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"
17 #include "Utils/X86ShuffleDecode.h"
19 #include "X86InstrBuilder.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/ADT/SmallSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/ADT/VariadicFunction.h"
26 #include "llvm/CodeGen/IntrinsicLowering.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineJumpTableInfo.h"
31 #include "llvm/CodeGen/MachineModuleInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DerivedTypes.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/GlobalAlias.h"
38 #include "llvm/IR/GlobalVariable.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/MC/MCAsmInfo.h"
43 #include "llvm/MC/MCContext.h"
44 #include "llvm/MC/MCExpr.h"
45 #include "llvm/MC/MCSymbol.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 // If the input is a buildvector just emit a smaller one.
89 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
90 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
91 Vec->op_begin()+NormalizedIdxVal, ElemsPerChunk);
93 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
94 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
100 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
101 /// sets things up to match to an AVX VINSERTF128 instruction or a
102 /// simple superregister reference. Idx is an index in the 128 bits
103 /// we want. It need not be aligned to a 128-bit bounday. That makes
104 /// lowering INSERT_VECTOR_ELT operations easier.
105 static SDValue Insert128BitVector(SDValue Result, SDValue Vec,
106 unsigned IdxVal, SelectionDAG &DAG,
108 // Inserting UNDEF is Result
109 if (Vec.getOpcode() == ISD::UNDEF)
112 EVT VT = Vec.getValueType();
113 assert(VT.is128BitVector() && "Unexpected vector size!");
115 EVT ElVT = VT.getVectorElementType();
116 EVT ResultVT = Result.getValueType();
118 // Insert the relevant 128 bits.
119 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
121 // This is the index of the first element of the 128-bit chunk
123 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
126 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
127 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
131 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
132 /// instructions. This is used because creating CONCAT_VECTOR nodes of
133 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
134 /// large BUILD_VECTORS.
135 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
136 unsigned NumElems, SelectionDAG &DAG,
138 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
139 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
142 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
143 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
144 bool is64Bit = Subtarget->is64Bit();
146 if (Subtarget->isTargetEnvMacho()) {
148 return new X86_64MachoTargetObjectFile();
149 return new TargetLoweringObjectFileMachO();
152 if (Subtarget->isTargetLinux())
153 return new X86LinuxTargetObjectFile();
154 if (Subtarget->isTargetELF())
155 return new TargetLoweringObjectFileELF();
156 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
157 return new TargetLoweringObjectFileCOFF();
158 llvm_unreachable("unknown subtarget type");
161 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<X86Subtarget>();
164 X86ScalarSSEf64 = Subtarget->hasSSE2();
165 X86ScalarSSEf32 = Subtarget->hasSSE1();
167 RegInfo = TM.getRegisterInfo();
168 TD = getDataLayout();
170 // Set up the TargetLowering object.
171 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
173 // X86 is weird, it always uses i8 for shift amounts and setcc results.
174 setBooleanContents(ZeroOrOneBooleanContent);
175 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
176 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
178 // For 64-bit since we have so many registers use the ILP scheduler, for
179 // 32-bit code use the register pressure specific scheduling.
180 // For Atom, always use ILP scheduling.
181 if (Subtarget->isAtom())
182 setSchedulingPreference(Sched::ILP);
183 else if (Subtarget->is64Bit())
184 setSchedulingPreference(Sched::ILP);
186 setSchedulingPreference(Sched::RegPressure);
187 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
189 // Bypass expensive divides on Atom when compiling with O2
190 if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default) {
191 addBypassSlowDiv(32, 8);
192 if (Subtarget->is64Bit())
193 addBypassSlowDiv(64, 16);
196 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
197 // Setup Windows compiler runtime calls.
198 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
199 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
200 setLibcallName(RTLIB::SREM_I64, "_allrem");
201 setLibcallName(RTLIB::UREM_I64, "_aullrem");
202 setLibcallName(RTLIB::MUL_I64, "_allmul");
203 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
204 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
205 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
206 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
207 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
209 // The _ftol2 runtime function has an unusual calling conv, which
210 // is modeled by a special pseudo-instruction.
211 setLibcallName(RTLIB::FPTOUINT_F64_I64, 0);
212 setLibcallName(RTLIB::FPTOUINT_F32_I64, 0);
213 setLibcallName(RTLIB::FPTOUINT_F64_I32, 0);
214 setLibcallName(RTLIB::FPTOUINT_F32_I32, 0);
217 if (Subtarget->isTargetDarwin()) {
218 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
219 setUseUnderscoreSetJmp(false);
220 setUseUnderscoreLongJmp(false);
221 } else if (Subtarget->isTargetMingw()) {
222 // MS runtime is weird: it exports _setjmp, but longjmp!
223 setUseUnderscoreSetJmp(true);
224 setUseUnderscoreLongJmp(false);
226 setUseUnderscoreSetJmp(true);
227 setUseUnderscoreLongJmp(true);
230 // Set up the register classes.
231 addRegisterClass(MVT::i8, &X86::GR8RegClass);
232 addRegisterClass(MVT::i16, &X86::GR16RegClass);
233 addRegisterClass(MVT::i32, &X86::GR32RegClass);
234 if (Subtarget->is64Bit())
235 addRegisterClass(MVT::i64, &X86::GR64RegClass);
237 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
239 // We don't accept any truncstore of integer registers.
240 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
241 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
242 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
243 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
244 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
245 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
247 // SETOEQ and SETUNE require checking two conditions.
248 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
249 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
250 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
251 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
252 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
253 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
255 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
257 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
258 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
259 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
261 if (Subtarget->is64Bit()) {
262 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
263 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
264 } else if (!TM.Options.UseSoftFloat) {
265 // We have an algorithm for SSE2->double, and we turn this into a
266 // 64-bit FILD followed by conditional FADD for other targets.
267 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
268 // We have an algorithm for SSE2, and we turn this into a 64-bit
269 // FILD for other targets.
270 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
273 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
275 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
276 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
278 if (!TM.Options.UseSoftFloat) {
279 // SSE has no i16 to fp conversion, only i32
280 if (X86ScalarSSEf32) {
281 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
282 // f32 and f64 cases are Legal, f80 case is not
283 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
285 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
286 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
289 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
290 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
293 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
294 // are Legal, f80 is custom lowered.
295 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
296 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
298 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
300 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
301 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
303 if (X86ScalarSSEf32) {
304 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
305 // f32 and f64 cases are Legal, f80 case is not
306 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
308 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
309 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
312 // Handle FP_TO_UINT by promoting the destination to a larger signed
314 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
315 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
316 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
318 if (Subtarget->is64Bit()) {
319 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
320 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
321 } else if (!TM.Options.UseSoftFloat) {
322 // Since AVX is a superset of SSE3, only check for SSE here.
323 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
324 // Expand FP_TO_UINT into a select.
325 // FIXME: We would like to use a Custom expander here eventually to do
326 // the optimal thing for SSE vs. the default expansion in the legalizer.
327 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
329 // With SSE3 we can use fisttpll to convert to a signed i64; without
330 // SSE, we're stuck with a fistpll.
331 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
334 if (isTargetFTOL()) {
335 // Use the _ftol2 runtime function, which has a pseudo-instruction
336 // to handle its weird calling convention.
337 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
340 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
341 if (!X86ScalarSSEf64) {
342 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
343 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
344 if (Subtarget->is64Bit()) {
345 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
346 // Without SSE, i64->f64 goes through memory.
347 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
351 // Scalar integer divide and remainder are lowered to use operations that
352 // produce two results, to match the available instructions. This exposes
353 // the two-result form to trivial CSE, which is able to combine x/y and x%y
354 // into a single instruction.
356 // Scalar integer multiply-high is also lowered to use two-result
357 // operations, to match the available instructions. However, plain multiply
358 // (low) operations are left as Legal, as there are single-result
359 // instructions for this in x86. Using the two-result multiply instructions
360 // when both high and low results are needed must be arranged by dagcombine.
361 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
363 setOperationAction(ISD::MULHS, VT, Expand);
364 setOperationAction(ISD::MULHU, VT, Expand);
365 setOperationAction(ISD::SDIV, VT, Expand);
366 setOperationAction(ISD::UDIV, VT, Expand);
367 setOperationAction(ISD::SREM, VT, Expand);
368 setOperationAction(ISD::UREM, VT, Expand);
370 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
371 setOperationAction(ISD::ADDC, VT, Custom);
372 setOperationAction(ISD::ADDE, VT, Custom);
373 setOperationAction(ISD::SUBC, VT, Custom);
374 setOperationAction(ISD::SUBE, VT, Custom);
377 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
378 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
379 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
380 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
381 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
382 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
383 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
384 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
385 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
386 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
387 if (Subtarget->is64Bit())
388 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
389 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
390 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
391 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
392 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
393 setOperationAction(ISD::FREM , MVT::f32 , Expand);
394 setOperationAction(ISD::FREM , MVT::f64 , Expand);
395 setOperationAction(ISD::FREM , MVT::f80 , Expand);
396 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
398 // Promote the i8 variants and force them on up to i32 which has a shorter
400 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
401 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
402 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
403 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
404 if (Subtarget->hasBMI()) {
405 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
406 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
407 if (Subtarget->is64Bit())
408 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
410 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
411 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
412 if (Subtarget->is64Bit())
413 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
416 if (Subtarget->hasLZCNT()) {
417 // When promoting the i8 variants, force them to i32 for a shorter
419 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
420 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
421 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
422 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
423 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
424 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
425 if (Subtarget->is64Bit())
426 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
428 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
429 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
430 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
431 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
432 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
433 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
434 if (Subtarget->is64Bit()) {
435 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
436 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
440 if (Subtarget->hasPOPCNT()) {
441 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
443 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
444 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
445 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
446 if (Subtarget->is64Bit())
447 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
450 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
451 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
453 // These should be promoted to a larger select which is supported.
454 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
455 // X86 wants to expand cmov itself.
456 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
457 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
458 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
459 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
460 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
461 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
462 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
463 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
464 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
465 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
466 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
467 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
468 if (Subtarget->is64Bit()) {
469 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
470 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
472 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
473 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
474 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
475 // support continuation, user-level threading, and etc.. As a result, no
476 // other SjLj exception interfaces are implemented and please don't build
477 // your own exception handling based on them.
478 // LLVM/Clang supports zero-cost DWARF exception handling.
479 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
480 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
483 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
484 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
485 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
486 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
487 if (Subtarget->is64Bit())
488 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
489 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
490 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
491 if (Subtarget->is64Bit()) {
492 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
493 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
494 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
495 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
496 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
498 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
499 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
500 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
501 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
502 if (Subtarget->is64Bit()) {
503 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
504 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
505 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
508 if (Subtarget->hasSSE1())
509 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
511 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
512 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
514 // On X86 and X86-64, atomic operations are lowered to locked instructions.
515 // Locked instructions, in turn, have implicit fence semantics (all memory
516 // operations are flushed before issuing the locked instruction, and they
517 // are not buffered), so we can fold away the common pattern of
518 // fence-atomic-fence.
519 setShouldFoldAtomicFences(true);
521 // Expand certain atomics
522 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
524 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
525 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
526 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
529 if (!Subtarget->is64Bit()) {
530 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
531 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
532 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
533 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
534 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
535 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
536 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
537 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
538 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i64, Custom);
539 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i64, Custom);
540 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
541 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
544 if (Subtarget->hasCmpxchg16b()) {
545 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
548 // FIXME - use subtarget debug flags
549 if (!Subtarget->isTargetDarwin() &&
550 !Subtarget->isTargetELF() &&
551 !Subtarget->isTargetCygMing()) {
552 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
555 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
556 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
557 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
558 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
559 if (Subtarget->is64Bit()) {
560 setExceptionPointerRegister(X86::RAX);
561 setExceptionSelectorRegister(X86::RDX);
563 setExceptionPointerRegister(X86::EAX);
564 setExceptionSelectorRegister(X86::EDX);
566 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
567 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
569 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
570 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
572 setOperationAction(ISD::TRAP, MVT::Other, Legal);
573 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
575 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
576 setOperationAction(ISD::VASTART , MVT::Other, Custom);
577 setOperationAction(ISD::VAEND , MVT::Other, Expand);
578 if (Subtarget->is64Bit()) {
579 setOperationAction(ISD::VAARG , MVT::Other, Custom);
580 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
582 setOperationAction(ISD::VAARG , MVT::Other, Expand);
583 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
586 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
587 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
589 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
590 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
591 MVT::i64 : MVT::i32, Custom);
592 else if (TM.Options.EnableSegmentedStacks)
593 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
594 MVT::i64 : MVT::i32, Custom);
596 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
597 MVT::i64 : MVT::i32, Expand);
599 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
600 // f32 and f64 use SSE.
601 // Set up the FP register classes.
602 addRegisterClass(MVT::f32, &X86::FR32RegClass);
603 addRegisterClass(MVT::f64, &X86::FR64RegClass);
605 // Use ANDPD to simulate FABS.
606 setOperationAction(ISD::FABS , MVT::f64, Custom);
607 setOperationAction(ISD::FABS , MVT::f32, Custom);
609 // Use XORP to simulate FNEG.
610 setOperationAction(ISD::FNEG , MVT::f64, Custom);
611 setOperationAction(ISD::FNEG , MVT::f32, Custom);
613 // Use ANDPD and ORPD to simulate FCOPYSIGN.
614 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
615 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
617 // Lower this to FGETSIGNx86 plus an AND.
618 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
619 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
621 // We don't support sin/cos/fmod
622 setOperationAction(ISD::FSIN , MVT::f64, Expand);
623 setOperationAction(ISD::FCOS , MVT::f64, Expand);
624 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
625 setOperationAction(ISD::FSIN , MVT::f32, Expand);
626 setOperationAction(ISD::FCOS , MVT::f32, Expand);
627 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
629 // Expand FP immediates into loads from the stack, except for the special
631 addLegalFPImmediate(APFloat(+0.0)); // xorpd
632 addLegalFPImmediate(APFloat(+0.0f)); // xorps
633 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
634 // Use SSE for f32, x87 for f64.
635 // Set up the FP register classes.
636 addRegisterClass(MVT::f32, &X86::FR32RegClass);
637 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
639 // Use ANDPS to simulate FABS.
640 setOperationAction(ISD::FABS , MVT::f32, Custom);
642 // Use XORP to simulate FNEG.
643 setOperationAction(ISD::FNEG , MVT::f32, Custom);
645 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
647 // Use ANDPS and ORPS to simulate FCOPYSIGN.
648 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
649 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
651 // We don't support sin/cos/fmod
652 setOperationAction(ISD::FSIN , MVT::f32, Expand);
653 setOperationAction(ISD::FCOS , MVT::f32, Expand);
654 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
656 // Special cases we handle for FP constants.
657 addLegalFPImmediate(APFloat(+0.0f)); // xorps
658 addLegalFPImmediate(APFloat(+0.0)); // FLD0
659 addLegalFPImmediate(APFloat(+1.0)); // FLD1
660 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
661 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
663 if (!TM.Options.UnsafeFPMath) {
664 setOperationAction(ISD::FSIN , MVT::f64, Expand);
665 setOperationAction(ISD::FCOS , MVT::f64, Expand);
666 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
668 } else if (!TM.Options.UseSoftFloat) {
669 // f32 and f64 in x87.
670 // Set up the FP register classes.
671 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
672 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
674 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
675 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
676 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
677 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
679 if (!TM.Options.UnsafeFPMath) {
680 setOperationAction(ISD::FSIN , MVT::f64, Expand);
681 setOperationAction(ISD::FSIN , MVT::f32, Expand);
682 setOperationAction(ISD::FCOS , MVT::f64, Expand);
683 setOperationAction(ISD::FCOS , MVT::f32, Expand);
684 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
685 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
687 addLegalFPImmediate(APFloat(+0.0)); // FLD0
688 addLegalFPImmediate(APFloat(+1.0)); // FLD1
689 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
690 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
691 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
692 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
693 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
694 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
697 // We don't support FMA.
698 setOperationAction(ISD::FMA, MVT::f64, Expand);
699 setOperationAction(ISD::FMA, MVT::f32, Expand);
701 // Long double always uses X87.
702 if (!TM.Options.UseSoftFloat) {
703 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
704 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
705 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
707 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
708 addLegalFPImmediate(TmpFlt); // FLD0
710 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
713 APFloat TmpFlt2(+1.0);
714 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
716 addLegalFPImmediate(TmpFlt2); // FLD1
717 TmpFlt2.changeSign();
718 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
721 if (!TM.Options.UnsafeFPMath) {
722 setOperationAction(ISD::FSIN , MVT::f80, Expand);
723 setOperationAction(ISD::FCOS , MVT::f80, Expand);
724 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
727 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
728 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
729 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
730 setOperationAction(ISD::FRINT, MVT::f80, Expand);
731 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
732 setOperationAction(ISD::FMA, MVT::f80, Expand);
735 // Always use a library call for pow.
736 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
737 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
738 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
740 setOperationAction(ISD::FLOG, MVT::f80, Expand);
741 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
742 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
743 setOperationAction(ISD::FEXP, MVT::f80, Expand);
744 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
746 // First set operation action for all vector types to either promote
747 // (for widening) or expand (for scalarization). Then we will selectively
748 // turn on ones that can be effectively codegen'd.
