't custom lower most intrinsics.
9978 // Comparison intrinsics.
9979 case Intrinsic::x86_sse_comieq_ss:
9980 case Intrinsic::x86_sse_comilt_ss:
9981 case Intrinsic::x86_sse_comile_ss:
9982 case Intrinsic::x86_sse_comigt_ss:
9983 case Intrinsic::x86_sse_comige_ss:
9984 case Intrinsic::x86_sse_comineq_ss:
9985 case Intrinsic::x86_sse_ucomieq_ss:
9986 case Intrinsic::x86_sse_ucomilt_ss:
9987 case Intrinsic::x86_sse_ucomile_ss:
9988 case Intrinsic::x86_sse_ucomigt_ss:
9989 case Intrinsic::x86_sse_ucomige_ss:
9990 case Intrinsic::x86_sse_ucomineq_ss:
9991 case Intrinsic::x86_sse2_comieq_sd:
9992 case Intrinsic::x86_sse2_comilt_sd:
9993 case Intrinsic::x86_sse2_comile_sd:
9994 case Intrinsic::x86_sse2_comigt_sd:
9995 case Intrinsic::x86_sse2_comige_sd:
9996 case Intrinsic::x86_sse2_comineq_sd:
9997 case Intrinsic::x86_sse2_ucomieq_sd:
9998 case Intrinsic::x86_sse2_ucomilt_sd:
9999 case Intrinsic::x86_sse2_ucomile_sd:
10000 case Intrinsic::x86_sse2_ucomigt_sd:
10001 case Intrinsic::x86_sse2_ucomige_sd:
10002 case Intrinsic::x86_sse2_ucomineq_sd: {
10006 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10007 case Intrinsic::x86_sse_comieq_ss:
10008 case Intrinsic::x86_sse2_comieq_sd:
10009 Opc = X86ISD::COMI;
10012 case Intrinsic::x86_sse_comilt_ss:
10013 case Intrinsic::x86_sse2_comilt_sd:
10014 Opc = X86ISD::COMI;
10017 case Intrinsic::x86_sse_comile_ss:
10018 case Intrinsic::x86_sse2_comile_sd:
10019 Opc = X86ISD::COMI;
10022 case Intrinsic::x86_sse_comigt_ss:
10023 case Intrinsic::x86_sse2_comigt_sd:
10024 Opc = X86ISD::COMI;
10027 case Intrinsic::x86_sse_comige_ss:
10028 case Intrinsic::x86_sse2_comige_sd:
10029 Opc = X86ISD::COMI;
10032 case Intrinsic::x86_sse_comineq_ss:
10033 case Intrinsic::x86_sse2_comineq_sd:
10034 Opc = X86ISD::COMI;
10037 case Intrinsic::x86_sse_ucomieq_ss:
10038 case Intrinsic::x86_sse2_ucomieq_sd:
10039 Opc = X86ISD::UCOMI;
10042 case Intrinsic::x86_sse_ucomilt_ss:
10043 case Intrinsic::x86_sse2_ucomilt_sd:
10044 Opc = X86ISD::UCOMI;
10047 case Intrinsic::x86_sse_ucomile_ss:
10048 case Intrinsic::x86_sse2_ucomile_sd:
10049 Opc = X86ISD::UCOMI;
10052 case Intrinsic::x86_sse_ucomigt_ss:
10053 case Intrinsic::x86_sse2_ucomigt_sd:
10054 Opc = X86ISD::UCOMI;
10057 case Intrinsic::x86_sse_ucomige_ss:
10058 case Intrinsic::x86_sse2_ucomige_sd:
10059 Opc = X86ISD::UCOMI;
10062 case Intrinsic::x86_sse_ucomineq_ss:
10063 case Intrinsic::x86_sse2_ucomineq_sd:
10064 Opc = X86ISD::UCOMI;
10069 SDValue LHS = Op.getOperand(1);
10070 SDValue RHS = Op.getOperand(2);
10071 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
10072 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
10073 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
10074 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10075 DAG.getConstant(X86CC, MVT::i8), Cond);
10076 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10079 // Arithmetic intrinsics.
10080 case Intrinsic::x86_sse2_pmulu_dq:
10081 case Intrinsic::x86_avx2_pmulu_dq:
10082 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
10083 Op.getOperand(1), Op.getOperand(2));
10085 // SSE3/AVX horizontal add/sub intrinsics
10086 case Intrinsic::x86_sse3_hadd_ps:
10087 case Intrinsic::x86_sse3_hadd_pd:
10088 case Intrinsic::x86_avx_hadd_ps_256:
10089 case Intrinsic::x86_avx_hadd_pd_256:
10090 case Intrinsic::x86_sse3_hsub_ps:
10091 case Intrinsic::x86_sse3_hsub_pd:
10092 case Intrinsic::x86_avx_hsub_ps_256:
10093 case Intrinsic::x86_avx_hsub_pd_256:
10094 case Intrinsic::x86_ssse3_phadd_w_128:
10095 case Intrinsic::x86_ssse3_phadd_d_128:
10096 case Intrinsic::x86_avx2_phadd_w:
10097 case Intrinsic::x86_avx2_phadd_d:
10098 case Intrinsic::x86_ssse3_phsub_w_128:
10099 case Intrinsic::x86_ssse3_phsub_d_128:
10100 case Intrinsic::x86_avx2_phsub_w:
10101 case Intrinsic::x86_avx2_phsub_d: {
10104 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10105 case Intrinsic::x86_sse3_hadd_ps:
10106 case Intrinsic::x86_sse3_hadd_pd:
10107 case Intrinsic::x86_avx_hadd_ps_256:
10108 case Intrinsic::x86_avx_hadd_pd_256:
10109 Opcode = X86ISD::FHADD;
10111 case Intrinsic::x86_sse3_hsub_ps:
10112 case Intrinsic::x86_sse3_hsub_pd:
10113 case Intrinsic::x86_avx_hsub_ps_256:
10114 case Intrinsic::x86_avx_hsub_pd_256:
10115 Opcode = X86ISD::FHSUB;
10117 case Intrinsic::x86_ssse3_phadd_w_128:
10118 case Intrinsic::x86_ssse3_phadd_d_128:
10119 case Intrinsic::x86_avx2_phadd_w:
10120 case Intrinsic::x86_avx2_phadd_d:
10121 Opcode = X86ISD::HADD;
10123 case Intrinsic::x86_ssse3_phsub_w_128:
10124 case Intrinsic::x86_ssse3_phsub_d_128:
10125 case Intrinsic::x86_avx2_phsub_w:
10126 case Intrinsic::x86_avx2_phsub_d:
10127 Opcode = X86ISD::HSUB;
10130 return DAG.getNode(Opcode, dl, Op.getValueType(),
10131 Op.getOperand(1), Op.getOperand(2));
10134 // AVX2 variable shift intrinsics
10135 case Intrinsic::x86_avx2_psllv_d:
10136 case Intrinsic::x86_avx2_psllv_q:
10137 case Intrinsic::x86_avx2_psllv_d_256:
10138 case Intrinsic::x86_avx2_psllv_q_256:
10139 case Intrinsic::x86_avx2_psrlv_d:
10140 case Intrinsic::x86_avx2_psrlv_q:
10141 case Intrinsic::x86_avx2_psrlv_d_256:
10142 case Intrinsic::x86_avx2_psrlv_q_256:
10143 case Intrinsic::x86_avx2_psrav_d:
10144 case Intrinsic::x86_avx2_psrav_d_256: {
10147 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10148 case Intrinsic::x86_avx2_psllv_d:
10149 case Intrinsic::x86_avx2_psllv_q:
10150 case Intrinsic::x86_avx2_psllv_d_256:
10151 case Intrinsic::x86_avx2_psllv_q_256:
10154 case Intrinsic::x86_avx2_psrlv_d:
10155 case Intrinsic::x86_avx2_psrlv_q:
10156 case Intrinsic::x86_avx2_psrlv_d_256:
10157 case Intrinsic::x86_avx2_psrlv_q_256:
10160 case Intrinsic::x86_avx2_psrav_d:
10161 case Intrinsic::x86_avx2_psrav_d_256:
10165 return DAG.getNode(Opcode, dl, Op.getValueType(),
10166 Op.getOperand(1), Op.getOperand(2));
10169 case Intrinsic::x86_ssse3_pshuf_b_128:
10170 case Intrinsic::x86_avx2_pshuf_b:
10171 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
10172 Op.getOperand(1), Op.getOperand(2));
10174 case Intrinsic::x86_ssse3_psign_b_128:
10175 case Intrinsic::x86_ssse3_psign_w_128:
10176 case Intrinsic::x86_ssse3_psign_d_128:
10177 case Intrinsic::x86_avx2_psign_b:
10178 case Intrinsic::x86_avx2_psign_w:
10179 case Intrinsic::x86_avx2_psign_d:
10180 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
10181 Op.getOperand(1), Op.getOperand(2));
10183 case Intrinsic::x86_sse41_insertps:
10184 return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
10185 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
10187 case Intrinsic::x86_avx_vperm2f128_ps_256:
10188 case Intrinsic::x86_avx_vperm2f128_pd_256:
10189 case Intrinsic::x86_avx_vperm2f128_si_256:
10190 case Intrinsic::x86_avx2_vperm2i128:
10191 return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
10192 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
10194 case Intrinsic::x86_avx2_permd:
10195 case Intrinsic::x86_avx2_permps:
10196 // Operands intentionally swapped. Mask is last operand to intrinsic,
10197 // but second operand for node/intruction.
10198 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
10199 Op.getOperand(2), Op.getOperand(1));
10201 // ptest and testp intrinsics. The intrinsic these come from are designed to
10202 // return an integer value, not just an instruction so lower it to the ptest
10203 // or testp pattern and a setcc for the result.
10204 case Intrinsic::x86_sse41_ptestz:
10205 case Intrinsic::x86_sse41_ptestc:
10206 case Intrinsic::x86_sse41_ptestnzc:
10207 case Intrinsic::x86_avx_ptestz_256:
10208 case Intrinsic::x86_avx_ptestc_256:
10209 case Intrinsic::x86_avx_ptestnzc_256:
10210 case Intrinsic::x86_avx_vtestz_ps:
10211 case Intrinsic::x86_avx_vtestc_ps:
10212 case Intrinsic::x86_avx_vtestnzc_ps:
10213 case Intrinsic::x86_avx_vtestz_pd:
10214 case Intrinsic::x86_avx_vtestc_pd:
10215 case Intrinsic::x86_avx_vtestnzc_pd:
10216 case Intrinsic::x86_avx_vtestz_ps_256:
10217 case Intrinsic::x86_avx_vtestc_ps_256:
10218 case Intrinsic::x86_avx_vtestnzc_ps_256:
10219 case Intrinsic::x86_avx_vtestz_pd_256:
10220 case Intrinsic::x86_avx_vtestc_pd_256:
10221 case Intrinsic::x86_avx_vtestnzc_pd_256: {
10222 bool IsTestPacked = false;
10225 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
10226 case Intrinsic::x86_avx_vtestz_ps:
10227 case Intrinsic::x86_avx_vtestz_pd:
10228 case Intrinsic::x86_avx_vtestz_ps_256:
10229 case Intrinsic::x86_avx_vtestz_pd_256:
10230 IsTestPacked = true; // Fallthrough
10231 case Intrinsic::x86_sse41_ptestz:
10232 case Intrinsic::x86_avx_ptestz_256:
10234 X86CC = X86::COND_E;
10236 case Intrinsic::x86_avx_vtestc_ps:
10237 case Intrinsic::x86_avx_vtestc_pd:
10238 case Intrinsic::x86_avx_vtestc_ps_256:
10239 case Intrinsic::x86_avx_vtestc_pd_256:
10240 IsTestPacked = true; // Fallthrough
10241 case Intrinsic::x86_sse41_ptestc:
10242 case Intrinsic::x86_avx_ptestc_256:
10244 X86CC = X86::COND_B;
10246 case Intrinsic::x86_avx_vtestnzc_ps:
10247 case Intrinsic::x86_avx_vtestnzc_pd:
10248 case Intrinsic::x86_avx_vtestnzc_ps_256:
10249 case Intrinsic::x86_avx_vtestnzc_pd_256:
10250 IsTestPacked = true; // Fallthrough
10251 case Intrinsic::x86_sse41_ptestnzc:
10252 case Intrinsic::x86_avx_ptestnzc_256:
10254 X86CC = X86::COND_A;
10258 SDValue LHS = Op.getOperand(1);
10259 SDValue RHS = Op.getOperand(2);
10260 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
10261 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
10262 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
10263 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
10264 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10267 // SSE/AVX shift intrinsics
10268 case Intrinsic::x86_sse2_psll_w:
10269 case Intrinsic::x86_sse2_psll_d:
10270 case Intrinsic::x86_sse2_psll_q:
10271 case Intrinsic::x86_avx2_psll_w:
10272 case Intrinsic::x86_avx2_psll_d:
10273 case Intrinsic::x86_avx2_psll_q:
10274 case Intrinsic::x86_sse2_psrl_w:
10275 case Intrinsic::x86_sse2_psrl_d:
10276 case Intrinsic::x86_sse2_psrl_q:
10277 case Intrinsic::x86_avx2_psrl_w:
10278 case Intrinsic::x86_avx2_psrl_d:
10279 case Intrinsic::x86_avx2_psrl_q:
10280 case Intrinsic::x86_sse2_psra_w:
10281 case Intrinsic::x86_sse2_psra_d:
10282 case Intrinsic::x86_avx2_psra_w:
10283 case Intrinsic::x86_avx2_psra_d: {
10286 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10287 case Intrinsic::x86_sse2_psll_w:
10288 case Intrinsic::x86_sse2_psll_d:
10289 case Intrinsic::x86_sse2_psll_q:
10290 case Intrinsic::x86_avx2_psll_w:
10291 case Intrinsic::x86_avx2_psll_d:
10292 case Intrinsic::x86_avx2_psll_q:
10293 Opcode = X86ISD::VSHL;
10295 case Intrinsic::x86_sse2_psrl_w:
10296 case Intrinsic::x86_sse2_psrl_d:
10297 case Intrinsic::x86_sse2_psrl_q:
10298 case Intrinsic::x86_avx2_psrl_w:
10299 case Intrinsic::x86_avx2_psrl_d:
10300 case Intrinsic::x86_avx2_psrl_q:
10301 Opcode = X86ISD::VSRL;
10303 case Intrinsic::x86_sse2_psra_w:
10304 case Intrinsic::x86_sse2_psra_d:
10305 case Intrinsic::x86_avx2_psra_w:
10306 case Intrinsic::x86_avx2_psra_d:
10307 Opcode = X86ISD::VSRA;
10310 return DAG.getNode(Opcode, dl, Op.getValueType(),
10311 Op.getOperand(1), Op.getOperand(2));
10314 // SSE/AVX immediate shift intrinsics
10315 case Intrinsic::x86_sse2_pslli_w:
10316 case Intrinsic::x86_sse2_pslli_d:
10317 case Intrinsic::x86_sse2_pslli_q:
10318 case Intrinsic::x86_avx2_pslli_w:
10319 case Intrinsic::x86_avx2_pslli_d:
10320 case Intrinsic::x86_avx2_pslli_q:
10321 case Intrinsic::x86_sse2_psrli_w:
10322 case Intrinsic::x86_sse2_psrli_d:
10323 case Intrinsic::x86_sse2_psrli_q:
10324 case Intrinsic::x86_avx2_psrli_w:
10325 case Intrinsic::x86_avx2_psrli_d:
10326 case Intrinsic::x86_avx2_psrli_q:
10327 case Intrinsic::x86_sse2_psrai_w:
10328 case Intrinsic::x86_sse2_psrai_d:
10329 case Intrinsic::x86_avx2_psrai_w:
10330 case Intrinsic::x86_avx2_psrai_d: {
10333 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10334 case Intrinsic::x86_sse2_pslli_w:
10335 case Intrinsic::x86_sse2_pslli_d:
10336 case Intrinsic::x86_sse2_pslli_q:
10337 case Intrinsic::x86_avx2_pslli_w:
10338 case Intrinsic::x86_avx2_pslli_d:
10339 case Intrinsic::x86_avx2_pslli_q:
10340 Opcode = X86ISD::VSHLI;
10342 case Intrinsic::x86_sse2_psrli_w:
10343 case Intrinsic::x86_sse2_psrli_d:
10344 case Intrinsic::x86_sse2_psrli_q:
10345 case Intrinsic::x86_avx2_psrli_w:
10346 case Intrinsic::x86_avx2_psrli_d:
10347 case Intrinsic::x86_avx2_psrli_q:
10348 Opcode = X86ISD::VSRLI;
10350 case Intrinsic::x86_sse2_psrai_w:
10351 case Intrinsic::x86_sse2_psrai_d:
10352 case Intrinsic::x86_avx2_psrai_w:
10353 case Intrinsic::x86_avx2_psrai_d:
10354 Opcode = X86ISD::VSRAI;
10357 return getTargetVShiftNode(Opcode, dl, Op.getValueType(),
10358 Op.getOperand(1), Op.getOperand(2), DAG);
10361 case Intrinsic::x86_sse42_pcmpistria128:
10362 case Intrinsic::x86_sse42_pcmpestria128:
10363 case Intrinsic::x86_sse42_pcmpistric128:
10364 case Intrinsic::x86_sse42_pcmpestric128:
10365 case Intrinsic::x86_sse42_pcmpistrio128:
10366 case Intrinsic::x86_sse42_pcmpestrio128:
10367 case Intrinsic::x86_sse42_pcmpistris128:
10368 case Intrinsic::x86_sse42_pcmpestris128:
10369 case Intrinsic::x86_sse42_pcmpistriz128:
10370 case Intrinsic::x86_sse42_pcmpestriz128: {
10374 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10375 case Intrinsic::x86_sse42_pcmpistria128:
10376 Opcode = X86ISD::PCMPISTRI;
10377 X86CC = X86::COND_A;
10379 case Intrinsic::x86_sse42_pcmpestria128:
10380 Opcode = X86ISD::PCMPESTRI;
10381 X86CC = X86::COND_A;
10383 case Intrinsic::x86_sse42_pcmpistric128:
10384 Opcode = X86ISD::PCMPISTRI;
10385 X86CC = X86::COND_B;
10387 case Intrinsic::x86_sse42_pcmpestric128:
10388 Opcode = X86ISD::PCMPESTRI;
10389 X86CC = X86::COND_B;
10391 case Intrinsic::x86_sse42_pcmpistrio128:
10392 Opcode = X86ISD::PCMPISTRI;
10393 X86CC = X86::COND_O;
10395 case Intrinsic::x86_sse42_pcmpestrio128:
10396 Opcode = X86ISD::PCMPESTRI;
10397 X86CC = X86::COND_O;
10399 case Intrinsic::x86_sse42_pcmpistris128:
10400 Opcode = X86ISD::PCMPISTRI;
10401 X86CC = X86::COND_S;
10403 case Intrinsic::x86_sse42_pcmpestris128:
10404 Opcode = X86ISD::PCMPESTRI;
10405 X86CC = X86::COND_S;
10407 case Intrinsic::x86_sse42_pcmpistriz128:
10408 Opcode = X86ISD::PCMPISTRI;
10409 X86CC = X86::COND_E;
10411 case Intrinsic::x86_sse42_pcmpestriz128:
10412 Opcode = X86ISD::PCMPESTRI;
10413 X86CC = X86::COND_E;
10416 SmallVector<SDValue, 5> NewOps;
10417 NewOps.append(Op->op_begin()+1, Op->op_end());
10418 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
10419 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
10420 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10421 DAG.getConstant(X86CC, MVT::i8),
10422 SDValue(PCMP.getNode(), 1));
10423 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10426 case Intrinsic::x86_sse42_pcmpistri128:
10427 case Intrinsic::x86_sse42_pcmpestri128: {
10429 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
10430 Opcode = X86ISD::PCMPISTRI;
10432 Opcode = X86ISD::PCMPESTRI;
10434 SmallVector<SDValue, 5> NewOps;
10435 NewOps.append(Op->op_begin()+1, Op->op_end());
10436 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
10437 return DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
10439 case Intrinsic::x86_fma_vfmadd_ps:
10440 case Intrinsic::x86_fma_vfmadd_pd:
10441 case Intrinsic::x86_fma_vfmsub_ps:
10442 case Intrinsic::x86_fma_vfmsub_pd:
10443 case Intrinsic::x86_fma_vfnmadd_ps:
10444 case Intrinsic::x86_fma_vfnmadd_pd:
10445 case Intrinsic::x86_fma_vfnmsub_ps:
10446 case Intrinsic::x86_fma_vfnmsub_pd:
10447 case Intrinsic::x86_fma_vfmaddsub_ps:
10448 case Intrinsic::x86_fma_vfmaddsub_pd:
10449 case Intrinsic::x86_fma_vfmsubadd_ps:
10450 case Intrinsic::x86_fma_vfmsubadd_pd:
10451 case Intrinsic::x86_fma_vfmadd_ps_256:
10452 case Intrinsic::x86_fma_vfmadd_pd_256:
10453 case Intrinsic::x86_fma_vfmsub_ps_256:
10454 case Intrinsic::x86_fma_vfmsub_pd_256:
10455 case Intrinsic::x86_fma_vfnmadd_ps_256:
10456 case Intrinsic::x86_fma_vfnmadd_pd_256:
10457 case Intrinsic::x86_fma_vfnmsub_ps_256:
10458 case Intrinsic::x86_fma_vfnmsub_pd_256:
10459 case Intrinsic::x86_fma_vfmaddsub_ps_256:
10460 case Intrinsic::x86_fma_vfmaddsub_pd_256:
10461 case Intrinsic::x86_fma_vfmsubadd_ps_256:
10462 case Intrinsic::x86_fma_vfmsubadd_pd_256: {
10465 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10466 case Intrinsic::x86_fma_vfmadd_ps:
10467 case Intrinsic::x86_fma_vfmadd_pd:
10468 case Intrinsic::x86_fma_vfmadd_ps_256:
10469 case Intrinsic::x86_fma_vfmadd_pd_256:
10470 Opc = X86ISD::FMADD;
10472 case Intrinsic::x86_fma_vfmsub_ps:
10473 case Intrinsic::x86_fma_vfmsub_pd:
10474 case Intrinsic::x86_fma_vfmsub_ps_256:
10475 case Intrinsic::x86_fma_vfmsub_pd_256:
10476 Opc = X86ISD::FMSUB;
10478 case Intrinsic::x86_fma_vfnmadd_ps:
10479 case Intrinsic::x86_fma_vfnmadd_pd:
10480 case Intrinsic::x86_fma_vfnmadd_ps_256:
10481 case Intrinsic::x86_fma_vfnmadd_pd_256:
10482 Opc = X86ISD::FNMADD;
10484 case Intrinsic::x86_fma_vfnmsub_ps:
10485 case Intrinsic::x86_fma_vfnmsub_pd:
10486 case Intrinsic::x86_fma_vfnmsub_ps_256:
10487 case Intrinsic::x86_fma_vfnmsub_pd_256:
10488 Opc = X86ISD::FNMSUB;
10490 case Intrinsic::x86_fma_vfmaddsub_ps:
10491 case Intrinsic::x86_fma_vfmaddsub_pd:
10492 case Intrinsic::x86_fma_vfmaddsub_ps_256:
10493 case Intrinsic::x86_fma_vfmaddsub_pd_256:
10494 Opc = X86ISD::FMADDSUB;
10496 case Intrinsic::x86_fma_vfmsubadd_ps:
10497 case Intrinsic::x86_fma_vfmsubadd_pd:
10498 case Intrinsic::x86_fma_vfmsubadd_ps_256:
10499 case Intrinsic::x86_fma_vfmsubadd_pd_256:
10500 Opc = X86ISD::FMSUBADD;
10504 return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
10505 Op.getOperand(2), Op.getOperand(3));
10510 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
10511 DebugLoc dl = Op.getDebugLoc();
10512 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10514 default: return SDValue(); // Don't custom lower most intrinsics.
10516 // RDRAND intrinsics.
10517 case Intrinsic::x86_rdrand_16:
10518 case Intrinsic::x86_rdrand_32:
10519 case Intrinsic::x86_rdrand_64: {
10520 // Emit the node with the right value type.
10521 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
10522 SDValue Result = DAG.getNode(X86ISD::RDRAND, dl, VTs, Op.getOperand(0));
10524 // If the value returned by RDRAND was valid (CF=1), return 1. Otherwise
10525 // return the value from Rand, which is always 0, casted to i32.
10526 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
10527 DAG.getConstant(1, Op->getValueType(1)),
10528 DAG.getConstant(X86::COND_B, MVT::i32),
10529 SDValue(Result.getNode(), 1) };
10530 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
10531 DAG.getVTList(Op->getValueType(1), MVT::Glue),
10534 // Return { result, isValid, chain }.
10535 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
10536 SDValue(Result.getNode(), 2));
10541 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
10542 SelectionDAG &DAG) const {
10543 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10544 MFI->setReturnAddressIsTaken(true);
10546 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10547 DebugLoc dl = Op.getDebugLoc();
10548 EVT PtrVT = getPointerTy();
10551 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
10553 DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
10554 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
10555 DAG.getNode(ISD::ADD, dl, PtrVT,
10556 FrameAddr, Offset),
10557 MachinePointerInfo(), false, false, false, 0);
10560 // Just load the return address.
10561 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
10562 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
10563 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
10566 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
10567 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10568 MFI->setFrameAddressIsTaken(true);
10570 EVT VT = Op.getValueType();
10571 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
10572 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10573 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
10574 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
10576 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
10577 MachinePointerInfo(),
10578 false, false, false, 0);
10582 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
10583 SelectionDAG &DAG) const {
10584 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
10587 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
10588 SDValue Chain = Op.getOperand(0);
10589 SDValue Offset = Op.getOperand(1);
10590 SDValue Handler = Op.getOperand(2);
10591 DebugLoc dl = Op.getDebugLoc();
10593 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
10594 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
10596 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
10598 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
10599 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
10600 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
10601 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
10603 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
10605 return DAG.getNode(X86ISD::EH_RETURN, dl,
10607 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
10610 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
10611 SelectionDAG &DAG) const {
10612 DebugLoc DL = Op.getDebugLoc();
10613 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
10614 DAG.getVTList(MVT::i32, MVT::Other),
10615 Op.getOperand(0), Op.getOperand(1));
10618 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
10619 SelectionDAG &DAG) const {
10620 DebugLoc DL = Op.getDebugLoc();
10621 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
10622 Op.getOperand(0), Op.getOperand(1));
10625 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
10626 return Op.getOperand(0);
10629 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
10630 SelectionDAG &DAG) const {
10631 SDValue Root = Op.getOperand(0);
10632 SDValue Trmp = Op.getOperand(1); // trampoline
10633 SDValue FPtr = Op.getOperand(2); // nested function
10634 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
10635 DebugLoc dl = Op.getDebugLoc();
10637 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
10638 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
10640 if (Subtarget->is64Bit()) {
10641 SDValue OutChains[6];
10643 // Large code-model.
10644 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
10645 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
10647 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
10648 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
10650 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
10652 // Load the pointer to the nested function into R11.
10653 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
10654 SDValue Addr = Trmp;
10655 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
10656 Addr, MachinePointerInfo(TrmpAddr),
10659 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
10660 DAG.getConstant(2, MVT::i64));
10661 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
10662 MachinePointerInfo(TrmpAddr, 2),
10665 // Load the 'nest' parameter value into R10.
10666 // R10 is specified in X86CallingConv.td
10667 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
10668 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
10669 DAG.getConstant(10, MVT::i64));
10670 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
10671 Addr, MachinePointerInfo(TrmpAddr, 10),
10674 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
10675 DAG.getConstant(12, MVT::i64));
10676 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
10677 MachinePointerInfo(TrmpAddr, 12),
10680 // Jump to the nested function.
10681 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
10682 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
10683 DAG.getConstant(20, MVT::i64));
10684 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
10685 Addr, MachinePointerInfo(TrmpAddr, 20),
10688 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
10689 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
10690 DAG.getConstant(22, MVT::i64));
10691 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
10692 MachinePointerInfo(TrmpAddr, 22),
10695 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6);
10697 const Function *Func =
10698 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
10699 CallingConv::ID CC = Func->getCallingConv();
10704 llvm_unreachable("Unsupported calling convention");
10705 case CallingConv::C:
10706 case CallingConv::X86_StdCall: {
10707 // Pass 'nest' parameter in ECX.