749 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
750 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
751 MVT VT = (MVT::SimpleValueType)i;
752 setOperationAction(ISD::ADD , VT, Expand);
753 setOperationAction(ISD::SUB , VT, Expand);
754 setOperationAction(ISD::FADD, VT, Expand);
755 setOperationAction(ISD::FNEG, VT, Expand);
756 setOperationAction(ISD::FSUB, VT, Expand);
757 setOperationAction(ISD::MUL , VT, Expand);
758 setOperationAction(ISD::FMUL, VT, Expand);
759 setOperationAction(ISD::SDIV, VT, Expand);
760 setOperationAction(ISD::UDIV, VT, Expand);
761 setOperationAction(ISD::FDIV, VT, Expand);
762 setOperationAction(ISD::SREM, VT, Expand);
763 setOperationAction(ISD::UREM, VT, Expand);
764 setOperationAction(ISD::LOAD, VT, Expand);
765 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
766 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
767 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
768 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
769 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
770 setOperationAction(ISD::FABS, VT, Expand);
771 setOperationAction(ISD::FSIN, VT, Expand);
772 setOperationAction(ISD::FSINCOS, VT, Expand);
773 setOperationAction(ISD::FCOS, VT, Expand);
774 setOperationAction(ISD::FSINCOS, VT, Expand);
775 setOperationAction(ISD::FREM, VT, Expand);
776 setOperationAction(ISD::FMA, VT, Expand);
777 setOperationAction(ISD::FPOWI, VT, Expand);
778 setOperationAction(ISD::FSQRT, VT, Expand);
779 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
780 setOperationAction(ISD::FFLOOR, VT, Expand);
781 setOperationAction(ISD::FCEIL, VT, Expand);
782 setOperationAction(ISD::FTRUNC, VT, Expand);
783 setOperationAction(ISD::FRINT, VT, Expand);
784 setOperationAction(ISD::FNEARBYINT, VT, Expand);
785 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
786 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
787 setOperationAction(ISD::SDIVREM, VT, Expand);
788 setOperationAction(ISD::UDIVREM, VT, Expand);
789 setOperationAction(ISD::FPOW, VT, Expand);
790 setOperationAction(ISD::CTPOP, VT, Expand);
791 setOperationAction(ISD::CTTZ, VT, Expand);
792 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
793 setOperationAction(ISD::CTLZ, VT, Expand);
794 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
795 setOperationAction(ISD::SHL, VT, Expand);
796 setOperationAction(ISD::SRA, VT, Expand);
797 setOperationAction(ISD::SRL, VT, Expand);
798 setOperationAction(ISD::ROTL, VT, Expand);
799 setOperationAction(ISD::ROTR, VT, Expand);
800 setOperationAction(ISD::BSWAP, VT, Expand);
801 setOperationAction(ISD::SETCC, VT, Expand);
802 setOperationAction(ISD::FLOG, VT, Expand);
803 setOperationAction(ISD::FLOG2, VT, Expand);
804 setOperationAction(ISD::FLOG10, VT, Expand);
805 setOperationAction(ISD::FEXP, VT, Expand);
806 setOperationAction(ISD::FEXP2, VT, Expand);
807 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
808 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
809 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
810 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
811 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
812 setOperationAction(ISD::TRUNCATE, VT, Expand);
813 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
814 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
815 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
816 setOperationAction(ISD::VSELECT, VT, Expand);
817 for (int InnerVT = MVT::FIRST_VECTOR_VALUETYPE;
818 InnerVT <= MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
819 setTruncStoreAction(VT,
820 (MVT::SimpleValueType)InnerVT, Expand);
821 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
822 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
823 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
826 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
827 // with -msoft-float, disable use of MMX as well.
828 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
829 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
830 // No operations on x86mmx supported, everything uses intrinsics.
833 // MMX-sized vectors (other than x86mmx) are expected to be expanded
834 // into smaller operations.
835 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
836 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
837 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
838 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
839 setOperationAction(ISD::AND, MVT::v8i8, Expand);
840 setOperationAction(ISD::AND, MVT::v4i16, Expand);
841 setOperationAction(ISD::AND, MVT::v2i32, Expand);
842 setOperationAction(ISD::AND, MVT::v1i64, Expand);
843 setOperationAction(ISD::OR, MVT::v8i8, Expand);
844 setOperationAction(ISD::OR, MVT::v4i16, Expand);
845 setOperationAction(ISD::OR, MVT::v2i32, Expand);
846 setOperationAction(ISD::OR, MVT::v1i64, Expand);
847 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
848 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
849 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
850 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
851 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
852 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
853 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
854 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
856 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
857 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
858 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
859 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
860 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
861 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
862 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
863 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
865 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
866 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
868 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
869 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
870 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
871 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
873 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
874 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
875 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
876 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
877 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
878 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
879 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
882 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
883 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
885 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
886 // registers cannot be used even for integer operations.
887 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
888 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
889 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
890 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
892 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
893 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
894 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
895 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
896 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
897 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
898 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
899 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
900 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
901 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
902 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
903 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
904 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
905 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
906 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
907 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
908 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
909 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
911 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
912 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
913 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
914 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
916 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
917 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
919 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
922 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
923 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
924 MVT VT = (MVT::SimpleValueType)i;
925 // Do not attempt to custom lower non-power-of-2 vectors
926 if (!isPowerOf2_32(VT.getVectorNumElements()))
928 // Do not attempt to custom lower non-128-bit vectors
929 if (!VT.is128BitVector())
931 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
932 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
933 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
936 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
937 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
938 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
939 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
943 if (Subtarget->is64Bit()) {
944 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
945 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
948 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
949 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
950 MVT VT = (MVT::SimpleValueType)i;
952 // Do not attempt to promote non-128-bit vectors
953 if (!VT.is128BitVector())
956 setOperationAction(ISD::AND, VT, Promote);
957 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
958 setOperationAction(ISD::OR, VT, Promote);
959 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
960 setOperationAction(ISD::XOR, VT, Promote);
961 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
962 setOperationAction(ISD::LOAD, VT, Promote);
963 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
964 setOperationAction(ISD::SELECT, VT, Promote);
965 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
968 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
970 // Custom lower v2i64 and v2f64 selects.
971 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
972 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
973 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
974 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
976 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
977 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
979 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
980 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
981 // As there is no 64-bit GPR available, we need build a special custom
982 // sequence to convert from v2i32 to v2f32.
983 if (!Subtarget->is64Bit())
984 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
986 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
987 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
989 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
992 if (Subtarget->hasSSE41()) {
993 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
994 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
995 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
996 setOperationAction(ISD::FRINT, MVT::f32, Legal);
997 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
998 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
999 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1000 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1001 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1002 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1004 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1005 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1006 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1007 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1008 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1009 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1010 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1011 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1012 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1013 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1015 // FIXME: Do we need to handle scalar-to-vector here?
1016 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1018 setOperationAction(ISD::VSELECT, MVT::v2f64, Legal);
1019 setOperationAction(ISD::VSELECT, MVT::v2i64, Legal);
1020 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1021 setOperationAction(ISD::VSELECT, MVT::v4i32, Legal);
1022 setOperationAction(ISD::VSELECT, MVT::v4f32, Legal);
1024 // i8 and i16 vectors are custom , because the source register and source
1025 // source memory operand types are not the same width. f32 vectors are
1026 // custom since the immediate controlling the insert encodes additional
1028 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1029 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1030 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1031 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1033 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1034 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1035 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1036 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1038 // FIXME: these should be Legal but thats only for the case where
1039 // the index is constant. For now custom expand to deal with that.
1040 if (Subtarget->is64Bit()) {
1041 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1042 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1046 if (Subtarget->hasSSE2()) {
1047 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1048 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1050 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1051 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1053 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1054 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1056 // In the customized shift lowering, the legal cases in AVX2 will be
1058 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1059 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1061 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1062 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1064 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1066 setOperationAction(ISD::SDIV, MVT::v8i16, Custom);
1067 setOperationAction(ISD::SDIV, MVT::v4i32, Custom);
1070 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1071 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1072 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1073 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1074 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1075 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1076 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1078 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1079 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1080 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1082 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1087 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1088 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1089 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1090 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1091 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1092 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1093 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1095 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1100 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1101 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1102 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1103 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1104 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1105 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1106 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1108 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1109 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1111 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
1113 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1114 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1115 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1117 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1118 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1119 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1121 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
1123 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1124 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1126 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1127 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1129 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1130 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1132 setOperationAction(ISD::SDIV, MVT::v16i16, Custom);
1134 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1135 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1136 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1137 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1139 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1140 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1141 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1143 setOperationAction(ISD::VSELECT, MVT::v4f64, Legal);
1144 setOperationAction(ISD::VSELECT, MVT::v4i64, Legal);
1145 setOperationAction(ISD::VSELECT, MVT::v8i32, Legal);
1146 setOperationAction(ISD::VSELECT, MVT::v8f32, Legal);
1148 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1149 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1150 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1151 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1152 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1153 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1155 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1156 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1157 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1158 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1159 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1160 setOperationAction(ISD::FMA, MVT::f32, Legal);
1161 setOperationAction(ISD::FMA, MVT::f64, Legal);
1164 if (Subtarget->hasInt256()) {
1165 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1166 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1167 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1168 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1170 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1171 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1172 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1173 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1175 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1176 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1177 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1178 // Don't lower v32i8 because there is no 128-bit byte mul
1180 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1182 setOperationAction(ISD::SDIV, MVT::v8i32, Custom);
1184 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1185 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1186 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1187 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1189 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1190 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1191 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1192 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1194 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1195 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1196 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1197 // Don't lower v32i8 because there is no 128-bit byte mul
1200 // In the customized shift lowering, the legal cases in AVX2 will be
1202 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1203 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1205 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1206 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1208 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1210 // Custom lower several nodes for 256-bit types.
1211 for (int i = MVT::FIRST_VECTOR_VALUETYPE;
1212 i <= MVT::LAST_VECTOR_VALUETYPE; ++i) {
1213 MVT VT = (MVT::SimpleValueType)i;
1215 // Extract subvector is special because the value type
1216 // (result) is 128-bit but the source is 256-bit wide.
1217 if (VT.is128BitVector())
1218 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1220 // Do not attempt to custom lower other non-256-bit vectors
1221 if (!VT.is256BitVector())
1224 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1225 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1226 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1227 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1228 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1229 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1230 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1233 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1234 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1235 MVT VT = (MVT::SimpleValueType)i;
1237 // Do not attempt to promote non-256-bit vectors
1238 if (!VT.is256BitVector())
1241 setOperationAction(ISD::AND, VT, Promote);
1242 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1243 setOperationAction(ISD::OR, VT, Promote);
1244 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1245 setOperationAction(ISD::XOR, VT, Promote);
1246 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1247 setOperationAction(ISD::LOAD, VT, Promote);
1248 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1249 setOperationAction(ISD::SELECT, VT, Promote);
1250 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1254 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1255 // of this type with custom code.
1256 for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
1257 VT != MVT::LAST_VECTOR_VALUETYPE; VT++) {
1258 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
1262 // We want to custom lower some of our intrinsics.
1263 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1264 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1266 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1267 // handle type legalization for these operations here.
1269 // FIXME: We really should do custom legalization for addition and
1270 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1271 // than generic legalization for 64-bit multiplication-with-overflow, though.
1272 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1273 // Add/Sub/Mul with overflow operations are custom lowered.
1275 setOperationAction(ISD::SADDO, VT, Custom);
1276 setOperationAction(ISD::UADDO, VT, Custom);
1277 setOperationAction(ISD::SSUBO, VT, Custom);
1278 setOperationAction(ISD::USUBO, VT, Custom);
1279 setOperationAction(ISD::SMULO, VT, Custom);
1280 setOperationAction(ISD::UMULO, VT, Custom);
1283 // There are no 8-bit 3-address imul/mul instructions
1284 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1285 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1287 if (!Subtarget->is64Bit()) {
1288 // These libcalls are not available in 32-bit.
1289 setLibcallName(RTLIB::SHL_I128, 0);
1290 setLibcallName(RTLIB::SRL_I128, 0);
1291 setLibcallName(RTLIB::SRA_I128, 0);
1294 // Combine sin / cos into one node or libcall if possible.
1295 if (Subtarget->hasSinCos()) {
1296 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1297 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1298 if (Subtarget->isTargetDarwin()) {
1299 // For MacOSX, we don't want to the normal expansion of a libcall to
1300 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
1302 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1303 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1307 // We have target-specific dag combine patterns for the following nodes:
1308 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1309 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1310 setTargetDAGCombine(ISD::VSELECT);
1311 setTargetDAGCombine(ISD::SELECT);
1312 setTargetDAGCombine(ISD::SHL);
1313 setTargetDAGCombine(ISD::SRA);
1314 setTargetDAGCombine(ISD::SRL);
1315 setTargetDAGCombine(ISD::OR);
1316 setTargetDAGCombine(ISD::AND);
1317 setTargetDAGCombine(ISD::ADD);
1318 setTargetDAGCombine(ISD::FADD);
1319 setTargetDAGCombine(ISD::FSUB);
1320 setTargetDAGCombine(ISD::FMA);
1321 setTargetDAGCombine(ISD::SUB);
1322 setTargetDAGCombine(ISD::LOAD);
1323 setTargetDAGCombine(ISD::STORE);
1324 setTargetDAGCombine(ISD::ZERO_EXTEND);
1325 setTargetDAGCombine(ISD::ANY_EXTEND);
1326 setTargetDAGCombine(ISD::SIGN_EXTEND);
1327 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1328 setTargetDAGCombine(ISD::TRUNCATE);
1329 setTargetDAGCombine(ISD::SINT_TO_FP);
1330 setTargetDAGCombine(ISD::SETCC);
1331 if (Subtarget->is64Bit())
1332 setTargetDAGCombine(ISD::MUL);
1333 setTargetDAGCombine(ISD::XOR);
1335 computeRegisterProperties();
1337 // On Darwin, -Os means optimize for size without hurting performance,
1338 // do not reduce the limit.
1339 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1340 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1341 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1342 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1343 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1344 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1345 setPrefLoopAlignment(4); // 2^4 bytes.
1346 BenefitFromCodePlacementOpt = true;
1348 // Predictable cmov don't hurt on atom because it's in-order.
1349 PredictableSelectIsExpensive = !Subtarget->isAtom();
1351 setPrefFunctionAlignment(4); // 2^4 bytes.
1354 EVT X86TargetLowering::getSetCCResultType(EVT VT) const {
1355 if (!VT.isVector()) return MVT::i8;
1356 return VT.changeVectorElementTypeToInteger();
1359 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1360 /// the desired ByVal argument alignment.
1361 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1364 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1365 if (VTy->getBitWidth() == 128)
1367 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1368 unsigned EltAlign = 0;
1369 getMaxByValAlign(ATy->getElementType(), EltAlign);
1370 if (EltAlign > MaxAlign)
1371 MaxAlign = EltAlign;
1372 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1373 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1374 unsigned EltAlign = 0;
1375 getMaxByValAlign(STy->getElementType(i), EltAlign);
1376 if (EltAlign > MaxAlign)
1377 MaxAlign = EltAlign;
1384 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1385 /// function arguments in the caller parameter area. For X86, aggregates
1386 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1387 /// are at 4-byte boundaries.
1388 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1389 if (Subtarget->is64Bit()) {
1390 // Max of 8 and alignment of type.
1391 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1398 if (Subtarget->hasSSE1())
1399 getMaxByValAlign(Ty, Align);
1403 /// getOptimalMemOpType - Returns the target specific optimal type for load
1404 /// and store operations as a result of memset, memcpy, and memmove
1405 /// lowering. If DstAlign is zero that means it's safe to destination
1406 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1407 /// means there isn't a need to check it against alignment requirement,
1408 /// probably because the source does not need to be loaded. If 'IsMemset' is
1409 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1410 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1411 /// source is constant so it does not need to be loaded.
1412 /// It returns EVT::Other if the type should be determined using generic
1413 /// target-independent logic.
1415 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1416 unsigned DstAlign, unsigned SrcAlign,
1417 bool IsMemset, bool ZeroMemset,
1419 MachineFunction &MF) const {
1420 const Function *F = MF.getFunction();
1421 if ((!IsMemset || ZeroMemset) &&
1422 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1423 Attribute::NoImplicitFloat)) {
1425 (Subtarget->isUnalignedMemAccessFast() ||
1426 ((DstAlign == 0 || DstAlign >= 16) &&
1427 (SrcAlign == 0 || SrcAlign >= 16)))) {
1429 if (Subtarget->hasInt256())
1431 if (Subtarget->hasFp256())
1434 if (Subtarget->hasSSE2())
1436 if (Subtarget->hasSSE1())
1438 } else if (!MemcpyStrSrc && Size >= 8 &&
1439 !Subtarget->is64Bit() &&
1440 Subtarget->hasSSE2()) {
1441 // Do not use f64 to lower memcpy if source is string constant. It's
1442 // better to use i32 to avoid the loads.
1446 if (Subtarget->is64Bit() && Size >= 8)
1451 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1453 return X86ScalarSSEf32;
1454 else if (VT == MVT::f64)
1455 return X86ScalarSSEf64;
1460 X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
1462 *Fast = Subtarget->isUnalignedMemAccessFast();
1466 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1467 /// current function. The returned value is a member of the
1468 /// MachineJumpTableInfo::JTEntryKind enum.
1469 unsigned X86TargetLowering::getJumpTableEncoding() const {
1470 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1472 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1473 Subtarget->isPICStyleGOT())
1474 return MachineJumpTableInfo::EK_Custom32;
1476 // Otherwise, use the normal jump table encoding heuristics.
1477 return TargetLowering::getJumpTableEncoding();
1481 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1482 const MachineBasicBlock *MBB,
1483 unsigned uid,MCContext &Ctx) const{
1484 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1485 Subtarget->isPICStyleGOT());
1486 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1488 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1489 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1492 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1494 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1495 SelectionDAG &DAG) const {
1496 if (!Subtarget->is64Bit())
1497 // This doesn't have DebugLoc associated with it, but is not really the
1498 // same as a Register.
1499 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1503 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1504 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1506 const MCExpr *X86TargetLowering::
1507 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1508 MCContext &Ctx) const {
1509 // X86-64 uses RIP relative addressing based on the jump table label.
1510 if (Subtarget->isPICStyleRIPRel())
1511 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1513 // Otherwise, the reference is relative to the PIC base.
1514 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1517 // FIXME: Why this routine is here? Move to RegInfo!
1518 std::pair<const TargetRegisterClass*, uint8_t>
1519 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1520 const TargetRegisterClass *RRC = 0;
1522 switch (VT.SimpleTy) {
1524 return TargetLowering::findRepresentativeClass(VT);
1525 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1526 RRC = Subtarget->is64Bit() ?
1527 (const TargetRegisterClass*)&X86::GR64RegClass :
1528 (const TargetRegisterClass*)&X86::GR32RegClass;
1531 RRC = &X86::VR64RegClass;
1533 case MVT::f32: case MVT::f64:
1534 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1535 case MVT::v4f32: case MVT::v2f64:
1536 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1538 RRC = &X86::VR128RegClass;
1541 return std::make_pair(RRC, Cost);
1544 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1545 unsigned &Offset) const {
1546 if (!Subtarget->isTargetLinux())
1549 if (Subtarget->is64Bit()) {
1550 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1552 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1564 //===----------------------------------------------------------------------===//
1565 // Return Value Calling Convention Implementation
1566 //===----------------------------------------------------------------------===//
1568 #include "X86GenCallingConv.inc"
1571 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1572 MachineFunction &MF, bool isVarArg,
1573 const SmallVectorImpl<ISD::OutputArg> &Outs,
1574 LLVMContext &Context) const {
1575 SmallVector<CCValAssign, 16> RVLocs;
1576 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1578 return CCInfo.CheckReturn(Outs, RetCC_X86);
1582 X86TargetLowering::LowerReturn(SDValue Chain,
1583 CallingConv::ID CallConv, bool isVarArg,
1584 const SmallVectorImpl<ISD::OutputArg> &Outs,
1585 const SmallVectorImpl<SDValue> &OutVals,
1586 DebugLoc dl, SelectionDAG &DAG) const {
1587 MachineFunction &MF = DAG.getMachineFunction();
1588 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1590 SmallVector<CCValAssign, 16> RVLocs;
1591 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1592 RVLocs, *DAG.getContext());
1593 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1596 SmallVector<SDValue, 6> RetOps;
1597 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1598 // Operand #1 = Bytes To Pop
1599 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1602 // Copy the result values into the output registers.
1603 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1604 CCValAssign &VA = RVLocs[i];
1605 assert(VA.isRegLoc() && "Can only return in registers!");
1606 SDValue ValToCopy = OutVals[i];
1607 EVT ValVT = ValToCopy.getValueType();
1609 // Promote values to the appropriate types
1610 if (VA.getLocInfo() == CCValAssign::SExt)
1611 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
1612 else if (VA.getLocInfo() == CCValAssign::ZExt)
1613 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
1614 else if (VA.getLocInfo() == CCValAssign::AExt)
1615 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
1616 else if (VA.getLocInfo() == CCValAssign::BCvt)
1617 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
1619 // If this is x86-64, and we disabled SSE, we can't return FP values,
1620 // or SSE or MMX vectors.
1621 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1622 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1623 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1624 report_fatal_error("SSE register return with SSE disabled");
1626 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1627 // llvm-gcc has never done it right and no one has noticed, so this
1628 // should be OK for now.
1629 if (ValVT == MVT::f64 &&
1630 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1631 report_fatal_error("SSE2 register return with SSE2 disabled");
1633 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1634 // the RET instruction and handled by the FP Stackifier.
1635 if (VA.getLocReg() == X86::ST0 ||
1636 VA.getLocReg() == X86::ST1) {
1637 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1638 // change the value to the FP stack register class.
1639 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1640 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1641 RetOps.push_back(ValToCopy);
1642 // Don't emit a copytoreg.
1646 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1647 // which is returned in RAX / RDX.
1648 if (Subtarget->is64Bit()) {
1649 if (ValVT == MVT::x86mmx) {
1650 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1651 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1652 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1654 // If we don't have SSE2 available, convert to v4f32 so the generated
1655 // register is legal.
1656 if (!Subtarget->hasSSE2())
1657 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1662 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1663 Flag = Chain.getValue(1);
1664 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1667 // The x86-64 ABIs require that for returning structs by value we copy
1668 // the sret argument into %rax/%eax (depending on ABI) for the return.
1669 // Win32 requires us to put the sret argument to %eax as well.
1670 // We saved the argument into a virtual register in the entry block,
1671 // so now we copy the value out and into %rax/%eax.
1672 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
1673 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
1674 MachineFunction &MF = DAG.getMachineFunction();
1675 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1676 unsigned Reg = FuncInfo->getSRetReturnReg();
1678 "SRetReturnReg should have been set in LowerFormalArguments().");
1679 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1682 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
1683 X86::RAX : X86::EAX;
1684 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
1685 Flag = Chain.getValue(1);
1687 // RAX/EAX now acts like a return value.
1688 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
1691 RetOps[0] = Chain; // Update chain.
1693 // Add the flag if we have it.
1695 RetOps.push_back(Flag);
1697 return DAG.getNode(X86ISD::RET_FLAG, dl,
1698 MVT::Other, &RetOps[0], RetOps.size());
1701 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
1702 if (N->getNumValues() != 1)
1704 if (!N->hasNUsesOfValue(1, 0))
1707 SDValue TCChain = Chain;
1708 SDNode *Copy = *N->use_begin();
1709 if (Copy->getOpcode() == ISD::CopyToReg) {
1710 // If the copy has a glue operand, we conservatively assume it isn't safe to
1711 // perform a tail call.
1712 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
1714 TCChain = Copy->getOperand(0);
1715 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
1718 bool HasRet = false;
1719 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1721 if (UI->getOpcode() != X86ISD::RET_FLAG)
1734 X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
1735 ISD::NodeType ExtendKind) const {
1737 // TODO: Is this also valid on 32-bit?
1738 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1739 ReturnMVT = MVT::i8;
1741 ReturnMVT = MVT::i32;
1743 MVT MinVT = getRegisterType(ReturnMVT);
1744 return VT.bitsLT(MinVT) ? MinVT : VT;
1747 /// LowerCallResult - Lower the result values of a call into the
1748 /// appropriate copies out of appropriate physical registers.
1751 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1752 CallingConv::ID CallConv, bool isVarArg,
1753 const SmallVectorImpl<ISD::InputArg> &Ins,
1754 DebugLoc dl, SelectionDAG &DAG,
1755 SmallVectorImpl<SDValue> &InVals) const {
1757 // Assign locations to each value returned by this call.
1758 SmallVector<CCValAssign, 16> RVLocs;
1759 bool Is64Bit = Subtarget->is64Bit();
1760 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1761 getTargetMachine(), RVLocs, *DAG.getContext());
1762 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1764 // Copy all of the result registers out of their specified physreg.
1765 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1766 CCValAssign &VA = RVLocs[i];
1767 EVT CopyVT = VA.getValVT();
1769 // If this is x86-64, and we disabled SSE, we can't return FP values
1770 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1771 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1772 report_fatal_error("SSE register return with SSE disabled");
1777 // If this is a call to a function that returns an fp value on the floating
1778 // point stack, we must guarantee the value is popped from the stack, so
1779 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1780 // if the return value is not used. We use the FpPOP_RETVAL instruction
1782 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1783 // If we prefer to use the value in xmm registers, copy it out as f80 and
1784 // use a truncate to move it from fp stack reg to xmm reg.
1785 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1786 SDValue Ops[] = { Chain, InFlag };
1787 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1788 MVT::Other, MVT::Glue, Ops, 2), 1);
1789 Val = Chain.getValue(0);
1791 // Round the f80 to the right size, which also moves it to the appropriate
1793 if (CopyVT != VA.getValVT())
1794 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1795 // This truncation won't change the value.
1796 DAG.getIntPtrConstant(1));
1798 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1799 CopyVT, InFlag).getValue(1);
1800 Val = Chain.getValue(0);
1802 InFlag = Chain.getValue(2);
1803 InVals.push_back(Val);
1809 //===----------------------------------------------------------------------===//
1810 // C & StdCall & Fast Calling Convention implementation
1811 //===----------------------------------------------------------------------===//
1812 // StdCall calling convention seems to be standard for many Windows' API
1813 // routines and around. It differs from C calling convention just a little:
1814 // callee should clean up the stack, not caller. Symbols should be also
1815 // decorated in some fancy way :) It doesn't support any vector arguments.
1816 // For info on fast calling convention see Fast Calling Convention (tail call)
1817 // implementation LowerX86_32FastCCCallTo.
1819 /// CallIsStructReturn - Determines whether a call uses struct return
1821 enum StructReturnType {
1826 static StructReturnType
1827 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1829 return NotStructReturn;
1831 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
1832 if (!Flags.isSRet())
1833 return NotStructReturn;
1834 if (Flags.isInReg())
1835 return RegStructReturn;
1836 return StackStructReturn;
1839 /// ArgsAreStructReturn - Determines whether a function uses struct
1840 /// return semantics.
1841 static StructReturnType
1842 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1844 return NotStructReturn;
1846 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
1847 if (!Flags.isSRet())
1848 return NotStructReturn;
1849 if (Flags.isInReg())
1850 return RegStructReturn;
1851 return StackStructReturn;
1854 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1855 /// by "Src" to address "Dst" with size and alignment information specified by
1856 /// the specific parameter attribute. The copy will be passed as a byval
1857 /// function parameter.
1859 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1860 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1862 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1864 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1865 /*isVolatile*/false, /*AlwaysInline=*/true,
1866 MachinePointerInfo(), MachinePointerInfo());
1869 /// IsTailCallConvention - Return true if the calling convention is one that
1870 /// supports tail call optimization.
1871 static bool IsTailCallConvention(CallingConv::ID CC) {
1872 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
1873 CC == CallingConv::HiPE);
1876 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1877 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
1881 CallingConv::ID CalleeCC = CS.getCallingConv();
1882 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1888 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1889 /// a tailcall target by changing its ABI.
1890 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
1891 bool GuaranteedTailCallOpt) {
1892 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1896 X86TargetLowering::LowerMemArgument(SDValue Chain,
1897 CallingConv::ID CallConv,
1898 const SmallVectorImpl<ISD::InputArg> &Ins,
1899 DebugLoc dl, SelectionDAG &DAG,
1900 const CCValAssign &VA,
1901 MachineFrameInfo *MFI,
1903 // Create the nodes corresponding to a load from this parameter slot.
1904 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1905 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv,
1906 getTargetMachine().Options.GuaranteedTailCallOpt);
1907 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1910 // If value is passed by pointer we have address passed instead of the value
1912 if (VA.getLocInfo() == CCValAssign::Indirect)
1913 ValVT = VA.getLocVT();
1915 ValVT = VA.getValVT();
1917 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1918 // changed with more analysis.
1919 // In case of tail call optimization mark all arguments mutable. Since they
1920 // could be overwritten by lowering of arguments in case of a tail call.
1921 if (Flags.isByVal()) {
1922 unsigned Bytes = Flags.getByValSize();
1923 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1924 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1925 return DAG.getFrameIndex(FI, getPointerTy());
1927 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1928 VA.getLocMemOffset(), isImmutable);
1929 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1930 return DAG.getLoad(ValVT, dl, Chain, FIN,
1931 MachinePointerInfo::getFixedStack(FI),
1932 false, false, false, 0);
1937 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1938 CallingConv::ID CallConv,
1940 const SmallVectorImpl<ISD::InputArg> &Ins,
1943 SmallVectorImpl<SDValue> &InVals)
1945 MachineFunction &MF = DAG.getMachineFunction();
1946 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1948 const Function* Fn = MF.getFunction();
1949 if (Fn->hasExternalLinkage() &&
1950 Subtarget->isTargetCygMing() &&
1951 Fn->getName() == "main")
1952 FuncInfo->setForceFramePointer(true);
1954 MachineFrameInfo *MFI = MF.getFrameInfo();
1955 bool Is64Bit = Subtarget->is64Bit();
1956 bool IsWindows = Subtarget->isTargetWindows();
1957 bool IsWin64 = Subtarget->isTargetWin64();
1959 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1960 "Var args not supported with calling convention fastcc, ghc or hipe");
1962 // Assign locations to all of the incoming arguments.
1963 SmallVector<CCValAssign, 16> ArgLocs;
1964 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1965 ArgLocs, *DAG.getContext());
1967 // Allocate shadow area for Win64
1969 CCInfo.AllocateStack(32, 8);
1972 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1974 unsigned LastVal = ~0U;
1976 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1977 CCValAssign &VA = ArgLocs[i];
1978 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1980 assert(VA.getValNo() != LastVal &&
1981 "Don't support value assigned to multiple locs yet");
1983 LastVal = VA.getValNo();
1985 if (VA.isRegLoc()) {
1986 EVT RegVT = VA.getLocVT();
1987 const TargetRegisterClass *RC;
1988 if (RegVT == MVT::i32)
1989 RC = &X86::GR32RegClass;
1990 else if (Is64Bit && RegVT == MVT::i64)
1991 RC = &X86::GR64RegClass;
1992 else if (RegVT == MVT::f32)
1993 RC = &X86::FR32RegClass;
1994 else if (RegVT == MVT::f64)
1995 RC = &X86::FR64RegClass;
1996 else if (RegVT.is256BitVector())
1997 RC = &X86::VR256RegClass;
1998 else if (RegVT.is128BitVector())
1999 RC = &X86::VR128RegClass;
2000 else if (RegVT == MVT::x86mmx)
2001 RC = &X86::VR64RegClass;
2003 llvm_unreachable("Unknown argument type!");
2005 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2006 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2008 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2009 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2011 if (VA.getLocInfo() == CCValAssign::SExt)
2012 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2013 DAG.getValueType(VA.getValVT()));
2014 else if (VA.getLocInfo() == CCValAssign::ZExt)
2015 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2016 DAG.getValueType(VA.getValVT()));
2017 else if (VA.getLocInfo() == CCValAssign::BCvt)
2018 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2020 if (VA.isExtInLoc()) {
2021 // Handle MMX values passed in XMM regs.
2022 if (RegVT.isVector())
2023 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2025 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2028 assert(VA.isMemLoc());
2029 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2032 // If value is passed via pointer - do a load.
2033 if (VA.getLocInfo() == CCValAssign::Indirect)
2034 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2035 MachinePointerInfo(), false, false, false, 0);
2037 InVals.push_back(ArgValue);
2040 // The x86-64 ABIs require that for returning structs by value we copy
2041 // the sret argument into %rax/%eax (depending on ABI) for the return.
2042 // Win32 requires us to put the sret argument to %eax as well.
2043 // Save the argument into a virtual register so that we can access it
2044 // from the return points.
2045 if (MF.getFunction()->hasStructRetAttr() &&
2046 (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
2047 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2048 unsigned Reg = FuncInfo->getSRetReturnReg();
2050 MVT PtrTy = getPointerTy();
2051 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2052 FuncInfo->setSRetReturnReg(Reg);
2054 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
2055 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2058 unsigned StackSize = CCInfo.getNextStackOffset();
2059 // Align stack specially for tail calls.
2060 if (FuncIsMadeTailCallSafe(CallConv,
2061 MF.getTarget().Options.GuaranteedTailCallOpt))
2062 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2064 // If the function takes variable number of arguments, make a frame index for
2065 // the start of the first vararg value... for expansion of llvm.va_start.
2067 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2068 CallConv != CallingConv::X86_ThisCall)) {
2069 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
2072 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
2074 // FIXME: We should really autogenerate these arrays
2075 static const uint16_t GPR64ArgRegsWin64[] = {
2076 X86::RCX, X86::RDX, X86::R8, X86::R9
2078 static const uint16_t GPR64ArgRegs64Bit[] = {
2079 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2081 static const uint16_t XMMArgRegs64Bit[] = {
2082 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2083 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2085 const uint16_t *GPR64ArgRegs;
2086 unsigned NumXMMRegs = 0;
2089 // The XMM registers which might contain var arg parameters are shadowed
2090 // in their paired GPR. So we only need to save the GPR to their home
2092 TotalNumIntRegs = 4;
2093 GPR64ArgRegs = GPR64ArgRegsWin64;
2095 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
2096 GPR64ArgRegs = GPR64ArgRegs64Bit;
2098 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
2101 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
2104 bool NoImplicitFloatOps = Fn->getAttributes().
2105 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2106 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2107 "SSE register cannot be used when SSE is disabled!");
2108 assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
2109 NoImplicitFloatOps) &&
2110 "SSE register cannot be used when SSE is disabled!");
2111 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2112 !Subtarget->hasSSE1())
2113 // Kernel mode asks for SSE to be disabled, so don't push them
2115 TotalNumXMMRegs = 0;
2118 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
2119 // Get to the caller-allocated home save location. Add 8 to account
2120 // for the return address.
2121 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2122 FuncInfo->setRegSaveFrameIndex(
2123 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2124 // Fixup to set vararg frame on shadow area (4 x i64).
2126 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2128 // For X86-64, if there are vararg parameters that are passed via
2129 // registers, then we must store them to their spots on the stack so
2130 // they may be loaded by deferencing the result of va_next.
2131 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2132 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
2133 FuncInfo->setRegSaveFrameIndex(
2134 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
2138 // Store the integer parameter registers.
2139 SmallVector<SDValue, 8> MemOps;
2140 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2142 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2143 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
2144 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2145 DAG.getIntPtrConstant(Offset));
2146 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
2147 &X86::GR64RegClass);
2148 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2150 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2151 MachinePointerInfo::getFixedStack(
2152 FuncInfo->getRegSaveFrameIndex(), Offset),
2154 MemOps.push_back(Store);
2158 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
2159 // Now store the XMM (fp + vector) parameter registers.