10708 // Must be kept in sync with X86CallingConv.td
10709 NestReg = X86::ECX;
10711 // Check that ECX wasn't needed by an 'inreg' parameter.
10712 FunctionType *FTy = Func->getFunctionType();
10713 const AttrListPtr &Attrs = Func->getAttributes();
10715 if (!Attrs.isEmpty() && !Func->isVarArg()) {
10716 unsigned InRegCount = 0;
10719 for (FunctionType::param_iterator I = FTy->param_begin(),
10720 E = FTy->param_end(); I != E; ++I, ++Idx)
10721 if (Attrs.getParamAttributes(Idx).hasAttribute(Attributes::InReg))
10722 // FIXME: should only count parameters that are lowered to integers.
10723 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
10725 if (InRegCount > 2) {
10726 report_fatal_error("Nest register in use - reduce number of inreg"
10732 case CallingConv::X86_FastCall:
10733 case CallingConv::X86_ThisCall:
10734 case CallingConv::Fast:
10735 // Pass 'nest' parameter in EAX.
10736 // Must be kept in sync with X86CallingConv.td
10737 NestReg = X86::EAX;
10741 SDValue OutChains[4];
10742 SDValue Addr, Disp;
10744 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
10745 DAG.getConstant(10, MVT::i32));
10746 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
10748 // This is storing the opcode for MOV32ri.
10749 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
10750 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
10751 OutChains[0] = DAG.getStore(Root, dl,
10752 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
10753 Trmp, MachinePointerInfo(TrmpAddr),
10756 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
10757 DAG.getConstant(1, MVT::i32));
10758 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
10759 MachinePointerInfo(TrmpAddr, 1),
10762 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
10763 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
10764 DAG.getConstant(5, MVT::i32));
10765 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
10766 MachinePointerInfo(TrmpAddr, 5),
10769 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
10770 DAG.getConstant(6, MVT::i32));
10771 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
10772 MachinePointerInfo(TrmpAddr, 6),
10775 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4);
10779 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
10780 SelectionDAG &DAG) const {
10782 The rounding mode is in bits 11:10 of FPSR, and has the following
10784 00 Round to nearest
10789 FLT_ROUNDS, on the other hand, expects the following:
10796 To perform the conversion, we do:
10797 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
10800 MachineFunction &MF = DAG.getMachineFunction();
10801 const TargetMachine &TM = MF.getTarget();
10802 const TargetFrameLowering &TFI = *TM.getFrameLowering();
10803 unsigned StackAlignment = TFI.getStackAlignment();
10804 EVT VT = Op.getValueType();
10805 DebugLoc DL = Op.getDebugLoc();
10807 // Save FP Control Word to stack slot
10808 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
10809 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
10812 MachineMemOperand *MMO =
10813 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
10814 MachineMemOperand::MOStore, 2, 2);
10816 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
10817 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
10818 DAG.getVTList(MVT::Other),
10819 Ops, 2, MVT::i16, MMO);
10821 // Load FP Control Word from stack slot
10822 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
10823 MachinePointerInfo(), false, false, false, 0);
10825 // Transform as necessary
10827 DAG.getNode(ISD::SRL, DL, MVT::i16,
10828 DAG.getNode(ISD::AND, DL, MVT::i16,
10829 CWD, DAG.getConstant(0x800, MVT::i16)),
10830 DAG.getConstant(11, MVT::i8));
10832 DAG.getNode(ISD::SRL, DL, MVT::i16,
10833 DAG.getNode(ISD::AND, DL, MVT::i16,
10834 CWD, DAG.getConstant(0x400, MVT::i16)),
10835 DAG.getConstant(9, MVT::i8));
10838 DAG.getNode(ISD::AND, DL, MVT::i16,
10839 DAG.getNode(ISD::ADD, DL, MVT::i16,
10840 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
10841 DAG.getConstant(1, MVT::i16)),
10842 DAG.getConstant(3, MVT::i16));
10845 return DAG.getNode((VT.getSizeInBits() < 16 ?
10846 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
10849 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
10850 EVT VT = Op.getValueType();
10852 unsigned NumBits = VT.getSizeInBits();
10853 DebugLoc dl = Op.getDebugLoc();
10855 Op = Op.getOperand(0);
10856 if (VT == MVT::i8) {
10857 // Zero extend to i32 since there is not an i8 bsr.
10859 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
10862 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
10863 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
10864 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
10866 // If src is zero (i.e. bsr sets ZF), returns NumBits.
10869 DAG.getConstant(NumBits+NumBits-1, OpVT),
10870 DAG.getConstant(X86::COND_E, MVT::i8),
10873 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
10875 // Finally xor with NumBits-1.
10876 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
10879 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
10883 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
10884 EVT VT = Op.getValueType();
10886 unsigned NumBits = VT.getSizeInBits();
10887 DebugLoc dl = Op.getDebugLoc();
10889 Op = Op.getOperand(0);
10890 if (VT == MVT::i8) {
10891 // Zero extend to i32 since there is not an i8 bsr.
10893 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
10896 // Issue a bsr (scan bits in reverse).
10897 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
10898 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
10900 // And xor with NumBits-1.
10901 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
10904 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
10908 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
10909 EVT VT = Op.getValueType();
10910 unsigned NumBits = VT.getSizeInBits();
10911 DebugLoc dl = Op.getDebugLoc();
10912 Op = Op.getOperand(0);
10914 // Issue a bsf (scan bits forward) which also sets EFLAGS.
10915 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
10916 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
10918 // If src is zero (i.e. bsf sets ZF), returns NumBits.
10921 DAG.getConstant(NumBits, VT),
10922 DAG.getConstant(X86::COND_E, MVT::i8),
10925 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops, array_lengthof(Ops));
10928 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
10929 // ones, and then concatenate the result back.
10930 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
10931 EVT VT = Op.getValueType();
10933 assert(VT.is256BitVector() && VT.isInteger() &&
10934 "Unsupported value type for operation");
10936 unsigned NumElems = VT.getVectorNumElements();
10937 DebugLoc dl = Op.getDebugLoc();
10939 // Extract the LHS vectors
10940 SDValue LHS = Op.getOperand(0);
10941 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
10942 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
10944 // Extract the RHS vectors
10945 SDValue RHS = Op.getOperand(1);
10946 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
10947 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
10949 MVT EltVT = VT.getVectorElementType().getSimpleVT();
10950 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
10952 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10953 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
10954 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
10957 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
10958 assert(Op.getValueType().is256BitVector() &&
10959 Op.getValueType().isInteger() &&
10960 "Only handle AVX 256-bit vector integer operation");
10961 return Lower256IntArith(Op, DAG);
10964 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
10965 assert(Op.getValueType().is256BitVector() &&
10966 Op.getValueType().isInteger() &&
10967 "Only handle AVX 256-bit vector integer operation");
10968 return Lower256IntArith(Op, DAG);
10971 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
10972 SelectionDAG &DAG) {
10973 EVT VT = Op.getValueType();
10975 // Decompose 256-bit ops into smaller 128-bit ops.
10976 if (VT.is256BitVector() && !Subtarget->hasAVX2())
10977 return Lower256IntArith(Op, DAG);
10979 assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
10980 "Only know how to lower V2I64/V4I64 multiply");
10982 DebugLoc dl = Op.getDebugLoc();
10984 // Ahi = psrlqi(a, 32);
10985 // Bhi = psrlqi(b, 32);
10987 // AloBlo = pmuludq(a, b);
10988 // AloBhi = pmuludq(a, Bhi);
10989 // AhiBlo = pmuludq(Ahi, b);
10991 // AloBhi = psllqi(AloBhi, 32);
10992 // AhiBlo = psllqi(AhiBlo, 32);
10993 // return AloBlo + AloBhi + AhiBlo;
10995 SDValue A = Op.getOperand(0);
10996 SDValue B = Op.getOperand(1);
10998 SDValue ShAmt = DAG.getConstant(32, MVT::i32);
11000 SDValue Ahi = DAG.getNode(X86ISD::VSRLI, dl, VT, A, ShAmt);
11001 SDValue Bhi = DAG.getNode(X86ISD::VSRLI, dl, VT, B, ShAmt);
11003 // Bit cast to 32-bit vectors for MULUDQ
11004 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 : MVT::v8i32;
11005 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
11006 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
11007 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
11008 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
11010 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
11011 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
11012 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
11014 AloBhi = DAG.getNode(X86ISD::VSHLI, dl, VT, AloBhi, ShAmt);
11015 AhiBlo = DAG.getNode(X86ISD::VSHLI, dl, VT, AhiBlo, ShAmt);
11017 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
11018 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
11021 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
11023 EVT VT = Op.getValueType();
11024 DebugLoc dl = Op.getDebugLoc();
11025 SDValue R = Op.getOperand(0);
11026 SDValue Amt = Op.getOperand(1);
11027 LLVMContext *Context = DAG.getContext();
11029 if (!Subtarget->hasSSE2())
11032 // Optimize shl/srl/sra with constant shift amount.
11033 if (isSplatVector(Amt.getNode())) {
11034 SDValue SclrAmt = Amt->getOperand(0);
11035 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
11036 uint64_t ShiftAmt = C->getZExtValue();
11038 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
11039 (Subtarget->hasAVX2() &&
11040 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16))) {
11041 if (Op.getOpcode() == ISD::SHL)
11042 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
11043 DAG.getConstant(ShiftAmt, MVT::i32));
11044 if (Op.getOpcode() == ISD::SRL)
11045 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
11046 DAG.getConstant(ShiftAmt, MVT::i32));
11047 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
11048 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
11049 DAG.getConstant(ShiftAmt, MVT::i32));
11052 if (VT == MVT::v16i8) {
11053 if (Op.getOpcode() == ISD::SHL) {
11054 // Make a large shift.
11055 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, R,
11056 DAG.getConstant(ShiftAmt, MVT::i32));
11057 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
11058 // Zero out the rightmost bits.
11059 SmallVector<SDValue, 16> V(16,
11060 DAG.getConstant(uint8_t(-1U << ShiftAmt),
11062 return DAG.getNode(ISD::AND, dl, VT, SHL,
11063 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
11065 if (Op.getOpcode() == ISD::SRL) {
11066 // Make a large shift.
11067 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v8i16, R,
11068 DAG.getConstant(ShiftAmt, MVT::i32));
11069 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
11070 // Zero out the leftmost bits.
11071 SmallVector<SDValue, 16> V(16,
11072 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
11074 return DAG.getNode(ISD::AND, dl, VT, SRL,
11075 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
11077 if (Op.getOpcode() == ISD::SRA) {
11078 if (ShiftAmt == 7) {
11079 // R s>> 7 === R s< 0
11080 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
11081 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
11084 // R s>> a === ((R u>> a) ^ m) - m
11085 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
11086 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
11088 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16);
11089 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
11090 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
11093 llvm_unreachable("Unknown shift opcode.");
11096 if (Subtarget->hasAVX2() && VT == MVT::v32i8) {
11097 if (Op.getOpcode() == ISD::SHL) {
11098 // Make a large shift.
11099 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v16i16, R,
11100 DAG.getConstant(ShiftAmt, MVT::i32));
11101 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
11102 // Zero out the rightmost bits.
11103 SmallVector<SDValue, 32> V(32,
11104 DAG.getConstant(uint8_t(-1U << ShiftAmt),
11106 return DAG.getNode(ISD::AND, dl, VT, SHL,
11107 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
11109 if (Op.getOpcode() == ISD::SRL) {
11110 // Make a large shift.
11111 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v16i16, R,
11112 DAG.getConstant(ShiftAmt, MVT::i32));
11113 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
11114 // Zero out the leftmost bits.
11115 SmallVector<SDValue, 32> V(32,
11116 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
11118 return DAG.getNode(ISD::AND, dl, VT, SRL,
11119 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
11121 if (Op.getOpcode() == ISD::SRA) {
11122 if (ShiftAmt == 7) {
11123 // R s>> 7 === R s< 0
11124 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
11125 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
11128 // R s>> a === ((R u>> a) ^ m) - m
11129 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
11130 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
11132 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32);
11133 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
11134 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
11137 llvm_unreachable("Unknown shift opcode.");
11142 // Lower SHL with variable shift amount.
11143 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
11144 Op = DAG.getNode(X86ISD::VSHLI, dl, VT, Op.getOperand(1),
11145 DAG.getConstant(23, MVT::i32));
11147 const uint32_t CV[] = { 0x3f800000U, 0x3f800000U, 0x3f800000U, 0x3f800000U};
11148 Constant *C = ConstantDataVector::get(*Context, CV);
11149 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
11150 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11151 MachinePointerInfo::getConstantPool(),
11152 false, false, false, 16);
11154 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
11155 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
11156 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
11157 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
11159 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
11160 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
11163 Op = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, Op.getOperand(1),
11164 DAG.getConstant(5, MVT::i32));
11165 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
11167 // Turn 'a' into a mask suitable for VSELECT
11168 SDValue VSelM = DAG.getConstant(0x80, VT);
11169 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11170 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11172 SDValue CM1 = DAG.getConstant(0x0f, VT);
11173 SDValue CM2 = DAG.getConstant(0x3f, VT);
11175 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
11176 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
11177 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
11178 DAG.getConstant(4, MVT::i32), DAG);
11179 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
11180 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
11183 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
11184 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11185 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11187 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
11188 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
11189 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
11190 DAG.getConstant(2, MVT::i32), DAG);
11191 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
11192 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
11195 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
11196 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11197 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11199 // return VSELECT(r, r+r, a);
11200 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
11201 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
11205 // Decompose 256-bit shifts into smaller 128-bit shifts.
11206 if (VT.is256BitVector()) {
11207 unsigned NumElems = VT.getVectorNumElements();
11208 MVT EltVT = VT.getVectorElementType().getSimpleVT();
11209 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
11211 // Extract the two vectors
11212 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
11213 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
11215 // Recreate the shift amount vectors
11216 SDValue Amt1, Amt2;
11217 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
11218 // Constant shift amount
11219 SmallVector<SDValue, 4> Amt1Csts;
11220 SmallVector<SDValue, 4> Amt2Csts;
11221 for (unsigned i = 0; i != NumElems/2; ++i)
11222 Amt1Csts.push_back(Amt->getOperand(i));
11223 for (unsigned i = NumElems/2; i != NumElems; ++i)
11224 Amt2Csts.push_back(Amt->getOperand(i));
11226 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
11227 &Amt1Csts[0], NumElems/2);
11228 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
11229 &Amt2Csts[0], NumElems/2);
11231 // Variable shift amount
11232 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
11233 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
11236 // Issue new vector shifts for the smaller types
11237 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
11238 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
11240 // Concatenate the result back
11241 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
11247 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
11248 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
11249 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
11250 // looks for this combo and may remove the "setcc" instruction if the "setcc"
11251 // has only one use.
11252 SDNode *N = Op.getNode();
11253 SDValue LHS = N->getOperand(0);
11254 SDValue RHS = N->getOperand(1);
11255 unsigned BaseOp = 0;
11257 DebugLoc DL = Op.getDebugLoc();
11258 switch (Op.getOpcode()) {
11259 default: llvm_unreachable("Unknown ovf instruction!");
11261 // A subtract of one will be selected as a INC. Note that INC doesn't
11262 // set CF, so we can't do this for UADDO.
11263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
11265 BaseOp = X86ISD::INC;
11266 Cond = X86::COND_O;
11269 BaseOp = X86ISD::ADD;
11270 Cond = X86::COND_O;
11273 BaseOp = X86ISD::ADD;
11274 Cond = X86::COND_B;
11277 // A subtract of one will be selected as a DEC. Note that DEC doesn't
11278 // set CF, so we can't do this for USUBO.
11279 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
11281 BaseOp = X86ISD::DEC;
11282 Cond = X86::COND_O;
11285 BaseOp = X86ISD::SUB;
11286 Cond = X86::COND_O;
11289 BaseOp = X86ISD::SUB;
11290 Cond = X86::COND_B;
11293 BaseOp = X86ISD::SMUL;
11294 Cond = X86::COND_O;
11296 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
11297 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
11299 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
11302 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11303 DAG.getConstant(X86::COND_O, MVT::i32),
11304 SDValue(Sum.getNode(), 2));
11306 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
11310 // Also sets EFLAGS.
11311 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
11312 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
11315 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
11316 DAG.getConstant(Cond, MVT::i32),
11317 SDValue(Sum.getNode(), 1));
11319 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
11322 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
11323 SelectionDAG &DAG) const {
11324 DebugLoc dl = Op.getDebugLoc();
11325 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
11326 EVT VT = Op.getValueType();
11328 if (!Subtarget->hasSSE2() || !VT.isVector())
11331 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
11332 ExtraVT.getScalarType().getSizeInBits();
11333 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
11335 switch (VT.getSimpleVT().SimpleTy) {
11336 default: return SDValue();
11339 if (!Subtarget->hasAVX())
11341 if (!Subtarget->hasAVX2()) {
11342 // needs to be split
11343 unsigned NumElems = VT.getVectorNumElements();
11345 // Extract the LHS vectors
11346 SDValue LHS = Op.getOperand(0);
11347 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
11348 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
11350 MVT EltVT = VT.getVectorElementType().getSimpleVT();
11351 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
11353 EVT ExtraEltVT = ExtraVT.getVectorElementType();
11354 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
11355 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
11357 SDValue Extra = DAG.getValueType(ExtraVT);
11359 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
11360 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
11362 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
11367 SDValue Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT,
11368 Op.getOperand(0), ShAmt, DAG);
11369 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, Tmp1, ShAmt, DAG);
11375 static SDValue LowerMEMBARRIER(SDValue Op, const X86Subtarget *Subtarget,
11376 SelectionDAG &DAG) {
11377 DebugLoc dl = Op.getDebugLoc();
11379 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
11380 // There isn't any reason to disable it if the target processor supports it.
11381 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
11382 SDValue Chain = Op.getOperand(0);
11383 SDValue Zero = DAG.getConstant(0, MVT::i32);
11385 DAG.getRegister(X86::ESP, MVT::i32), // Base
11386 DAG.getTargetConstant(1, MVT::i8), // Scale
11387 DAG.getRegister(0, MVT::i32), // Index
11388 DAG.getTargetConstant(0, MVT::i32), // Disp
11389 DAG.getRegister(0, MVT::i32), // Segment.
11394 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
11395 array_lengthof(Ops));
11396 return SDValue(Res, 0);
11399 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
11401 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
11403 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11404 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
11405 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
11406 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
11408 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
11409 if (!Op1 && !Op2 && !Op3 && Op4)
11410 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
11412 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
11413 if (Op1 && !Op2 && !Op3 && !Op4)
11414 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
11416 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
11418 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
11421 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
11422 SelectionDAG &DAG) {
11423 DebugLoc dl = Op.getDebugLoc();
11424 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
11425 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
11426 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
11427 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
11429 // The only fence that needs an instruction is a sequentially-consistent
11430 // cross-thread fence.
11431 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
11432 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
11433 // no-sse2). There isn't any reason to disable it if the target processor
11435 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
11436 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
11438 SDValue Chain = Op.getOperand(0);
11439 SDValue Zero = DAG.getConstant(0, MVT::i32);
11441 DAG.getRegister(X86::ESP, MVT::i32), // Base
11442 DAG.getTargetConstant(1, MVT::i8), // Scale
11443 DAG.getRegister(0, MVT::i32), // Index
11444 DAG.getTargetConstant(0, MVT::i32), // Disp
11445 DAG.getRegister(0, MVT::i32), // Segment.
11450 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
11451 array_lengthof(Ops));
11452 return SDValue(Res, 0);
11455 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
11456 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
11460 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
11461 SelectionDAG &DAG) {
11462 EVT T = Op.getValueType();
11463 DebugLoc DL = Op.getDebugLoc();
11466 switch(T.getSimpleVT().SimpleTy) {
11467 default: llvm_unreachable("Invalid value type!");
11468 case MVT::i8: Reg = X86::AL; size = 1; break;
11469 case MVT::i16: Reg = X86::AX; size = 2; break;
11470 case MVT::i32: Reg = X86::EAX; size = 4; break;
11472 assert(Subtarget->is64Bit() && "Node not type legal!");
11473 Reg = X86::RAX; size = 8;
11476 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
11477 Op.getOperand(2), SDValue());
11478 SDValue Ops[] = { cpIn.getValue(0),
11481 DAG.getTargetConstant(size, MVT::i8),
11482 cpIn.getValue(1) };
11483 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
11484 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
11485 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
11488 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
11492 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
11493 SelectionDAG &DAG) {
11494 assert(Subtarget->is64Bit() && "Result not type legalized?");
11495 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
11496 SDValue TheChain = Op.getOperand(0);
11497 DebugLoc dl = Op.getDebugLoc();
11498 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
11499 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
11500 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
11502 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
11503 DAG.getConstant(32, MVT::i8));
11505 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
11508 return DAG.getMergeValues(Ops, 2, dl);
11511 SDValue X86TargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
11512 EVT SrcVT = Op.getOperand(0).getValueType();
11513 EVT DstVT = Op.getValueType();
11514 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
11515 Subtarget->hasMMX() && "Unexpected custom BITCAST");
11516 assert((DstVT == MVT::i64 ||
11517 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
11518 "Unexpected custom BITCAST");
11519 // i64 <=> MMX conversions are Legal.
11520 if (SrcVT==MVT::i64 && DstVT.isVector())
11522 if (DstVT==MVT::i64 && SrcVT.isVector())
11524 // MMX <=> MMX conversions are Legal.
11525 if (SrcVT.isVector() && DstVT.isVector())
11527 // All other conversions need to be expanded.
11531 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
11532 SDNode *Node = Op.getNode();
11533 DebugLoc dl = Node->getDebugLoc();
11534 EVT T = Node->getValueType(0);
11535 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
11536 DAG.getConstant(0, T), Node->getOperand(2));
11537 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
11538 cast<AtomicSDNode>(Node)->getMemoryVT(),
11539 Node->getOperand(0),
11540 Node->getOperand(1), negOp,
11541 cast<AtomicSDNode>(Node)->getSrcValue(),
11542 cast<AtomicSDNode>(Node)->getAlignment(),
11543 cast<AtomicSDNode>(Node)->getOrdering(),
11544 cast<AtomicSDNode>(Node)->getSynchScope());
11547 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
11548 SDNode *Node = Op.getNode();
11549 DebugLoc dl = Node->getDebugLoc();
11550 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
11552 // Convert seq_cst store -> xchg
11553 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
11554 // FIXME: On 32-bit, store -> fist or movq would be more efficient
11555 // (The only way to get a 16-byte store is cmpxchg16b)
11556 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
11557 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
11558 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
11559 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
11560 cast<AtomicSDNode>(Node)->getMemoryVT(),
11561 Node->getOperand(0),
11562 Node->getOperand(1), Node->getOperand(2),
11563 cast<AtomicSDNode>(Node)->getMemOperand(),
11564 cast<AtomicSDNode>(Node)->getOrdering(),
11565 cast<AtomicSDNode>(Node)->getSynchScope());
11566 return Swap.getValue(1);
11568 // Other atomic stores have a simple pattern.
11572 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
11573 EVT VT = Op.getNode()->getValueType(0);
11575 // Let legalize expand this if it isn't a legal type yet.
11576 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
11579 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
11582 bool ExtraOp = false;
11583 switch (Op.getOpcode()) {
11584 default: llvm_unreachable("Invalid code");
11585 case ISD::ADDC: Opc = X86ISD::ADD; break;
11586 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
11587 case ISD::SUBC: Opc = X86ISD::SUB; break;
11588 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
11592 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
11594 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
11595 Op.getOperand(1), Op.getOperand(2));
11598 /// LowerOperation - Provide custom lowering hooks for some operations.
11600 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
11601 switch (Op.getOpcode()) {
11602 default: llvm_unreachable("Should not custom lower this!");
11603 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
11604 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, Subtarget, DAG);
11605 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
11606 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op, Subtarget, DAG);
11607 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
11608 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
11609 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
11610 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
11611 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
11612 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
11613 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
11614 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
11615 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
11616 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
11617 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
11618 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
11619 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
11620 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
11621 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
11622 case ISD::SHL_PARTS:
11623 case ISD::SRA_PARTS:
11624 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
11625 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
11626 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
11627 case ISD::TRUNCATE: return lowerTRUNCATE(Op, DAG);
11628 case ISD::ZERO_EXTEND: return lowerZERO_EXTEND(Op, DAG);
11629 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
11630 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
11631 case ISD::FP_EXTEND: return lowerFP_EXTEND(Op, DAG);
11632 case ISD::FABS: return LowerFABS(Op, DAG);
11633 case ISD::FNEG: return LowerFNEG(Op, DAG);
11634 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
11635 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
11636 case ISD::SETCC: return LowerSETCC(Op, DAG);
11637 case ISD::SELECT: return LowerSELECT(Op, DAG);
11638 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
11639 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
11640 case ISD::VASTART: return LowerVASTART(Op, DAG);
11641 case ISD::VAARG: return LowerVAARG(Op, DAG);
11642 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
11643 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
11644 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
11645 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
11646 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
11647 case ISD::FRAME_TO_ARGS_OFFSET:
11648 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
11649 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
11650 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
11651 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
11652 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
11653 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
11654 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
11655 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
11656 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
11657 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
11658 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
11659 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
11662 case ISD::SHL: return LowerShift(Op, DAG);
11668 case ISD::UMULO: return LowerXALUO(Op, DAG);
11669 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
11670 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
11674 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
11675 case ISD::ADD: return LowerADD(Op, DAG);
11676 case ISD::SUB: return LowerSUB(Op, DAG);
11680 static void ReplaceATOMIC_LOAD(SDNode *Node,
11681 SmallVectorImpl<SDValue> &Results,
11682 SelectionDAG &DAG) {
11683 DebugLoc dl = Node->getDebugLoc();
11684 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
11686 // Convert wide load -> cmpxchg8b/cmpxchg16b
11687 // FIXME: On 32-bit, load -> fild or movq would be more efficient
11688 // (The only way to get a 16-byte load is cmpxchg16b)
11689 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
11690 SDValue Zero = DAG.getConstant(0, VT);
11691 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
11692 Node->getOperand(0),
11693 Node->getOperand(1), Zero, Zero,
11694 cast<AtomicSDNode>(Node)->getMemOperand(),
11695 cast<AtomicSDNode>(Node)->getOrdering(),
11696 cast<AtomicSDNode>(Node)->getSynchScope());
11697 Results.push_back(Swap.getValue(0));
11698 Results.push_back(Swap.getValue(1));
11702 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
11703 SelectionDAG &DAG, unsigned NewOp) {
11704 DebugLoc dl = Node->getDebugLoc();
11705 assert (Node->getValueType(0) == MVT::i64 &&
11706 "Only know how to expand i64 atomics");
11708 SDValue Chain = Node->getOperand(0);
11709 SDValue In1 = Node->getOperand(1);
11710 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
11711 Node->getOperand(2), DAG.getIntPtrConstant(0));
11712 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
11713 Node->getOperand(2), DAG.getIntPtrConstant(1));
11714 SDValue Ops[] = { Chain, In1, In2L, In2H };
11715 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
11717 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
11718 cast<MemSDNode>(Node)->getMemOperand());
11719 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
11720 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
11721 Results.push_back(Result.getValue(2));
11724 /// ReplaceNodeResults - Replace a node with an illegal result type
11725 /// with a new node built out of custom code.
11726 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
11727 SmallVectorImpl<SDValue>&Results,
11728 SelectionDAG &DAG) const {
11729 DebugLoc dl = N->getDebugLoc();
11730 switch (N->getOpcode()) {
11732 llvm_unreachable("Do not know how to custom type legalize this operation!");
11733 case ISD::SIGN_EXTEND_INREG:
11738 // We don't want to expand or promote these.
11740 case ISD::FP_TO_SINT:
11741 case ISD::FP_TO_UINT: {
11742 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
11744 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
11747 std::pair<SDValue,SDValue> Vals =
11748 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
11749 SDValue FIST = Vals.first, StackSlot = Vals.second;
11750 if (FIST.getNode() != 0) {
11751 EVT VT = N->getValueType(0);
11752 // Return a load from the stack slot.