2160 SmallVector<SDValue, 11> SaveXMMOps;
2161 SaveXMMOps.push_back(Chain);
2163 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2164 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
2165 SaveXMMOps.push_back(ALVal);
2167 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2168 FuncInfo->getRegSaveFrameIndex()));
2169 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2170 FuncInfo->getVarArgsFPOffset()));
2172 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
2173 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
2174 &X86::VR128RegClass);
2175 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
2176 SaveXMMOps.push_back(Val);
2178 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2180 &SaveXMMOps[0], SaveXMMOps.size()));
2183 if (!MemOps.empty())
2184 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2185 &MemOps[0], MemOps.size());
2189 // Some CCs need callee pop.
2190 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2191 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2192 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2194 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2195 // If this is an sret function, the return should pop the hidden pointer.
2196 if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2197 argsAreStructReturn(Ins) == StackStructReturn)
2198 FuncInfo->setBytesToPopOnReturn(4);
2202 // RegSaveFrameIndex is X86-64 only.
2203 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2204 if (CallConv == CallingConv::X86_FastCall ||
2205 CallConv == CallingConv::X86_ThisCall)
2206 // fastcc functions can't have varargs.
2207 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2210 FuncInfo->setArgumentStackSize(StackSize);
2216 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2217 SDValue StackPtr, SDValue Arg,
2218 DebugLoc dl, SelectionDAG &DAG,
2219 const CCValAssign &VA,
2220 ISD::ArgFlagsTy Flags) const {
2221 unsigned LocMemOffset = VA.getLocMemOffset();
2222 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2223 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2224 if (Flags.isByVal())
2225 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2227 return DAG.getStore(Chain, dl, Arg, PtrOff,
2228 MachinePointerInfo::getStack(LocMemOffset),
2232 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
2233 /// optimization is performed and it is required.
2235 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2236 SDValue &OutRetAddr, SDValue Chain,
2237 bool IsTailCall, bool Is64Bit,
2238 int FPDiff, DebugLoc dl) const {
2239 // Adjust the Return address stack slot.
2240 EVT VT = getPointerTy();
2241 OutRetAddr = getReturnAddressFrameIndex(DAG);
2243 // Load the "old" Return address.
2244 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2245 false, false, false, 0);
2246 return SDValue(OutRetAddr.getNode(), 1);
2249 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
2250 /// optimization is performed and it is required (FPDiff!=0).
2252 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
2253 SDValue Chain, SDValue RetAddrFrIdx, EVT PtrVT,
2254 unsigned SlotSize, int FPDiff, DebugLoc dl) {
2255 // Store the return address to the appropriate stack slot.
2256 if (!FPDiff) return Chain;
2257 // Calculate the new stack slot for the return address.
2258 int NewReturnAddrFI =
2259 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
2260 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2261 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2262 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2268 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2269 SmallVectorImpl<SDValue> &InVals) const {
2270 SelectionDAG &DAG = CLI.DAG;
2271 DebugLoc &dl = CLI.DL;
2272 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2273 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2274 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2275 SDValue Chain = CLI.Chain;
2276 SDValue Callee = CLI.Callee;
2277 CallingConv::ID CallConv = CLI.CallConv;
2278 bool &isTailCall = CLI.IsTailCall;
2279 bool isVarArg = CLI.IsVarArg;
2281 MachineFunction &MF = DAG.getMachineFunction();
2282 bool Is64Bit = Subtarget->is64Bit();
2283 bool IsWin64 = Subtarget->isTargetWin64();
2284 bool IsWindows = Subtarget->isTargetWindows();
2285 StructReturnType SR = callIsStructReturn(Outs);
2286 bool IsSibcall = false;
2288 if (MF.getTarget().Options.DisableTailCalls)
2292 // Check if it's really possible to do a tail call.
2293 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2294 isVarArg, SR != NotStructReturn,
2295 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2296 Outs, OutVals, Ins, DAG);
2298 // Sibcalls are automatically detected tailcalls which do not require
2300 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2307 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2308 "Var args not supported with calling convention fastcc, ghc or hipe");
2310 // Analyze operands of the call, assigning locations to each operand.
2311 SmallVector<CCValAssign, 16> ArgLocs;
2312 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2313 ArgLocs, *DAG.getContext());
2315 // Allocate shadow area for Win64
2317 CCInfo.AllocateStack(32, 8);
2320 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2322 // Get a count of how many bytes are to be pushed on the stack.
2323 unsigned NumBytes = CCInfo.getNextStackOffset();
2325 // This is a sibcall. The memory operands are available in caller's
2326 // own caller's stack.
2328 else if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2329 IsTailCallConvention(CallConv))
2330 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2333 if (isTailCall && !IsSibcall) {
2334 // Lower arguments at fp - stackoffset + fpdiff.
2335 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2336 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2338 FPDiff = NumBytesCallerPushed - NumBytes;
2340 // Set the delta of movement of the returnaddr stackslot.
2341 // But only set if delta is greater than previous delta.
2342 if (FPDiff < X86Info->getTCReturnAddrDelta())
2343 X86Info->setTCReturnAddrDelta(FPDiff);
2347 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2349 SDValue RetAddrFrIdx;
2350 // Load return address for tail calls.
2351 if (isTailCall && FPDiff)
2352 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2353 Is64Bit, FPDiff, dl);
2355 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2356 SmallVector<SDValue, 8> MemOpChains;
2359 // Walk the register/memloc assignments, inserting copies/loads. In the case
2360 // of tail call optimization arguments are handle later.
2361 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2362 CCValAssign &VA = ArgLocs[i];
2363 EVT RegVT = VA.getLocVT();
2364 SDValue Arg = OutVals[i];
2365 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2366 bool isByVal = Flags.isByVal();
2368 // Promote the value if needed.
2369 switch (VA.getLocInfo()) {
2370 default: llvm_unreachable("Unknown loc info!");
2371 case CCValAssign::Full: break;
2372 case CCValAssign::SExt:
2373 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2375 case CCValAssign::ZExt:
2376 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2378 case CCValAssign::AExt:
2379 if (RegVT.is128BitVector()) {
2380 // Special case: passing MMX values in XMM registers.
2381 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2382 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2383 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2385 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2387 case CCValAssign::BCvt:
2388 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2390 case CCValAssign::Indirect: {
2391 // Store the argument.
2392 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2393 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2394 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2395 MachinePointerInfo::getFixedStack(FI),
2402 if (VA.isRegLoc()) {
2403 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2404 if (isVarArg && IsWin64) {
2405 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2406 // shadow reg if callee is a varargs function.
2407 unsigned ShadowReg = 0;
2408 switch (VA.getLocReg()) {
2409 case X86::XMM0: ShadowReg = X86::RCX; break;
2410 case X86::XMM1: ShadowReg = X86::RDX; break;
2411 case X86::XMM2: ShadowReg = X86::R8; break;
2412 case X86::XMM3: ShadowReg = X86::R9; break;
2415 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2417 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2418 assert(VA.isMemLoc());
2419 if (StackPtr.getNode() == 0)
2420 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2422 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2423 dl, DAG, VA, Flags));
2427 if (!MemOpChains.empty())
2428 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2429 &MemOpChains[0], MemOpChains.size());
2431 if (Subtarget->isPICStyleGOT()) {
2432 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2435 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2436 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy())));
2438 // If we are tail calling and generating PIC/GOT style code load the
2439 // address of the callee into ECX. The value in ecx is used as target of
2440 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2441 // for tail calls on PIC/GOT architectures. Normally we would just put the
2442 // address of GOT into ebx and then call target@PLT. But for tail calls
2443 // ebx would be restored (since ebx is callee saved) before jumping to the
2446 // Note: The actual moving to ECX is done further down.
2447 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2448 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2449 !G->getGlobal()->hasProtectedVisibility())
2450 Callee = LowerGlobalAddress(Callee, DAG);
2451 else if (isa<ExternalSymbolSDNode>(Callee))
2452 Callee = LowerExternalSymbol(Callee, DAG);
2456 if (Is64Bit && isVarArg && !IsWin64) {
2457 // From AMD64 ABI document:
2458 // For calls that may call functions that use varargs or stdargs
2459 // (prototype-less calls or calls to functions containing ellipsis (...) in
2460 // the declaration) %al is used as hidden argument to specify the number
2461 // of SSE registers used. The contents of %al do not need to match exactly
2462 // the number of registers, but must be an ubound on the number of SSE
2463 // registers used and is in the range 0 - 8 inclusive.
2465 // Count the number of XMM registers allocated.
2466 static const uint16_t XMMArgRegs[] = {
2467 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2468 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2470 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2471 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2472 && "SSE registers cannot be used when SSE is disabled");
2474 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
2475 DAG.getConstant(NumXMMRegs, MVT::i8)));
2478 // For tail calls lower the arguments to the 'real' stack slot.
2480 // Force all the incoming stack arguments to be loaded from the stack
2481 // before any new outgoing arguments are stored to the stack, because the
2482 // outgoing stack slots may alias the incoming argument stack slots, and
2483 // the alias isn't otherwise explicit. This is slightly more conservative
2484 // than necessary, because it means that each store effectively depends
2485 // on every argument instead of just those arguments it would clobber.
2486 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2488 SmallVector<SDValue, 8> MemOpChains2;
2491 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2492 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2493 CCValAssign &VA = ArgLocs[i];
2496 assert(VA.isMemLoc());
2497 SDValue Arg = OutVals[i];
2498 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2499 // Create frame index.
2500 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2501 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2502 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2503 FIN = DAG.getFrameIndex(FI, getPointerTy());
2505 if (Flags.isByVal()) {
2506 // Copy relative to framepointer.
2507 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2508 if (StackPtr.getNode() == 0)
2509 StackPtr = DAG.getCopyFromReg(Chain, dl,
2510 RegInfo->getStackRegister(),
2512 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2514 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2518 // Store relative to framepointer.
2519 MemOpChains2.push_back(
2520 DAG.getStore(ArgChain, dl, Arg, FIN,
2521 MachinePointerInfo::getFixedStack(FI),
2527 if (!MemOpChains2.empty())
2528 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2529 &MemOpChains2[0], MemOpChains2.size());
2531 // Store the return address to the appropriate stack slot.
2532 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
2533 getPointerTy(), RegInfo->getSlotSize(),
2537 // Build a sequence of copy-to-reg nodes chained together with token chain
2538 // and flag operands which copy the outgoing args into registers.
2540 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2541 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2542 RegsToPass[i].second, InFlag);
2543 InFlag = Chain.getValue(1);
2546 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2547 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2548 // In the 64-bit large code model, we have to make all calls
2549 // through a register, since the call instruction's 32-bit
2550 // pc-relative offset may not be large enough to hold the whole
2552 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2553 // If the callee is a GlobalAddress node (quite common, every direct call
2554 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2557 // We should use extra load for direct calls to dllimported functions in
2559 const GlobalValue *GV = G->getGlobal();
2560 if (!GV->hasDLLImportLinkage()) {
2561 unsigned char OpFlags = 0;
2562 bool ExtraLoad = false;
2563 unsigned WrapperKind = ISD::DELETED_NODE;
2565 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2566 // external symbols most go through the PLT in PIC mode. If the symbol
2567 // has hidden or protected visibility, or if it is static or local, then
2568 // we don't need to use the PLT - we can directly call it.
2569 if (Subtarget->isTargetELF() &&
2570 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2571 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2572 OpFlags = X86II::MO_PLT;
2573 } else if (Subtarget->isPICStyleStubAny() &&
2574 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2575 (!Subtarget->getTargetTriple().isMacOSX() ||
2576 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2577 // PC-relative references to external symbols should go through $stub,
2578 // unless we're building with the leopard linker or later, which
2579 // automatically synthesizes these stubs.
2580 OpFlags = X86II::MO_DARWIN_STUB;
2581 } else if (Subtarget->isPICStyleRIPRel() &&
2582 isa<Function>(GV) &&
2583 cast<Function>(GV)->getAttributes().
2584 hasAttribute(AttributeSet::FunctionIndex,
2585 Attribute::NonLazyBind)) {
2586 // If the function is marked as non-lazy, generate an indirect call
2587 // which loads from the GOT directly. This avoids runtime overhead
2588 // at the cost of eager binding (and one extra byte of encoding).
2589 OpFlags = X86II::MO_GOTPCREL;
2590 WrapperKind = X86ISD::WrapperRIP;
2594 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2595 G->getOffset(), OpFlags);
2597 // Add a wrapper if needed.
2598 if (WrapperKind != ISD::DELETED_NODE)
2599 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2600 // Add extra indirection if needed.
2602 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2603 MachinePointerInfo::getGOT(),
2604 false, false, false, 0);
2606 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2607 unsigned char OpFlags = 0;
2609 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2610 // external symbols should go through the PLT.
2611 if (Subtarget->isTargetELF() &&
2612 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2613 OpFlags = X86II::MO_PLT;
2614 } else if (Subtarget->isPICStyleStubAny() &&
2615 (!Subtarget->getTargetTriple().isMacOSX() ||
2616 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2617 // PC-relative references to external symbols should go through $stub,
2618 // unless we're building with the leopard linker or later, which
2619 // automatically synthesizes these stubs.
2620 OpFlags = X86II::MO_DARWIN_STUB;
2623 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2627 // Returns a chain & a flag for retval copy to use.
2628 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2629 SmallVector<SDValue, 8> Ops;
2631 if (!IsSibcall && isTailCall) {
2632 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2633 DAG.getIntPtrConstant(0, true), InFlag);
2634 InFlag = Chain.getValue(1);
2637 Ops.push_back(Chain);
2638 Ops.push_back(Callee);
2641 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2643 // Add argument registers to the end of the list so that they are known live
2645 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2646 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2647 RegsToPass[i].second.getValueType()));
2649 // Add a register mask operand representing the call-preserved registers.
2650 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2651 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2652 assert(Mask && "Missing call preserved mask for calling convention");
2653 Ops.push_back(DAG.getRegisterMask(Mask));
2655 if (InFlag.getNode())
2656 Ops.push_back(InFlag);
2660 //// If this is the first return lowered for this function, add the regs
2661 //// to the liveout set for the function.
2662 // This isn't right, although it's probably harmless on x86; liveouts
2663 // should be computed from returns not tail calls. Consider a void
2664 // function making a tail call to a function returning int.
2665 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
2668 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2669 InFlag = Chain.getValue(1);
2671 // Create the CALLSEQ_END node.
2672 unsigned NumBytesForCalleeToPush;
2673 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2674 getTargetMachine().Options.GuaranteedTailCallOpt))
2675 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2676 else if (!Is64Bit && !IsTailCallConvention(CallConv) && !IsWindows &&
2677 SR == StackStructReturn)
2678 // If this is a call to a struct-return function, the callee
2679 // pops the hidden struct pointer, so we have to push it back.
2680 // This is common for Darwin/X86, Linux & Mingw32 targets.
2681 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
2682 NumBytesForCalleeToPush = 4;
2684 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2686 // Returns a flag for retval copy to use.
2688 Chain = DAG.getCALLSEQ_END(Chain,
2689 DAG.getIntPtrConstant(NumBytes, true),
2690 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2693 InFlag = Chain.getValue(1);
2696 // Handle result values, copying them out of physregs into vregs that we
2698 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2699 Ins, dl, DAG, InVals);
2702 //===----------------------------------------------------------------------===//
2703 // Fast Calling Convention (tail call) implementation
2704 //===----------------------------------------------------------------------===//
2706 // Like std call, callee cleans arguments, convention except that ECX is
2707 // reserved for storing the tail called function address. Only 2 registers are
2708 // free for argument passing (inreg). Tail call optimization is performed
2710 // * tailcallopt is enabled
2711 // * caller/callee are fastcc
2712 // On X86_64 architecture with GOT-style position independent code only local
2713 // (within module) calls are supported at the moment.
2714 // To keep the stack aligned according to platform abi the function
2715 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2716 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2717 // If a tail called function callee has more arguments than the caller the
2718 // caller needs to make sure that there is room to move the RETADDR to. This is
2719 // achieved by reserving an area the size of the argument delta right after the
2720 // original REtADDR, but before the saved framepointer or the spilled registers
2721 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2733 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2734 /// for a 16 byte align requirement.
2736 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2737 SelectionDAG& DAG) const {
2738 MachineFunction &MF = DAG.getMachineFunction();
2739 const TargetMachine &TM = MF.getTarget();
2740 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2741 unsigned StackAlignment = TFI.getStackAlignment();
2742 uint64_t AlignMask = StackAlignment - 1;
2743 int64_t Offset = StackSize;
2744 unsigned SlotSize = RegInfo->getSlotSize();
2745 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2746 // Number smaller than 12 so just add the difference.
2747 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2749 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2750 Offset = ((~AlignMask) & Offset) + StackAlignment +
2751 (StackAlignment-SlotSize);
2756 /// MatchingStackOffset - Return true if the given stack call argument is
2757 /// already available in the same position (relatively) of the caller's
2758 /// incoming argument stack.
2760 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2761 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2762 const X86InstrInfo *TII) {
2763 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2765 if (Arg.getOpcode() == ISD::CopyFromReg) {
2766 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2767 if (!TargetRegisterInfo::isVirtualRegister(VR))
2769 MachineInstr *Def = MRI->getVRegDef(VR);
2772 if (!Flags.isByVal()) {
2773 if (!TII->isLoadFromStackSlot(Def, FI))
2776 unsigned Opcode = Def->getOpcode();
2777 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2778 Def->getOperand(1).isFI()) {
2779 FI = Def->getOperand(1).getIndex();
2780 Bytes = Flags.getByValSize();
2784 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2785 if (Flags.isByVal())
2786 // ByVal argument is passed in as a pointer but it's now being
2787 // dereferenced. e.g.