11753 if (StackSlot.getNode() != 0)
11754 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
11755 MachinePointerInfo(),
11756 false, false, false, 0));
11758 Results.push_back(FIST);
11762 case ISD::UINT_TO_FP: {
11763 if (N->getOperand(0).getValueType() != MVT::v2i32 &&
11764 N->getValueType(0) != MVT::v2f32)
11766 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
11768 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
11770 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
11771 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
11772 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
11773 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
11774 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
11775 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
11778 case ISD::FP_ROUND: {
11779 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
11780 Results.push_back(V);
11783 case ISD::READCYCLECOUNTER: {
11784 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
11785 SDValue TheChain = N->getOperand(0);
11786 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
11787 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
11789 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
11791 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
11792 SDValue Ops[] = { eax, edx };
11793 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
11794 Results.push_back(edx.getValue(1));
11797 case ISD::ATOMIC_CMP_SWAP: {
11798 EVT T = N->getValueType(0);
11799 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
11800 bool Regs64bit = T == MVT::i128;
11801 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
11802 SDValue cpInL, cpInH;
11803 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
11804 DAG.getConstant(0, HalfT));
11805 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
11806 DAG.getConstant(1, HalfT));
11807 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
11808 Regs64bit ? X86::RAX : X86::EAX,
11810 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
11811 Regs64bit ? X86::RDX : X86::EDX,
11812 cpInH, cpInL.getValue(1));
11813 SDValue swapInL, swapInH;
11814 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
11815 DAG.getConstant(0, HalfT));
11816 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
11817 DAG.getConstant(1, HalfT));
11818 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
11819 Regs64bit ? X86::RBX : X86::EBX,
11820 swapInL, cpInH.getValue(1));
11821 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
11822 Regs64bit ? X86::RCX : X86::ECX,
11823 swapInH, swapInL.getValue(1));
11824 SDValue Ops[] = { swapInH.getValue(0),
11826 swapInH.getValue(1) };
11827 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
11828 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
11829 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
11830 X86ISD::LCMPXCHG8_DAG;
11831 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
11833 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
11834 Regs64bit ? X86::RAX : X86::EAX,
11835 HalfT, Result.getValue(1));
11836 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
11837 Regs64bit ? X86::RDX : X86::EDX,
11838 HalfT, cpOutL.getValue(2));
11839 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
11840 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
11841 Results.push_back(cpOutH.getValue(1));
11844 case ISD::ATOMIC_LOAD_ADD:
11845 case ISD::ATOMIC_LOAD_AND:
11846 case ISD::ATOMIC_LOAD_NAND:
11847 case ISD::ATOMIC_LOAD_OR:
11848 case ISD::ATOMIC_LOAD_SUB:
11849 case ISD::ATOMIC_LOAD_XOR:
11850 case ISD::ATOMIC_LOAD_MAX:
11851 case ISD::ATOMIC_LOAD_MIN:
11852 case ISD::ATOMIC_LOAD_UMAX:
11853 case ISD::ATOMIC_LOAD_UMIN:
11854 case ISD::ATOMIC_SWAP: {
11856 switch (N->getOpcode()) {
11857 default: llvm_unreachable("Unexpected opcode");
11858 case ISD::ATOMIC_LOAD_ADD:
11859 Opc = X86ISD::ATOMADD64_DAG;
11861 case ISD::ATOMIC_LOAD_AND:
11862 Opc = X86ISD::ATOMAND64_DAG;
11864 case ISD::ATOMIC_LOAD_NAND:
11865 Opc = X86ISD::ATOMNAND64_DAG;
11867 case ISD::ATOMIC_LOAD_OR:
11868 Opc = X86ISD::ATOMOR64_DAG;
11870 case ISD::ATOMIC_LOAD_SUB:
11871 Opc = X86ISD::ATOMSUB64_DAG;
11873 case ISD::ATOMIC_LOAD_XOR:
11874 Opc = X86ISD::ATOMXOR64_DAG;
11876 case ISD::ATOMIC_LOAD_MAX:
11877 Opc = X86ISD::ATOMMAX64_DAG;
11879 case ISD::ATOMIC_LOAD_MIN:
11880 Opc = X86ISD::ATOMMIN64_DAG;
11882 case ISD::ATOMIC_LOAD_UMAX:
11883 Opc = X86ISD::ATOMUMAX64_DAG;
11885 case ISD::ATOMIC_LOAD_UMIN:
11886 Opc = X86ISD::ATOMUMIN64_DAG;
11888 case ISD::ATOMIC_SWAP:
11889 Opc = X86ISD::ATOMSWAP64_DAG;
11892 ReplaceATOMIC_BINARY_64(N, Results, DAG, Opc);
11895 case ISD::ATOMIC_LOAD:
11896 ReplaceATOMIC_LOAD(N, Results, DAG);
11900 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
11902 default: return NULL;
11903 case X86ISD::BSF: return "X86ISD::BSF";
11904 case X86ISD::BSR: return "X86ISD::BSR";
11905 case X86ISD::SHLD: return "X86ISD::SHLD";
11906 case X86ISD::SHRD: return "X86ISD::SHRD";
11907 case X86ISD::FAND: return "X86ISD::FAND";
11908 case X86ISD::FOR: return "X86ISD::FOR";
11909 case X86ISD::FXOR: return "X86ISD::FXOR";
11910 case X86ISD::FSRL: return "X86ISD::FSRL";
11911 case X86ISD::FILD: return "X86ISD::FILD";
11912 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
11913 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
11914 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
11915 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
11916 case X86ISD::FLD: return "X86ISD::FLD";
11917 case X86ISD::FST: return "X86ISD::FST";
11918 case X86ISD::CALL: return "X86ISD::CALL";
11919 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
11920 case X86ISD::BT: return "X86ISD::BT";
11921 case X86ISD::CMP: return "X86ISD::CMP";
11922 case X86ISD::COMI: return "X86ISD::COMI";
11923 case X86ISD::UCOMI: return "X86ISD::UCOMI";
11924 case X86ISD::SETCC: return "X86ISD::SETCC";
11925 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
11926 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
11927 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
11928 case X86ISD::CMOV: return "X86ISD::CMOV";
11929 case X86ISD::BRCOND: return "X86ISD::BRCOND";
11930 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
11931 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
11932 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
11933 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
11934 case X86ISD::Wrapper: return "X86ISD::Wrapper";
11935 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
11936 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
11937 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
11938 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
11939 case X86ISD::PINSRB: return "X86ISD::PINSRB";
11940 case X86ISD::PINSRW: return "X86ISD::PINSRW";
11941 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
11942 case X86ISD::ANDNP: return "X86ISD::ANDNP";
11943 case X86ISD::PSIGN: return "X86ISD::PSIGN";
11944 case X86ISD::BLENDV: return "X86ISD::BLENDV";
11945 case X86ISD::BLENDPW: return "X86ISD::BLENDPW";
11946 case X86ISD::BLENDPS: return "X86ISD::BLENDPS";
11947 case X86ISD::BLENDPD: return "X86ISD::BLENDPD";
11948 case X86ISD::HADD: return "X86ISD::HADD";
11949 case X86ISD::HSUB: return "X86ISD::HSUB";
11950 case X86ISD::FHADD: return "X86ISD::FHADD";
11951 case X86ISD::FHSUB: return "X86ISD::FHSUB";
11952 case X86ISD::FMAX: return "X86ISD::FMAX";
11953 case X86ISD::FMIN: return "X86ISD::FMIN";
11954 case X86ISD::FMAXC: return "X86ISD::FMAXC";
11955 case X86ISD::FMINC: return "X86ISD::FMINC";
11956 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
11957 case X86ISD::FRCP: return "X86ISD::FRCP";
11958 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
11959 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
11960 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
11961 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
11962 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
11963 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
11964 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
11965 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
11966 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
11967 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
11968 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
11969 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
11970 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
11971 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
11972 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
11973 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
11974 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
11975 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
11976 case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
11977 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
11978 case X86ISD::VZEXT: return "X86ISD::VZEXT";
11979 case X86ISD::VSEXT: return "X86ISD::VSEXT";
11980 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
11981 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
11982 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
11983 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
11984 case X86ISD::VSHL: return "X86ISD::VSHL";
11985 case X86ISD::VSRL: return "X86ISD::VSRL";
11986 case X86ISD::VSRA: return "X86ISD::VSRA";
11987 case X86ISD::VSHLI: return "X86ISD::VSHLI";
11988 case X86ISD::VSRLI: return "X86ISD::VSRLI";
11989 case X86ISD::VSRAI: return "X86ISD::VSRAI";
11990 case X86ISD::CMPP: return "X86ISD::CMPP";
11991 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
11992 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
11993 case X86ISD::ADD: return "X86ISD::ADD";
11994 case X86ISD::SUB: return "X86ISD::SUB";
11995 case X86ISD::ADC: return "X86ISD::ADC";
11996 case X86ISD::SBB: return "X86ISD::SBB";
11997 case X86ISD::SMUL: return "X86ISD::SMUL";
11998 case X86ISD::UMUL: return "X86ISD::UMUL";
11999 case X86ISD::INC: return "X86ISD::INC";
12000 case X86ISD::DEC: return "X86ISD::DEC";
12001 case X86ISD::OR: return "X86ISD::OR";
12002 case X86ISD::XOR: return "X86ISD::XOR";
12003 case X86ISD::AND: return "X86ISD::AND";
12004 case X86ISD::ANDN: return "X86ISD::ANDN";
12005 case X86ISD::BLSI: return "X86ISD::BLSI";
12006 case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
12007 case X86ISD::BLSR: return "X86ISD::BLSR";
12008 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
12009 case X86ISD::PTEST: return "X86ISD::PTEST";
12010 case X86ISD::TESTP: return "X86ISD::TESTP";
12011 case X86ISD::PALIGN: return "X86ISD::PALIGN";
12012 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
12013 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
12014 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
12015 case X86ISD::SHUFP: return "X86ISD::SHUFP";
12016 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
12017 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
12018 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
12019 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
12020 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
12021 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
12022 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
12023 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
12024 case X86ISD::MOVSD: return "X86ISD::MOVSD";
12025 case X86ISD::MOVSS: return "X86ISD::MOVSS";
12026 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
12027 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
12028 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
12029 case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
12030 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
12031 case X86ISD::VPERMV: return "X86ISD::VPERMV";
12032 case X86ISD::VPERMI: return "X86ISD::VPERMI";
12033 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
12034 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
12035 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
12036 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
12037 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
12038 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
12039 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
12040 case X86ISD::SAHF: return "X86ISD::SAHF";
12041 case X86ISD::RDRAND: return "X86ISD::RDRAND";
12042 case X86ISD::FMADD: return "X86ISD::FMADD";
12043 case X86ISD::FMSUB: return "X86ISD::FMSUB";
12044 case X86ISD::FNMADD: return "X86ISD::FNMADD";
12045 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
12046 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
12047 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
12051 // isLegalAddressingMode - Return true if the addressing mode represented
12052 // by AM is legal for this target, for a load/store of the specified type.
12053 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
12055 // X86 supports extremely general addressing modes.
12056 CodeModel::Model M = getTargetMachine().getCodeModel();
12057 Reloc::Model R = getTargetMachine().getRelocationModel();
12059 // X86 allows a sign-extended 32-bit immediate field as a displacement.
12060 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
12065 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
12067 // If a reference to this global requires an extra load, we can't fold it.
12068 if (isGlobalStubReference(GVFlags))
12071 // If BaseGV requires a register for the PIC base, we cannot also have a
12072 // BaseReg specified.
12073 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
12076 // If lower 4G is not available, then we must use rip-relative addressing.
12077 if ((M != CodeModel::Small || R != Reloc::Static) &&
12078 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
12082 switch (AM.Scale) {
12088 // These scales always work.
12093 // These scales are formed with basereg+scalereg. Only accept if there is
12098 default: // Other stuff never works.
12106 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
12107 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
12109 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
12110 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
12111 if (NumBits1 <= NumBits2)
12116 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
12117 return Imm == (int32_t)Imm;
12120 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
12121 // Can also use sub to handle negated immediates.
12122 return Imm == (int32_t)Imm;
12125 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
12126 if (!VT1.isInteger() || !VT2.isInteger())
12128 unsigned NumBits1 = VT1.getSizeInBits();
12129 unsigned NumBits2 = VT2.getSizeInBits();
12130 if (NumBits1 <= NumBits2)
12135 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
12136 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
12137 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
12140 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
12141 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
12142 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
12145 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
12146 // i16 instructions are longer (0x66 prefix) and potentially slower.
12147 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
12150 /// isShuffleMaskLegal - Targets can use this to indicate that they only
12151 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
12152 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
12153 /// are assumed to be legal.
12155 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
12157 // Very little shuffling can be done for 64-bit vectors right now.
12158 if (VT.getSizeInBits() == 64)
12161 // FIXME: pshufb, blends, shifts.
12162 return (VT.getVectorNumElements() == 2 ||
12163 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
12164 isMOVLMask(M, VT) ||
12165 isSHUFPMask(M, VT, Subtarget->hasAVX()) ||
12166 isPSHUFDMask(M, VT) ||
12167 isPSHUFHWMask(M, VT, Subtarget->hasAVX2()) ||
12168 isPSHUFLWMask(M, VT, Subtarget->hasAVX2()) ||
12169 isPALIGNRMask(M, VT, Subtarget) ||
12170 isUNPCKLMask(M, VT, Subtarget->hasAVX2()) ||
12171 isUNPCKHMask(M, VT, Subtarget->hasAVX2()) ||
12172 isUNPCKL_v_undef_Mask(M, VT, Subtarget->hasAVX2()) ||
12173 isUNPCKH_v_undef_Mask(M, VT, Subtarget->hasAVX2()));
12177 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
12179 unsigned NumElts = VT.getVectorNumElements();
12180 // FIXME: This collection of masks seems suspect.
12183 if (NumElts == 4 && VT.is128BitVector()) {
12184 return (isMOVLMask(Mask, VT) ||
12185 isCommutedMOVLMask(Mask, VT, true) ||
12186 isSHUFPMask(Mask, VT, Subtarget->hasAVX()) ||
12187 isSHUFPMask(Mask, VT, Subtarget->hasAVX(), /* Commuted */ true));
12192 //===----------------------------------------------------------------------===//
12193 // X86 Scheduler Hooks
12194 //===----------------------------------------------------------------------===//
12196 // private utility function
12198 /// Utility function to emit xbegin specifying the start of an RTM region.
12199 MachineBasicBlock *
12200 X86TargetLowering::EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB) const {
12201 DebugLoc DL = MI->getDebugLoc();
12202 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12204 const BasicBlock *BB = MBB->getBasicBlock();
12205 MachineFunction::iterator I = MBB;
12208 // For the v = xbegin(), we generate
12219 MachineBasicBlock *thisMBB = MBB;
12220 MachineFunction *MF = MBB->getParent();
12221 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
12222 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
12223 MF->insert(I, mainMBB);
12224 MF->insert(I, sinkMBB);
12226 // Transfer the remainder of BB and its successor edges to sinkMBB.
12227 sinkMBB->splice(sinkMBB->begin(), MBB,
12228 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
12229 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
12233 // # fallthrough to mainMBB
12234 // # abortion to sinkMBB
12235 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
12236 thisMBB->addSuccessor(mainMBB);
12237 thisMBB->addSuccessor(sinkMBB);
12241 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
12242 mainMBB->addSuccessor(sinkMBB);
12245 // EAX is live into the sinkMBB
12246 sinkMBB->addLiveIn(X86::EAX);
12247 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
12248 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
12251 MI->eraseFromParent();
12255 // Get CMPXCHG opcode for the specified data type.
12256 static unsigned getCmpXChgOpcode(EVT VT) {
12257 switch (VT.getSimpleVT().SimpleTy) {
12258 case MVT::i8: return X86::LCMPXCHG8;
12259 case MVT::i16: return X86::LCMPXCHG16;
12260 case MVT::i32: return X86::LCMPXCHG32;
12261 case MVT::i64: return X86::LCMPXCHG64;
12265 llvm_unreachable("Invalid operand size!");
12268 // Get LOAD opcode for the specified data type.
12269 static unsigned getLoadOpcode(EVT VT) {
12270 switch (VT.getSimpleVT().SimpleTy) {
12271 case MVT::i8: return X86::MOV8rm;
12272 case MVT::i16: return X86::MOV16rm;
12273 case MVT::i32: return X86::MOV32rm;
12274 case MVT::i64: return X86::MOV64rm;
12278 llvm_unreachable("Invalid operand size!");
12281 // Get opcode of the non-atomic one from the specified atomic instruction.
12282 static unsigned getNonAtomicOpcode(unsigned Opc) {
12284 case X86::ATOMAND8: return X86::AND8rr;
12285 case X86::ATOMAND16: return X86::AND16rr;
12286 case X86::ATOMAND32: return X86::AND32rr;
12287 case X86::ATOMAND64: return X86::AND64rr;
12288 case X86::ATOMOR8: return X86::OR8rr;
12289 case X86::ATOMOR16: return X86::OR16rr;
12290 case X86::ATOMOR32: return X86::OR32rr;
12291 case X86::ATOMOR64: return X86::OR64rr;
12292 case X86::ATOMXOR8: return X86::XOR8rr;
12293 case X86::ATOMXOR16: return X86::XOR16rr;
12294 case X86::ATOMXOR32: return X86::XOR32rr;
12295 case X86::ATOMXOR64: return X86::XOR64rr;
12297 llvm_unreachable("Unhandled atomic-load-op opcode!");
12300 // Get opcode of the non-atomic one from the specified atomic instruction with
12302 static unsigned getNonAtomicOpcodeWithExtraOpc(unsigned Opc,
12303 unsigned &ExtraOpc) {
12305 case X86::ATOMNAND8: ExtraOpc = X86::NOT8r; return X86::AND8rr;
12306 case X86::ATOMNAND16: ExtraOpc = X86::NOT16r; return X86::AND16rr;
12307 case X86::ATOMNAND32: ExtraOpc = X86::NOT32r; return X86::AND32rr;
12308 case X86::ATOMNAND64: ExtraOpc = X86::NOT64r; return X86::AND64rr;
12309 case X86::ATOMMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVL32rr;
12310 case X86::ATOMMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVL16rr;
12311 case X86::ATOMMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVL32rr;
12312 case X86::ATOMMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVL64rr;
12313 case X86::ATOMMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVG32rr;
12314 case X86::ATOMMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVG16rr;
12315 case X86::ATOMMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVG32rr;
12316 case X86::ATOMMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVG64rr;
12317 case X86::ATOMUMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVB32rr;
12318 case X86::ATOMUMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVB16rr;
12319 case X86::ATOMUMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVB32rr;
12320 case X86::ATOMUMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVB64rr;
12321 case X86::ATOMUMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVA32rr;
12322 case X86::ATOMUMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVA16rr;
12323 case X86::ATOMUMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVA32rr;
12324 case X86::ATOMUMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVA64rr;
12326 llvm_unreachable("Unhandled atomic-load-op opcode!");
12329 // Get opcode of the non-atomic one from the specified atomic instruction for
12330 // 64-bit data type on 32-bit target.
12331 static unsigned getNonAtomic6432Opcode(unsigned Opc, unsigned &HiOpc) {
12333 case X86::ATOMAND6432: HiOpc = X86::AND32rr; return X86::AND32rr;
12334 case X86::ATOMOR6432: HiOpc = X86::OR32rr; return X86::OR32rr;
12335 case X86::ATOMXOR6432: HiOpc = X86::XOR32rr; return X86::XOR32rr;
12336 case X86::ATOMADD6432: HiOpc = X86::ADC32rr; return X86::ADD32rr;
12337 case X86::ATOMSUB6432: HiOpc = X86::SBB32rr; return X86::SUB32rr;
12338 case X86::ATOMSWAP6432: HiOpc = X86::MOV32rr; return X86::MOV32rr;
12339 case X86::ATOMMAX6432: HiOpc = X86::SETLr; return X86::SETLr;
12340 case X86::ATOMMIN6432: HiOpc = X86::SETGr; return X86::SETGr;
12341 case X86::ATOMUMAX6432: HiOpc = X86::SETBr; return X86::SETBr;
12342 case X86::ATOMUMIN6432: HiOpc = X86::SETAr; return X86::SETAr;
12344 llvm_unreachable("Unhandled atomic-load-op opcode!");
12347 // Get opcode of the non-atomic one from the specified atomic instruction for
12348 // 64-bit data type on 32-bit target with extra opcode.
12349 static unsigned getNonAtomic6432OpcodeWithExtraOpc(unsigned Opc,
12351 unsigned &ExtraOpc) {
12353 case X86::ATOMNAND6432:
12354 ExtraOpc = X86::NOT32r;
12355 HiOpc = X86::AND32rr;
12356 return X86::AND32rr;
12358 llvm_unreachable("Unhandled atomic-load-op opcode!");
12361 // Get pseudo CMOV opcode from the specified data type.
12362 static unsigned getPseudoCMOVOpc(EVT VT) {
12363 switch (VT.getSimpleVT().SimpleTy) {
12364 case MVT::i8: return X86::CMOV_GR8;
12365 case MVT::i16: return X86::CMOV_GR16;
12366 case MVT::i32: return X86::CMOV_GR32;
12370 llvm_unreachable("Unknown CMOV opcode!");
12373 // EmitAtomicLoadArith - emit the code sequence for pseudo atomic instructions.
12374 // They will be translated into a spin-loop or compare-exchange loop from
12377 // dst = atomic-fetch-op MI.addr, MI.val
12383 // EAX = LOAD MI.addr
12385 // t1 = OP MI.val, EAX
12386 // LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
12391 MachineBasicBlock *
12392 X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
12393 MachineBasicBlock *MBB) const {
12394 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12395 DebugLoc DL = MI->getDebugLoc();
12397 MachineFunction *MF = MBB->getParent();
12398 MachineRegisterInfo &MRI = MF->getRegInfo();
12400 const BasicBlock *BB = MBB->getBasicBlock();
12401 MachineFunction::iterator I = MBB;
12404 assert(MI->getNumOperands() <= X86::AddrNumOperands + 2 &&
12405 "Unexpected number of operands");
12407 assert(MI->hasOneMemOperand() &&
12408 "Expected atomic-load-op to have one memoperand");
12410 // Memory Reference
12411 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
12412 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
12414 unsigned DstReg, SrcReg;
12415 unsigned MemOpndSlot;
12417 unsigned CurOp = 0;
12419 DstReg = MI->getOperand(CurOp++).getReg();
12420 MemOpndSlot = CurOp;
12421 CurOp += X86::AddrNumOperands;
12422 SrcReg = MI->getOperand(CurOp++).getReg();
12424 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
12425 MVT::SimpleValueType VT = *RC->vt_begin();
12426 unsigned AccPhyReg = getX86SubSuperRegister(X86::EAX, VT);
12428 unsigned LCMPXCHGOpc = getCmpXChgOpcode(VT);
12429 unsigned LOADOpc = getLoadOpcode(VT);
12431 // For the atomic load-arith operator, we generate
12434 // EAX = LOAD [MI.addr]
12436 // t1 = OP MI.val, EAX
12437 // LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
12441 MachineBasicBlock *thisMBB = MBB;
12442 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
12443 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
12444 MF->insert(I, mainMBB);
12445 MF->insert(I, sinkMBB);
12447 MachineInstrBuilder MIB;
12449 // Transfer the remainder of BB and its successor edges to sinkMBB.
12450 sinkMBB->splice(sinkMBB->begin(), MBB,
12451 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
12452 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
12455 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), AccPhyReg);
12456 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
12457 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
12458 MIB.setMemRefs(MMOBegin, MMOEnd);
12460 thisMBB->addSuccessor(mainMBB);
12463 MachineBasicBlock *origMainMBB = mainMBB;
12464 mainMBB->addLiveIn(AccPhyReg);
12466 // Copy AccPhyReg as it is used more than once.
12467 unsigned AccReg = MRI.createVirtualRegister(RC);
12468 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), AccReg)
12469 .addReg(AccPhyReg);
12471 unsigned t1 = MRI.createVirtualRegister(RC);
12472 unsigned Opc = MI->getOpcode();
12475 llvm_unreachable("Unhandled atomic-load-op opcode!");
12476 case X86::ATOMAND8:
12477 case X86::ATOMAND16:
12478 case X86::ATOMAND32:
12479 case X86::ATOMAND64:
12481 case X86::ATOMOR16:
12482 case X86::ATOMOR32:
12483 case X86::ATOMOR64:
12484 case X86::ATOMXOR8:
12485 case X86::ATOMXOR16:
12486 case X86::ATOMXOR32:
12487 case X86::ATOMXOR64: {
12488 unsigned ARITHOpc = getNonAtomicOpcode(Opc);
12489 BuildMI(mainMBB, DL, TII->get(ARITHOpc), t1).addReg(SrcReg)
12493 case X86::ATOMNAND8:
12494 case X86::ATOMNAND16:
12495 case X86::ATOMNAND32:
12496 case X86::ATOMNAND64: {
12497 unsigned t2 = MRI.createVirtualRegister(RC);
12499 unsigned ANDOpc = getNonAtomicOpcodeWithExtraOpc(Opc, NOTOpc);
12500 BuildMI(mainMBB, DL, TII->get(ANDOpc), t2).addReg(SrcReg)
12502 BuildMI(mainMBB, DL, TII->get(NOTOpc), t1).addReg(t2);
12505 case X86::ATOMMAX8:
12506 case X86::ATOMMAX16:
12507 case X86::ATOMMAX32:
12508 case X86::ATOMMAX64:
12509 case X86::ATOMMIN8:
12510 case X86::ATOMMIN16:
12511 case X86::ATOMMIN32:
12512 case X86::ATOMMIN64:
12513 case X86::ATOMUMAX8:
12514 case X86::ATOMUMAX16:
12515 case X86::ATOMUMAX32:
12516 case X86::ATOMUMAX64:
12517 case X86::ATOMUMIN8:
12518 case X86::ATOMUMIN16:
12519 case X86::ATOMUMIN32:
12520 case X86::ATOMUMIN64: {
12522 unsigned CMOVOpc = getNonAtomicOpcodeWithExtraOpc(Opc, CMPOpc);
12524 BuildMI(mainMBB, DL, TII->get(CMPOpc))
12528 if (Subtarget->hasCMov()) {
12529 if (VT != MVT::i8) {
12531 BuildMI(mainMBB, DL, TII->get(CMOVOpc), t1)
12535 // Promote i8 to i32 to use CMOV32
12536 const TargetRegisterClass *RC32 = getRegClassFor(MVT::i32);
12537 unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
12538 unsigned AccReg32 = MRI.createVirtualRegister(RC32);
12539 unsigned t2 = MRI.createVirtualRegister(RC32);
12541 unsigned Undef = MRI.createVirtualRegister(RC32);
12542 BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
12544 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), SrcReg32)
12547 .addImm(X86::sub_8bit);
12548 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
12551 .addImm(X86::sub_8bit);
12553 BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
12557 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t1)
12558 .addReg(t2, 0, X86::sub_8bit);
12561 // Use pseudo select and lower them.
12562 assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
12563 "Invalid atomic-load-op transformation!");
12564 unsigned SelOpc = getPseudoCMOVOpc(VT);
12565 X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
12566 assert(CC != X86::COND_INVALID && "Invalid atomic-load-op transformation!");
12567 MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t1)
12568 .addReg(SrcReg).addReg(AccReg)
12570 mainMBB = EmitLoweredSelect(MIB, mainMBB);
12576 // Copy AccPhyReg back from virtual register.
12577 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), AccPhyReg)
12580 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
12581 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
12582 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
12584 MIB.setMemRefs(MMOBegin, MMOEnd);
12586 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
12588 mainMBB->addSuccessor(origMainMBB);
12589 mainMBB->addSuccessor(sinkMBB);
12592 sinkMBB->addLiveIn(AccPhyReg);
12594 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
12595 TII->get(TargetOpcode::COPY), DstReg)
12596 .addReg(AccPhyReg);
12598 MI->eraseFromParent();
12602 // EmitAtomicLoadArith6432 - emit the code sequence for pseudo atomic
12603 // instructions. They will be translated into a spin-loop or compare-exchange
12607 // dst = atomic-fetch-op MI.addr, MI.val
12613 // EAX = LOAD [MI.addr + 0]
12614 // EDX = LOAD [MI.addr + 4]
12616 // EBX = OP MI.val.lo, EAX
12617 // ECX = OP MI.val.hi, EDX
12618 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
12623 MachineBasicBlock *
12624 X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
12625 MachineBasicBlock *MBB) const {
12626 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12627 DebugLoc DL = MI->getDebugLoc();
12629 MachineFunction *MF = MBB->getParent();
12630 MachineRegisterInfo &MRI = MF->getRegInfo();
12632 const BasicBlock *BB = MBB->getBasicBlock();
12633 MachineFunction::iterator I = MBB;
12636 assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
12637 "Unexpected number of operands");
12639 assert(MI->hasOneMemOperand() &&
12640 "Expected atomic-load-op32 to have one memoperand");
12642 // Memory Reference
12643 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
12644 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
12646 unsigned DstLoReg, DstHiReg;
12647 unsigned SrcLoReg, SrcHiReg;
12648 unsigned MemOpndSlot;
12650 unsigned CurOp = 0;
12652 DstLoReg = MI->getOperand(CurOp++).getReg();
12653 DstHiReg = MI->getOperand(CurOp++).getReg();
12654 MemOpndSlot = CurOp;
12655 CurOp += X86::AddrNumOperands;
12656 SrcLoReg = MI->getOperand(CurOp++).getReg();
12657 SrcHiReg = MI->getOperand(CurOp++).getReg();
12659 const TargetRegisterClass *RC = &X86::GR32RegClass;
12660 const TargetRegisterClass *RC8 = &X86::GR8RegClass;
12662 unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
12663 unsigned LOADOpc = X86::MOV32rm;
12665 // For the atomic load-arith operator, we generate
12668 // EAX = LOAD [MI.addr + 0]
12669 // EDX = LOAD [MI.addr + 4]
12671 // EBX = OP MI.vallo, EAX
12672 // ECX = OP MI.valhi, EDX
12673 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
12677 MachineBasicBlock *thisMBB = MBB;
12678 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
12679 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
12680 MF->insert(I, mainMBB);
12681 MF->insert(I, sinkMBB);
12683 MachineInstrBuilder MIB;
12685 // Transfer the remainder of BB and its successor edges to sinkMBB.
12686 sinkMBB->splice(sinkMBB->begin(), MBB,
12687 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
12688 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
12692 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), X86::EAX);
12693 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
12694 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
12695 MIB.setMemRefs(MMOBegin, MMOEnd);
12697 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), X86::EDX);
12698 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
12699 if (i == X86::AddrDisp)
12700 MIB.addDisp(MI->getOperand(MemOpndSlot + i), 4); // 4 == sizeof(i32)
12702 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
12704 MIB.setMemRefs(MMOBegin, MMOEnd);
12706 thisMBB->addSuccessor(mainMBB);
12709 MachineBasicBlock *origMainMBB = mainMBB;
12710 mainMBB->addLiveIn(X86::EAX);
12711 mainMBB->addLiveIn(X86::EDX);
12713 // Copy EDX:EAX as they are used more than once.
12714 unsigned LoReg = MRI.createVirtualRegister(RC);
12715 unsigned HiReg = MRI.createVirtualRegister(RC);
12716 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), LoReg).addReg(X86::EAX);
12717 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), HiReg).addReg(X86::EDX);
12719 unsigned t1L = MRI.createVirtualRegister(RC);
12720 unsigned t1H = MRI.createVirtualRegister(RC);
12722 unsigned Opc = MI->getOpcode();
12725 llvm_unreachable("Unhandled atomic-load-op6432 opcode!");
12726 case X86::ATOMAND6432:
12727 case X86::ATOMOR6432:
12728 case X86::ATOMXOR6432:
12729 case X86::ATOMADD6432:
12730 case X86::ATOMSUB6432: {
12732 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
12733 BuildMI(mainMBB, DL, TII->get(LoOpc), t1L).addReg(SrcLoReg).addReg(LoReg);
12734 BuildMI(mainMBB, DL, TII->get(HiOpc), t1H).addReg(SrcHiReg).addReg(HiReg);
12737 case X86::ATOMNAND6432: {
12738 unsigned HiOpc, NOTOpc;
12739 unsigned LoOpc = getNonAtomic6432OpcodeWithExtraOpc(Opc, HiOpc, NOTOpc);
12740 unsigned t2L = MRI.createVirtualRegister(RC);
12741 unsigned t2H = MRI.createVirtualRegister(RC);
12742 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg).addReg(LoReg);
12743 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg).addReg(HiReg);
12744 BuildMI(mainMBB, DL, TII->get(NOTOpc), t1L).addReg(t2L);
12745 BuildMI(mainMBB, DL, TII->get(NOTOpc), t1H).addReg(t2H);
12748 case X86::ATOMMAX6432:
12749 case X86::ATOMMIN6432:
12750 case X86::ATOMUMAX6432:
12751 case X86::ATOMUMIN6432: {
12753 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
12754 unsigned cL = MRI.createVirtualRegister(RC8);
12755 unsigned cH = MRI.createVirtualRegister(RC8);
12756 unsigned cL32 = MRI.createVirtualRegister(RC);
12757 unsigned cH32 = MRI.createVirtualRegister(RC);
12758 unsigned cc = MRI.createVirtualRegister(RC);
12759 // cl := cmp src_lo, lo
12760 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
12761 .addReg(SrcLoReg).addReg(LoReg);
12762 BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
12763 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
12764 // ch := cmp src_hi, hi
12765 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
12766 .addReg(SrcHiReg).addReg(HiReg);
12767 BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
12768 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
12769 // cc := if (src_hi == hi) ? cl : ch;
12770 if (Subtarget->hasCMov()) {
12771 BuildMI(mainMBB, DL, TII->get(X86::CMOVE32rr), cc)
12772 .addReg(cH32).addReg(cL32);
12774 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), cc)
12775 .addReg(cH32).addReg(cL32)
12776 .addImm(X86::COND_E);
12777 mainMBB = EmitLoweredSelect(MIB, mainMBB);
12779 BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
12780 if (Subtarget->hasCMov()) {
12781 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1L)
12782 .addReg(SrcLoReg).addReg(LoReg);
12783 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1H)
12784 .addReg(SrcHiReg).addReg(HiReg);
12786 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1L)
12787 .addReg(SrcLoReg).addReg(LoReg)
12788 .addImm(X86::COND_NE);
12789 mainMBB = EmitLoweredSelect(MIB, mainMBB);
12790 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1H)
12791 .addReg(SrcHiReg).addReg(HiReg)
12792 .addImm(X86::COND_NE);
12793 mainMBB = EmitLoweredSelect(MIB, mainMBB);
12797 case X86::ATOMSWAP6432: {
12799 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
12800 BuildMI(mainMBB, DL, TII->get(LoOpc), t1L).addReg(SrcLoReg);
12801 BuildMI(mainMBB, DL, TII->get(HiOpc), t1H).addReg(SrcHiReg);
12806 // Copy EDX:EAX back from HiReg:LoReg
12807 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(LoReg);
12808 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(HiReg);
12809 // Copy ECX:EBX from t1H:t1L
12810 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t1L);
12811 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t1H);
12813 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
12814 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
12815 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
12816 MIB.setMemRefs(MMOBegin, MMOEnd);
12818 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
12820 mainMBB->addSuccessor(origMainMBB);
12821 mainMBB->addSuccessor(sinkMBB);
12824 sinkMBB->addLiveIn(X86::EAX);
12825 sinkMBB->addLiveIn(X86::EDX);
12827 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
12828 TII->get(TargetOpcode::COPY), DstLoReg)
12830 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
12831 TII->get(TargetOpcode::COPY), DstHiReg)
12834 MI->eraseFromParent();
12838 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
12839 // or XMM0_V32I8 in AVX all of this code can be replaced with that
12840 // in the .td file.
12841 MachineBasicBlock *
12842 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
12843 unsigned numArgs, bool memArg) const {
12844 assert(Subtarget->hasSSE42() &&
12845 "Target must have SSE4.2 or AVX features enabled");
12847 DebugLoc dl = MI->getDebugLoc();
12848 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12850 if (!Subtarget->hasAVX()) {
12852 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
12854 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
12857 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
12859 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
12862 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
12863 for (unsigned i = 0; i < numArgs; ++i) {
12864 MachineOperand &Op = MI->getOperand(i+1);
12865 if (!(Op.isReg() && Op.isImplicit()))
12866 MIB.addOperand(Op);
12868 BuildMI(*BB, MI, dl,
12869 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
12870 .addReg(X86::XMM0);
12872 MI->eraseFromParent();
12876 MachineBasicBlock *
12877 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
12878 DebugLoc dl = MI->getDebugLoc();
12879 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12881 // Address into RAX/EAX, other two args into ECX, EDX.
12882 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
12883 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12884 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
12885 for (int i = 0; i < X86::AddrNumOperands; ++i)
12886 MIB.addOperand(MI->getOperand(i));
12888 unsigned ValOps = X86::AddrNumOperands;
12889 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
12890 .addReg(MI->getOperand(ValOps).getReg());
12891 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
12892 .addReg(MI->getOperand(ValOps+1).getReg());
12894 // The instruction doesn't actually take any operands though.
12895 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
12897 MI->eraseFromParent(); // The pseudo is gone now.
12901 MachineBasicBlock *
12902 X86TargetLowering::EmitVAARG64WithCustomInserter(
12904 MachineBasicBlock *MBB) const {
12905 // Emit va_arg instruction on X86-64.
12907 // Operands to this pseudo-instruction:
12908 // 0 ) Output : destination address (reg)
12909 // 1-5) Input : va_list address (addr, i64mem)
12910 // 6 ) ArgSize : Size (in bytes) of vararg type
12911 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
12912 // 8 ) Align : Alignment of type
12913 // 9 ) EFLAGS (implicit-def)
12915 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
12916 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
12918 unsigned DestReg = MI->getOperand(0).getReg();
12919 MachineOperand &Base = MI->getOperand(1);
12920 MachineOperand &Scale = MI->getOperand(2);
12921 MachineOperand &Index = MI->getOperand(3);
12922 MachineOperand &Disp = MI->getOperand(4);
12923 MachineOperand &Segment = MI->getOperand(5);
12924 unsigned ArgSize = MI->getOperand(6).getImm();
12925 unsigned ArgMode = MI->getOperand(7).getImm();
12926 unsigned Align = MI->getOperand(8).getImm();
12928 // Memory Reference
12929 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
12930 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
12931 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
12933 // Machine Information
12934 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
12935 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
12936 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
12937 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
12938 DebugLoc DL = MI->getDebugLoc();
12940 // struct va_list {
12943 // i64 overflow_area (address)
12944 // i64 reg_save_area (address)
12946 // sizeof(va_list) = 24
12947 // alignment(va_list) = 8
12949 unsigned TotalNumIntRegs = 6;
12950 unsigned TotalNumXMMRegs = 8;
12951 bool UseGPOffset = (ArgMode == 1);
12952 bool UseFPOffset = (ArgMode == 2);
12953 unsigned MaxOffset = TotalNumIntRegs * 8 +
12954 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
12956 /* Align ArgSize to a multiple of 8 */
12957 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
12958 bool NeedsAlign = (Align > 8);
12960 MachineBasicBlock *thisMBB = MBB;
12961 MachineBasicBlock *overflowMBB;
12962 MachineBasicBlock *offsetMBB;
12963 MachineBasicBlock *endMBB;
12965 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
12966 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
12967 unsigned OffsetReg = 0;
12969 if (!UseGPOffset && !UseFPOffset) {
12970 // If we only pull from the overflow region, we don't create a branch.
12971 // We don't need to alter control flow.
12972 OffsetDestReg = 0; // unused
12973 OverflowDestReg = DestReg;
12976 overflowMBB = thisMBB;
12979 // First emit code to check if gp_offset (or fp_offset) is below the bound.
12980 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
12981 // If not, pull from overflow_area. (branch to overflowMBB)
12986 // offsetMBB overflowMBB
12991 // Registers for the PHI in endMBB
12992 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
12993 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
12995 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
12996 MachineFunction *MF = MBB->getParent();
12997 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
12998 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
12999 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13001 MachineFunction::iterator MBBIter = MBB;
13004 // Insert the new basic blocks
13005 MF->insert(MBBIter, offsetMBB);
13006 MF->insert(MBBIter, overflowMBB);
13007 MF->insert(MBBIter, endMBB);
13009 // Transfer the remainder of MBB and its successor edges to endMBB.
13010 endMBB->splice(endMBB->begin(), thisMBB,
13011 llvm::next(MachineBasicBlock::iterator(MI)),
13013 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
13015 // Make offsetMBB and overflowMBB successors of thisMBB
13016 thisMBB->addSuccessor(offsetMBB);
13017 thisMBB->addSuccessor(overflowMBB);
13019 // endMBB is a successor of both offsetMBB and overflowMBB
13020 offsetMBB->addSuccessor(endMBB);
13021 overflowMBB->addSuccessor(endMBB);
13023 // Load the offset value into a register
13024 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
13025 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
13029 .addDisp(Disp, UseFPOffset ? 4 : 0)
13030 .addOperand(Segment)
13031 .setMemRefs(MMOBegin, MMOEnd);
13033 // Check if there is enough room left to pull this argument.
13034 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
13036 .addImm(MaxOffset + 8 - ArgSizeA8);
13038 // Branch to "overflowMBB" if offset >= max
13039 // Fall through to "offsetMBB" otherwise
13040 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
13041 .addMBB(overflowMBB);
13044 // In offsetMBB, emit code to use the reg_save_area.
13046 assert(OffsetReg != 0);
13048 // Read the reg_save_area address.
13049 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
13050 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
13055 .addOperand(Segment)
13056 .setMemRefs(MMOBegin, MMOEnd);
13058 // Zero-extend the offset
13059 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
13060 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
13063 .addImm(X86::sub_32bit);
13065 // Add the offset to the reg_save_area to get the final address.
13066 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
13067 .addReg(OffsetReg64)
13068 .addReg(RegSaveReg);
13070 // Compute the offset for the next argument
13071 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
13072 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
13074 .addImm(UseFPOffset ? 16 : 8);
13076 // Store it back into the va_list.
13077 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
13081 .addDisp(Disp, UseFPOffset ? 4 : 0)
13082 .addOperand(Segment)
13083 .addReg(NextOffsetReg)
13084 .setMemRefs(MMOBegin, MMOEnd);
13087 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
13092 // Emit code to use overflow area
13095 // Load the overflow_area address into a register.
13096 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
13097 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
13102 .addOperand(Segment)
13103 .setMemRefs(MMOBegin, MMOEnd);
13105 // If we need to align it, do so. Otherwise, just copy the address
13106 // to OverflowDestReg.
13108 // Align the overflow address
13109 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
13110 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
13112 // aligned_addr = (addr + (align-1)) & ~(align-1)
13113 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
13114 .addReg(OverflowAddrReg)
13117 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
13119 .addImm(~(uint64_t)(Align-1));
13121 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
13122 .addReg(OverflowAddrReg);
13125 // Compute the next overflow address after this argument.
13126 // (the overflow address should be kept 8-byte aligned)
13127 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
13128 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
13129 .addReg(OverflowDestReg)
13130 .addImm(ArgSizeA8);
13132 // Store the new overflow address.
13133 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
13138 .addOperand(Segment)
13139 .addReg(NextAddrReg)
13140 .setMemRefs(MMOBegin, MMOEnd);
13142 // If we branched, emit the PHI to the front of endMBB.
13144 BuildMI(*endMBB, endMBB->begin(), DL,
13145 TII->get(X86::PHI), DestReg)
13146 .addReg(OffsetDestReg).addMBB(offsetMBB)
13147 .addReg(OverflowDestReg).addMBB(overflowMBB);
13150 // Erase the pseudo instruction
13151 MI->eraseFromParent();
13156 MachineBasicBlock *
13157 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
13159 MachineBasicBlock *MBB) const {
13160 // Emit code to save XMM registers to the stack. The ABI says that the
13161 // number of registers to save is given in %al, so it's theoretically
13162 // possible to do an indirect jump trick to avoid saving all of them,
13163 // however this code takes a simpler approach and just executes all
13164 // of the stores if %al is non-zero. It's less code, and it's probably
13165 // easier on the hardware branch predictor, and stores aren't all that
13166 // expensive anyway.
13168 // Create the new basic blocks. One block contains all the XMM stores,
13169 // and one block is the final destination regardless of whether any
13170 // stores were performed.
13171 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
13172 MachineFunction *F = MBB->getParent();
13173 MachineFunction::iterator MBBIter = MBB;
13175 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
13176 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
13177 F->insert(MBBIter, XMMSaveMBB);
13178 F->insert(MBBIter, EndMBB);
13180 // Transfer the remainder of MBB and its successor edges to EndMBB.
13181 EndMBB->splice(EndMBB->begin(), MBB,
13182 llvm::next(MachineBasicBlock::iterator(MI)),
13184 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
13186 // The original block will now fall through to the XMM save block.
13187 MBB->addSuccessor(XMMSaveMBB);
13188 // The XMMSaveMBB will fall through to the end block.
13189 XMMSaveMBB->addSuccessor(EndMBB);
13191 // Now add the instructions.
13192 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13193 DebugLoc DL = MI->getDebugLoc();
13195 unsigned CountReg = MI->getOperand(0).getReg();
13196 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
13197 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
13199 if (!Subtarget->isTargetWin64()) {
13200 // If %al is 0, branch around the XMM save block.
13201 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
13202 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
13203 MBB->addSuccessor(EndMBB);
13206 unsigned MOVOpc = Subtarget->hasAVX() ? X86::VMOVAPSmr : X86::MOVAPSmr;
13207 // In the XMM save block, save all the XMM argument registers.
13208 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
13209 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
13210 MachineMemOperand *MMO =
13211 F->getMachineMemOperand(
13212 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
13213 MachineMemOperand::MOStore,
13214 /*Size=*/16, /*Align=*/16);
13215 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
13216 .addFrameIndex(RegSaveFrameIndex)
13217 .addImm(/*Scale=*/1)
13218 .addReg(/*IndexReg=*/0)
13219 .addImm(/*Disp=*/Offset)
13220 .addReg(/*Segment=*/0)
13221 .addReg(MI->getOperand(i).getReg())
13222 .addMemOperand(MMO);
13225 MI->eraseFromParent(); // The pseudo instruction is gone now.
13230 // The EFLAGS operand of SelectItr might be missing a kill marker
13231 // because there were multiple uses of EFLAGS, and ISel didn't know
13232 // which to mark. Figure out whether SelectItr should have had a
13233 // kill marker, and set it if it should. Returns the correct kill
13235 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
13236 MachineBasicBlock* BB,
13237 const TargetRegisterInfo* TRI) {
13238 // Scan forward through BB for a use/def of EFLAGS.
13239 MachineBasicBlock::iterator miI(llvm::next(SelectItr));
13240 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
13241 const MachineInstr& mi = *miI;
13242 if (mi.readsRegister(X86::EFLAGS))
13244 if (mi.definesRegister(X86::EFLAGS))
13245 break; // Should have kill-flag - update below.
13248 // If we hit the end of the block, check whether EFLAGS is live into a
13250 if (miI == BB->end()) {
13251 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
13252 sEnd = BB->succ_end();
13253 sItr != sEnd; ++sItr) {
13254 MachineBasicBlock* succ = *sItr;
13255 if (succ->isLiveIn(X86::EFLAGS))
13260 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
13261 // out. SelectMI should have a kill flag on EFLAGS.
13262 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
13266 MachineBasicBlock *
13267 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
13268 MachineBasicBlock *BB) const {
13269 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13270 DebugLoc DL = MI->getDebugLoc();
13272 // To "insert" a SELECT_CC instruction, we actually have to insert the
13273 // diamond control-flow pattern. The incoming instruction knows the
13274 // destination vreg to set, the condition code register to branch on, the
13275 // true/false values to select between, and a branch opcode to use.
13276 const BasicBlock *LLVM_BB = BB->getBasicBlock();
13277 MachineFunction::iterator It = BB;
13283 // cmpTY ccX, r1, r2
13285 // fallthrough --> copy0MBB
13286 MachineBasicBlock *thisMBB = BB;
13287 MachineFunction *F = BB->getParent();
13288 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
13289 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
13290 F->insert(It, copy0MBB);
13291 F->insert(It, sinkMBB);
13293 // If the EFLAGS register isn't dead in the terminator, then claim that it's
13294 // live into the sink and copy blocks.
13295 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
13296 if (!MI->killsRegister(X86::EFLAGS) &&
13297 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
13298 copy0MBB->addLiveIn(X86::EFLAGS);
13299 sinkMBB->addLiveIn(X86::EFLAGS);
13302 // Transfer the remainder of BB and its successor edges to sinkMBB.
13303 sinkMBB->splice(sinkMBB->begin(), BB,
13304 llvm::next(MachineBasicBlock::iterator(MI)),
13306 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
13308 // Add the true and fallthrough blocks as its successors.
13309 BB->addSuccessor(copy0MBB);
13310 BB->addSuccessor(sinkMBB);
13312 // Create the conditional branch instruction.
13314 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
13315 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
13318 // %FalseValue = ...
13319 // # fallthrough to sinkMBB
13320 copy0MBB->addSuccessor(sinkMBB);
13323 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
13325 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13326 TII->get(X86::PHI), MI->getOperand(0).getReg())
13327 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
13328 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
13330 MI->eraseFromParent(); // The pseudo instruction is gone now.
13334 MachineBasicBlock *
13335 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
13336 bool Is64Bit) const {
13337 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13338 DebugLoc DL = MI->getDebugLoc();
13339 MachineFunction *MF = BB->getParent();
13340 const BasicBlock *LLVM_BB = BB->getBasicBlock();
13342 assert(getTargetMachine().Options.EnableSegmentedStacks);
13344 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
13345 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
13348 // ... [Till the alloca]
13349 // If stacklet is not large enough, jump to mallocMBB
13352 // Allocate by subtracting from RSP
13353 // Jump to continueMBB
13356 // Allocate by call to runtime
13360 // [rest of original BB]
13363 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13364 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13365 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13367 MachineRegisterInfo &MRI = MF->getRegInfo();
13368 const TargetRegisterClass *AddrRegClass =
13369 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
13371 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
13372 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
13373 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
13374 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
13375 sizeVReg = MI->getOperand(1).getReg(),
13376 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
13378 MachineFunction::iterator MBBIter = BB;
13381 MF->insert(MBBIter, bumpMBB);
13382 MF->insert(MBBIter, mallocMBB);
13383 MF->insert(MBBIter, continueMBB);
13385 continueMBB->splice(continueMBB->begin(), BB, llvm::next
13386 (MachineBasicBlock::iterator(MI)), BB->end());
13387 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
13389 // Add code to the main basic block to check if the stack limit has been hit,
13390 // and if so, jump to mallocMBB otherwise to bumpMBB.
13391 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
13392 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
13393 .addReg(tmpSPVReg).addReg(sizeVReg);
13394 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
13395 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
13396 .addReg(SPLimitVReg);
13397 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
13399 // bumpMBB simply decreases the stack pointer, since we know the current
13400 // stacklet has enough space.
13401 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
13402 .addReg(SPLimitVReg);
13403 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
13404 .addReg(SPLimitVReg);
13405 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
13407 // Calls into a routine in libgcc to allocate more space from the heap.
13408 const uint32_t *RegMask =
13409 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
13411 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
13413 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
13414 .addExternalSymbol("__morestack_allocate_stack_space")
13415 .addRegMask(RegMask)
13416 .addReg(X86::RDI, RegState::Implicit)
13417 .addReg(X86::RAX, RegState::ImplicitDefine);
13419 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
13421 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
13422 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
13423 .addExternalSymbol("__morestack_allocate_stack_space")
13424 .addRegMask(RegMask)
13425 .addReg(X86::EAX, RegState::ImplicitDefine);
13429 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
13432 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
13433 .addReg(Is64Bit ? X86::RAX : X86::EAX);
13434 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
13436 // Set up the CFG correctly.
13437 BB->addSuccessor(bumpMBB);
13438 BB->addSuccessor(mallocMBB);
13439 mallocMBB->addSuccessor(continueMBB);
13440 bumpMBB->addSuccessor(continueMBB);
13442 // Take care of the PHI nodes.
13443 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
13444 MI->getOperand(0).getReg())
13445 .addReg(mallocPtrVReg).addMBB(mallocMBB)
13446 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
13448 // Delete the original pseudo instruction.
13449 MI->eraseFromParent();
13452 return continueMBB;
13455 MachineBasicBlock *
13456 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
13457 MachineBasicBlock *BB) const {
13458 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13459 DebugLoc DL = MI->getDebugLoc();
13461 assert(!Subtarget->isTargetEnvMacho());
13463 // The lowering is pretty easy: we're just emitting the call to _alloca. The
13464 // non-trivial part is impdef of ESP.
13466 if (Subtarget->isTargetWin64()) {
13467 if (Subtarget->isTargetCygMing()) {
13468 // ___chkstk(Mingw64):
13469 // Clobbers R10, R11, RAX and EFLAGS.
13471 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
13472 .addExternalSymbol("___chkstk")
13473 .addReg(X86::RAX, RegState::Implicit)
13474 .addReg(X86::RSP, RegState::Implicit)
13475 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
13476 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
13477 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
13479 // __chkstk(MSVCRT): does not update stack pointer.
13480 // Clobbers R10, R11 and EFLAGS.
13481 // FIXME: RAX(allocated size) might be reused and not killed.
13482 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
13483 .addExternalSymbol("__chkstk")
13484 .addReg(X86::RAX, RegState::Implicit)
13485 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
13486 // RAX has the offset to subtracted from RSP.
13487 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
13492 const char *StackProbeSymbol =
13493 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
13495 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
13496 .addExternalSymbol(StackProbeSymbol)
13497 .addReg(X86::EAX, RegState::Implicit)
13498 .addReg(X86::ESP, RegState::Implicit)
13499 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
13500 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
13501 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
13504 MI->eraseFromParent(); // The pseudo instruction is gone now.
13508 MachineBasicBlock *
13509 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
13510 MachineBasicBlock *BB) const {
13511 // This is pretty easy. We're taking the value that we received from
13512 // our load from the relocation, sticking it in either RDI (x86-64)
13513 // or EAX and doing an indirect call. The return value will then
13514 // be in the normal return register.
13515 const X86InstrInfo *TII
13516 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
13517 DebugLoc DL = MI->getDebugLoc();
13518 MachineFunction *F = BB->getParent();
13520 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
13521 assert(MI->getOperand(3).isGlobal() && "This should be a global");
13523 // Get a register mask for the lowered call.
13524 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
13525 // proper register mask.
13526 const uint32_t *RegMask =
13527 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
13528 if (Subtarget->is64Bit()) {
13529 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
13530 TII->get(X86::MOV64rm), X86::RDI)
13532 .addImm(0).addReg(0)
13533 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
13534 MI->getOperand(3).getTargetFlags())
13536 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
13537 addDirectMem(MIB, X86::RDI);
13538 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
13539 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
13540 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
13541 TII->get(X86::MOV32rm), X86::EAX)
13543 .addImm(0).addReg(0)
13544 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
13545 MI->getOperand(3).getTargetFlags())
13547 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
13548 addDirectMem(MIB, X86::EAX);
13549 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
13551 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
13552 TII->get(X86::MOV32rm), X86::EAX)
13553 .addReg(TII->getGlobalBaseReg(F))
13554 .addImm(0).addReg(0)
13555 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
13556 MI->getOperand(3).getTargetFlags())
13558 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
13559 addDirectMem(MIB, X86::EAX);
13560 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
13563 MI->eraseFromParent(); // The pseudo instruction is gone now.
13567 MachineBasicBlock *
13568 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
13569 MachineBasicBlock *MBB) const {
13570 DebugLoc DL = MI->getDebugLoc();
13571 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13573 MachineFunction *MF = MBB->getParent();
13574 MachineRegisterInfo &MRI = MF->getRegInfo();
13576 const BasicBlock *BB = MBB->getBasicBlock();
13577 MachineFunction::iterator I = MBB;
13580 // Memory Reference
13581 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13582 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13585 unsigned MemOpndSlot = 0;
13587 unsigned CurOp = 0;
13589 DstReg = MI->getOperand(CurOp++).getReg();
13590 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
13591 assert(RC->hasType(MVT::i32) && "Invalid destination!");
13592 unsigned mainDstReg = MRI.createVirtualRegister(RC);
13593 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
13595 MemOpndSlot = CurOp;
13597 MVT PVT = getPointerTy();
13598 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
13599 "Invalid Pointer Size!");
13601 // For v = setjmp(buf), we generate
13604 // buf[LabelOffset] = restoreMBB
13605 // SjLjSetup restoreMBB
13611 // v = phi(main, restore)
13616 MachineBasicBlock *thisMBB = MBB;
13617 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13618 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13619 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
13620 MF->insert(I, mainMBB);
13621 MF->insert(I, sinkMBB);
13622 MF->push_back(restoreMBB);
13624 MachineInstrBuilder MIB;
13626 // Transfer the remainder of BB and its successor edges to sinkMBB.