2788 // define @foo(%struct.X* %A) {
2789 // tail call @bar(%struct.X* byval %A)
2792 SDValue Ptr = Ld->getBasePtr();
2793 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2796 FI = FINode->getIndex();
2797 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2798 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2799 FI = FINode->getIndex();
2800 Bytes = Flags.getByValSize();
2804 assert(FI != INT_MAX);
2805 if (!MFI->isFixedObjectIndex(FI))
2807 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2810 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2811 /// for tail call optimization. Targets which want to do tail call
2812 /// optimization should implement this function.
2814 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2815 CallingConv::ID CalleeCC,
2817 bool isCalleeStructRet,
2818 bool isCallerStructRet,
2820 const SmallVectorImpl<ISD::OutputArg> &Outs,
2821 const SmallVectorImpl<SDValue> &OutVals,
2822 const SmallVectorImpl<ISD::InputArg> &Ins,
2823 SelectionDAG &DAG) const {
2824 if (!IsTailCallConvention(CalleeCC) &&
2825 CalleeCC != CallingConv::C)
2828 // If -tailcallopt is specified, make fastcc functions tail-callable.
2829 const MachineFunction &MF = DAG.getMachineFunction();
2830 const Function *CallerF = DAG.getMachineFunction().getFunction();
2832 // If the function return type is x86_fp80 and the callee return type is not,
2833 // then the FP_EXTEND of the call result is not a nop. It's not safe to
2834 // perform a tailcall optimization here.
2835 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
2838 CallingConv::ID CallerCC = CallerF->getCallingConv();
2839 bool CCMatch = CallerCC == CalleeCC;
2841 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
2842 if (IsTailCallConvention(CalleeCC) && CCMatch)
2847 // Look for obvious safe cases to perform tail call optimization that do not
2848 // require ABI changes. This is what gcc calls sibcall.
2850 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2851 // emit a special epilogue.
2852 if (RegInfo->needsStackRealignment(MF))
2855 // Also avoid sibcall optimization if either caller or callee uses struct
2856 // return semantics.
2857 if (isCalleeStructRet || isCallerStructRet)
2860 // An stdcall caller is expected to clean up its arguments; the callee
2861 // isn't going to do that.
2862 if (!CCMatch && CallerCC == CallingConv::X86_StdCall)
2865 // Do not sibcall optimize vararg calls unless all arguments are passed via
2867 if (isVarArg && !Outs.empty()) {
2869 // Optimizing for varargs on Win64 is unlikely to be safe without
2870 // additional testing.
2871 if (Subtarget->isTargetWin64())
2874 SmallVector<CCValAssign, 16> ArgLocs;
2875 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2876 getTargetMachine(), ArgLocs, *DAG.getContext());
2878 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2879 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2880 if (!ArgLocs[i].isRegLoc())
2884 // If the call result is in ST0 / ST1, it needs to be popped off the x87
2885 // stack. Therefore, if it's not used by the call it is not safe to optimize
2886 // this into a sibcall.
2887 bool Unused = false;
2888 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2895 SmallVector<CCValAssign, 16> RVLocs;
2896 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2897 getTargetMachine(), RVLocs, *DAG.getContext());
2898 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2899 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2900 CCValAssign &VA = RVLocs[i];
2901 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2906 // If the calling conventions do not match, then we'd better make sure the
2907 // results are returned in the same way as what the caller expects.
2909 SmallVector<CCValAssign, 16> RVLocs1;
2910 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2911 getTargetMachine(), RVLocs1, *DAG.getContext());
2912 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2914 SmallVector<CCValAssign, 16> RVLocs2;
2915 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2916 getTargetMachine(), RVLocs2, *DAG.getContext());
2917 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2919 if (RVLocs1.size() != RVLocs2.size())
2921 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2922 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2924 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2926 if (RVLocs1[i].isRegLoc()) {
2927 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2930 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2936 // If the callee takes no arguments then go on to check the results of the
2938 if (!Outs.empty()) {
2939 // Check if stack adjustment is needed. For now, do not do this if any
2940 // argument is passed on the stack.
2941 SmallVector<CCValAssign, 16> ArgLocs;
2942 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2943 getTargetMachine(), ArgLocs, *DAG.getContext());
2945 // Allocate shadow area for Win64
2946 if (Subtarget->isTargetWin64()) {
2947 CCInfo.AllocateStack(32, 8);
2950 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2951 if (CCInfo.getNextStackOffset()) {
2952 MachineFunction &MF = DAG.getMachineFunction();
2953 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2956 // Check if the arguments are already laid out in the right way as
2957 // the caller's fixed stack objects.
2958 MachineFrameInfo *MFI = MF.getFrameInfo();
2959 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2960 const X86InstrInfo *TII =
2961 ((const X86TargetMachine&)getTargetMachine()).getInstrInfo();
2962 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2963 CCValAssign &VA = ArgLocs[i];
2964 SDValue Arg = OutVals[i];
2965 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2966 if (VA.getLocInfo() == CCValAssign::Indirect)
2968 if (!VA.isRegLoc()) {
2969 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2976 // If the tailcall address may be in a register, then make sure it's
2977 // possible to register allocate for it. In 32-bit, the call address can
2978 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2979 // callee-saved registers are restored. These happen to be the same
2980 // registers used to pass 'inreg' arguments so watch out for those.
2981 if (!Subtarget->is64Bit() &&
2982 ((!isa<GlobalAddressSDNode>(Callee) &&
2983 !isa<ExternalSymbolSDNode>(Callee)) ||
2984 getTargetMachine().getRelocationModel() == Reloc::PIC_)) {
2985 unsigned NumInRegs = 0;
2986 // In PIC we need an extra register to formulate the address computation
2988 unsigned MaxInRegs =
2989 (getTargetMachine().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
2991 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2992 CCValAssign &VA = ArgLocs[i];
2995 unsigned Reg = VA.getLocReg();
2998 case X86::EAX: case X86::EDX: case X86::ECX:
2999 if (++NumInRegs == MaxInRegs)
3011 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3012 const TargetLibraryInfo *libInfo) const {
3013 return X86::createFastISel(funcInfo, libInfo);
3016 //===----------------------------------------------------------------------===//
3017 // Other Lowering Hooks
3018 //===----------------------------------------------------------------------===//
3020 static bool MayFoldLoad(SDValue Op) {
3021 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3024 static bool MayFoldIntoStore(SDValue Op) {
3025 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3028 static bool isTargetShuffle(unsigned Opcode) {
3030 default: return false;
3031 case X86ISD::PSHUFD:
3032 case X86ISD::PSHUFHW:
3033 case X86ISD::PSHUFLW:
3035 case X86ISD::PALIGNR:
3036 case X86ISD::MOVLHPS:
3037 case X86ISD::MOVLHPD:
3038 case X86ISD::MOVHLPS:
3039 case X86ISD::MOVLPS:
3040 case X86ISD::MOVLPD:
3041 case X86ISD::MOVSHDUP:
3042 case X86ISD::MOVSLDUP:
3043 case X86ISD::MOVDDUP:
3046 case X86ISD::UNPCKL:
3047 case X86ISD::UNPCKH:
3048 case X86ISD::VPERMILP:
3049 case X86ISD::VPERM2X128:
3050 case X86ISD::VPERMI:
3055 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3056 SDValue V1, SelectionDAG &DAG) {
3058 default: llvm_unreachable("Unknown x86 shuffle node");
3059 case X86ISD::MOVSHDUP:
3060 case X86ISD::MOVSLDUP:
3061 case X86ISD::MOVDDUP:
3062 return DAG.getNode(Opc, dl, VT, V1);
3066 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3067 SDValue V1, unsigned TargetMask,
3068 SelectionDAG &DAG) {
3070 default: llvm_unreachable("Unknown x86 shuffle node");
3071 case X86ISD::PSHUFD:
3072 case X86ISD::PSHUFHW:
3073 case X86ISD::PSHUFLW:
3074 case X86ISD::VPERMILP:
3075 case X86ISD::VPERMI:
3076 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3080 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3081 SDValue V1, SDValue V2, unsigned TargetMask,
3082 SelectionDAG &DAG) {
3084 default: llvm_unreachable("Unknown x86 shuffle node");
3085 case X86ISD::PALIGNR:
3087 case X86ISD::VPERM2X128:
3088 return DAG.getNode(Opc, dl, VT, V1, V2,
3089 DAG.getConstant(TargetMask, MVT::i8));
3093 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
3094 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3096 default: llvm_unreachable("Unknown x86 shuffle node");
3097 case X86ISD::MOVLHPS:
3098 case X86ISD::MOVLHPD:
3099 case X86ISD::MOVHLPS:
3100 case X86ISD::MOVLPS:
3101 case X86ISD::MOVLPD:
3104 case X86ISD::UNPCKL:
3105 case X86ISD::UNPCKH:
3106 return DAG.getNode(Opc, dl, VT, V1, V2);
3110 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3111 MachineFunction &MF = DAG.getMachineFunction();
3112 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3113 int ReturnAddrIndex = FuncInfo->getRAIndex();
3115 if (ReturnAddrIndex == 0) {
3116 // Set up a frame object for the return address.
3117 unsigned SlotSize = RegInfo->getSlotSize();
3118 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
3120 FuncInfo->setRAIndex(ReturnAddrIndex);
3123 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3126 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3127 bool hasSymbolicDisplacement) {
3128 // Offset should fit into 32 bit immediate field.
3129 if (!isInt<32>(Offset))
3132 // If we don't have a symbolic displacement - we don't have any extra
3134 if (!hasSymbolicDisplacement)
3137 // FIXME: Some tweaks might be needed for medium code model.
3138 if (M != CodeModel::Small && M != CodeModel::Kernel)
3141 // For small code model we assume that latest object is 16MB before end of 31
3142 // bits boundary. We may also accept pretty large negative constants knowing
3143 // that all objects are in the positive half of address space.
3144 if (M == CodeModel::Small && Offset < 16*1024*1024)
3147 // For kernel code model we know that all object resist in the negative half
3148 // of 32bits address space. We may not accept negative offsets, since they may
3149 // be just off and we may accept pretty large positive ones.
3150 if (M == CodeModel::Kernel && Offset > 0)
3156 /// isCalleePop - Determines whether the callee is required to pop its
3157 /// own arguments. Callee pop is necessary to support tail calls.
3158 bool X86::isCalleePop(CallingConv::ID CallingConv,
3159 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3163 switch (CallingConv) {
3166 case CallingConv::X86_StdCall:
3168 case CallingConv::X86_FastCall:
3170 case CallingConv::X86_ThisCall:
3172 case CallingConv::Fast:
3174 case CallingConv::GHC:
3176 case CallingConv::HiPE:
3181 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3182 /// specific condition code, returning the condition code and the LHS/RHS of the
3183 /// comparison to make.
3184 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3185 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3187 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3188 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3189 // X > -1 -> X == 0, jump !sign.
3190 RHS = DAG.getConstant(0, RHS.getValueType());
3191 return X86::COND_NS;
3193 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3194 // X < 0 -> X == 0, jump on sign.
3197 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3199 RHS = DAG.getConstant(0, RHS.getValueType());
3200 return X86::COND_LE;
3204 switch (SetCCOpcode) {
3205 default: llvm_unreachable("Invalid integer condition!");
3206 case ISD::SETEQ: return X86::COND_E;
3207 case ISD::SETGT: return X86::COND_G;
3208 case ISD::SETGE: return X86::COND_GE;
3209 case ISD::SETLT: return X86::COND_L;
3210 case ISD::SETLE: return X86::COND_LE;
3211 case ISD::SETNE: return X86::COND_NE;
3212 case ISD::SETULT: return X86::COND_B;
3213 case ISD::SETUGT: return X86::COND_A;
3214 case ISD::SETULE: return X86::COND_BE;
3215 case ISD::SETUGE: return X86::COND_AE;
3219 // First determine if it is required or is profitable to flip the operands.
3221 // If LHS is a foldable load, but RHS is not, flip the condition.
3222 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3223 !ISD::isNON_EXTLoad(RHS.getNode())) {
3224 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3225 std::swap(LHS, RHS);
3228 switch (SetCCOpcode) {
3234 std::swap(LHS, RHS);
3238 // On a floating point condition, the flags are set as follows:
3240 // 0 | 0 | 0 | X > Y
3241 // 0 | 0 | 1 | X < Y
3242 // 1 | 0 | 0 | X == Y
3243 // 1 | 1 | 1 | unordered
3244 switch (SetCCOpcode) {
3245 default: llvm_unreachable("Condcode should be pre-legalized away");
3247 case ISD::SETEQ: return X86::COND_E;
3248 case ISD::SETOLT: // flipped
3250 case ISD::SETGT: return X86::COND_A;
3251 case ISD::SETOLE: // flipped
3253 case ISD::SETGE: return X86::COND_AE;
3254 case ISD::SETUGT: // flipped
3256 case ISD::SETLT: return X86::COND_B;
3257 case ISD::SETUGE: // flipped
3259 case ISD::SETLE: return X86::COND_BE;
3261 case ISD::SETNE: return X86::COND_NE;
3262 case ISD::SETUO: return X86::COND_P;
3263 case ISD::SETO: return X86::COND_NP;
3265 case ISD::SETUNE: return X86::COND_INVALID;
3269 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3270 /// code. Current x86 isa includes the following FP cmov instructions:
3271 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3272 static bool hasFPCMov(unsigned X86CC) {
3288 /// isFPImmLegal - Returns true if the target can instruction select the
3289 /// specified FP immediate natively. If false, the legalizer will
3290 /// materialize the FP immediate as a load from a constant pool.
3291 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3292 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3293 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3299 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3300 /// the specified range (L, H].
3301 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3302 return (Val < 0) || (Val >= Low && Val < Hi);
3305 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3306 /// specified value.
3307 static bool isUndefOrEqual(int Val, int CmpVal) {
3308 return (Val < 0 || Val == CmpVal);
3311 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3312 /// from position Pos and ending in Pos+Size, falls within the specified
3313 /// sequential range (L, L+Pos]. or is undef.
3314 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3315 unsigned Pos, unsigned Size, int Low) {
3316 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3317 if (!isUndefOrEqual(Mask[i], Low))
3322 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3323 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3324 /// the second operand.
3325 static bool isPSHUFDMask(ArrayRef<int> Mask, EVT VT) {
3326 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3327 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3328 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3329 return (Mask[0] < 2 && Mask[1] < 2);
3333 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3334 /// is suitable for input to PSHUFHW.
3335 static bool isPSHUFHWMask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3336 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3339 // Lower quadword copied in order or undef.
3340 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3343 // Upper quadword shuffled.
3344 for (unsigned i = 4; i != 8; ++i)
3345 if (!isUndefOrInRange(Mask[i], 4, 8))
3348 if (VT == MVT::v16i16) {
3349 // Lower quadword copied in order or undef.
3350 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3353 // Upper quadword shuffled.
3354 for (unsigned i = 12; i != 16; ++i)
3355 if (!isUndefOrInRange(Mask[i], 12, 16))
3362 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3363 /// is suitable for input to PSHUFLW.
3364 static bool isPSHUFLWMask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3365 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3368 // Upper quadword copied in order.
3369 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3372 // Lower quadword shuffled.
3373 for (unsigned i = 0; i != 4; ++i)
3374 if (!isUndefOrInRange(Mask[i], 0, 4))
3377 if (VT == MVT::v16i16) {
3378 // Upper quadword copied in order.
3379 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
3382 // Lower quadword shuffled.
3383 for (unsigned i = 8; i != 12; ++i)
3384 if (!isUndefOrInRange(Mask[i], 8, 12))
3391 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3392 /// is suitable for input to PALIGNR.
3393 static bool isPALIGNRMask(ArrayRef<int> Mask, EVT VT,
3394 const X86Subtarget *Subtarget) {
3395 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
3396 (VT.is256BitVector() && !Subtarget->hasInt256()))
3399 unsigned NumElts = VT.getVectorNumElements();
3400 unsigned NumLanes = VT.getSizeInBits()/128;
3401 unsigned NumLaneElts = NumElts/NumLanes;
3403 // Do not handle 64-bit element shuffles with palignr.
3404 if (NumLaneElts == 2)
3407 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
3409 for (i = 0; i != NumLaneElts; ++i) {
3414 // Lane is all undef, go to next lane
3415 if (i == NumLaneElts)
3418 int Start = Mask[i+l];
3420 // Make sure its in this lane in one of the sources
3421 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
3422 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
3425 // If not lane 0, then we must match lane 0
3426 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
3429 // Correct second source to be contiguous with first source
3430 if (Start >= (int)NumElts)
3431 Start -= NumElts - NumLaneElts;
3433 // Make sure we're shifting in the right direction.
3434 if (Start <= (int)(i+l))
3439 // Check the rest of the elements to see if they are consecutive.
3440 for (++i; i != NumLaneElts; ++i) {
3441 int Idx = Mask[i+l];
3443 // Make sure its in this lane
3444 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
3445 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
3448 // If not lane 0, then we must match lane 0
3449 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
3452 if (Idx >= (int)NumElts)
3453 Idx -= NumElts - NumLaneElts;
3455 if (!isUndefOrEqual(Idx, Start+i))
3464 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3465 /// the two vector operands have swapped position.
3466 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3467 unsigned NumElems) {
3468 for (unsigned i = 0; i != NumElems; ++i) {
3472 else if (idx < (int)NumElems)
3473 Mask[i] = idx + NumElems;
3475 Mask[i] = idx - NumElems;
3479 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3480 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
3481 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
3482 /// reverse of what x86 shuffles want.