13627 sinkMBB->splice(sinkMBB->begin(), MBB,
13628 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13629 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13632 unsigned PtrStoreOpc = 0;
13633 unsigned LabelReg = 0;
13634 const int64_t LabelOffset = 1 * PVT.getStoreSize();
13635 Reloc::Model RM = getTargetMachine().getRelocationModel();
13636 bool UseImmLabel = (getTargetMachine().getCodeModel() == CodeModel::Small) &&
13637 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
13639 // Prepare IP either in reg or imm.
13640 if (!UseImmLabel) {
13641 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
13642 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
13643 LabelReg = MRI.createVirtualRegister(PtrRC);
13644 if (Subtarget->is64Bit()) {
13645 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
13649 .addMBB(restoreMBB)
13652 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
13653 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
13654 .addReg(XII->getGlobalBaseReg(MF))
13657 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
13661 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
13663 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
13664 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13665 if (i == X86::AddrDisp)
13666 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
13668 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
13671 MIB.addReg(LabelReg);
13673 MIB.addMBB(restoreMBB);
13674 MIB.setMemRefs(MMOBegin, MMOEnd);
13676 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
13677 .addMBB(restoreMBB);
13678 MIB.addRegMask(RegInfo->getNoPreservedMask());
13679 thisMBB->addSuccessor(mainMBB);
13680 thisMBB->addSuccessor(restoreMBB);
13684 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
13685 mainMBB->addSuccessor(sinkMBB);
13688 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13689 TII->get(X86::PHI), DstReg)
13690 .addReg(mainDstReg).addMBB(mainMBB)
13691 .addReg(restoreDstReg).addMBB(restoreMBB);
13694 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
13695 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
13696 restoreMBB->addSuccessor(sinkMBB);
13698 MI->eraseFromParent();
13702 MachineBasicBlock *
13703 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
13704 MachineBasicBlock *MBB) const {
13705 DebugLoc DL = MI->getDebugLoc();
13706 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13708 MachineFunction *MF = MBB->getParent();
13709 MachineRegisterInfo &MRI = MF->getRegInfo();
13711 // Memory Reference
13712 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13713 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13715 MVT PVT = getPointerTy();
13716 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
13717 "Invalid Pointer Size!");
13719 const TargetRegisterClass *RC =
13720 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
13721 unsigned Tmp = MRI.createVirtualRegister(RC);
13722 // Since FP is only updated here but NOT referenced, it's treated as GPR.
13723 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
13724 unsigned SP = RegInfo->getStackRegister();
13726 MachineInstrBuilder MIB;
13728 const int64_t LabelOffset = 1 * PVT.getStoreSize();
13729 const int64_t SPOffset = 2 * PVT.getStoreSize();
13731 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
13732 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
13735 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
13736 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
13737 MIB.addOperand(MI->getOperand(i));
13738 MIB.setMemRefs(MMOBegin, MMOEnd);
13740 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
13741 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13742 if (i == X86::AddrDisp)
13743 MIB.addDisp(MI->getOperand(i), LabelOffset);
13745 MIB.addOperand(MI->getOperand(i));
13747 MIB.setMemRefs(MMOBegin, MMOEnd);
13749 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
13750 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13751 if (i == X86::AddrDisp)
13752 MIB.addDisp(MI->getOperand(i), SPOffset);
13754 MIB.addOperand(MI->getOperand(i));
13756 MIB.setMemRefs(MMOBegin, MMOEnd);
13758 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
13760 MI->eraseFromParent();
13764 MachineBasicBlock *
13765 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
13766 MachineBasicBlock *BB) const {
13767 switch (MI->getOpcode()) {
13768 default: llvm_unreachable("Unexpected instr type to insert");
13769 case X86::TAILJMPd64:
13770 case X86::TAILJMPr64:
13771 case X86::TAILJMPm64:
13772 llvm_unreachable("TAILJMP64 would not be touched here.");
13773 case X86::TCRETURNdi64:
13774 case X86::TCRETURNri64:
13775 case X86::TCRETURNmi64:
13777 case X86::WIN_ALLOCA:
13778 return EmitLoweredWinAlloca(MI, BB);
13779 case X86::SEG_ALLOCA_32:
13780 return EmitLoweredSegAlloca(MI, BB, false);
13781 case X86::SEG_ALLOCA_64:
13782 return EmitLoweredSegAlloca(MI, BB, true);
13783 case X86::TLSCall_32:
13784 case X86::TLSCall_64:
13785 return EmitLoweredTLSCall(MI, BB);
13786 case X86::CMOV_GR8:
13787 case X86::CMOV_FR32:
13788 case X86::CMOV_FR64:
13789 case X86::CMOV_V4F32:
13790 case X86::CMOV_V2F64:
13791 case X86::CMOV_V2I64:
13792 case X86::CMOV_V8F32:
13793 case X86::CMOV_V4F64:
13794 case X86::CMOV_V4I64:
13795 case X86::CMOV_GR16:
13796 case X86::CMOV_GR32:
13797 case X86::CMOV_RFP32:
13798 case X86::CMOV_RFP64:
13799 case X86::CMOV_RFP80:
13800 return EmitLoweredSelect(MI, BB);
13802 case X86::FP32_TO_INT16_IN_MEM:
13803 case X86::FP32_TO_INT32_IN_MEM:
13804 case X86::FP32_TO_INT64_IN_MEM:
13805 case X86::FP64_TO_INT16_IN_MEM:
13806 case X86::FP64_TO_INT32_IN_MEM:
13807 case X86::FP64_TO_INT64_IN_MEM:
13808 case X86::FP80_TO_INT16_IN_MEM:
13809 case X86::FP80_TO_INT32_IN_MEM:
13810 case X86::FP80_TO_INT64_IN_MEM: {
13811 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13812 DebugLoc DL = MI->getDebugLoc();
13814 // Change the floating point control register to use "round towards zero"
13815 // mode when truncating to an integer value.
13816 MachineFunction *F = BB->getParent();
13817 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
13818 addFrameReference(BuildMI(*BB, MI, DL,
13819 TII->get(X86::FNSTCW16m)), CWFrameIdx);
13821 // Load the old value of the high byte of the control word...
13823 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
13824 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
13827 // Set the high part to be round to zero...
13828 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
13831 // Reload the modified control word now...
13832 addFrameReference(BuildMI(*BB, MI, DL,
13833 TII->get(X86::FLDCW16m)), CWFrameIdx);
13835 // Restore the memory image of control word to original value
13836 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
13839 // Get the X86 opcode to use.
13841 switch (MI->getOpcode()) {
13842 default: llvm_unreachable("illegal opcode!");
13843 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
13844 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
13845 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
13846 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
13847 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
13848 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
13849 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
13850 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
13851 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
13855 MachineOperand &Op = MI->getOperand(0);
13857 AM.BaseType = X86AddressMode::RegBase;
13858 AM.Base.Reg = Op.getReg();
13860 AM.BaseType = X86AddressMode::FrameIndexBase;
13861 AM.Base.FrameIndex = Op.getIndex();
13863 Op = MI->getOperand(1);
13865 AM.Scale = Op.getImm();
13866 Op = MI->getOperand(2);
13868 AM.IndexReg = Op.getImm();
13869 Op = MI->getOperand(3);
13870 if (Op.isGlobal()) {
13871 AM.GV = Op.getGlobal();
13873 AM.Disp = Op.getImm();
13875 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
13876 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
13878 // Reload the original control word now.
13879 addFrameReference(BuildMI(*BB, MI, DL,
13880 TII->get(X86::FLDCW16m)), CWFrameIdx);
13882 MI->eraseFromParent(); // The pseudo instruction is gone now.
13885 // String/text processing lowering.
13886 case X86::PCMPISTRM128REG:
13887 case X86::VPCMPISTRM128REG:
13888 case X86::PCMPISTRM128MEM:
13889 case X86::VPCMPISTRM128MEM:
13890 case X86::PCMPESTRM128REG:
13891 case X86::VPCMPESTRM128REG:
13892 case X86::PCMPESTRM128MEM:
13893 case X86::VPCMPESTRM128MEM: {
13896 switch (MI->getOpcode()) {
13897 default: llvm_unreachable("illegal opcode!");
13898 case X86::PCMPISTRM128REG:
13899 case X86::VPCMPISTRM128REG:
13900 NumArgs = 3; MemArg = false; break;
13901 case X86::PCMPISTRM128MEM:
13902 case X86::VPCMPISTRM128MEM:
13903 NumArgs = 3; MemArg = true; break;
13904 case X86::PCMPESTRM128REG:
13905 case X86::VPCMPESTRM128REG:
13906 NumArgs = 5; MemArg = false; break;
13907 case X86::PCMPESTRM128MEM:
13908 case X86::VPCMPESTRM128MEM:
13909 NumArgs = 5; MemArg = true; break;
13911 return EmitPCMP(MI, BB, NumArgs, MemArg);
13914 // Thread synchronization.
13916 return EmitMonitor(MI, BB);
13920 return EmitXBegin(MI, BB);
13922 // Atomic Lowering.
13923 case X86::ATOMAND8:
13924 case X86::ATOMAND16:
13925 case X86::ATOMAND32:
13926 case X86::ATOMAND64:
13929 case X86::ATOMOR16:
13930 case X86::ATOMOR32:
13931 case X86::ATOMOR64:
13933 case X86::ATOMXOR16:
13934 case X86::ATOMXOR8:
13935 case X86::ATOMXOR32:
13936 case X86::ATOMXOR64:
13938 case X86::ATOMNAND8:
13939 case X86::ATOMNAND16:
13940 case X86::ATOMNAND32:
13941 case X86::ATOMNAND64:
13943 case X86::ATOMMAX8:
13944 case X86::ATOMMAX16:
13945 case X86::ATOMMAX32:
13946 case X86::ATOMMAX64:
13948 case X86::ATOMMIN8:
13949 case X86::ATOMMIN16:
13950 case X86::ATOMMIN32:
13951 case X86::ATOMMIN64:
13953 case X86::ATOMUMAX8:
13954 case X86::ATOMUMAX16:
13955 case X86::ATOMUMAX32:
13956 case X86::ATOMUMAX64:
13958 case X86::ATOMUMIN8:
13959 case X86::ATOMUMIN16:
13960 case X86::ATOMUMIN32:
13961 case X86::ATOMUMIN64:
13962 return EmitAtomicLoadArith(MI, BB);
13964 // This group does 64-bit operations on a 32-bit host.
13965 case X86::ATOMAND6432:
13966 case X86::ATOMOR6432:
13967 case X86::ATOMXOR6432:
13968 case X86::ATOMNAND6432:
13969 case X86::ATOMADD6432:
13970 case X86::ATOMSUB6432:
13971 case X86::ATOMMAX6432:
13972 case X86::ATOMMIN6432:
13973 case X86::ATOMUMAX6432:
13974 case X86::ATOMUMIN6432:
13975 case X86::ATOMSWAP6432:
13976 return EmitAtomicLoadArith6432(MI, BB);
13978 case X86::VASTART_SAVE_XMM_REGS:
13979 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
13981 case X86::VAARG_64:
13982 return EmitVAARG64WithCustomInserter(MI, BB);
13984 case X86::EH_SjLj_SetJmp32:
13985 case X86::EH_SjLj_SetJmp64:
13986 return emitEHSjLjSetJmp(MI, BB);
13988 case X86::EH_SjLj_LongJmp32:
13989 case X86::EH_SjLj_LongJmp64:
13990 return emitEHSjLjLongJmp(MI, BB);
13994 //===----------------------------------------------------------------------===//
13995 // X86 Optimization Hooks
13996 //===----------------------------------------------------------------------===//
13998 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
14001 const SelectionDAG &DAG,
14002 unsigned Depth) const {
14003 unsigned BitWidth = KnownZero.getBitWidth();
14004 unsigned Opc = Op.getOpcode();
14005 assert((Opc >= ISD::BUILTIN_OP_END ||
14006 Opc == ISD::INTRINSIC_WO_CHAIN ||
14007 Opc == ISD::INTRINSIC_W_CHAIN ||
14008 Opc == ISD::INTRINSIC_VOID) &&
14009 "Should use MaskedValueIsZero if you don't know whether Op"
14010 " is a target node!");
14012 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
14026 // These nodes' second result is a boolean.
14027 if (Op.getResNo() == 0)
14030 case X86ISD::SETCC:
14031 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
14033 case ISD::INTRINSIC_WO_CHAIN: {
14034 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14035 unsigned NumLoBits = 0;
14038 case Intrinsic::x86_sse_movmsk_ps:
14039 case Intrinsic::x86_avx_movmsk_ps_256:
14040 case Intrinsic::x86_sse2_movmsk_pd:
14041 case Intrinsic::x86_avx_movmsk_pd_256:
14042 case Intrinsic::x86_mmx_pmovmskb:
14043 case Intrinsic::x86_sse2_pmovmskb_128:
14044 case Intrinsic::x86_avx2_pmovmskb: {
14045 // High bits of movmskp{s|d}, pmovmskb are known zero.
14047 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14048 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
14049 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
14050 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
14051 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
14052 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
14053 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
14054 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
14056 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
14065 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
14066 unsigned Depth) const {
14067 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
14068 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
14069 return Op.getValueType().getScalarType().getSizeInBits();
14075 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
14076 /// node is a GlobalAddress + offset.
14077 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
14078 const GlobalValue* &GA,
14079 int64_t &Offset) const {
14080 if (N->getOpcode() == X86ISD::Wrapper) {
14081 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
14082 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
14083 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
14087 return TargetLowering::isGAPlusOffset(N, GA, Offset);
14090 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
14091 /// same as extracting the high 128-bit part of 256-bit vector and then
14092 /// inserting the result into the low part of a new 256-bit vector
14093 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
14094 EVT VT = SVOp->getValueType(0);
14095 unsigned NumElems = VT.getVectorNumElements();
14097 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
14098 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
14099 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
14100 SVOp->getMaskElt(j) >= 0)
14106 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
14107 /// same as extracting the low 128-bit part of 256-bit vector and then
14108 /// inserting the result into the high part of a new 256-bit vector
14109 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
14110 EVT VT = SVOp->getValueType(0);
14111 unsigned NumElems = VT.getVectorNumElements();
14113 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
14114 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
14115 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
14116 SVOp->getMaskElt(j) >= 0)
14122 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
14123 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
14124 TargetLowering::DAGCombinerInfo &DCI,
14125 const X86Subtarget* Subtarget) {
14126 DebugLoc dl = N->getDebugLoc();
14127 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
14128 SDValue V1 = SVOp->getOperand(0);
14129 SDValue V2 = SVOp->getOperand(1);
14130 EVT VT = SVOp->getValueType(0);
14131 unsigned NumElems = VT.getVectorNumElements();
14133 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
14134 V2.getOpcode() == ISD::CONCAT_VECTORS) {
14138 // V UNDEF BUILD_VECTOR UNDEF
14140 // CONCAT_VECTOR CONCAT_VECTOR
14143 // RESULT: V + zero extended
14145 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
14146 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
14147 V1.getOperand(1).getOpcode() != ISD::UNDEF)
14150 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
14153 // To match the shuffle mask, the first half of the mask should
14154 // be exactly the first vector, and all the rest a splat with the
14155 // first element of the second one.
14156 for (unsigned i = 0; i != NumElems/2; ++i)
14157 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
14158 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
14161 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
14162 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
14163 if (Ld->hasNUsesOfValue(1, 0)) {
14164 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
14165 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
14167 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2,
14169 Ld->getPointerInfo(),
14170 Ld->getAlignment(),
14171 false/*isVolatile*/, true/*ReadMem*/,
14172 false/*WriteMem*/);
14173 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
14177 // Emit a zeroed vector and insert the desired subvector on its
14179 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
14180 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
14181 return DCI.CombineTo(N, InsV);
14184 //===--------------------------------------------------------------------===//
14185 // Combine some shuffles into subvector extracts and inserts:
14188 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
14189 if (isShuffleHigh128VectorInsertLow(SVOp)) {
14190 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
14191 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
14192 return DCI.CombineTo(N, InsV);
14195 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
14196 if (isShuffleLow128VectorInsertHigh(SVOp)) {
14197 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
14198 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
14199 return DCI.CombineTo(N, InsV);
14205 /// PerformShuffleCombine - Performs several different shuffle combines.
14206 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
14207 TargetLowering::DAGCombinerInfo &DCI,
14208 const X86Subtarget *Subtarget) {
14209 DebugLoc dl = N->getDebugLoc();
14210 EVT VT = N->getValueType(0);
14212 // Don't create instructions with illegal types after legalize types has run.
14213 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14214 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
14217 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
14218 if (Subtarget->hasAVX() && VT.is256BitVector() &&
14219 N->getOpcode() == ISD::VECTOR_SHUFFLE)
14220 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
14222 // Only handle 128 wide vector from here on.
14223 if (!VT.is128BitVector())
14226 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
14227 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
14228 // consecutive, non-overlapping, and in the right order.
14229 SmallVector<SDValue, 16> Elts;
14230 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
14231 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
14233 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
14237 /// PerformTruncateCombine - Converts truncate operation to
14238 /// a sequence of vector shuffle operations.
14239 /// It is possible when we truncate 256-bit vector to 128-bit vector
14240 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
14241 TargetLowering::DAGCombinerInfo &DCI,
14242 const X86Subtarget *Subtarget) {
14243 if (!DCI.isBeforeLegalizeOps())
14246 if (!Subtarget->hasAVX())
14249 EVT VT = N->getValueType(0);
14250 SDValue Op = N->getOperand(0);
14251 EVT OpVT = Op.getValueType();
14252 DebugLoc dl = N->getDebugLoc();
14254 if ((VT == MVT::v4i32) && (OpVT == MVT::v4i64)) {
14256 if (Subtarget->hasAVX2()) {
14257 // AVX2: v4i64 -> v4i32
14260 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
14262 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v8i32, Op);
14263 Op = DAG.getVectorShuffle(MVT::v8i32, dl, Op, DAG.getUNDEF(MVT::v8i32),
14266 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Op,
14267 DAG.getIntPtrConstant(0));
14270 // AVX: v4i64 -> v4i32
14271 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
14272 DAG.getIntPtrConstant(0));
14274 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
14275 DAG.getIntPtrConstant(2));
14277 OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpLo);
14278 OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpHi);
14281 static const int ShufMask1[] = {0, 2, 0, 0};
14283 SDValue Undef = DAG.getUNDEF(VT);
14284 OpLo = DAG.getVectorShuffle(VT, dl, OpLo, Undef, ShufMask1);
14285 OpHi = DAG.getVectorShuffle(VT, dl, OpHi, Undef, ShufMask1);
14288 static const int ShufMask2[] = {0, 1, 4, 5};
14290 return DAG.getVectorShuffle(VT, dl, OpLo, OpHi, ShufMask2);
14293 if ((VT == MVT::v8i16) && (OpVT == MVT::v8i32)) {
14295 if (Subtarget->hasAVX2()) {
14296 // AVX2: v8i32 -> v8i16
14298 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v32i8, Op);
14301 SmallVector<SDValue,32> pshufbMask;
14302 for (unsigned i = 0; i < 2; ++i) {
14303 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
14304 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
14305 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
14306 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
14307 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
14308 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
14309 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
14310 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
14311 for (unsigned j = 0; j < 8; ++j)
14312 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
14314 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v32i8,
14315 &pshufbMask[0], 32);
14316 Op = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, Op, BV);
14318 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4i64, Op);
14320 static const int ShufMask[] = {0, 2, -1, -1};
14321 Op = DAG.getVectorShuffle(MVT::v4i64, dl, Op, DAG.getUNDEF(MVT::v4i64),
14324 Op = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
14325 DAG.getIntPtrConstant(0));
14327 return DAG.getNode(ISD::BITCAST, dl, VT, Op);
14330 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i32, Op,
14331 DAG.getIntPtrConstant(0));
14333 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i32, Op,
14334 DAG.getIntPtrConstant(4));
14336 OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLo);
14337 OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpHi);
14340 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
14341 -1, -1, -1, -1, -1, -1, -1, -1};
14343 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
14344 OpLo = DAG.getVectorShuffle(MVT::v16i8, dl, OpLo, Undef, ShufMask1);
14345 OpHi = DAG.getVectorShuffle(MVT::v16i8, dl, OpHi, Undef, ShufMask1);
14347 OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpLo);
14348 OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpHi);
14351 static const int ShufMask2[] = {0, 1, 4, 5};
14353 SDValue res = DAG.getVectorShuffle(MVT::v4i32, dl, OpLo, OpHi, ShufMask2);
14354 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, res);
14360 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
14361 /// specific shuffle of a load can be folded into a single element load.
14362 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
14363 /// shuffles have been customed lowered so we need to handle those here.
14364 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
14365 TargetLowering::DAGCombinerInfo &DCI) {
14366 if (DCI.isBeforeLegalizeOps())
14369 SDValue InVec = N->getOperand(0);
14370 SDValue EltNo = N->getOperand(1);
14372 if (!isa<ConstantSDNode>(EltNo))
14375 EVT VT = InVec.getValueType();
14377 bool HasShuffleIntoBitcast = false;
14378 if (InVec.getOpcode() == ISD::BITCAST) {
14379 // Don't duplicate a load with other uses.
14380 if (!InVec.hasOneUse())
14382 EVT BCVT = InVec.getOperand(0).getValueType();
14383 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
14385 InVec = InVec.getOperand(0);
14386 HasShuffleIntoBitcast = true;
14389 if (!isTargetShuffle(InVec.getOpcode()))
14392 // Don't duplicate a load with other uses.
14393 if (!InVec.hasOneUse())
14396 SmallVector<int, 16> ShuffleMask;
14398 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
14402 // Select the input vector, guarding against out of range extract vector.
14403 unsigned NumElems = VT.getVectorNumElements();
14404 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
14405 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
14406 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
14407 : InVec.getOperand(1);
14409 // If inputs to shuffle are the same for both ops, then allow 2 uses
14410 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
14412 if (LdNode.getOpcode() == ISD::BITCAST) {
14413 // Don't duplicate a load with other uses.
14414 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
14417 AllowedUses = 1; // only allow 1 load use if we have a bitcast
14418 LdNode = LdNode.getOperand(0);
14421 if (!ISD::isNormalLoad(LdNode.getNode()))
14424 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
14426 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
14429 if (HasShuffleIntoBitcast) {
14430 // If there's a bitcast before the shuffle, check if the load type and
14431 // alignment is valid.
14432 unsigned Align = LN0->getAlignment();
14433 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14434 unsigned NewAlign = TLI.getDataLayout()->
14435 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
14437 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
14441 // All checks match so transform back to vector_shuffle so that DAG combiner
14442 // can finish the job
14443 DebugLoc dl = N->getDebugLoc();
14445 // Create shuffle node taking into account the case that its a unary shuffle
14446 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
14447 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
14448 InVec.getOperand(0), Shuffle,
14450 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
14451 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
14455 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
14456 /// generation and convert it from being a bunch of shuffles and extracts
14457 /// to a simple store and scalar loads to extract the elements.
14458 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
14459 TargetLowering::DAGCombinerInfo &DCI) {
14460 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
14461 if (NewOp.getNode())
14464 SDValue InputVector = N->getOperand(0);
14465 // Detect whether we are trying to convert from mmx to i32 and the bitcast
14466 // from mmx to v2i32 has a single usage.
14467 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
14468 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
14469 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
14470 return DAG.getNode(X86ISD::MMX_MOVD2W, InputVector.getDebugLoc(),
14471 N->getValueType(0),
14472 InputVector.getNode()->getOperand(0));
14474 // Only operate on vectors of 4 elements, where the alternative shuffling
14475 // gets to be more expensive.
14476 if (InputVector.getValueType() != MVT::v4i32)
14479 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
14480 // single use which is a sign-extend or zero-extend, and all elements are
14482 SmallVector<SDNode *, 4> Uses;
14483 unsigned ExtractedElements = 0;
14484 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
14485 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
14486 if (UI.getUse().getResNo() != InputVector.getResNo())
14489 SDNode *Extract = *UI;
14490 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
14493 if (Extract->getValueType(0) != MVT::i32)
14495 if (!Extract->hasOneUse())
14497 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
14498 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
14500 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
14503 // Record which element was extracted.
14504 ExtractedElements |=
14505 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
14507 Uses.push_back(Extract);
14510 // If not all the elements were used, this may not be worthwhile.
14511 if (ExtractedElements != 15)
14514 // Ok, we've now decided to do the transformation.
14515 DebugLoc dl = InputVector.getDebugLoc();
14517 // Store the value to a temporary stack slot.
14518 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
14519 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
14520 MachinePointerInfo(), false, false, 0);
14522 // Replace each use (extract) with a load of the appropriate element.
14523 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
14524 UE = Uses.end(); UI != UE; ++UI) {
14525 SDNode *Extract = *UI;
14527 // cOMpute the element's address.
14528 SDValue Idx = Extract->getOperand(1);
14530 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
14531 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
14532 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14533 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
14535 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
14536 StackPtr, OffsetVal);
14538 // Load the scalar.
14539 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
14540 ScalarAddr, MachinePointerInfo(),
14541 false, false, false, 0);
14543 // Replace the exact with the load.
14544 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
14547 // The replacement was made in place; don't return anything.
14551 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
14553 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
14554 TargetLowering::DAGCombinerInfo &DCI,
14555 const X86Subtarget *Subtarget) {
14556 DebugLoc DL = N->getDebugLoc();
14557 SDValue Cond = N->getOperand(0);
14558 // Get the LHS/RHS of the select.
14559 SDValue LHS = N->getOperand(1);
14560 SDValue RHS = N->getOperand(2);
14561 EVT VT = LHS.getValueType();
14563 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
14564 // instructions match the semantics of the common C idiom x<y?x:y but not
14565 // x<=y?x:y, because of how they handle negative zero (which can be
14566 // ignored in unsafe-math mode).
14567 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
14568 VT != MVT::f80 && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
14569 (Subtarget->hasSSE2() ||
14570 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
14571 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
14573 unsigned Opcode = 0;
14574 // Check for x CC y ? x : y.
14575 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
14576 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
14580 // Converting this to a min would handle NaNs incorrectly, and swapping
14581 // the operands would cause it to handle comparisons between positive
14582 // and negative zero incorrectly.
14583 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
14584 if (!DAG.getTarget().Options.UnsafeFPMath &&
14585 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
14587 std::swap(LHS, RHS);
14589 Opcode = X86ISD::FMIN;
14592 // Converting this to a min would handle comparisons between positive
14593 // and negative zero incorrectly.
14594 if (!DAG.getTarget().Options.UnsafeFPMath &&
14595 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
14597 Opcode = X86ISD::FMIN;
14600 // Converting this to a min would handle both negative zeros and NaNs
14601 // incorrectly, but we can swap the operands to fix both.
14602 std::swap(LHS, RHS);
14606 Opcode = X86ISD::FMIN;
14610 // Converting this to a max would handle comparisons between positive
14611 // and negative zero incorrectly.
14612 if (!DAG.getTarget().Options.UnsafeFPMath &&
14613 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
14615 Opcode = X86ISD::FMAX;
14618 // Converting this to a max would handle NaNs incorrectly, and swapping
14619 // the operands would cause it to handle comparisons between positive
14620 // and negative zero incorrectly.
14621 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
14622 if (!DAG.getTarget().Options.UnsafeFPMath &&
14623 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
14625 std::swap(LHS, RHS);
14627 Opcode = X86ISD::FMAX;
14630 // Converting this to a max would handle both negative zeros and NaNs
14631 // incorrectly, but we can swap the operands to fix both.
14632 std::swap(LHS, RHS);
14636 Opcode = X86ISD::FMAX;
14639 // Check for x CC y ? y : x -- a min/max with reversed arms.
14640 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
14641 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
14645 // Converting this to a min would handle comparisons between positive
14646 // and negative zero incorrectly, and swapping the operands would
14647 // cause it to handle NaNs incorrectly.
14648 if (!DAG.getTarget().Options.UnsafeFPMath &&
14649 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
14650 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
14652 std::swap(LHS, RHS);
14654 Opcode = X86ISD::FMIN;
14657 // Converting this to a min would handle NaNs incorrectly.
14658 if (!DAG.getTarget().Options.UnsafeFPMath &&
14659 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
14661 Opcode = X86ISD::FMIN;
14664 // Converting this to a min would handle both negative zeros and NaNs
14665 // incorrectly, but we can swap the operands to fix both.
14666 std::swap(LHS, RHS);
14670 Opcode = X86ISD::FMIN;
14674 // Converting this to a max would handle NaNs incorrectly.
14675 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
14677 Opcode = X86ISD::FMAX;
14680 // Converting this to a max would handle comparisons between positive
14681 // and negative zero incorrectly, and swapping the operands would
14682 // cause it to handle NaNs incorrectly.
14683 if (!DAG.getTarget().Options.UnsafeFPMath &&
14684 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
14685 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
14687 std::swap(LHS, RHS);
14689 Opcode = X86ISD::FMAX;
14692 // Converting this to a max would handle both negative zeros and NaNs
14693 // incorrectly, but we can swap the operands to fix both.
14694 std::swap(LHS, RHS);
14698 Opcode = X86ISD::FMAX;
14704 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
14707 // If this is a select between two integer constants, try to do some
14709 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
14710 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
14711 // Don't do this for crazy integer types.
14712 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
14713 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
14714 // so that TrueC (the true value) is larger than FalseC.
14715 bool NeedsCondInvert = false;
14717 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
14718 // Efficiently invertible.
14719 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
14720 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
14721 isa<ConstantSDNode>(Cond.getOperand(1))))) {
14722 NeedsCondInvert = true;
14723 std::swap(TrueC, FalseC);
14726 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
14727 if (FalseC->getAPIntValue() == 0 &&
14728 TrueC->getAPIntValue().isPowerOf2()) {
14729 if (NeedsCondInvert) // Invert the condition if needed.
14730 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
14731 DAG.getConstant(1, Cond.getValueType()));
14733 // Zero extend the condition if needed.
14734 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
14736 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
14737 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
14738 DAG.getConstant(ShAmt, MVT::i8));
14741 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
14742 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
14743 if (NeedsCondInvert) // Invert the condition if needed.
14744 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
14745 DAG.getConstant(1, Cond.getValueType()));
14747 // Zero extend the condition if needed.
14748 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
14749 FalseC->getValueType(0), Cond);
14750 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
14751 SDValue(FalseC, 0));
14754 // Optimize cases that will turn into an LEA instruction. This requires
14755 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
14756 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
14757 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
14758 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
14760 bool isFastMultiplier = false;
14762 switch ((unsigned char)Diff) {
14764 case 1: // result = add base, cond
14765 case 2: // result = lea base( , cond*2)
14766 case 3: // result = lea base(cond, cond*2)
14767 case 4: // result = lea base( , cond*4)
14768 case 5: // result = lea base(cond, cond*4)
14769 case 8: // result = lea base( , cond*8)
14770 case 9: // result = lea base(cond, cond*8)
14771 isFastMultiplier = true;
14776 if (isFastMultiplier) {
14777 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
14778 if (NeedsCondInvert) // Invert the condition if needed.
14779 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
14780 DAG.getConstant(1, Cond.getValueType()));
14782 // Zero extend the condition if needed.
14783 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
14785 // Scale the condition by the difference.
14787 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
14788 DAG.getConstant(Diff, Cond.getValueType()));
14790 // Add the base if non-zero.
14791 if (FalseC->getAPIntValue() != 0)
14792 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
14793 SDValue(FalseC, 0));
14800 // Canonicalize max and min:
14801 // (x > y) ? x : y -> (x >= y) ? x : y
14802 // (x < y) ? x : y -> (x <= y) ? x : y
14803 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
14804 // the need for an extra compare
14805 // against zero. e.g.
14806 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
14808 // testl %edi, %edi
14810 // cmovgl %edi, %eax
14814 // cmovsl %eax, %edi
14815 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
14816 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
14817 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
14818 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
14823 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
14824 Cond = DAG.getSetCC(Cond.getDebugLoc(), Cond.getValueType(),
14825 Cond.getOperand(0), Cond.getOperand(1), NewCC);
14826 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
14831 // If we know that this node is legal then we know that it is going to be
14832 // matched by one of the SSE/AVX BLEND instructions. These instructions only
14833 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
14834 // to simplify previous instructions.
14835 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14836 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
14837 !DCI.isBeforeLegalize() && TLI.isOperationLegal(ISD::VSELECT, VT)) {
14838 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
14840 // Don't optimize vector selects that map to mask-registers.
14844 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
14845 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
14847 APInt KnownZero, KnownOne;
14848 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
14849 DCI.isBeforeLegalizeOps());
14850 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
14851 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
14852 DCI.CommitTargetLoweringOpt(TLO);
14858 // Check whether a boolean test is testing a boolean value generated by
14859 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
14862 // Simplify the following patterns:
14863 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
14864 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
14865 // to (Op EFLAGS Cond)
14867 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
14868 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
14869 // to (Op EFLAGS !Cond)
14871 // where Op could be BRCOND or CMOV.
14873 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
14874 // Quit if not CMP and SUB with its value result used.
14875 if (Cmp.getOpcode() != X86ISD::CMP &&
14876 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
14879 // Quit if not used as a boolean value.
14880 if (CC != X86::COND_E && CC != X86::COND_NE)
14883 // Check CMP operands. One of them should be 0 or 1 and the other should be
14884 // an SetCC or extended from it.
14885 SDValue Op1 = Cmp.getOperand(0);
14886 SDValue Op2 = Cmp.getOperand(1);
14889 const ConstantSDNode* C = 0;
14890 bool needOppositeCond = (CC == X86::COND_E);
14892 if ((C = dyn_cast<ConstantSDNode>(Op1)))
14894 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
14896 else // Quit if all operands are not constants.
14899 if (C->getZExtValue() == 1)
14900 needOppositeCond = !needOppositeCond;
14901 else if (C->getZExtValue() != 0)
14902 // Quit if the constant is neither 0 or 1.
14905 // Skip 'zext' node.
14906 if (SetCC.getOpcode() == ISD::ZERO_EXTEND)
14907 SetCC = SetCC.getOperand(0);
14909 switch (SetCC.getOpcode()) {
14910 case X86ISD::SETCC:
14911 // Set the condition code or opposite one if necessary.
14912 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
14913 if (needOppositeCond)
14914 CC = X86::GetOppositeBranchCondition(CC);
14915 return SetCC.getOperand(1);
14916 case X86ISD::CMOV: {
14917 // Check whether false/true value has canonical one, i.e. 0 or 1.
14918 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
14919 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
14920 // Quit if true value is not a constant.
14923 // Quit if false value is not a constant.
14925 // A special case for rdrand, where 0 is set if false cond is found.
14926 SDValue Op = SetCC.getOperand(0);
14927 if (Op.getOpcode() != X86ISD::RDRAND)
14930 // Quit if false value is not the constant 0 or 1.
14931 bool FValIsFalse = true;
14932 if (FVal && FVal->getZExtValue() != 0) {
14933 if (FVal->getZExtValue() != 1)
14935 // If FVal is 1, opposite cond is needed.
14936 needOppositeCond = !needOppositeCond;
14937 FValIsFalse = false;
14939 // Quit if TVal is not the constant opposite of FVal.
14940 if (FValIsFalse && TVal->getZExtValue() != 1)
14942 if (!FValIsFalse && TVal->getZExtValue() != 0)
14944 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
14945 if (needOppositeCond)
14946 CC = X86::GetOppositeBranchCondition(CC);
14947 return SetCC.getOperand(3);
14954 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
14955 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
14956 TargetLowering::DAGCombinerInfo &DCI,
14957 const X86Subtarget *Subtarget) {
14958 DebugLoc DL = N->getDebugLoc();
14960 // If the flag operand isn't dead, don't touch this CMOV.
14961 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
14964 SDValue FalseOp = N->getOperand(0);
14965 SDValue TrueOp = N->getOperand(1);
14966 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
14967 SDValue Cond = N->getOperand(3);
14969 if (CC == X86::COND_E || CC == X86::COND_NE) {
14970 switch (Cond.getOpcode()) {
14974 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
14975 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
14976 return (CC == X86::COND_E) ? FalseOp : TrueOp;
14982 Flags = checkBoolTestSetCCCombine(Cond, CC);
14983 if (Flags.getNode() &&
14984 // Extra check as FCMOV only supports a subset of X86 cond.
14985 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
14986 SDValue Ops[] = { FalseOp, TrueOp,
14987 DAG.getConstant(CC, MVT::i8), Flags };
14988 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(),
14989 Ops, array_lengthof(Ops));
14992 // If this is a select between two integer constants, try to do some
14993 // optimizations. Note that the operands are ordered the opposite of SELECT
14995 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
14996 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
14997 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
14998 // larger than FalseC (the false value).
14999 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
15000 CC = X86::GetOppositeBranchCondition(CC);
15001 std::swap(TrueC, FalseC);
15002 std::swap(TrueOp, FalseOp);
15005 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
15006 // This is efficient for any integer data type (including i8/i16) and
15008 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
15009 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15010 DAG.getConstant(CC, MVT::i8), Cond);
15012 // Zero extend the condition if needed.
15013 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
15015 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
15016 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
15017 DAG.getConstant(ShAmt, MVT::i8));
15018 if (N->getNumValues() == 2) // Dead flag value?
15019 return DCI.CombineTo(N, Cond, SDValue());
15023 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
15024 // for any integer data type, including i8/i16.
15025 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
15026 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15027 DAG.getConstant(CC, MVT::i8), Cond);
15029 // Zero extend the condition if needed.
15030 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
15031 FalseC->getValueType(0), Cond);
15032 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15033 SDValue(FalseC, 0));
15035 if (N->getNumValues() == 2) // Dead flag value?
15036 return DCI.CombineTo(N, Cond, SDValue());
15040 // Optimize cases that will turn into an LEA instruction. This requires
15041 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
15042 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
15043 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
15044 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
15046 bool isFastMultiplier = false;
15048 switch ((unsigned char)Diff) {
15050 case 1: // result = add base, cond
15051 case 2: // result = lea base( , cond*2)
15052 case 3: // result = lea base(cond, cond*2)
15053 case 4: // result = lea base( , cond*4)
15054 case 5: // result = lea base(cond, cond*4)
15055 case 8: // result = lea base( , cond*8)
15056 case 9: // result = lea base(cond, cond*8)
15057 isFastMultiplier = true;
15062 if (isFastMultiplier) {
15063 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
15064 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15065 DAG.getConstant(CC, MVT::i8), Cond);
15066 // Zero extend the condition if needed.
15067 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
15069 // Scale the condition by the difference.
15071 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
15072 DAG.getConstant(Diff, Cond.getValueType()));
15074 // Add the base if non-zero.
15075 if (FalseC->getAPIntValue() != 0)
15076 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15077 SDValue(FalseC, 0));
15078 if (N->getNumValues() == 2) // Dead flag value?
15079 return DCI.CombineTo(N, Cond, SDValue());
15086 // Handle these cases:
15087 // (select (x != c), e, c) -> select (x != c), e, x),
15088 // (select (x == c), c, e) -> select (x == c), x, e)
15089 // where the c is an integer constant, and the "select" is the combination
15090 // of CMOV and CMP.
15092 // The rationale for this change is that the conditional-move from a constant
15093 // needs two instructions, however, conditional-move from a register needs
15094 // only one instruction.
15096 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
15097 // some instruction-combining opportunities. This opt needs to be
15098 // postponed as late as possible.
15100 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
15101 // the DCI.xxxx conditions are provided to postpone the optimization as
15102 // late as possible.
15104 ConstantSDNode *CmpAgainst = 0;
15105 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
15106 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
15107 dyn_cast<ConstantSDNode>(Cond.getOperand(0)) == 0) {
15109 if (CC == X86::COND_NE &&
15110 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
15111 CC = X86::GetOppositeBranchCondition(CC);
15112 std::swap(TrueOp, FalseOp);
15115 if (CC == X86::COND_E &&
15116 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
15117 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
15118 DAG.getConstant(CC, MVT::i8), Cond };
15119 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops,
15120 array_lengthof(Ops));
15129 /// PerformMulCombine - Optimize a single multiply with constant into two
15130 /// in order to implement it with two cheaper instructions, e.g.
15131 /// LEA + SHL, LEA + LEA.
15132 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
15133 TargetLowering::DAGCombinerInfo &DCI) {
15134 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
15137 EVT VT = N->getValueType(0);
15138 if (VT != MVT::i64)
15141 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
15144 uint64_t MulAmt = C->getZExtValue();
15145 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
15148 uint64_t MulAmt1 = 0;
15149 uint64_t MulAmt2 = 0;
15150 if ((MulAmt % 9) == 0) {
15152 MulAmt2 = MulAmt / 9;
15153 } else if ((MulAmt % 5) == 0) {
15155 MulAmt2 = MulAmt / 5;
15156 } else if ((MulAmt % 3) == 0) {
15158 MulAmt2 = MulAmt / 3;
15161 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
15162 DebugLoc DL = N->getDebugLoc();
15164 if (isPowerOf2_64(MulAmt2) &&
15165 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
15166 // If second multiplifer is pow2, issue it first. We want the multiply by
15167 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
15169 std::swap(MulAmt1, MulAmt2);
15172 if (isPowerOf2_64(MulAmt1))
15173 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
15174 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
15176 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
15177 DAG.getConstant(MulAmt1, VT));
15179 if (isPowerOf2_64(MulAmt2))
15180 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
15181 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
15183 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
15184 DAG.getConstant(MulAmt2, VT));
15186 // Do not add new nodes to DAG combiner worklist.
15187 DCI.CombineTo(N, NewMul, false);
15192 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
15193 SDValue N0 = N->getOperand(0);
15194 SDValue N1 = N->getOperand(1);
15195 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
15196 EVT VT = N0.getValueType();
15198 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
15199 // since the result of setcc_c is all zero's or all ones.
15200 if (VT.isInteger() && !VT.isVector() &&
15201 N1C && N0.getOpcode() == ISD::AND &&
15202 N0.getOperand(1).getOpcode() == ISD::Constant) {
15203 SDValue N00 = N0.getOperand(0);
15204 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
15205 ((N00.getOpcode() == ISD::ANY_EXTEND ||
15206 N00.getOpcode() == ISD::ZERO_EXTEND) &&
15207 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
15208 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
15209 APInt ShAmt = N1C->getAPIntValue();
15210 Mask = Mask.shl(ShAmt);
15212 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
15213 N00, DAG.getConstant(Mask, VT));
15218 // Hardware support for vector shifts is sparse which makes us scalarize the
15219 // vector operations in many cases. Also, on sandybridge ADD is faster than
15221 // (shl V, 1) -> add V,V
15222 if (isSplatVector(N1.getNode())) {
15223 assert(N0.getValueType().isVector() && "Invalid vector shift type");
15224 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(0));
15225 // We shift all of the values by one. In many cases we do not have
15226 // hardware support for this operation. This is better expressed as an ADD
15228 if (N1C && (1 == N1C->getZExtValue())) {
15229 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N0);
15236 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
15238 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
15239 TargetLowering::DAGCombinerInfo &DCI,
15240 const X86Subtarget *Subtarget) {
15241 EVT VT = N->getValueType(0);
15242 if (N->getOpcode() == ISD::SHL) {
15243 SDValue V = PerformSHLCombine(N, DAG);
15244 if (V.getNode()) return V;
15247 // On X86 with SSE2 support, we can transform this to a vector shift if
15248 // all elements are shifted by the same amount. We can't do this in legalize
15249 // because the a constant vector is typically transformed to a constant pool
15250 // so we have no knowledge of the shift amount.
15251 if (!Subtarget->hasSSE2())
15254 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
15255 (!Subtarget->hasAVX2() ||
15256 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
15259 SDValue ShAmtOp = N->getOperand(1);
15260 EVT EltVT = VT.getVectorElementType();
15261 DebugLoc DL = N->getDebugLoc();
15262 SDValue BaseShAmt = SDValue();
15263 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
15264 unsigned NumElts = VT.getVectorNumElements();
15266 for (; i != NumElts; ++i) {
15267 SDValue Arg = ShAmtOp.getOperand(i);
15268 if (Arg.getOpcode() == ISD::UNDEF) continue;
15272 // Handle the case where the build_vector is all undef
15273 // FIXME: Should DAG allow this?
15277 for (; i != NumElts; ++i) {
15278 SDValue Arg = ShAmtOp.getOperand(i);
15279 if (Arg.getOpcode() == ISD::UNDEF) continue;
15280 if (Arg != BaseShAmt) {
15284 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
15285 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
15286 SDValue InVec = ShAmtOp.getOperand(0);
15287 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
15288 unsigned NumElts = InVec.getValueType().getVectorNumElements();
15290 for (; i != NumElts; ++i) {
15291 SDValue Arg = InVec.getOperand(i);
15292 if (Arg.getOpcode() == ISD::UNDEF) continue;
15296 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
15297 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
15298 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
15299 if (C->getZExtValue() == SplatIdx)
15300 BaseShAmt = InVec.getOperand(1);
15303 if (BaseShAmt.getNode() == 0) {
15304 // Don't create instructions with illegal types after legalize
15306 if (!DAG.getTargetLoweringInfo().isTypeLegal(EltVT) &&
15307 !DCI.isBeforeLegalize())
15310 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
15311 DAG.getIntPtrConstant(0));
15316 // The shift amount is an i32.
15317 if (EltVT.bitsGT(MVT::i32))
15318 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
15319 else if (EltVT.bitsLT(MVT::i32))
15320 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
15322 // The shift amount is identical so we can do a vector shift.
15323 SDValue ValOp = N->getOperand(0);
15324 switch (N->getOpcode()) {
15326 llvm_unreachable("Unknown shift opcode!");
15328 switch (VT.getSimpleVT().SimpleTy) {
15329 default: return SDValue();
15336 return getTargetVShiftNode(X86ISD::VSHLI, DL, VT, ValOp, BaseShAmt, DAG);
15339 switch (VT.getSimpleVT().SimpleTy) {
15340 default: return SDValue();
15345 return getTargetVShiftNode(X86ISD::VSRAI, DL, VT, ValOp, BaseShAmt, DAG);
15348 switch (VT.getSimpleVT().SimpleTy) {
15349 default: return SDValue();
15356 return getTargetVShiftNode(X86ISD::VSRLI, DL, VT, ValOp, BaseShAmt, DAG);
15362 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
15363 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
15364 // and friends. Likewise for OR -> CMPNEQSS.
15365 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
15366 TargetLowering::DAGCombinerInfo &DCI,
15367 const X86Subtarget *Subtarget) {
15370 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
15371 // we're requiring SSE2 for both.
15372 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
15373 SDValue N0 = N->getOperand(0);
15374 SDValue N1 = N->getOperand(1);
15375 SDValue CMP0 = N0->getOperand(1);
15376 SDValue CMP1 = N1->getOperand(1);
15377 DebugLoc DL = N->getDebugLoc();
15379 // The SETCCs should both refer to the same CMP.
15380 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
15383 SDValue CMP00 = CMP0->getOperand(0);
15384 SDValue CMP01 = CMP0->getOperand(1);
15385 EVT VT = CMP00.getValueType();
15387 if (VT == MVT::f32 || VT == MVT::f64) {
15388 bool ExpectingFlags = false;
15389 // Check for any users that want flags:
15390 for (SDNode::use_iterator UI = N->use_begin(),
15392 !ExpectingFlags && UI != UE; ++UI)
15393 switch (UI->getOpcode()) {
15398 ExpectingFlags = true;
15400 case ISD::CopyToReg:
15401 case ISD::SIGN_EXTEND:
15402 case ISD::ZERO_EXTEND:
15403 case ISD::ANY_EXTEND:
15407 if (!ExpectingFlags) {
15408 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
15409 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
15411 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
15412 X86::CondCode tmp = cc0;
15417 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
15418 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
15419 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
15420 X86ISD::NodeType NTOperator = is64BitFP ?
15421 X86ISD::FSETCCsd : X86ISD::FSETCCss;
15422 // FIXME: need symbolic constants for these magic numbers.
15423 // See X86ATTInstPrinter.cpp:printSSECC().
15424 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
15425 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
15426 DAG.getConstant(x86cc, MVT::i8));
15427 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
15429 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
15430 DAG.getConstant(1, MVT::i32));
15431 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
15432 return OneBitOfTruth;
15440 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
15441 /// so it can be folded inside ANDNP.
15442 static bool CanFoldXORWithAllOnes(const SDNode *N) {
15443 EVT VT = N->getValueType(0);
15445 // Match direct AllOnes for 128 and 256-bit vectors
15446 if (ISD::isBuildVectorAllOnes(N))
15449 // Look through a bit convert.
15450 if (N->getOpcode() == ISD::BITCAST)
15451 N = N->getOperand(0).getNode();
15453 // Sometimes the operand may come from a insert_subvector building a 256-bit
15455 if (VT.is256BitVector() &&
15456 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
15457 SDValue V1 = N->getOperand(0);
15458 SDValue V2 = N->getOperand(1);
15460 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
15461 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
15462 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
15463 ISD::isBuildVectorAllOnes(V2.getNode()))
15470 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
15471 TargetLowering::DAGCombinerInfo &DCI,
15472 const X86Subtarget *Subtarget) {
15473 if (DCI.isBeforeLegalizeOps())
15476 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
15480 EVT VT = N->getValueType(0);
15482 // Create ANDN, BLSI, and BLSR instructions
15483 // BLSI is X & (-X)
15484 // BLSR is X & (X-1)
15485 if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
15486 SDValue N0 = N->getOperand(0);
15487 SDValue N1 = N->getOperand(1);
15488 DebugLoc DL = N->getDebugLoc();
15490 // Check LHS for not
15491 if (N0.getOpcode() == ISD::XOR && isAllOnes(N0.getOperand(1)))
15492 return DAG.getNode(X86ISD::ANDN, DL, VT, N0.getOperand(0), N1);
15493 // Check RHS for not
15494 if (N1.getOpcode() == ISD::XOR && isAllOnes(N1.getOperand(1)))
15495 return DAG.getNode(X86ISD::ANDN, DL, VT, N1.getOperand(0), N0);
15497 // Check LHS for neg
15498 if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
15499 isZero(N0.getOperand(0)))
15500 return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
15502 // Check RHS for neg
15503 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
15504 isZero(N1.getOperand(0)))
15505 return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
15507 // Check LHS for X-1
15508 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
15509 isAllOnes(N0.getOperand(1)))
15510 return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
15512 // Check RHS for X-1
15513 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
15514 isAllOnes(N1.getOperand(1)))
15515 return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
15520 // Want to form ANDNP nodes:
15521 // 1) In the hopes of then easily combining them with OR and AND nodes
15522 // to form PBLEND/PSIGN.
15523 // 2) To match ANDN packed intrinsics
15524 if (VT != MVT::v2i64 && VT != MVT::v4i64)
15527 SDValue N0 = N->getOperand(0);
15528 SDValue N1 = N->getOperand(1);
15529 DebugLoc DL = N->getDebugLoc();
15531 // Check LHS for vnot
15532 if (N0.getOpcode() == ISD::XOR &&
15533 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
15534 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
15535 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
15537 // Check RHS for vnot
15538 if (N1.getOpcode() == ISD::XOR &&
15539 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
15540 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
15541 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
15546 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
15547 TargetLowering::DAGCombinerInfo &DCI,
15548 const X86Subtarget *Subtarget) {
15549 if (DCI.isBeforeLegalizeOps())
15552 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
15556 EVT VT = N->getValueType(0);
15558 SDValue N0 = N->getOperand(0);
15559 SDValue N1 = N->getOperand(1);
15561 // look for psign/blend
15562 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
15563 if (!Subtarget->hasSSSE3() ||
15564 (VT == MVT::v4i64 && !Subtarget->hasAVX2()))
15567 // Canonicalize pandn to RHS
15568 if (N0.getOpcode() == X86ISD::ANDNP)
15570 // or (and (m, y), (pandn m, x))
15571 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
15572 SDValue Mask = N1.getOperand(0);
15573 SDValue X = N1.getOperand(1);
15575 if (N0.getOperand(0) == Mask)
15576 Y = N0.getOperand(1);
15577 if (N0.getOperand(1) == Mask)
15578 Y = N0.getOperand(0);
15580 // Check to see if the mask appeared in both the AND and ANDNP and
15584 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
15585 // Look through mask bitcast.
15586 if (Mask.getOpcode() == ISD::BITCAST)
15587 Mask = Mask.getOperand(0);
15588 if (X.getOpcode() == ISD::BITCAST)
15589 X = X.getOperand(0);
15590 if (Y.getOpcode() == ISD::BITCAST)
15591 Y = Y.getOperand(0);
15593 EVT MaskVT = Mask.getValueType();
15595 // Validate that the Mask operand is a vector sra node.
15596 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
15597 // there is no psrai.b
15598 if (Mask.getOpcode() != X86ISD::VSRAI)
15601 // Check that the SRA is all signbits.
15602 SDValue SraC = Mask.getOperand(1);
15603 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
15604 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
15605 if ((SraAmt + 1) != EltBits)
15608 DebugLoc DL = N->getDebugLoc();
15610 // Now we know we at least have a plendvb with the mask val. See if
15611 // we can form a psignb/w/d.
15612 // psign = x.type == y.type == mask.type && y = sub(0, x);
15613 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
15614 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
15615 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
15616 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
15617 "Unsupported VT for PSIGN");
15618 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
15619 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
15621 // PBLENDVB only available on SSE 4.1
15622 if (!Subtarget->hasSSE41())
15625 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
15627 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
15628 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
15629 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
15630 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
15631 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
15635 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
15638 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
15639 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
15641 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
15643 if (!N0.hasOneUse() || !N1.hasOneUse())
15646 SDValue ShAmt0 = N0.getOperand(1);
15647 if (ShAmt0.getValueType() != MVT::i8)
15649 SDValue ShAmt1 = N1.getOperand(1);
15650 if (ShAmt1.getValueType() != MVT::i8)
15652 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
15653 ShAmt0 = ShAmt0.getOperand(0);
15654 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
15655 ShAmt1 = ShAmt1.getOperand(0);
15657 DebugLoc DL = N->getDebugLoc();
15658 unsigned Opc = X86ISD::SHLD;
15659 SDValue Op0 = N0.getOperand(0);
15660 SDValue Op1 = N1.getOperand(0);
15661 if (ShAmt0.getOpcode() == ISD::SUB) {
15662 Opc = X86ISD::SHRD;
15663 std::swap(Op0, Op1);
15664 std::swap(ShAmt0, ShAmt1);
15667 unsigned Bits = VT.getSizeInBits();
15668 if (ShAmt1.getOpcode() == ISD::SUB) {
15669 SDValue Sum = ShAmt1.getOperand(0);
15670 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
15671 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
15672 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
15673 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
15674 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
15675 return DAG.getNode(Opc, DL, VT,
15677 DAG.getNode(ISD::TRUNCATE, DL,
15680 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
15681 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
15683 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
15684 return DAG.getNode(Opc, DL, VT,
15685 N0.getOperand(0), N1.getOperand(0),
15686 DAG.getNode(ISD::TRUNCATE, DL,
15693 // Generate NEG and CMOV for integer abs.
15694 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
15695 EVT VT = N->getValueType(0);
15697 // Since X86 does not have CMOV for 8-bit integer, we don't convert
15698 // 8-bit integer abs to NEG and CMOV.
15699 if (VT.isInteger() && VT.getSizeInBits() == 8)
15702 SDValue N0 = N->getOperand(0);
15703 SDValue N1 = N->getOperand(1);
15704 DebugLoc DL = N->getDebugLoc();
15706 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
15707 // and change it to SUB and CMOV.
15708 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
15709 N0.getOpcode() == ISD::ADD &&
15710 N0.getOperand(1) == N1 &&
15711 N1.getOpcode() == ISD::SRA &&
15712 N1.getOperand(0) == N0.getOperand(0))
15713 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
15714 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
15715 // Generate SUB & CMOV.
15716 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
15717 DAG.getConstant(0, VT), N0.getOperand(0));
15719 SDValue Ops[] = { N0.getOperand(0), Neg,
15720 DAG.getConstant(X86::COND_GE, MVT::i8),
15721 SDValue(Neg.getNode(), 1) };
15722 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue),
15723 Ops, array_lengthof(Ops));
15728 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
15729 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
15730 TargetLowering::DAGCombinerInfo &DCI,
15731 const X86Subtarget *Subtarget) {
15732 if (DCI.isBeforeLegalizeOps())
15735 if (Subtarget->hasCMov()) {
15736 SDValue RV = performIntegerAbsCombine(N, DAG);
15741 // Try forming BMI if it is available.