3483 static bool isSHUFPMask(ArrayRef<int> Mask, EVT VT, bool HasFp256,
3484 bool Commuted = false) {
3485 if (!HasFp256 && VT.is256BitVector())
3488 unsigned NumElems = VT.getVectorNumElements();
3489 unsigned NumLanes = VT.getSizeInBits()/128;
3490 unsigned NumLaneElems = NumElems/NumLanes;
3492 if (NumLaneElems != 2 && NumLaneElems != 4)
3495 // VSHUFPSY divides the resulting vector into 4 chunks.
3496 // The sources are also splitted into 4 chunks, and each destination
3497 // chunk must come from a different source chunk.
3499 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3500 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3502 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3503 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3505 // VSHUFPDY divides the resulting vector into 4 chunks.
3506 // The sources are also splitted into 4 chunks, and each destination
3507 // chunk must come from a different source chunk.
3509 // SRC1 => X3 X2 X1 X0
3510 // SRC2 => Y3 Y2 Y1 Y0
3512 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
3514 unsigned HalfLaneElems = NumLaneElems/2;
3515 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
3516 for (unsigned i = 0; i != NumLaneElems; ++i) {
3517 int Idx = Mask[i+l];
3518 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
3519 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
3521 // For VSHUFPSY, the mask of the second half must be the same as the
3522 // first but with the appropriate offsets. This works in the same way as
3523 // VPERMILPS works with masks.
3524 if (NumElems != 8 || l == 0 || Mask[i] < 0)
3526 if (!isUndefOrEqual(Idx, Mask[i]+l))
3534 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3535 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3536 static bool isMOVHLPSMask(ArrayRef<int> Mask, EVT VT) {
3537 if (!VT.is128BitVector())
3540 unsigned NumElems = VT.getVectorNumElements();
3545 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3546 return isUndefOrEqual(Mask[0], 6) &&
3547 isUndefOrEqual(Mask[1], 7) &&
3548 isUndefOrEqual(Mask[2], 2) &&
3549 isUndefOrEqual(Mask[3], 3);
3552 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3553 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3555 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, EVT VT) {
3556 if (!VT.is128BitVector())
3559 unsigned NumElems = VT.getVectorNumElements();
3564 return isUndefOrEqual(Mask[0], 2) &&
3565 isUndefOrEqual(Mask[1], 3) &&
3566 isUndefOrEqual(Mask[2], 2) &&
3567 isUndefOrEqual(Mask[3], 3);
3570 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3571 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3572 static bool isMOVLPMask(ArrayRef<int> Mask, EVT VT) {
3573 if (!VT.is128BitVector())
3576 unsigned NumElems = VT.getVectorNumElements();
3578 if (NumElems != 2 && NumElems != 4)
3581 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3582 if (!isUndefOrEqual(Mask[i], i + NumElems))
3585 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
3586 if (!isUndefOrEqual(Mask[i], i))
3592 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3593 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3594 static bool isMOVLHPSMask(ArrayRef<int> Mask, EVT VT) {
3595 if (!VT.is128BitVector())
3598 unsigned NumElems = VT.getVectorNumElements();
3600 if (NumElems != 2 && NumElems != 4)
3603 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3604 if (!isUndefOrEqual(Mask[i], i))
3607 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3608 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
3615 // Some special combinations that can be optimized.
3618 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
3619 SelectionDAG &DAG) {
3620 MVT VT = SVOp->getValueType(0).getSimpleVT();
3621 DebugLoc dl = SVOp->getDebugLoc();
3623 if (VT != MVT::v8i32 && VT != MVT::v8f32)
3626 ArrayRef<int> Mask = SVOp->getMask();
3628 // These are the special masks that may be optimized.
3629 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
3630 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
3631 bool MatchEvenMask = true;
3632 bool MatchOddMask = true;
3633 for (int i=0; i<8; ++i) {
3634 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
3635 MatchEvenMask = false;
3636 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
3637 MatchOddMask = false;
3640 if (!MatchEvenMask && !MatchOddMask)
3643 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
3645 SDValue Op0 = SVOp->getOperand(0);
3646 SDValue Op1 = SVOp->getOperand(1);
3648 if (MatchEvenMask) {
3649 // Shift the second operand right to 32 bits.
3650 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
3651 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
3653 // Shift the first operand left to 32 bits.
3654 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
3655 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
3657 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
3658 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
3661 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3662 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3663 static bool isUNPCKLMask(ArrayRef<int> Mask, EVT VT,
3664 bool HasInt256, bool V2IsSplat = false) {
3665 unsigned NumElts = VT.getVectorNumElements();
3667 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3668 "Unsupported vector type for unpckh");
3670 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3671 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3674 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3675 // independently on 128-bit lanes.
3676 unsigned NumLanes = VT.getSizeInBits()/128;
3677 unsigned NumLaneElts = NumElts/NumLanes;
3679 for (unsigned l = 0; l != NumLanes; ++l) {
3680 for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
3681 i != (l+1)*NumLaneElts;
3684 int BitI1 = Mask[i+1];
3685 if (!isUndefOrEqual(BitI, j))
3688 if (!isUndefOrEqual(BitI1, NumElts))
3691 if (!isUndefOrEqual(BitI1, j + NumElts))
3700 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3701 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3702 static bool isUNPCKHMask(ArrayRef<int> Mask, EVT VT,
3703 bool HasInt256, bool V2IsSplat = false) {
3704 unsigned NumElts = VT.getVectorNumElements();
3706 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3707 "Unsupported vector type for unpckh");
3709 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3710 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3713 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3714 // independently on 128-bit lanes.
3715 unsigned NumLanes = VT.getSizeInBits()/128;
3716 unsigned NumLaneElts = NumElts/NumLanes;
3718 for (unsigned l = 0; l != NumLanes; ++l) {
3719 for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
3720 i != (l+1)*NumLaneElts; i += 2, ++j) {
3722 int BitI1 = Mask[i+1];
3723 if (!isUndefOrEqual(BitI, j))
3726 if (isUndefOrEqual(BitI1, NumElts))
3729 if (!isUndefOrEqual(BitI1, j+NumElts))
3737 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3738 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3740 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3741 unsigned NumElts = VT.getVectorNumElements();
3742 bool Is256BitVec = VT.is256BitVector();
3744 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3745 "Unsupported vector type for unpckh");
3747 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
3748 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3751 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3752 // FIXME: Need a better way to get rid of this, there's no latency difference
3753 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3754 // the former later. We should also remove the "_undef" special mask.
3755 if (NumElts == 4 && Is256BitVec)
3758 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3759 // independently on 128-bit lanes.
3760 unsigned NumLanes = VT.getSizeInBits()/128;
3761 unsigned NumLaneElts = NumElts/NumLanes;
3763 for (unsigned l = 0; l != NumLanes; ++l) {
3764 for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
3765 i != (l+1)*NumLaneElts;
3768 int BitI1 = Mask[i+1];
3770 if (!isUndefOrEqual(BitI, j))
3772 if (!isUndefOrEqual(BitI1, j))
3780 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3781 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3783 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, EVT VT, bool HasInt256) {
3784 unsigned NumElts = VT.getVectorNumElements();
3786 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3787 "Unsupported vector type for unpckh");
3789 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
3790 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
3793 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3794 // independently on 128-bit lanes.
3795 unsigned NumLanes = VT.getSizeInBits()/128;
3796 unsigned NumLaneElts = NumElts/NumLanes;
3798 for (unsigned l = 0; l != NumLanes; ++l) {
3799 for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
3800 i != (l+1)*NumLaneElts; i += 2, ++j) {
3802 int BitI1 = Mask[i+1];
3803 if (!isUndefOrEqual(BitI, j))
3805 if (!isUndefOrEqual(BitI1, j))
3812 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3813 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3814 /// MOVSD, and MOVD, i.e. setting the lowest element.
3815 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
3816 if (VT.getVectorElementType().getSizeInBits() < 32)
3818 if (!VT.is128BitVector())
3821 unsigned NumElts = VT.getVectorNumElements();
3823 if (!isUndefOrEqual(Mask[0], NumElts))
3826 for (unsigned i = 1; i != NumElts; ++i)
3827 if (!isUndefOrEqual(Mask[i], i))
3833 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
3834 /// as permutations between 128-bit chunks or halves. As an example: this
3836 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
3837 /// The first half comes from the second half of V1 and the second half from the
3838 /// the second half of V2.
3839 static bool isVPERM2X128Mask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
3840 if (!HasFp256 || !VT.is256BitVector())
3843 // The shuffle result is divided into half A and half B. In total the two
3844 // sources have 4 halves, namely: C, D, E, F. The final values of A and
3845 // B must come from C, D, E or F.
3846 unsigned HalfSize = VT.getVectorNumElements()/2;
3847 bool MatchA = false, MatchB = false;
3849 // Check if A comes from one of C, D, E, F.
3850 for (unsigned Half = 0; Half != 4; ++Half) {
3851 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
3857 // Check if B comes from one of C, D, E, F.
3858 for (unsigned Half = 0; Half != 4; ++Half) {
3859 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
3865 return MatchA && MatchB;
3868 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
3869 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
3870 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
3871 MVT VT = SVOp->getValueType(0).getSimpleVT();
3873 unsigned HalfSize = VT.getVectorNumElements()/2;
3875 unsigned FstHalf = 0, SndHalf = 0;
3876 for (unsigned i = 0; i < HalfSize; ++i) {
3877 if (SVOp->getMaskElt(i) > 0) {
3878 FstHalf = SVOp->getMaskElt(i)/HalfSize;
3882 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
3883 if (SVOp->getMaskElt(i) > 0) {
3884 SndHalf = SVOp->getMaskElt(i)/HalfSize;
3889 return (FstHalf | (SndHalf << 4));
3892 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
3893 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3894 /// Note that VPERMIL mask matching is different depending whether theunderlying
3895 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3896 /// to the same elements of the low, but to the higher half of the source.
3897 /// In VPERMILPD the two lanes could be shuffled independently of each other
3898 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
3899 static bool isVPERMILPMask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
3903 unsigned NumElts = VT.getVectorNumElements();
3904 // Only match 256-bit with 32/64-bit types
3905 if (!VT.is256BitVector() || (NumElts != 4 && NumElts != 8))
3908 unsigned NumLanes = VT.getSizeInBits()/128;
3909 unsigned LaneSize = NumElts/NumLanes;
3910 for (unsigned l = 0; l != NumElts; l += LaneSize) {
3911 for (unsigned i = 0; i != LaneSize; ++i) {
3912 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
3914 if (NumElts != 8 || l == 0)
3916 // VPERMILPS handling
3919 if (!isUndefOrEqual(Mask[i+l], Mask[i]+l))
3927 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
3928 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3929 /// element of vector 2 and the other elements to come from vector 1 in order.
3930 static bool isCommutedMOVLMask(ArrayRef<int> Mask, EVT VT,
3931 bool V2IsSplat = false, bool V2IsUndef = false) {
3932 if (!VT.is128BitVector())
3935 unsigned NumOps = VT.getVectorNumElements();
3936 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3939 if (!isUndefOrEqual(Mask[0], 0))
3942 for (unsigned i = 1; i != NumOps; ++i)
3943 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3944 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3945 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3951 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3952 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3953 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3954 static bool isMOVSHDUPMask(ArrayRef<int> Mask, EVT VT,
3955 const X86Subtarget *Subtarget) {
3956 if (!Subtarget->hasSSE3())
3959 unsigned NumElems = VT.getVectorNumElements();
3961 if ((VT.is128BitVector() && NumElems != 4) ||
3962 (VT.is256BitVector() && NumElems != 8))
3965 // "i+1" is the value the indexed mask element must have
3966 for (unsigned i = 0; i != NumElems; i += 2)
3967 if (!isUndefOrEqual(Mask[i], i+1) ||
3968 !isUndefOrEqual(Mask[i+1], i+1))
3974 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3975 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3976 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3977 static bool isMOVSLDUPMask(ArrayRef<int> Mask, EVT VT,
3978 const X86Subtarget *Subtarget) {
3979 if (!Subtarget->hasSSE3())
3982 unsigned NumElems = VT.getVectorNumElements();
3984 if ((VT.is128BitVector() && NumElems != 4) ||
3985 (VT.is256BitVector() && NumElems != 8))
3988 // "i" is the value the indexed mask element must have
3989 for (unsigned i = 0; i != NumElems; i += 2)
3990 if (!isUndefOrEqual(Mask[i], i) ||
3991 !isUndefOrEqual(Mask[i+1], i))
3997 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
3998 /// specifies a shuffle of elements that is suitable for input to 256-bit
3999 /// version of MOVDDUP.
4000 static bool isMOVDDUPYMask(ArrayRef<int> Mask, EVT VT, bool HasFp256) {
4001 if (!HasFp256 || !VT.is256BitVector())
4004 unsigned NumElts = VT.getVectorNumElements();
4008 for (unsigned i = 0; i != NumElts/2; ++i)
4009 if (!isUndefOrEqual(Mask[i], 0))
4011 for (unsigned i = NumElts/2; i != NumElts; ++i)
4012 if (!isUndefOrEqual(Mask[i], NumElts/2))
4017 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4018 /// specifies a shuffle of elements that is suitable for input to 128-bit
4019 /// version of MOVDDUP.
4020 static bool isMOVDDUPMask(ArrayRef<int> Mask, EVT VT) {
4021 if (!VT.is128BitVector())
4024 unsigned e = VT.getVectorNumElements() / 2;
4025 for (unsigned i = 0; i != e; ++i)
4026 if (!isUndefOrEqual(Mask[i], i))
4028 for (unsigned i = 0; i != e; ++i)
4029 if (!isUndefOrEqual(Mask[e+i], i))
4034 /// isVEXTRACTF128Index - Return true if the specified
4035 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4036 /// suitable for input to VEXTRACTF128.
4037 bool X86::isVEXTRACTF128Index(SDNode *N) {
4038 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4041 // The index should be aligned on a 128-bit boundary.
4043 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4045 MVT VT = N->getValueType(0).getSimpleVT();
4046 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4047 bool Result = (Index * ElSize) % 128 == 0;
4052 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
4053 /// operand specifies a subvector insert that is suitable for input to
4055 bool X86::isVINSERTF128Index(SDNode *N) {
4056 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4059 // The index should be aligned on a 128-bit boundary.
4061 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4063 MVT VT = N->getValueType(0).getSimpleVT();
4064 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4065 bool Result = (Index * ElSize) % 128 == 0;
4070 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4071 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4072 /// Handles 128-bit and 256-bit.
4073 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4074 MVT VT = N->getValueType(0).getSimpleVT();
4076 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4077 "Unsupported vector type for PSHUF/SHUFP");
4079 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4080 // independently on 128-bit lanes.
4081 unsigned NumElts = VT.getVectorNumElements();
4082 unsigned NumLanes = VT.getSizeInBits()/128;
4083 unsigned NumLaneElts = NumElts/NumLanes;
4085 assert((NumLaneElts == 2 || NumLaneElts == 4) &&
4086 "Only supports 2 or 4 elements per lane");
4088 unsigned Shift = (NumLaneElts == 4) ? 1 : 0;
4090 for (unsigned i = 0; i != NumElts; ++i) {
4091 int Elt = N->getMaskElt(i);
4092 if (Elt < 0) continue;
4093 Elt &= NumLaneElts - 1;
4094 unsigned ShAmt = (i << Shift) % 8;
4095 Mask |= Elt << ShAmt;
4101 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4102 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4103 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4104 MVT VT = N->getValueType(0).getSimpleVT();
4106 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4107 "Unsupported vector type for PSHUFHW");
4109 unsigned NumElts = VT.getVectorNumElements();
4112 for (unsigned l = 0; l != NumElts; l += 8) {
4113 // 8 nodes per lane, but we only care about the last 4.
4114 for (unsigned i = 0; i < 4; ++i) {
4115 int Elt = N->getMaskElt(l+i+4);
4116 if (Elt < 0) continue;
4117 Elt &= 0x3; // only 2-bits.
4118 Mask |= Elt << (i * 2);
4125 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4126 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4127 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4128 MVT VT = N->getValueType(0).getSimpleVT();
4130 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4131 "Unsupported vector type for PSHUFHW");
4133 unsigned NumElts = VT.getVectorNumElements();
4136 for (unsigned l = 0; l != NumElts; l += 8) {
4137 // 8 nodes per lane, but we only care about the first 4.
4138 for (unsigned i = 0; i < 4; ++i) {
4139 int Elt = N->getMaskElt(l+i);
4140 if (Elt < 0) continue;
4141 Elt &= 0x3; // only 2-bits
4142 Mask |= Elt << (i * 2);
4149 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4150 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4151 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4152 MVT VT = SVOp->getValueType(0).getSimpleVT();
4153 unsigned EltSize = VT.getVectorElementType().getSizeInBits() >> 3;
4155 unsigned NumElts = VT.getVectorNumElements();
4156 unsigned NumLanes = VT.getSizeInBits()/128;
4157 unsigned NumLaneElts = NumElts/NumLanes;
4161 for (i = 0; i != NumElts; ++i) {
4162 Val = SVOp->getMaskElt(i);
4166 if (Val >= (int)NumElts)
4167 Val -= NumElts - NumLaneElts;
4169 assert(Val - i > 0 && "PALIGNR imm should be positive");
4170 return (Val - i) * EltSize;
4173 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
4174 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4176 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
4177 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4178 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
4181 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4183 MVT VecVT = N->getOperand(0).getValueType().getSimpleVT();
4184 MVT ElVT = VecVT.getVectorElementType();
4186 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4187 return Index / NumElemsPerChunk;
4190 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
4191 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4193 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
4194 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4195 llvm_unreachable("Illegal insert subvector for VINSERTF128");
4198 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4200 MVT VecVT = N->getValueType(0).getSimpleVT();
4201 MVT ElVT = VecVT.getVectorElementType();
4203 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4204 return Index / NumElemsPerChunk;
4207 /// getShuffleCLImmediate - Return the appropriate immediate to shuffle
4208 /// the specified VECTOR_SHUFFLE mask with VPERMQ and VPERMPD instructions.