15742 if (!Subtarget->hasBMI())
15745 EVT VT = N->getValueType(0);
15747 if (VT != MVT::i32 && VT != MVT::i64)
15750 assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
15752 // Create BLSMSK instructions by finding X ^ (X-1)
15753 SDValue N0 = N->getOperand(0);
15754 SDValue N1 = N->getOperand(1);
15755 DebugLoc DL = N->getDebugLoc();
15757 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
15758 isAllOnes(N0.getOperand(1)))
15759 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
15761 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
15762 isAllOnes(N1.getOperand(1)))
15763 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
15768 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
15769 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
15770 TargetLowering::DAGCombinerInfo &DCI,
15771 const X86Subtarget *Subtarget) {
15772 LoadSDNode *Ld = cast<LoadSDNode>(N);
15773 EVT RegVT = Ld->getValueType(0);
15774 EVT MemVT = Ld->getMemoryVT();
15775 DebugLoc dl = Ld->getDebugLoc();
15776 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15778 ISD::LoadExtType Ext = Ld->getExtensionType();
15780 // If this is a vector EXT Load then attempt to optimize it using a
15781 // shuffle. We need SSSE3 shuffles.
15782 // TODO: It is possible to support ZExt by zeroing the undef values
15783 // during the shuffle phase or after the shuffle.
15784 if (RegVT.isVector() && RegVT.isInteger() &&
15785 Ext == ISD::EXTLOAD && Subtarget->hasSSSE3()) {
15786 assert(MemVT != RegVT && "Cannot extend to the same type");
15787 assert(MemVT.isVector() && "Must load a vector from memory");
15789 unsigned NumElems = RegVT.getVectorNumElements();
15790 unsigned RegSz = RegVT.getSizeInBits();
15791 unsigned MemSz = MemVT.getSizeInBits();
15792 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15794 // All sizes must be a power of two.
15795 if (!isPowerOf2_32(RegSz * MemSz * NumElems))
15798 // Attempt to load the original value using scalar loads.
15799 // Find the largest scalar type that divides the total loaded size.
15800 MVT SclrLoadTy = MVT::i8;
15801 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
15802 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
15803 MVT Tp = (MVT::SimpleValueType)tp;
15804 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15809 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15810 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15812 SclrLoadTy = MVT::f64;
15814 // Calculate the number of scalar loads that we need to perform
15815 // in order to load our vector from memory.
15816 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15818 // Represent our vector as a sequence of elements which are the
15819 // largest scalar that we can load.
15820 EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
15821 RegSz/SclrLoadTy.getSizeInBits());
15823 // Represent the data using the same element type that is stored in
15824 // memory. In practice, we ''widen'' MemVT.
15825 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15826 RegSz/MemVT.getScalarType().getSizeInBits());
15828 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15829 "Invalid vector type");
15831 // We can't shuffle using an illegal type.
15832 if (!TLI.isTypeLegal(WideVecVT))
15835 SmallVector<SDValue, 8> Chains;
15836 SDValue Ptr = Ld->getBasePtr();
15837 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
15838 TLI.getPointerTy());
15839 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15841 for (unsigned i = 0; i < NumLoads; ++i) {
15842 // Perform a single load.
15843 SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
15844 Ptr, Ld->getPointerInfo(),
15845 Ld->isVolatile(), Ld->isNonTemporal(),
15846 Ld->isInvariant(), Ld->getAlignment());
15847 Chains.push_back(ScalarLoad.getValue(1));
15848 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15849 // another round of DAGCombining.
15851 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15853 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15854 ScalarLoad, DAG.getIntPtrConstant(i));
15856 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15859 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
15862 // Bitcast the loaded value to a vector of the original element type, in
15863 // the size of the target vector type.
15864 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
15865 unsigned SizeRatio = RegSz/MemSz;
15867 // Redistribute the loaded elements into the different locations.
15868 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
15869 for (unsigned i = 0; i != NumElems; ++i)
15870 ShuffleVec[i*SizeRatio] = i;
15872 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15873 DAG.getUNDEF(WideVecVT),
15876 // Bitcast to the requested type.
15877 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
15878 // Replace the original load with the new sequence
15879 // and return the new chain.
15880 return DCI.CombineTo(N, Shuff, TF, true);
15886 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
15887 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
15888 const X86Subtarget *Subtarget) {
15889 StoreSDNode *St = cast<StoreSDNode>(N);
15890 EVT VT = St->getValue().getValueType();
15891 EVT StVT = St->getMemoryVT();
15892 DebugLoc dl = St->getDebugLoc();
15893 SDValue StoredVal = St->getOperand(1);
15894 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15896 // If we are saving a concatenation of two XMM registers, perform two stores.
15897 // On Sandy Bridge, 256-bit memory operations are executed by two
15898 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
15899 // memory operation.
15900 if (VT.is256BitVector() && !Subtarget->hasAVX2() &&
15901 StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
15902 StoredVal.getNumOperands() == 2) {
15903 SDValue Value0 = StoredVal.getOperand(0);
15904 SDValue Value1 = StoredVal.getOperand(1);
15906 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
15907 SDValue Ptr0 = St->getBasePtr();
15908 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
15910 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
15911 St->getPointerInfo(), St->isVolatile(),
15912 St->isNonTemporal(), St->getAlignment());
15913 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
15914 St->getPointerInfo(), St->isVolatile(),
15915 St->isNonTemporal(), St->getAlignment());
15916 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
15919 // Optimize trunc store (of multiple scalars) to shuffle and store.
15920 // First, pack all of the elements in one place. Next, store to memory
15921 // in fewer chunks.
15922 if (St->isTruncatingStore() && VT.isVector()) {
15923 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15924 unsigned NumElems = VT.getVectorNumElements();
15925 assert(StVT != VT && "Cannot truncate to the same type");
15926 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
15927 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
15929 // From, To sizes and ElemCount must be pow of two
15930 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
15931 // We are going to use the original vector elt for storing.
15932 // Accumulated smaller vector elements must be a multiple of the store size.
15933 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
15935 unsigned SizeRatio = FromSz / ToSz;
15937 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
15939 // Create a type on which we perform the shuffle
15940 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
15941 StVT.getScalarType(), NumElems*SizeRatio);
15943 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
15945 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
15946 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
15947 for (unsigned i = 0; i != NumElems; ++i)
15948 ShuffleVec[i] = i * SizeRatio;
15950 // Can't shuffle using an illegal type.
15951 if (!TLI.isTypeLegal(WideVecVT))
15954 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
15955 DAG.getUNDEF(WideVecVT),
15957 // At this point all of the data is stored at the bottom of the
15958 // register. We now need to save it to mem.
15960 // Find the largest store unit
15961 MVT StoreType = MVT::i8;
15962 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
15963 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
15964 MVT Tp = (MVT::SimpleValueType)tp;
15965 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
15969 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15970 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
15971 (64 <= NumElems * ToSz))
15972 StoreType = MVT::f64;
15974 // Bitcast the original vector into a vector of store-size units
15975 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
15976 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
15977 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
15978 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
15979 SmallVector<SDValue, 8> Chains;
15980 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
15981 TLI.getPointerTy());
15982 SDValue Ptr = St->getBasePtr();
15984 // Perform one or more big stores into memory.
15985 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
15986 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
15987 StoreType, ShuffWide,
15988 DAG.getIntPtrConstant(i));
15989 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
15990 St->getPointerInfo(), St->isVolatile(),
15991 St->isNonTemporal(), St->getAlignment());
15992 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15993 Chains.push_back(Ch);
15996 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
16001 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
16002 // the FP state in cases where an emms may be missing.
16003 // A preferable solution to the general problem is to figure out the right
16004 // places to insert EMMS. This qualifies as a quick hack.
16006 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
16007 if (VT.getSizeInBits() != 64)
16010 const Function *F = DAG.getMachineFunction().getFunction();
16011 bool NoImplicitFloatOps = F->getFnAttributes().
16012 hasAttribute(Attributes::NoImplicitFloat);
16013 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
16014 && Subtarget->hasSSE2();
16015 if ((VT.isVector() ||
16016 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
16017 isa<LoadSDNode>(St->getValue()) &&
16018 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
16019 St->getChain().hasOneUse() && !St->isVolatile()) {
16020 SDNode* LdVal = St->getValue().getNode();
16021 LoadSDNode *Ld = 0;
16022 int TokenFactorIndex = -1;
16023 SmallVector<SDValue, 8> Ops;
16024 SDNode* ChainVal = St->getChain().getNode();
16025 // Must be a store of a load. We currently handle two cases: the load
16026 // is a direct child, and it's under an intervening TokenFactor. It is
16027 // possible to dig deeper under nested TokenFactors.
16028 if (ChainVal == LdVal)
16029 Ld = cast<LoadSDNode>(St->getChain());
16030 else if (St->getValue().hasOneUse() &&
16031 ChainVal->getOpcode() == ISD::TokenFactor) {
16032 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
16033 if (ChainVal->getOperand(i).getNode() == LdVal) {
16034 TokenFactorIndex = i;
16035 Ld = cast<LoadSDNode>(St->getValue());
16037 Ops.push_back(ChainVal->getOperand(i));
16041 if (!Ld || !ISD::isNormalLoad(Ld))
16044 // If this is not the MMX case, i.e. we are just turning i64 load/store
16045 // into f64 load/store, avoid the transformation if there are multiple
16046 // uses of the loaded value.
16047 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
16050 DebugLoc LdDL = Ld->getDebugLoc();
16051 DebugLoc StDL = N->getDebugLoc();
16052 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
16053 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
16055 if (Subtarget->is64Bit() || F64IsLegal) {
16056 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
16057 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
16058 Ld->getPointerInfo(), Ld->isVolatile(),
16059 Ld->isNonTemporal(), Ld->isInvariant(),
16060 Ld->getAlignment());
16061 SDValue NewChain = NewLd.getValue(1);
16062 if (TokenFactorIndex != -1) {
16063 Ops.push_back(NewChain);
16064 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
16067 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
16068 St->getPointerInfo(),
16069 St->isVolatile(), St->isNonTemporal(),
16070 St->getAlignment());
16073 // Otherwise, lower to two pairs of 32-bit loads / stores.
16074 SDValue LoAddr = Ld->getBasePtr();
16075 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
16076 DAG.getConstant(4, MVT::i32));
16078 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
16079 Ld->getPointerInfo(),
16080 Ld->isVolatile(), Ld->isNonTemporal(),
16081 Ld->isInvariant(), Ld->getAlignment());
16082 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
16083 Ld->getPointerInfo().getWithOffset(4),
16084 Ld->isVolatile(), Ld->isNonTemporal(),
16086 MinAlign(Ld->getAlignment(), 4));
16088 SDValue NewChain = LoLd.getValue(1);
16089 if (TokenFactorIndex != -1) {
16090 Ops.push_back(LoLd);
16091 Ops.push_back(HiLd);
16092 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
16096 LoAddr = St->getBasePtr();
16097 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
16098 DAG.getConstant(4, MVT::i32));
16100 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
16101 St->getPointerInfo(),
16102 St->isVolatile(), St->isNonTemporal(),
16103 St->getAlignment());
16104 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
16105 St->getPointerInfo().getWithOffset(4),
16107 St->isNonTemporal(),
16108 MinAlign(St->getAlignment(), 4));
16109 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
16114 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
16115 /// and return the operands for the horizontal operation in LHS and RHS. A
16116 /// horizontal operation performs the binary operation on successive elements
16117 /// of its first operand, then on successive elements of its second operand,
16118 /// returning the resulting values in a vector. For example, if
16119 /// A = < float a0, float a1, float a2, float a3 >
16121 /// B = < float b0, float b1, float b2, float b3 >
16122 /// then the result of doing a horizontal operation on A and B is
16123 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
16124 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
16125 /// A horizontal-op B, for some already available A and B, and if so then LHS is
16126 /// set to A, RHS to B, and the routine returns 'true'.
16127 /// Note that the binary operation should have the property that if one of the
16128 /// operands is UNDEF then the result is UNDEF.
16129 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
16130 // Look for the following pattern: if
16131 // A = < float a0, float a1, float a2, float a3 >
16132 // B = < float b0, float b1, float b2, float b3 >
16134 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
16135 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
16136 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
16137 // which is A horizontal-op B.
16139 // At least one of the operands should be a vector shuffle.
16140 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
16141 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
16144 EVT VT = LHS.getValueType();
16146 assert((VT.is128BitVector() || VT.is256BitVector()) &&
16147 "Unsupported vector type for horizontal add/sub");
16149 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
16150 // operate independently on 128-bit lanes.
16151 unsigned NumElts = VT.getVectorNumElements();
16152 unsigned NumLanes = VT.getSizeInBits()/128;
16153 unsigned NumLaneElts = NumElts / NumLanes;
16154 assert((NumLaneElts % 2 == 0) &&
16155 "Vector type should have an even number of elements in each lane");
16156 unsigned HalfLaneElts = NumLaneElts/2;
16158 // View LHS in the form
16159 // LHS = VECTOR_SHUFFLE A, B, LMask
16160 // If LHS is not a shuffle then pretend it is the shuffle
16161 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
16162 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
16165 SmallVector<int, 16> LMask(NumElts);
16166 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
16167 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
16168 A = LHS.getOperand(0);
16169 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
16170 B = LHS.getOperand(1);
16171 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
16172 std::copy(Mask.begin(), Mask.end(), LMask.begin());
16174 if (LHS.getOpcode() != ISD::UNDEF)
16176 for (unsigned i = 0; i != NumElts; ++i)
16180 // Likewise, view RHS in the form
16181 // RHS = VECTOR_SHUFFLE C, D, RMask
16183 SmallVector<int, 16> RMask(NumElts);
16184 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
16185 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
16186 C = RHS.getOperand(0);
16187 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
16188 D = RHS.getOperand(1);
16189 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
16190 std::copy(Mask.begin(), Mask.end(), RMask.begin());
16192 if (RHS.getOpcode() != ISD::UNDEF)
16194 for (unsigned i = 0; i != NumElts; ++i)
16198 // Check that the shuffles are both shuffling the same vectors.
16199 if (!(A == C && B == D) && !(A == D && B == C))
16202 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
16203 if (!A.getNode() && !B.getNode())
16206 // If A and B occur in reverse order in RHS, then "swap" them (which means
16207 // rewriting the mask).
16209 CommuteVectorShuffleMask(RMask, NumElts);
16211 // At this point LHS and RHS are equivalent to
16212 // LHS = VECTOR_SHUFFLE A, B, LMask
16213 // RHS = VECTOR_SHUFFLE A, B, RMask
16214 // Check that the masks correspond to performing a horizontal operation.
16215 for (unsigned i = 0; i != NumElts; ++i) {
16216 int LIdx = LMask[i], RIdx = RMask[i];
16218 // Ignore any UNDEF components.
16219 if (LIdx < 0 || RIdx < 0 ||
16220 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
16221 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
16224 // Check that successive elements are being operated on. If not, this is
16225 // not a horizontal operation.
16226 unsigned Src = (i/HalfLaneElts) % 2; // each lane is split between srcs
16227 unsigned LaneStart = (i/NumLaneElts) * NumLaneElts;
16228 int Index = 2*(i%HalfLaneElts) + NumElts*Src + LaneStart;
16229 if (!(LIdx == Index && RIdx == Index + 1) &&
16230 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
16234 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
16235 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
16239 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
16240 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
16241 const X86Subtarget *Subtarget) {
16242 EVT VT = N->getValueType(0);
16243 SDValue LHS = N->getOperand(0);
16244 SDValue RHS = N->getOperand(1);
16246 // Try to synthesize horizontal adds from adds of shuffles.
16247 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
16248 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
16249 isHorizontalBinOp(LHS, RHS, true))
16250 return DAG.getNode(X86ISD::FHADD, N->getDebugLoc(), VT, LHS, RHS);
16254 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
16255 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
16256 const X86Subtarget *Subtarget) {
16257 EVT VT = N->getValueType(0);
16258 SDValue LHS = N->getOperand(0);
16259 SDValue RHS = N->getOperand(1);
16261 // Try to synthesize horizontal subs from subs of shuffles.
16262 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
16263 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
16264 isHorizontalBinOp(LHS, RHS, false))
16265 return DAG.getNode(X86ISD::FHSUB, N->getDebugLoc(), VT, LHS, RHS);
16269 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
16270 /// X86ISD::FXOR nodes.
16271 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
16272 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
16273 // F[X]OR(0.0, x) -> x
16274 // F[X]OR(x, 0.0) -> x
16275 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
16276 if (C->getValueAPF().isPosZero())
16277 return N->getOperand(1);
16278 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
16279 if (C->getValueAPF().isPosZero())
16280 return N->getOperand(0);
16284 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
16285 /// X86ISD::FMAX nodes.
16286 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
16287 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
16289 // Only perform optimizations if UnsafeMath is used.
16290 if (!DAG.getTarget().Options.UnsafeFPMath)
16293 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
16294 // into FMINC and FMAXC, which are Commutative operations.
16295 unsigned NewOp = 0;
16296 switch (N->getOpcode()) {
16297 default: llvm_unreachable("unknown opcode");
16298 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
16299 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
16302 return DAG.getNode(NewOp, N->getDebugLoc(), N->getValueType(0),
16303 N->getOperand(0), N->getOperand(1));
16307 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
16308 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
16309 // FAND(0.0, x) -> 0.0
16310 // FAND(x, 0.0) -> 0.0
16311 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
16312 if (C->getValueAPF().isPosZero())
16313 return N->getOperand(0);
16314 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
16315 if (C->getValueAPF().isPosZero())
16316 return N->getOperand(1);
16320 static SDValue PerformBTCombine(SDNode *N,
16322 TargetLowering::DAGCombinerInfo &DCI) {
16323 // BT ignores high bits in the bit index operand.
16324 SDValue Op1 = N->getOperand(1);
16325 if (Op1.hasOneUse()) {
16326 unsigned BitWidth = Op1.getValueSizeInBits();
16327 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
16328 APInt KnownZero, KnownOne;
16329 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
16330 !DCI.isBeforeLegalizeOps());
16331 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16332 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
16333 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
16334 DCI.CommitTargetLoweringOpt(TLO);
16339 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
16340 SDValue Op = N->getOperand(0);
16341 if (Op.getOpcode() == ISD::BITCAST)
16342 Op = Op.getOperand(0);
16343 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
16344 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
16345 VT.getVectorElementType().getSizeInBits() ==
16346 OpVT.getVectorElementType().getSizeInBits()) {
16347 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
16352 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
16353 TargetLowering::DAGCombinerInfo &DCI,
16354 const X86Subtarget *Subtarget) {
16355 if (!DCI.isBeforeLegalizeOps())
16358 if (!Subtarget->hasAVX())
16361 EVT VT = N->getValueType(0);
16362 SDValue Op = N->getOperand(0);
16363 EVT OpVT = Op.getValueType();
16364 DebugLoc dl = N->getDebugLoc();
16366 if ((VT == MVT::v4i64 && OpVT == MVT::v4i32) ||
16367 (VT == MVT::v8i32 && OpVT == MVT::v8i16)) {
16369 if (Subtarget->hasAVX2())
16370 return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, Op);
16372 // Optimize vectors in AVX mode
16373 // Sign extend v8i16 to v8i32 and
16376 // Divide input vector into two parts
16377 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
16378 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
16379 // concat the vectors to original VT
16381 unsigned NumElems = OpVT.getVectorNumElements();
16382 SDValue Undef = DAG.getUNDEF(OpVT);
16384 SmallVector<int,8> ShufMask1(NumElems, -1);
16385 for (unsigned i = 0; i != NumElems/2; ++i)
16388 SDValue OpLo = DAG.getVectorShuffle(OpVT, dl, Op, Undef, &ShufMask1[0]);
16390 SmallVector<int,8> ShufMask2(NumElems, -1);
16391 for (unsigned i = 0; i != NumElems/2; ++i)
16392 ShufMask2[i] = i + NumElems/2;
16394 SDValue OpHi = DAG.getVectorShuffle(OpVT, dl, Op, Undef, &ShufMask2[0]);
16396 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(),
16397 VT.getVectorNumElements()/2);
16399 OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
16400 OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
16402 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
16407 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
16408 const X86Subtarget* Subtarget) {
16409 DebugLoc dl = N->getDebugLoc();
16410 EVT VT = N->getValueType(0);
16412 // Let legalize expand this if it isn't a legal type yet.
16413 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
16416 EVT ScalarVT = VT.getScalarType();
16417 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
16418 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
16421 SDValue A = N->getOperand(0);
16422 SDValue B = N->getOperand(1);
16423 SDValue C = N->getOperand(2);
16425 bool NegA = (A.getOpcode() == ISD::FNEG);
16426 bool NegB = (B.getOpcode() == ISD::FNEG);
16427 bool NegC = (C.getOpcode() == ISD::FNEG);
16429 // Negative multiplication when NegA xor NegB
16430 bool NegMul = (NegA != NegB);
16432 A = A.getOperand(0);
16434 B = B.getOperand(0);
16436 C = C.getOperand(0);
16440 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
16442 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
16444 return DAG.getNode(Opcode, dl, VT, A, B, C);
16447 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
16448 TargetLowering::DAGCombinerInfo &DCI,
16449 const X86Subtarget *Subtarget) {
16450 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
16451 // (and (i32 x86isd::setcc_carry), 1)
16452 // This eliminates the zext. This transformation is necessary because
16453 // ISD::SETCC is always legalized to i8.
16454 DebugLoc dl = N->getDebugLoc();
16455 SDValue N0 = N->getOperand(0);
16456 EVT VT = N->getValueType(0);
16457 EVT OpVT = N0.getValueType();
16459 if (N0.getOpcode() == ISD::AND &&
16461 N0.getOperand(0).hasOneUse()) {
16462 SDValue N00 = N0.getOperand(0);
16463 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
16465 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
16466 if (!C || C->getZExtValue() != 1)
16468 return DAG.getNode(ISD::AND, dl, VT,
16469 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
16470 N00.getOperand(0), N00.getOperand(1)),
16471 DAG.getConstant(1, VT));
16474 // Optimize vectors in AVX mode:
16477 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
16478 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
16479 // Concat upper and lower parts.
16482 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
16483 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
16484 // Concat upper and lower parts.
16486 if (!DCI.isBeforeLegalizeOps())
16489 if (!Subtarget->hasAVX())
16492 if (((VT == MVT::v8i32) && (OpVT == MVT::v8i16)) ||
16493 ((VT == MVT::v4i64) && (OpVT == MVT::v4i32))) {
16495 if (Subtarget->hasAVX2())
16496 return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, N0);
16498 SDValue ZeroVec = getZeroVector(OpVT, Subtarget, DAG, dl);
16499 SDValue OpLo = getUnpackl(DAG, dl, OpVT, N0, ZeroVec);
16500 SDValue OpHi = getUnpackh(DAG, dl, OpVT, N0, ZeroVec);
16502 EVT HVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
16503 VT.getVectorNumElements()/2);
16505 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
16506 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
16508 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
16514 // Optimize x == -y --> x+y == 0
16515 // x != -y --> x+y != 0
16516 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
16517 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
16518 SDValue LHS = N->getOperand(0);
16519 SDValue RHS = N->getOperand(1);
16521 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
16522 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
16523 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
16524 SDValue addV = DAG.getNode(ISD::ADD, N->getDebugLoc(),
16525 LHS.getValueType(), RHS, LHS.getOperand(1));
16526 return DAG.getSetCC(N->getDebugLoc(), N->getValueType(0),
16527 addV, DAG.getConstant(0, addV.getValueType()), CC);
16529 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
16530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
16531 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
16532 SDValue addV = DAG.getNode(ISD::ADD, N->getDebugLoc(),
16533 RHS.getValueType(), LHS, RHS.getOperand(1));
16534 return DAG.getSetCC(N->getDebugLoc(), N->getValueType(0),
16535 addV, DAG.getConstant(0, addV.getValueType()), CC);
16540 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
16541 // as "sbb reg,reg", since it can be extended without zext and produces
16542 // an all-ones bit which is more useful than 0/1 in some cases.
16543 static SDValue MaterializeSETB(DebugLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
16544 return DAG.getNode(ISD::AND, DL, MVT::i8,
16545 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
16546 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
16547 DAG.getConstant(1, MVT::i8));
16550 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
16551 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
16552 TargetLowering::DAGCombinerInfo &DCI,
16553 const X86Subtarget *Subtarget) {
16554 DebugLoc DL = N->getDebugLoc();
16555 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
16556 SDValue EFLAGS = N->getOperand(1);
16558 if (CC == X86::COND_A) {
16559 // Try to convert COND_A into COND_B in an attempt to facilitate
16560 // materializing "setb reg".
16562 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
16563 // cannot take an immediate as its first operand.
16565 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
16566 EFLAGS.getValueType().isInteger() &&
16567 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
16568 SDValue NewSub = DAG.getNode(X86ISD::SUB, EFLAGS.getDebugLoc(),
16569 EFLAGS.getNode()->getVTList(),
16570 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
16571 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
16572 return MaterializeSETB(DL, NewEFLAGS, DAG);
16576 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
16577 // a zext and produces an all-ones bit which is more useful than 0/1 in some
16579 if (CC == X86::COND_B)
16580 return MaterializeSETB(DL, EFLAGS, DAG);
16584 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
16585 if (Flags.getNode()) {
16586 SDValue Cond = DAG.getConstant(CC, MVT::i8);
16587 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
16593 // Optimize branch condition evaluation.
16595 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
16596 TargetLowering::DAGCombinerInfo &DCI,
16597 const X86Subtarget *Subtarget) {
16598 DebugLoc DL = N->getDebugLoc();
16599 SDValue Chain = N->getOperand(0);
16600 SDValue Dest = N->getOperand(1);
16601 SDValue EFLAGS = N->getOperand(3);
16602 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
16606 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
16607 if (Flags.getNode()) {
16608 SDValue Cond = DAG.getConstant(CC, MVT::i8);
16609 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
16616 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
16617 const X86TargetLowering *XTLI) {
16618 SDValue Op0 = N->getOperand(0);
16619 EVT InVT = Op0->getValueType(0);
16621 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
16622 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
16623 DebugLoc dl = N->getDebugLoc();
16624 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
16625 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
16626 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
16629 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
16630 // a 32-bit target where SSE doesn't support i64->FP operations.
16631 if (Op0.getOpcode() == ISD::LOAD) {
16632 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
16633 EVT VT = Ld->getValueType(0);
16634 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
16635 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
16636 !XTLI->getSubtarget()->is64Bit() &&
16637 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
16638 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
16639 Ld->getChain(), Op0, DAG);
16640 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
16647 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
16648 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
16649 X86TargetLowering::DAGCombinerInfo &DCI) {
16650 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
16651 // the result is either zero or one (depending on the input carry bit).
16652 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
16653 if (X86::isZeroNode(N->getOperand(0)) &&
16654 X86::isZeroNode(N->getOperand(1)) &&
16655 // We don't have a good way to replace an EFLAGS use, so only do this when
16657 SDValue(N, 1).use_empty()) {
16658 DebugLoc DL = N->getDebugLoc();
16659 EVT VT = N->getValueType(0);
16660 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
16661 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
16662 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
16663 DAG.getConstant(X86::COND_B,MVT::i8),
16665 DAG.getConstant(1, VT));
16666 return DCI.CombineTo(N, Res1, CarryOut);
16672 // fold (add Y, (sete X, 0)) -> adc 0, Y
16673 // (add Y, (setne X, 0)) -> sbb -1, Y
16674 // (sub (sete X, 0), Y) -> sbb 0, Y
16675 // (sub (setne X, 0), Y) -> adc -1, Y
16676 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
16677 DebugLoc DL = N->getDebugLoc();
16679 // Look through ZExts.
16680 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
16681 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
16684 SDValue SetCC = Ext.getOperand(0);
16685 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
16688 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
16689 if (CC != X86::COND_E && CC != X86::COND_NE)
16692 SDValue Cmp = SetCC.getOperand(1);
16693 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
16694 !X86::isZeroNode(Cmp.getOperand(1)) ||
16695 !Cmp.getOperand(0).getValueType().isInteger())
16698 SDValue CmpOp0 = Cmp.getOperand(0);
16699 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
16700 DAG.getConstant(1, CmpOp0.getValueType()));
16702 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
16703 if (CC == X86::COND_NE)
16704 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
16705 DL, OtherVal.getValueType(), OtherVal,
16706 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
16707 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
16708 DL, OtherVal.getValueType(), OtherVal,
16709 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
16712 /// PerformADDCombine - Do target-specific dag combines on integer adds.