4209 /// Handles 256-bit.
4210 static unsigned getShuffleCLImmediate(ShuffleVectorSDNode *N) {
4211 MVT VT = N->getValueType(0).getSimpleVT();
4213 unsigned NumElts = VT.getVectorNumElements();
4215 assert((VT.is256BitVector() && NumElts == 4) &&
4216 "Unsupported vector type for VPERMQ/VPERMPD");
4219 for (unsigned i = 0; i != NumElts; ++i) {
4220 int Elt = N->getMaskElt(i);
4223 Mask |= Elt << (i*2);
4228 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4230 bool X86::isZeroNode(SDValue Elt) {
4231 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Elt))
4232 return CN->isNullValue();
4233 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4234 return CFP->getValueAPF().isPosZero();
4238 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4239 /// their permute mask.
4240 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4241 SelectionDAG &DAG) {
4242 MVT VT = SVOp->getValueType(0).getSimpleVT();
4243 unsigned NumElems = VT.getVectorNumElements();
4244 SmallVector<int, 8> MaskVec;
4246 for (unsigned i = 0; i != NumElems; ++i) {
4247 int Idx = SVOp->getMaskElt(i);
4249 if (Idx < (int)NumElems)
4254 MaskVec.push_back(Idx);
4256 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
4257 SVOp->getOperand(0), &MaskVec[0]);
4260 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4261 /// match movhlps. The lower half elements should come from upper half of
4262 /// V1 (and in order), and the upper half elements should come from the upper
4263 /// half of V2 (and in order).
4264 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, EVT VT) {
4265 if (!VT.is128BitVector())
4267 if (VT.getVectorNumElements() != 4)
4269 for (unsigned i = 0, e = 2; i != e; ++i)
4270 if (!isUndefOrEqual(Mask[i], i+2))
4272 for (unsigned i = 2; i != 4; ++i)
4273 if (!isUndefOrEqual(Mask[i], i+4))
4278 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4279 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4281 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4282 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4284 N = N->getOperand(0).getNode();
4285 if (!ISD::isNON_EXTLoad(N))
4288 *LD = cast<LoadSDNode>(N);
4292 // Test whether the given value is a vector value which will be legalized
4294 static bool WillBeConstantPoolLoad(SDNode *N) {
4295 if (N->getOpcode() != ISD::BUILD_VECTOR)
4298 // Check for any non-constant elements.
4299 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4300 switch (N->getOperand(i).getNode()->getOpcode()) {
4302 case ISD::ConstantFP:
4309 // Vectors of all-zeros and all-ones are materialized with special
4310 // instructions rather than being loaded.
4311 return !ISD::isBuildVectorAllZeros(N) &&
4312 !ISD::isBuildVectorAllOnes(N);
4315 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4316 /// match movlp{s|d}. The lower half elements should come from lower half of
4317 /// V1 (and in order), and the upper half elements should come from the upper
4318 /// half of V2 (and in order). And since V1 will become the source of the
4319 /// MOVLP, it must be either a vector load or a scalar load to vector.
4320 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4321 ArrayRef<int> Mask, EVT VT) {
4322 if (!VT.is128BitVector())
4325 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4327 // Is V2 is a vector load, don't do this transformation. We will try to use
4328 // load folding shufps op.
4329 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
4332 unsigned NumElems = VT.getVectorNumElements();
4334 if (NumElems != 2 && NumElems != 4)
4336 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4337 if (!isUndefOrEqual(Mask[i], i))
4339 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4340 if (!isUndefOrEqual(Mask[i], i+NumElems))
4345 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4347 static bool isSplatVector(SDNode *N) {
4348 if (N->getOpcode() != ISD::BUILD_VECTOR)
4351 SDValue SplatValue = N->getOperand(0);
4352 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4353 if (N->getOperand(i) != SplatValue)
4358 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4359 /// to an zero vector.
4360 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4361 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4362 SDValue V1 = N->getOperand(0);
4363 SDValue V2 = N->getOperand(1);
4364 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4365 for (unsigned i = 0; i != NumElems; ++i) {
4366 int Idx = N->getMaskElt(i);
4367 if (Idx >= (int)NumElems) {
4368 unsigned Opc = V2.getOpcode();
4369 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4371 if (Opc != ISD::BUILD_VECTOR ||
4372 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4374 } else if (Idx >= 0) {
4375 unsigned Opc = V1.getOpcode();
4376 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4378 if (Opc != ISD::BUILD_VECTOR ||
4379 !X86::isZeroNode(V1.getOperand(Idx)))
4386 /// getZeroVector - Returns a vector of specified type with all zero elements.
4388 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4389 SelectionDAG &DAG, DebugLoc dl) {
4390 assert(VT.isVector() && "Expected a vector type");
4392 // Always build SSE zero vectors as <4 x i32> bitcasted
4393 // to their dest type. This ensures they get CSE'd.
4395 if (VT.is128BitVector()) { // SSE
4396 if (Subtarget->hasSSE2()) { // SSE2
4397 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4398 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4400 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4401 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4403 } else if (VT.is256BitVector()) { // AVX
4404 if (Subtarget->hasInt256()) { // AVX2
4405 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4406 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4407 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
4409 // 256-bit logic and arithmetic instructions in AVX are all
4410 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4411 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4412 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4413 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
4416 llvm_unreachable("Unexpected vector type");
4418 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4421 /// getOnesVector - Returns a vector of specified type with all bits set.
4422 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4423 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4424 /// Then bitcast to their original type, ensuring they get CSE'd.
4425 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4427 assert(VT.isVector() && "Expected a vector type");
4429 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4431 if (VT.is256BitVector()) {
4432 if (HasInt256) { // AVX2
4433 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4434 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
4436 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4437 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4439 } else if (VT.is128BitVector()) {
4440 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4442 llvm_unreachable("Unexpected vector type");
4444 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4447 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4448 /// that point to V2 points to its first element.
4449 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
4450 for (unsigned i = 0; i != NumElems; ++i) {
4451 if (Mask[i] > (int)NumElems) {
4457 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4458 /// operation of specified width.
4459 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4461 unsigned NumElems = VT.getVectorNumElements();
4462 SmallVector<int, 8> Mask;
4463 Mask.push_back(NumElems);
4464 for (unsigned i = 1; i != NumElems; ++i)
4466 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4469 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4470 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4472 unsigned NumElems = VT.getVectorNumElements();
4473 SmallVector<int, 8> Mask;
4474 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4476 Mask.push_back(i + NumElems);
4478 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4481 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4482 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4484 unsigned NumElems = VT.getVectorNumElements();
4485 SmallVector<int, 8> Mask;
4486 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4487 Mask.push_back(i + Half);
4488 Mask.push_back(i + NumElems + Half);
4490 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4493 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4494 // a generic shuffle instruction because the target has no such instructions.
4495 // Generate shuffles which repeat i16 and i8 several times until they can be
4496 // represented by v4f32 and then be manipulated by target suported shuffles.
4497 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4498 EVT VT = V.getValueType();
4499 int NumElems = VT.getVectorNumElements();
4500 DebugLoc dl = V.getDebugLoc();
4502 while (NumElems > 4) {
4503 if (EltNo < NumElems/2) {
4504 V = getUnpackl(DAG, dl, VT, V, V);
4506 V = getUnpackh(DAG, dl, VT, V, V);
4507 EltNo -= NumElems/2;
4514 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4515 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4516 EVT VT = V.getValueType();
4517 DebugLoc dl = V.getDebugLoc();
4519 if (VT.is128BitVector()) {
4520 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4521 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4522 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4524 } else if (VT.is256BitVector()) {
4525 // To use VPERMILPS to splat scalars, the second half of indicies must
4526 // refer to the higher part, which is a duplication of the lower one,
4527 // because VPERMILPS can only handle in-lane permutations.
4528 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4529 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4531 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4532 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4535 llvm_unreachable("Vector size not supported");
4537 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4540 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4541 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4542 EVT SrcVT = SV->getValueType(0);
4543 SDValue V1 = SV->getOperand(0);
4544 DebugLoc dl = SV->getDebugLoc();
4546 int EltNo = SV->getSplatIndex();
4547 int NumElems = SrcVT.getVectorNumElements();
4548 bool Is256BitVec = SrcVT.is256BitVector();
4550 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
4551 "Unknown how to promote splat for type");
4553 // Extract the 128-bit part containing the splat element and update
4554 // the splat element index when it refers to the higher register.
4556 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
4557 if (EltNo >= NumElems/2)
4558 EltNo -= NumElems/2;
4561 // All i16 and i8 vector types can't be used directly by a generic shuffle
4562 // instruction because the target has no such instruction. Generate shuffles
4563 // which repeat i16 and i8 several times until they fit in i32, and then can
4564 // be manipulated by target suported shuffles.
4565 EVT EltVT = SrcVT.getVectorElementType();
4566 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4567 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4569 // Recreate the 256-bit vector and place the same 128-bit vector
4570 // into the low and high part. This is necessary because we want
4571 // to use VPERM* to shuffle the vectors
4573 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
4576 return getLegalSplat(DAG, V1, EltNo);
4579 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4580 /// vector of zero or undef vector. This produces a shuffle where the low
4581 /// element of V2 is swizzled into the zero/undef vector, landing at element
4582 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4583 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4585 const X86Subtarget *Subtarget,
4586 SelectionDAG &DAG) {
4587 EVT VT = V2.getValueType();
4589 ? getZeroVector(VT, Subtarget, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4590 unsigned NumElems = VT.getVectorNumElements();
4591 SmallVector<int, 16> MaskVec;
4592 for (unsigned i = 0; i != NumElems; ++i)
4593 // If this is the insertion idx, put the low elt of V2 here.
4594 MaskVec.push_back(i == Idx ? NumElems : i);
4595 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4598 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4599 /// target specific opcode. Returns true if the Mask could be calculated.
4600 /// Sets IsUnary to true if only uses one source.
4601 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4602 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4603 unsigned NumElems = VT.getVectorNumElements();
4607 switch(N->getOpcode()) {
4609 ImmN = N->getOperand(N->getNumOperands()-1);
4610 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4612 case X86ISD::UNPCKH:
4613 DecodeUNPCKHMask(VT, Mask);
4615 case X86ISD::UNPCKL:
4616 DecodeUNPCKLMask(VT, Mask);
4618 case X86ISD::MOVHLPS:
4619 DecodeMOVHLPSMask(NumElems, Mask);
4621 case X86ISD::MOVLHPS:
4622 DecodeMOVLHPSMask(NumElems, Mask);
4624 case X86ISD::PALIGNR:
4625 ImmN = N->getOperand(N->getNumOperands()-1);
4626 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4628 case X86ISD::PSHUFD:
4629 case X86ISD::VPERMILP:
4630 ImmN = N->getOperand(N->getNumOperands()-1);
4631 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4634 case X86ISD::PSHUFHW:
4635 ImmN = N->getOperand(N->getNumOperands()-1);
4636 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4639 case X86ISD::PSHUFLW:
4640 ImmN = N->getOperand(N->getNumOperands()-1);
4641 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4644 case X86ISD::VPERMI:
4645 ImmN = N->getOperand(N->getNumOperands()-1);
4646 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4650 case X86ISD::MOVSD: {
4651 // The index 0 always comes from the first element of the second source,
4652 // this is why MOVSS and MOVSD are used in the first place. The other
4653 // elements come from the other positions of the first source vector
4654 Mask.push_back(NumElems);
4655 for (unsigned i = 1; i != NumElems; ++i) {
4660 case X86ISD::VPERM2X128:
4661 ImmN = N->getOperand(N->getNumOperands()-1);
4662 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4663 if (Mask.empty()) return false;
4665 case X86ISD::MOVDDUP:
4666 case X86ISD::MOVLHPD:
4667 case X86ISD::MOVLPD:
4668 case X86ISD::MOVLPS:
4669 case X86ISD::MOVSHDUP:
4670 case X86ISD::MOVSLDUP:
4671 // Not yet implemented
4673 default: llvm_unreachable("unknown target shuffle node");
4679 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4680 /// element of the result of the vector shuffle.
4681 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4684 return SDValue(); // Limit search depth.
4686 SDValue V = SDValue(N, 0);
4687 EVT VT = V.getValueType();
4688 unsigned Opcode = V.getOpcode();
4690 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4691 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4692 int Elt = SV->getMaskElt(Index);
4695 return DAG.getUNDEF(VT.getVectorElementType());
4697 unsigned NumElems = VT.getVectorNumElements();
4698 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4699 : SV->getOperand(1);
4700 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4703 // Recurse into target specific vector shuffles to find scalars.
4704 if (isTargetShuffle(Opcode)) {
4705 MVT ShufVT = V.getValueType().getSimpleVT();
4706 unsigned NumElems = ShufVT.getVectorNumElements();
4707 SmallVector<int, 16> ShuffleMask;
4710 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4713 int Elt = ShuffleMask[Index];
4715 return DAG.getUNDEF(ShufVT.getVectorElementType());
4717 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4719 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4723 // Actual nodes that may contain scalar elements
4724 if (Opcode == ISD::BITCAST) {
4725 V = V.getOperand(0);
4726 EVT SrcVT = V.getValueType();
4727 unsigned NumElems = VT.getVectorNumElements();
4729 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4733 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4734 return (Index == 0) ? V.getOperand(0)
4735 : DAG.getUNDEF(VT.getVectorElementType());
4737 if (V.getOpcode() == ISD::BUILD_VECTOR)
4738 return V.getOperand(Index);
4743 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4744 /// shuffle operation which come from a consecutively from a zero. The
4745 /// search can start in two different directions, from left or right.
4747 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, unsigned NumElems,
4748 bool ZerosFromLeft, SelectionDAG &DAG) {
4750 for (i = 0; i != NumElems; ++i) {
4751 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4752 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
4753 if (!(Elt.getNode() &&
4754 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4761 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
4762 /// correspond consecutively to elements from one of the vector operands,
4763 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4765 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
4766 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
4767 unsigned NumElems, unsigned &OpNum) {
4768 bool SeenV1 = false;
4769 bool SeenV2 = false;
4771 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
4772 int Idx = SVOp->getMaskElt(i);
4773 // Ignore undef indicies
4777 if (Idx < (int)NumElems)
4782 // Only accept consecutive elements from the same vector
4783 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4787 OpNum = SeenV1 ? 0 : 1;
4791 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4792 /// logical left shift of a vector.
4793 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4794 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4795 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4796 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4797 false /* check zeros from right */, DAG);
4803 // Considering the elements in the mask that are not consecutive zeros,
4804 // check if they consecutively come from only one of the source vectors.
4806 // V1 = {X, A, B, C} 0
4808 // vector_shuffle V1, V2 <1, 2, 3, X>
4810 if (!isShuffleMaskConsecutive(SVOp,
4811 0, // Mask Start Index
4812 NumElems-NumZeros, // Mask End Index(exclusive)
4813 NumZeros, // Where to start looking in the src vector
4814 NumElems, // Number of elements in vector
4815 OpSrc)) // Which source operand ?
4820 ShVal = SVOp->getOperand(OpSrc);
4824 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4825 /// logical left shift of a vector.
4826 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4827 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4828 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4829 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4830 true /* check zeros from left */, DAG);
4836 // Considering the elements in the mask that are not consecutive zeros,
4837 // check if they consecutively come from only one of the source vectors.
4839 // 0 { A, B, X, X } = V2
4841 // vector_shuffle V1, V2 <X, X, 4, 5>
4843 if (!isShuffleMaskConsecutive(SVOp,
4844 NumZeros, // Mask Start Index
4845 NumElems, // Mask End Index(exclusive)
4846 0, // Where to start looking in the src vector
4847 NumElems, // Number of elements in vector
4848 OpSrc)) // Which source operand ?
4853 ShVal = SVOp->getOperand(OpSrc);
4857 /// isVectorShift - Returns true if the shuffle can be implemented as a
4858 /// logical left or right shift of a vector.
4859 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4860 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4861 // Although the logic below support any bitwidth size, there are no
4862 // shift instructions which handle more than 128-bit vectors.
4863 if (!SVOp->getValueType(0).is128BitVector())
4866 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4867 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4873 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4875 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4876 unsigned NumNonZero, unsigned NumZero,
4878 const X86Subtarget* Subtarget,
4879 const TargetLowering &TLI) {
4883 DebugLoc dl = Op.getDebugLoc();
4886 for (unsigned i = 0; i < 16; ++i) {
4887 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4888 if (ThisIsNonZero && First) {
4890 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4892 V = DAG.getUNDEF(MVT::v8i16);
4897 SDValue ThisElt(0, 0), LastElt(0, 0);
4898 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4899 if (LastIsNonZero) {
4900 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4901 MVT::i16, Op.getOperand(i-1));
4903 if (ThisIsNonZero) {
4904 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4905 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4906 ThisElt, DAG.getConstant(8, MVT::i8));
4908 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4912 if (ThisElt.getNode())
4913 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4914 DAG.getIntPtrConstant(i/2));
4918 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4921 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4923 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4924 unsigned NumNonZero, unsigned NumZero,
4926 const X86Subtarget* Subtarget,
4927 const TargetLowering &TLI) {
4931 DebugLoc dl = Op.getDebugLoc();
4934 for (unsigned i = 0; i < 8; ++i) {
4935 bool isNonZero = (NonZeros & (1 << i)) != 0;
4939 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4941 V = DAG.getUNDEF(MVT::v8i16);
4944 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4945 MVT::v8i16, V, Op.getOperand(i),
4946 DAG.getIntPtrConstant(i));
4953 /// getVShift - Return a vector logical shift node.