16713 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
16714 const X86Subtarget *Subtarget) {
16715 EVT VT = N->getValueType(0);
16716 SDValue Op0 = N->getOperand(0);
16717 SDValue Op1 = N->getOperand(1);
16719 // Try to synthesize horizontal adds from adds of shuffles.
16720 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
16721 (Subtarget->hasAVX2() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
16722 isHorizontalBinOp(Op0, Op1, true))
16723 return DAG.getNode(X86ISD::HADD, N->getDebugLoc(), VT, Op0, Op1);
16725 return OptimizeConditionalInDecrement(N, DAG);
16728 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
16729 const X86Subtarget *Subtarget) {
16730 SDValue Op0 = N->getOperand(0);
16731 SDValue Op1 = N->getOperand(1);
16733 // X86 can't encode an immediate LHS of a sub. See if we can push the
16734 // negation into a preceding instruction.
16735 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
16736 // If the RHS of the sub is a XOR with one use and a constant, invert the
16737 // immediate. Then add one to the LHS of the sub so we can turn
16738 // X-Y -> X+~Y+1, saving one register.
16739 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
16740 isa<ConstantSDNode>(Op1.getOperand(1))) {
16741 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
16742 EVT VT = Op0.getValueType();
16743 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
16745 DAG.getConstant(~XorC, VT));
16746 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
16747 DAG.getConstant(C->getAPIntValue()+1, VT));
16751 // Try to synthesize horizontal adds from adds of shuffles.
16752 EVT VT = N->getValueType(0);
16753 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
16754 (Subtarget->hasAVX2() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
16755 isHorizontalBinOp(Op0, Op1, true))
16756 return DAG.getNode(X86ISD::HSUB, N->getDebugLoc(), VT, Op0, Op1);
16758 return OptimizeConditionalInDecrement(N, DAG);
16761 /// performVZEXTCombine - Performs build vector combines
16762 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
16763 TargetLowering::DAGCombinerInfo &DCI,
16764 const X86Subtarget *Subtarget) {
16765 // (vzext (bitcast (vzext (x)) -> (vzext x)
16766 SDValue In = N->getOperand(0);
16767 while (In.getOpcode() == ISD::BITCAST)
16768 In = In.getOperand(0);
16770 if (In.getOpcode() != X86ISD::VZEXT)
16773 return DAG.getNode(X86ISD::VZEXT, N->getDebugLoc(), N->getValueType(0), In.getOperand(0));
16776 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
16777 DAGCombinerInfo &DCI) const {
16778 SelectionDAG &DAG = DCI.DAG;
16779 switch (N->getOpcode()) {
16781 case ISD::EXTRACT_VECTOR_ELT:
16782 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
16784 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
16785 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
16786 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
16787 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
16788 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
16789 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
16792 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
16793 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
16794 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
16795 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
16796 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
16797 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
16798 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
16799 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
16800 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
16802 case X86ISD::FOR: return PerformFORCombine(N, DAG);
16804 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
16805 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
16806 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
16807 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
16808 case ISD::ANY_EXTEND:
16809 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
16810 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
16811 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
16812 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
16813 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
16814 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
16815 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
16816 case X86ISD::SHUFP: // Handle all target specific shuffles
16817 case X86ISD::PALIGN:
16818 case X86ISD::UNPCKH:
16819 case X86ISD::UNPCKL:
16820 case X86ISD::MOVHLPS:
16821 case X86ISD::MOVLHPS:
16822 case X86ISD::PSHUFD:
16823 case X86ISD::PSHUFHW:
16824 case X86ISD::PSHUFLW:
16825 case X86ISD::MOVSS:
16826 case X86ISD::MOVSD:
16827 case X86ISD::VPERMILP:
16828 case X86ISD::VPERM2X128:
16829 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
16830 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
16836 /// isTypeDesirableForOp - Return true if the target has native support for
16837 /// the specified value type and it is 'desirable' to use the type for the
16838 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
16839 /// instruction encodings are longer and some i16 instructions are slow.
16840 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
16841 if (!isTypeLegal(VT))
16843 if (VT != MVT::i16)
16850 case ISD::SIGN_EXTEND:
16851 case ISD::ZERO_EXTEND:
16852 case ISD::ANY_EXTEND:
16865 /// IsDesirableToPromoteOp - This method query the target whether it is
16866 /// beneficial for dag combiner to promote the specified node. If true, it
16867 /// should return the desired promotion type by reference.
16868 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
16869 EVT VT = Op.getValueType();
16870 if (VT != MVT::i16)
16873 bool Promote = false;
16874 bool Commute = false;
16875 switch (Op.getOpcode()) {
16878 LoadSDNode *LD = cast<LoadSDNode>(Op);
16879 // If the non-extending load has a single use and it's not live out, then it
16880 // might be folded.
16881 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
16882 Op.hasOneUse()*/) {
16883 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
16884 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
16885 // The only case where we'd want to promote LOAD (rather then it being
16886 // promoted as an operand is when it's only use is liveout.
16887 if (UI->getOpcode() != ISD::CopyToReg)
16894 case ISD::SIGN_EXTEND:
16895 case ISD::ZERO_EXTEND:
16896 case ISD::ANY_EXTEND:
16901 SDValue N0 = Op.getOperand(0);
16902 // Look out for (store (shl (load), x)).
16903 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
16916 SDValue N0 = Op.getOperand(0);
16917 SDValue N1 = Op.getOperand(1);
16918 if (!Commute && MayFoldLoad(N1))
16920 // Avoid disabling potential load folding opportunities.
16921 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
16923 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
16933 //===----------------------------------------------------------------------===//
16934 // X86 Inline Assembly Support
16935 //===----------------------------------------------------------------------===//
16938 // Helper to match a string separated by whitespace.
16939 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
16940 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
16942 for (unsigned i = 0, e = args.size(); i != e; ++i) {
16943 StringRef piece(*args[i]);
16944 if (!s.startswith(piece)) // Check if the piece matches.
16947 s = s.substr(piece.size());
16948 StringRef::size_type pos = s.find_first_not_of(" \t");
16949 if (pos == 0) // We matched a prefix.
16957 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
16960 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
16961 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
16963 std::string AsmStr = IA->getAsmString();
16965 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
16966 if (!Ty || Ty->getBitWidth() % 16 != 0)
16969 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
16970 SmallVector<StringRef, 4> AsmPieces;
16971 SplitString(AsmStr, AsmPieces, ";\n");
16973 switch (AsmPieces.size()) {
16974 default: return false;
16976 // FIXME: this should verify that we are targeting a 486 or better. If not,
16977 // we will turn this bswap into something that will be lowered to logical
16978 // ops instead of emitting the bswap asm. For now, we don't support 486 or
16979 // lower so don't worry about this.
16981 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
16982 matchAsm(AsmPieces[0], "bswapl", "$0") ||
16983 matchAsm(AsmPieces[0], "bswapq", "$0") ||
16984 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
16985 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
16986 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
16987 // No need to check constraints, nothing other than the equivalent of
16988 // "=r,0" would be valid here.
16989 return IntrinsicLowering::LowerToByteSwap(CI);
16992 // rorw $$8, ${0:w} --> llvm.bswap.i16
16993 if (CI->getType()->isIntegerTy(16) &&
16994 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
16995 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
16996 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
16998 const std::string &ConstraintsStr = IA->getConstraintString();
16999 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
17000 std::sort(AsmPieces.begin(), AsmPieces.end());
17001 if (AsmPieces.size() == 4 &&
17002 AsmPieces[0] == "~{cc}" &&
17003 AsmPieces[1] == "~{dirflag}" &&
17004 AsmPieces[2] == "~{flags}" &&
17005 AsmPieces[3] == "~{fpsr}")
17006 return IntrinsicLowering::LowerToByteSwap(CI);
17010 if (CI->getType()->isIntegerTy(32) &&
17011 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
17012 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
17013 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
17014 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
17016 const std::string &ConstraintsStr = IA->getConstraintString();
17017 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
17018 std::sort(AsmPieces.begin(), AsmPieces.end());
17019 if (AsmPieces.size() == 4 &&
17020 AsmPieces[0] == "~{cc}" &&
17021 AsmPieces[1] == "~{dirflag}" &&
17022 AsmPieces[2] == "~{flags}" &&
17023 AsmPieces[3] == "~{fpsr}")
17024 return IntrinsicLowering::LowerToByteSwap(CI);
17027 if (CI->getType()->isIntegerTy(64)) {
17028 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
17029 if (Constraints.size() >= 2 &&
17030 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
17031 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
17032 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
17033 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
17034 matchAsm(AsmPieces[1], "bswap", "%edx") &&
17035 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
17036 return IntrinsicLowering::LowerToByteSwap(CI);
17046 /// getConstraintType - Given a constraint letter, return the type of
17047 /// constraint it is for this target.
17048 X86TargetLowering::ConstraintType
17049 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
17050 if (Constraint.size() == 1) {
17051 switch (Constraint[0]) {
17062 return C_RegisterClass;
17086 return TargetLowering::getConstraintType(Constraint);
17089 /// Examine constraint type and operand type and determine a weight value.
17090 /// This object must already have been set up with the operand type
17091 /// and the current alternative constraint selected.
17092 TargetLowering::ConstraintWeight
17093 X86TargetLowering::getSingleConstraintMatchWeight(
17094 AsmOperandInfo &info, const char *constraint) const {
17095 ConstraintWeight weight = CW_Invalid;
17096 Value *CallOperandVal = info.CallOperandVal;
17097 // If we don't have a value, we can't do a match,
17098 // but allow it at the lowest weight.
17099 if (CallOperandVal == NULL)
17101 Type *type = CallOperandVal->getType();
17102 // Look at the constraint type.
17103 switch (*constraint) {
17105 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
17116 if (CallOperandVal->getType()->isIntegerTy())
17117 weight = CW_SpecificReg;
17122 if (type->isFloatingPointTy())
17123 weight = CW_SpecificReg;
17126 if (type->isX86_MMXTy() && Subtarget->hasMMX())
17127 weight = CW_SpecificReg;
17131 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
17132 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasAVX()))
17133 weight = CW_Register;
17136 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
17137 if (C->getZExtValue() <= 31)
17138 weight = CW_Constant;
17142 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17143 if (C->getZExtValue() <= 63)
17144 weight = CW_Constant;
17148 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17149 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
17150 weight = CW_Constant;
17154 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17155 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
17156 weight = CW_Constant;
17160 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17161 if (C->getZExtValue() <= 3)
17162 weight = CW_Constant;
17166 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17167 if (C->getZExtValue() <= 0xff)
17168 weight = CW_Constant;
17173 if (dyn_cast<ConstantFP>(CallOperandVal)) {
17174 weight = CW_Constant;
17178 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17179 if ((C->getSExtValue() >= -0x80000000LL) &&
17180 (C->getSExtValue() <= 0x7fffffffLL))
17181 weight = CW_Constant;
17185 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
17186 if (C->getZExtValue() <= 0xffffffff)
17187 weight = CW_Constant;
17194 /// LowerXConstraint - try to replace an X constraint, which matches anything,
17195 /// with another that has more specific requirements based on the type of the
17196 /// corresponding operand.
17197 const char *X86TargetLowering::
17198 LowerXConstraint(EVT ConstraintVT) const {
17199 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
17200 // 'f' like normal targets.
17201 if (ConstraintVT.isFloatingPoint()) {
17202 if (Subtarget->hasSSE2())
17204 if (Subtarget->hasSSE1())
17208 return TargetLowering::LowerXConstraint(ConstraintVT);
17211 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
17212 /// vector. If it is invalid, don't add anything to Ops.
17213 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
17214 std::string &Constraint,
17215 std::vector<SDValue>&Ops,
17216 SelectionDAG &DAG) const {
17217 SDValue Result(0, 0);
17219 // Only support length 1 constraints for now.
17220 if (Constraint.length() > 1) return;
17222 char ConstraintLetter = Constraint[0];
17223 switch (ConstraintLetter) {
17226 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17227 if (C->getZExtValue() <= 31) {
17228 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
17234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17235 if (C->getZExtValue() <= 63) {
17236 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
17242 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17243 if (isInt<8>(C->getSExtValue())) {
17244 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
17250 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17251 if (C->getZExtValue() <= 255) {
17252 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
17258 // 32-bit signed value
17259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17260 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
17261 C->getSExtValue())) {
17262 // Widen to 64 bits here to get it sign extended.
17263 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
17266 // FIXME gcc accepts some relocatable values here too, but only in certain
17267 // memory models; it's complicated.
17272 // 32-bit unsigned value
17273 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
17274 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
17275 C->getZExtValue())) {
17276 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
17280 // FIXME gcc accepts some relocatable values here too, but only in certain
17281 // memory models; it's complicated.
17285 // Literal immediates are always ok.
17286 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
17287 // Widen to 64 bits here to get it sign extended.
17288 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
17292 // In any sort of PIC mode addresses need to be computed at runtime by
17293 // adding in a register or some sort of table lookup. These can't
17294 // be used as immediates.
17295 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
17298 // If we are in non-pic codegen mode, we allow the address of a global (with
17299 // an optional displacement) to be used with 'i'.
17300 GlobalAddressSDNode *GA = 0;
17301 int64_t Offset = 0;
17303 // Match either (GA), (GA+C), (GA+C1+C2), etc.
17305 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
17306 Offset += GA->getOffset();
17308 } else if (Op.getOpcode() == ISD::ADD) {
17309 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
17310 Offset += C->getZExtValue();
17311 Op = Op.getOperand(0);
17314 } else if (Op.getOpcode() == ISD::SUB) {
17315 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
17316 Offset += -C->getZExtValue();
17317 Op = Op.getOperand(0);
17322 // Otherwise, this isn't something we can handle, reject it.
17326 const GlobalValue *GV = GA->getGlobal();
17327 // If we require an extra load to get this address, as in PIC mode, we
17328 // can't accept it.
17329 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
17330 getTargetMachine())))
17333 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
17334 GA->getValueType(0), Offset);
17339 if (Result.getNode()) {
17340 Ops.push_back(Result);
17343 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
17346 std::pair<unsigned, const TargetRegisterClass*>
17347 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
17349 // First, see if this is a constraint that directly corresponds to an LLVM
17351 if (Constraint.size() == 1) {
17352 // GCC Constraint Letters
17353 switch (Constraint[0]) {
17355 // TODO: Slight differences here in allocation order and leaving
17356 // RIP in the class. Do they matter any more here than they do
17357 // in the normal allocation?
17358 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
17359 if (Subtarget->is64Bit()) {
17360 if (VT == MVT::i32 || VT == MVT::f32)
17361 return std::make_pair(0U, &X86::GR32RegClass);
17362 if (VT == MVT::i16)
17363 return std::make_pair(0U, &X86::GR16RegClass);
17364 if (VT == MVT::i8 || VT == MVT::i1)
17365 return std::make_pair(0U, &X86::GR8RegClass);
17366 if (VT == MVT::i64 || VT == MVT::f64)
17367 return std::make_pair(0U, &X86::GR64RegClass);
17370 // 32-bit fallthrough
17371 case 'Q': // Q_REGS
17372 if (VT == MVT::i32 || VT == MVT::f32)
17373 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
17374 if (VT == MVT::i16)
17375 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
17376 if (VT == MVT::i8 || VT == MVT::i1)
17377 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
17378 if (VT == MVT::i64)
17379 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
17381 case 'r': // GENERAL_REGS
17382 case 'l': // INDEX_REGS
17383 if (VT == MVT::i8 || VT == MVT::i1)
17384 return std::make_pair(0U, &X86::GR8RegClass);
17385 if (VT == MVT::i16)
17386 return std::make_pair(0U, &X86::GR16RegClass);
17387 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
17388 return std::make_pair(0U, &X86::GR32RegClass);
17389 return std::make_pair(0U, &X86::GR64RegClass);
17390 case 'R': // LEGACY_REGS
17391 if (VT == MVT::i8 || VT == MVT::i1)
17392 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
17393 if (VT == MVT::i16)
17394 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
17395 if (VT == MVT::i32 || !Subtarget->is64Bit())
17396 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
17397 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
17398 case 'f': // FP Stack registers.
17399 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
17400 // value to the correct fpstack register class.
17401 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
17402 return std::make_pair(0U, &X86::RFP32RegClass);
17403 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
17404 return std::make_pair(0U, &X86::RFP64RegClass);
17405 return std::make_pair(0U, &X86::RFP80RegClass);
17406 case 'y': // MMX_REGS if MMX allowed.
17407 if (!Subtarget->hasMMX()) break;
17408 return std::make_pair(0U, &X86::VR64RegClass);
17409 case 'Y': // SSE_REGS if SSE2 allowed
17410 if (!Subtarget->hasSSE2()) break;
17412 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
17413 if (!Subtarget->hasSSE1()) break;
17415 switch (VT.getSimpleVT().SimpleTy) {
17417 // Scalar SSE types.
17420 return std::make_pair(0U, &X86::FR32RegClass);
17423 return std::make_pair(0U, &X86::FR64RegClass);
17431 return std::make_pair(0U, &X86::VR128RegClass);
17439 return std::make_pair(0U, &X86::VR256RegClass);
17445 // Use the default implementation in TargetLowering to convert the register
17446 // constraint into a member of a register class.
17447 std::pair<unsigned, const TargetRegisterClass*> Res;
17448 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
17450 // Not found as a standard register?
17451 if (Res.second == 0) {
17452 // Map st(0) -> st(7) -> ST0
17453 if (Constraint.size() == 7 && Constraint[0] == '{' &&
17454 tolower(Constraint[1]) == 's' &&
17455 tolower(Constraint[2]) == 't' &&
17456 Constraint[3] == '(' &&
17457 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
17458 Constraint[5] == ')' &&
17459 Constraint[6] == '}') {
17461 Res.first = X86::ST0+Constraint[4]-'0';
17462 Res.second = &X86::RFP80RegClass;
17466 // GCC allows "st(0)" to be called just plain "st".
17467 if (StringRef("{st}").equals_lower(Constraint)) {
17468 Res.first = X86::ST0;
17469 Res.second = &X86::RFP80RegClass;
17474 if (StringRef("{flags}").equals_lower(Constraint)) {
17475 Res.first = X86::EFLAGS;
17476 Res.second = &X86::CCRRegClass;
17480 // 'A' means EAX + EDX.
17481 if (Constraint == "A") {
17482 Res.first = X86::EAX;
17483 Res.second = &X86::GR32_ADRegClass;
17489 // Otherwise, check to see if this is a register class of the wrong value
17490 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
17491 // turn into {ax},{dx}.
17492 if (Res.second->hasType(VT))
17493 return Res; // Correct type already, nothing to do.
17495 // All of the single-register GCC register classes map their values onto
17496 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
17497 // really want an 8-bit or 32-bit register, map to the appropriate register
17498 // class and return the appropriate register.
17499 if (Res.second == &X86::GR16RegClass) {
17500 if (VT == MVT::i8) {
17501 unsigned DestReg = 0;
17502 switch (Res.first) {
17504 case X86::AX: DestReg = X86::AL; break;
17505 case X86::DX: DestReg = X86::DL; break;
17506 case X86::CX: DestReg = X86::CL; break;
17507 case X86::BX: DestReg = X86::BL; break;
17510 Res.first = DestReg;
17511 Res.second = &X86::GR8RegClass;
17513 } else if (VT == MVT::i32) {
17514 unsigned DestReg = 0;
17515 switch (Res.first) {
17517 case X86::AX: DestReg = X86::EAX; break;
17518 case X86::DX: DestReg = X86::EDX; break;
17519 case X86::CX: DestReg = X86::ECX; break;
17520 case X86::BX: DestReg = X86::EBX; break;
17521 case X86::SI: DestReg = X86::ESI; break;
17522 case X86::DI: DestReg = X86::EDI; break;
17523 case X86::BP: DestReg = X86::EBP; break;
17524 case X86::SP: DestReg = X86::ESP; break;
17527 Res.first = DestReg;
17528 Res.second = &X86::GR32RegClass;
17530 } else if (VT == MVT::i64) {
17531 unsigned DestReg = 0;
17532 switch (Res.first) {
17534 case X86::AX: DestReg = X86::RAX; break;
17535 case X86::DX: DestReg = X86::RDX; break;
17536 case X86::CX: DestReg = X86::RCX; break;
17537 case X86::BX: DestReg = X86::RBX; break;
17538 case X86::SI: DestReg = X86::RSI; break;
17539 case X86::DI: DestReg = X86::RDI; break;
17540 case X86::BP: DestReg = X86::RBP; break;
17541 case X86::SP: DestReg = X86::RSP; break;
17544 Res.first = DestReg;
17545 Res.second = &X86::GR64RegClass;
17548 } else if (Res.second == &X86::FR32RegClass ||
17549 Res.second == &X86::FR64RegClass ||
17550 Res.second == &X86::VR128RegClass) {
17551 // Handle references to XMM physical registers that got mapped into the
17552 // wrong class. This can happen with constraints like {xmm0} where the
17553 // target independent register mapper will just pick the first match it can
17554 // find, ignoring the required type.
17556 if (VT == MVT::f32 || VT == MVT::i32)
17557 Res.second = &X86::FR32RegClass;
17558 else if (VT == MVT::f64 || VT == MVT::i64)
17559 Res.second = &X86::FR64RegClass;
17560 else if (X86::VR128RegClass.hasType(VT))
17561 Res.second = &X86::VR128RegClass;
17562 else if (X86::VR256RegClass.hasType(VT))
17563 Res.second = &X86::VR256RegClass;
17569 //===----------------------------------------------------------------------===//
17573 //===----------------------------------------------------------------------===//
17575 struct X86CostTblEntry {
17582 FindInTable(const X86CostTblEntry *Tbl, unsigned len, int ISD, MVT Ty) {
17583 for (unsigned int i = 0; i < len; ++i)
17584 if (Tbl[i].ISD == ISD && Tbl[i].Type == Ty)
17587 // Could not find an entry.
17591 struct X86TypeConversionCostTblEntry {
17599 FindInConvertTable(const X86TypeConversionCostTblEntry *Tbl, unsigned len,
17600 int ISD, MVT Dst, MVT Src) {
17601 for (unsigned int i = 0; i < len; ++i)
17602 if (Tbl[i].ISD == ISD && Tbl[i].Src == Src && Tbl[i].Dst == Dst)
17605 // Could not find an entry.
17610 X86VectorTargetTransformInfo::getArithmeticInstrCost(unsigned Opcode,
17612 // Legalize the type.
17613 std::pair<unsigned, MVT> LT = getTypeLegalizationCost(Ty);
17615 int ISD = InstructionOpcodeToISD(Opcode);
17616 assert(ISD && "Invalid opcode");
17618 const X86Subtarget &ST = TLI->getTargetMachine().getSubtarget<X86Subtarget>();
17620 static const X86CostTblEntry AVX1CostTable[] = {
17621 // We don't have to scalarize unsupported ops. We can issue two half-sized
17622 // operations and we only need to extract the upper YMM half.
17623 // Two ops + 1 extract + 1 insert = 4.
17624 { ISD::MUL, MVT::v8i32, 4 },
17625 { ISD::SUB, MVT::v8i32, 4 },
17626 { ISD::ADD, MVT::v8i32, 4 },
17627 { ISD::MUL, MVT::v4i64, 4 },
17628 { ISD::SUB, MVT::v4i64, 4 },
17629 { ISD::ADD, MVT::v4i64, 4 },
17632 // Look for AVX1 lowering tricks.
17634 int Idx = FindInTable(AVX1CostTable, array_lengthof(AVX1CostTable), ISD,
17637 return LT.first * AVX1CostTable[Idx].Cost;
17639 // Fallback to the default implementation.
17640 return VectorTargetTransformImpl::getArithmeticInstrCost(Opcode, Ty);
17644 X86VectorTargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
17645 unsigned Index) const {
17646 assert(Val->isVectorTy() && "This must be a vector type");
17648 if (Index != -1U) {
17649 // Legalize the type.
17650 std::pair<unsigned, MVT> LT = getTypeLegalizationCost(Val);
17652 // This type is legalized to a scalar type.
17653 if (!LT.second.isVector())
17656 // The type may be split. Normalize the index to the new type.
17657 unsigned Width = LT.second.getVectorNumElements();
17658 Index = Index % Width;
17660 // Floating point scalars are already located in index #0.
17661 if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
17665 return VectorTargetTransformImpl::getVectorInstrCost(Opcode, Val, Index);
17668 unsigned X86VectorTargetTransformInfo::getCmpSelInstrCost(unsigned Opcode,
17670 Type *CondTy) const {
17671 // Legalize the type.
17672 std::pair<unsigned, MVT> LT = getTypeLegalizationCost(ValTy);
17674 MVT MTy = LT.second;
17676 int ISD = InstructionOpcodeToISD(Opcode);
17677 assert(ISD && "Invalid opcode");
17679 const X86Subtarget &ST =
17680 TLI->getTargetMachine().getSubtarget<X86Subtarget>();
17682 static const X86CostTblEntry SSE42CostTbl[] = {
17683 { ISD::SETCC, MVT::v2f64, 1 },
17684 { ISD::SETCC, MVT::v4f32, 1 },
17685 { ISD::SETCC, MVT::v2i64, 1 },
17686 { ISD::SETCC, MVT::v4i32, 1 },
17687 { ISD::SETCC, MVT::v8i16, 1 },
17688 { ISD::SETCC, MVT::v16i8, 1 },
17691 static const X86CostTblEntry AVX1CostTbl[] = {
17692 { ISD::SETCC, MVT::v4f64, 1 },
17693 { ISD::SETCC, MVT::v8f32, 1 },
17694 // AVX1 does not support 8-wide integer compare.
17695 { ISD::SETCC, MVT::v4i64, 4 },
17696 { ISD::SETCC, MVT::v8i32, 4 },
17697 { ISD::SETCC, MVT::v16i16, 4 },
17698 { ISD::SETCC, MVT::v32i8, 4 },
17701 static const X86CostTblEntry AVX2CostTbl[] = {
17702 { ISD::SETCC, MVT::v4i64, 1 },
17703 { ISD::SETCC, MVT::v8i32, 1 },
17704 { ISD::SETCC, MVT::v16i16, 1 },
17705 { ISD::SETCC, MVT::v32i8, 1 },
17708 if (ST.hasSSE42()) {
17709 int Idx = FindInTable(SSE42CostTbl, array_lengthof(SSE42CostTbl), ISD, MTy);
17711 return LT.first * SSE42CostTbl[Idx].Cost;
17715 int Idx = FindInTable(AVX1CostTbl, array_lengthof(AVX1CostTbl), ISD, MTy);
17717 return LT.first * AVX1CostTbl[Idx].Cost;
17720 if (ST.hasAVX2()) {
17721 int Idx = FindInTable(AVX2CostTbl, array_lengthof(AVX2CostTbl), ISD, MTy);
17723 return LT.first * AVX2CostTbl[Idx].Cost;
17726 return VectorTargetTransformImpl::getCmpSelInstrCost(Opcode, ValTy, CondTy);
17729 unsigned X86VectorTargetTransformInfo::getCastInstrCost(unsigned Opcode,
17732 int ISD = InstructionOpcodeToISD(Opcode);
17733 assert(ISD && "Invalid opcode");
17735 EVT SrcTy = TLI->getValueType(Src);
17736 EVT DstTy = TLI->getValueType(Dst);
17738 if (!SrcTy.isSimple() || !DstTy.isSimple())
17739 return VectorTargetTransformImpl::getCastInstrCost(Opcode, Dst, Src);
17741 const X86Subtarget &ST = TLI->getTargetMachine().getSubtarget<X86Subtarget>();
17743 static const X86TypeConversionCostTblEntry AVXConversionTbl[] = {
17744 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
17745 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
17746 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
17747 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
17748 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 },
17749 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1 },
17750 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
17751 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
17752 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 },
17753 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 },
17754 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 1 },
17755 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
17756 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 6 },
17757 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 9 },
17758 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 3 },
17762 int Idx = FindInConvertTable(AVXConversionTbl,
17763 array_lengthof(AVXConversionTbl),
17764 ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT());
17766 return AVXConversionTbl[Idx].Cost;
17769 return VectorTargetTransformImpl::getCastInstrCost(Opcode, Dst, Src);