4955 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4956 unsigned NumBits, SelectionDAG &DAG,
4957 const TargetLowering &TLI, DebugLoc dl) {
4958 assert(VT.is128BitVector() && "Unknown type for VShift");
4959 EVT ShVT = MVT::v2i64;
4960 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4961 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4962 return DAG.getNode(ISD::BITCAST, dl, VT,
4963 DAG.getNode(Opc, dl, ShVT, SrcOp,
4964 DAG.getConstant(NumBits,
4965 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
4969 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4970 SelectionDAG &DAG) const {
4972 // Check if the scalar load can be widened into a vector load. And if
4973 // the address is "base + cst" see if the cst can be "absorbed" into
4974 // the shuffle mask.
4975 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4976 SDValue Ptr = LD->getBasePtr();
4977 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4979 EVT PVT = LD->getValueType(0);
4980 if (PVT != MVT::i32 && PVT != MVT::f32)
4985 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4986 FI = FINode->getIndex();
4988 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4989 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4990 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4991 Offset = Ptr.getConstantOperandVal(1);
4992 Ptr = Ptr.getOperand(0);
4997 // FIXME: 256-bit vector instructions don't require a strict alignment,
4998 // improve this code to support it better.
4999 unsigned RequiredAlign = VT.getSizeInBits()/8;
5000 SDValue Chain = LD->getChain();
5001 // Make sure the stack object alignment is at least 16 or 32.
5002 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5003 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5004 if (MFI->isFixedObjectIndex(FI)) {
5005 // Can't change the alignment. FIXME: It's possible to compute
5006 // the exact stack offset and reference FI + adjust offset instead.
5007 // If someone *really* cares about this. That's the way to implement it.
5010 MFI->setObjectAlignment(FI, RequiredAlign);
5014 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5015 // Ptr + (Offset & ~15).
5018 if ((Offset % RequiredAlign) & 3)
5020 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5022 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
5023 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5025 int EltNo = (Offset - StartOffset) >> 2;
5026 unsigned NumElems = VT.getVectorNumElements();
5028 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5029 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5030 LD->getPointerInfo().getWithOffset(StartOffset),
5031 false, false, false, 0);
5033 SmallVector<int, 8> Mask;
5034 for (unsigned i = 0; i != NumElems; ++i)
5035 Mask.push_back(EltNo);
5037 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5043 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
5044 /// vector of type 'VT', see if the elements can be replaced by a single large
5045 /// load which has the same value as a build_vector whose operands are 'elts'.
5047 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5049 /// FIXME: we'd also like to handle the case where the last elements are zero
5050 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5051 /// There's even a handy isZeroNode for that purpose.
5052 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
5053 DebugLoc &DL, SelectionDAG &DAG) {
5054 EVT EltVT = VT.getVectorElementType();
5055 unsigned NumElems = Elts.size();
5057 LoadSDNode *LDBase = NULL;
5058 unsigned LastLoadedElt = -1U;
5060 // For each element in the initializer, see if we've found a load or an undef.
5061 // If we don't find an initial load element, or later load elements are
5062 // non-consecutive, bail out.
5063 for (unsigned i = 0; i < NumElems; ++i) {
5064 SDValue Elt = Elts[i];
5066 if (!Elt.getNode() ||
5067 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5070 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5072 LDBase = cast<LoadSDNode>(Elt.getNode());
5076 if (Elt.getOpcode() == ISD::UNDEF)
5079 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5080 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
5085 // If we have found an entire vector of loads and undefs, then return a large
5086 // load of the entire vector width starting at the base pointer. If we found
5087 // consecutive loads for the low half, generate a vzext_load node.
5088 if (LastLoadedElt == NumElems - 1) {
5089 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
5090 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5091 LDBase->getPointerInfo(),
5092 LDBase->isVolatile(), LDBase->isNonTemporal(),
5093 LDBase->isInvariant(), 0);
5094 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5095 LDBase->getPointerInfo(),
5096 LDBase->isVolatile(), LDBase->isNonTemporal(),
5097 LDBase->isInvariant(), LDBase->getAlignment());
5099 if (NumElems == 4 && LastLoadedElt == 1 &&
5100 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5101 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5102 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5104 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, 2, MVT::i64,
5105 LDBase->getPointerInfo(),
5106 LDBase->getAlignment(),
5107 false/*isVolatile*/, true/*ReadMem*/,
5110 // Make sure the newly-created LOAD is in the same position as LDBase in
5111 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5112 // update uses of LDBase's output chain to use the TokenFactor.
5113 if (LDBase->hasAnyUseOfValue(1)) {
5114 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5115 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5116 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5117 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5118 SDValue(ResNode.getNode(), 1));
5121 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5126 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5127 /// to generate a splat value for the following cases:
5128 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5129 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5130 /// a scalar load, or a constant.
5131 /// The VBROADCAST node is returned when a pattern is found,
5132 /// or SDValue() otherwise.
5134 X86TargetLowering::LowerVectorBroadcast(SDValue Op, SelectionDAG &DAG) const {
5135 if (!Subtarget->hasFp256())
5138 MVT VT = Op.getValueType().getSimpleVT();
5139 DebugLoc dl = Op.getDebugLoc();
5141 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5142 "Unsupported vector type for broadcast.");
5147 switch (Op.getOpcode()) {
5149 // Unknown pattern found.
5152 case ISD::BUILD_VECTOR: {
5153 // The BUILD_VECTOR node must be a splat.
5154 if (!isSplatVector(Op.getNode()))
5157 Ld = Op.getOperand(0);
5158 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5159 Ld.getOpcode() == ISD::ConstantFP);
5161 // The suspected load node has several users. Make sure that all
5162 // of its users are from the BUILD_VECTOR node.
5163 // Constants may have multiple users.
5164 if (!ConstSplatVal && !Ld->hasNUsesOfValue(VT.getVectorNumElements(), 0))
5169 case ISD::VECTOR_SHUFFLE: {
5170 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5172 // Shuffles must have a splat mask where the first element is
5174 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5177 SDValue Sc = Op.getOperand(0);
5178 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5179 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5181 if (!Subtarget->hasInt256())
5184 // Use the register form of the broadcast instruction available on AVX2.
5185 if (VT.is256BitVector())
5186 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5187 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5190 Ld = Sc.getOperand(0);
5191 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5192 Ld.getOpcode() == ISD::ConstantFP);
5194 // The scalar_to_vector node and the suspected
5195 // load node must have exactly one user.
5196 // Constants may have multiple users.
5197 if (!ConstSplatVal && (!Sc.hasOneUse() || !Ld.hasOneUse()))
5203 bool Is256 = VT.is256BitVector();
5205 // Handle the broadcasting a single constant scalar from the constant pool
5206 // into a vector. On Sandybridge it is still better to load a constant vector
5207 // from the constant pool and not to broadcast it from a scalar.
5208 if (ConstSplatVal && Subtarget->hasInt256()) {
5209 EVT CVT = Ld.getValueType();
5210 assert(!CVT.isVector() && "Must not broadcast a vector type");
5211 unsigned ScalarSize = CVT.getSizeInBits();
5213 if (ScalarSize == 32 || (Is256 && ScalarSize == 64)) {
5214 const Constant *C = 0;
5215 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5216 C = CI->getConstantIntValue();
5217 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5218 C = CF->getConstantFPValue();
5220 assert(C && "Invalid constant type");
5222 SDValue CP = DAG.getConstantPool(C, getPointerTy());
5223 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5224 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5225 MachinePointerInfo::getConstantPool(),
5226 false, false, false, Alignment);
5228 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5232 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5233 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5235 // Handle AVX2 in-register broadcasts.
5236 if (!IsLoad && Subtarget->hasInt256() &&
5237 (ScalarSize == 32 || (Is256 && ScalarSize == 64)))
5238 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5240 // The scalar source must be a normal load.
5244 if (ScalarSize == 32 || (Is256 && ScalarSize == 64))
5245 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5247 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5248 // double since there is no vbroadcastsd xmm
5249 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5250 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5251 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5254 // Unsupported broadcast.
5259 X86TargetLowering::buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const {
5260 EVT VT = Op.getValueType();
5262 // Skip if insert_vec_elt is not supported.
5263 if (!isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5266 DebugLoc DL = Op.getDebugLoc();
5267 unsigned NumElems = Op.getNumOperands();
5271 SmallVector<unsigned, 4> InsertIndices;
5272 SmallVector<int, 8> Mask(NumElems, -1);
5274 for (unsigned i = 0; i != NumElems; ++i) {
5275 unsigned Opc = Op.getOperand(i).getOpcode();
5277 if (Opc == ISD::UNDEF)
5280 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5281 // Quit if more than 1 elements need inserting.
5282 if (InsertIndices.size() > 1)
5285 InsertIndices.push_back(i);
5289 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5290 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5292 // Quit if extracted from vector of different type.
5293 if (ExtractedFromVec.getValueType() != VT)
5296 // Quit if non-constant index.
5297 if (!isa<ConstantSDNode>(ExtIdx))
5300 if (VecIn1.getNode() == 0)
5301 VecIn1 = ExtractedFromVec;
5302 else if (VecIn1 != ExtractedFromVec) {
5303 if (VecIn2.getNode() == 0)
5304 VecIn2 = ExtractedFromVec;
5305 else if (VecIn2 != ExtractedFromVec)
5306 // Quit if more than 2 vectors to shuffle
5310 unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5312 if (ExtractedFromVec == VecIn1)
5314 else if (ExtractedFromVec == VecIn2)
5315 Mask[i] = Idx + NumElems;
5318 if (VecIn1.getNode() == 0)
5321 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5322 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5323 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5324 unsigned Idx = InsertIndices[i];
5325 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5326 DAG.getIntPtrConstant(Idx));
5333 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5334 DebugLoc dl = Op.getDebugLoc();
5336 MVT VT = Op.getValueType().getSimpleVT();
5337 MVT ExtVT = VT.getVectorElementType();
5338 unsigned NumElems = Op.getNumOperands();
5340 // Vectors containing all zeros can be matched by pxor and xorps later
5341 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5342 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5343 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5344 if (VT == MVT::v4i32 || VT == MVT::v8i32)
5347 return getZeroVector(VT, Subtarget, DAG, dl);
5350 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5351 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5352 // vpcmpeqd on 256-bit vectors.
5353 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5354 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5357 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5360 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
5361 if (Broadcast.getNode())
5364 unsigned EVTBits = ExtVT.getSizeInBits();
5366 unsigned NumZero = 0;
5367 unsigned NumNonZero = 0;
5368 unsigned NonZeros = 0;
5369 bool IsAllConstants = true;
5370 SmallSet<SDValue, 8> Values;
5371 for (unsigned i = 0; i < NumElems; ++i) {
5372 SDValue Elt = Op.getOperand(i);
5373 if (Elt.getOpcode() == ISD::UNDEF)
5376 if (Elt.getOpcode() != ISD::Constant &&
5377 Elt.getOpcode() != ISD::ConstantFP)
5378 IsAllConstants = false;
5379 if (X86::isZeroNode(Elt))
5382 NonZeros |= (1 << i);
5387 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5388 if (NumNonZero == 0)
5389 return DAG.getUNDEF(VT);
5391 // Special case for single non-zero, non-undef, element.
5392 if (NumNonZero == 1) {
5393 unsigned Idx = CountTrailingZeros_32(NonZeros);
5394 SDValue Item = Op.getOperand(Idx);
5396 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5397 // the value are obviously zero, truncate the value to i32 and do the
5398 // insertion that way. Only do this if the value is non-constant or if the
5399 // value is a constant being inserted into element 0. It is cheaper to do
5400 // a constant pool load than it is to do a movd + shuffle.
5401 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5402 (!IsAllConstants || Idx == 0)) {
5403 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5405 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5406 EVT VecVT = MVT::v4i32;
5407 unsigned VecElts = 4;
5409 // Truncate the value (which may itself be a constant) to i32, and
5410 // convert it to a vector with movd (S2V+shuffle to zero extend).
5411 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5412 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5413 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5415 // Now we have our 32-bit value zero extended in the low element of
5416 // a vector. If Idx != 0, swizzle it into place.
5418 SmallVector<int, 4> Mask;
5419 Mask.push_back(Idx);
5420 for (unsigned i = 1; i != VecElts; ++i)
5422 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
5425 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5429 // If we have a constant or non-constant insertion into the low element of
5430 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5431 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5432 // depending on what the source datatype is.
5435 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5437 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5438 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5439 if (VT.is256BitVector()) {
5440 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5441 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5442 Item, DAG.getIntPtrConstant(0));
5444 assert(VT.is128BitVector() && "Expected an SSE value type!");
5445 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5446 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5447 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5450 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5451 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5452 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5453 if (VT.is256BitVector()) {
5454 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5455 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5457 assert(VT.is128BitVector() && "Expected an SSE value type!");
5458 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5460 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5464 // Is it a vector logical left shift?
5465 if (NumElems == 2 && Idx == 1 &&
5466 X86::isZeroNode(Op.getOperand(0)) &&
5467 !X86::isZeroNode(Op.getOperand(1))) {
5468 unsigned NumBits = VT.getSizeInBits();
5469 return getVShift(true, VT,
5470 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5471 VT, Op.getOperand(1)),
5472 NumBits/2, DAG, *this, dl);
5475 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5478 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5479 // is a non-constant being inserted into an element other than the low one,
5480 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5481 // movd/movss) to move this into the low element, then shuffle it into
5483 if (EVTBits == 32) {
5484 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5486 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5487 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
5488 SmallVector<int, 8> MaskVec;
5489 for (unsigned i = 0; i != NumElems; ++i)
5490 MaskVec.push_back(i == Idx ? 0 : 1);
5491 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5495 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5496 if (Values.size() == 1) {
5497 if (EVTBits == 32) {
5498 // Instead of a shuffle like this:
5499 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5500 // Check if it's possible to issue this instead.
5501 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5502 unsigned Idx = CountTrailingZeros_32(NonZeros);
5503 SDValue Item = Op.getOperand(Idx);
5504 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5505 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5510 // A vector full of immediates; various special cases are already
5511 // handled, so this is best done with a single constant-pool load.
5515 // For AVX-length vectors, build the individual 128-bit pieces and use
5516 // shuffles to put them in place.
5517 if (VT.is256BitVector()) {
5518 SmallVector<SDValue, 32> V;
5519 for (unsigned i = 0; i != NumElems; ++i)
5520 V.push_back(Op.getOperand(i));
5522 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5524 // Build both the lower and upper subvector.
5525 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5526 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5529 // Recreate the wider vector with the lower and upper part.
5530 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5533 // Let legalizer expand 2-wide build_vectors.
5534 if (EVTBits == 64) {
5535 if (NumNonZero == 1) {
5536 // One half is zero or undef.
5537 unsigned Idx = CountTrailingZeros_32(NonZeros);
5538 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5539 Op.getOperand(Idx));
5540 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5545 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5546 if (EVTBits == 8 && NumElems == 16) {
5547 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5549 if (V.getNode()) return V;
5552 if (EVTBits == 16 && NumElems == 8) {
5553 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5555 if (V.getNode()) return V;
5558 // If element VT is == 32 bits, turn it into a number of shuffles.
5559 SmallVector<SDValue, 8> V(NumElems);
5560 if (NumElems == 4 && NumZero > 0) {
5561 for (unsigned i = 0; i < 4; ++i) {
5562 bool isZero = !(NonZeros & (1 << i));
5564 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5566 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5569 for (unsigned i = 0; i < 2; ++i) {
5570 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5573 V[i] = V[i*2]; // Must be a zero vector.
5576 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5579 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5582 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5587 bool Reverse1 = (NonZeros & 0x3) == 2;
5588 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5592 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5593 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5595 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5598 if (Values.size() > 1 && VT.is128BitVector()) {
5599 // Check for a build vector of consecutive loads.
5600 for (unsigned i = 0; i < NumElems; ++i)
5601 V[i] = Op.getOperand(i);
5603 // Check for elements which are consecutive loads.
5604 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5608 // Check for a build vector from mostly shuffle plus few inserting.
5609 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5613 // For SSE 4.1, use insertps to put the high elements into the low element.
5614 if (getSubtarget()->hasSSE41()) {
5616 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5617 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5619 Result = DAG.getUNDEF(VT);
5621 for (unsigned i = 1; i < NumElems; ++i) {
5622 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5623 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5624 Op.getOperand(i), DAG.getIntPtrConstant(i));
5629 // Otherwise, expand into a number of unpckl*, start by extending each of
5630 // our (non-undef) elements to the full vector width with the element in the
5631 // bottom slot of the vector (which generates no code for SSE).
5632 for (unsigned i = 0; i < NumElems; ++i) {
5633 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5634 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5636 V[i] = DAG.getUNDEF(VT);
5639 // Next, we iteratively mix elements, e.g. for v4f32:
5640 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5641 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5642 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5643 unsigned EltStride = NumElems >> 1;
5644 while (EltStride != 0) {
5645 for (unsigned i = 0; i < EltStride; ++i) {
5646 // If V[i+EltStride] is undef and this is the first round of mixing,
5647 // then it is safe to just drop this shuffle: V[i] is already in the
5648 // right place, the one element (since it's the first round) being
5649 // inserted as undef can be dropped. This isn't safe for successive
5650 // rounds because they will permute elements within both vectors.
5651 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5652 EltStride == NumElems/2)
5655 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5664 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5665 // to create 256-bit vectors from two other 128-bit ones.
5666 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5667 DebugLoc dl = Op.getDebugLoc();
5668 MVT ResVT = Op.getValueType().getSimpleVT();
5670 assert(ResVT.is256BitVector() && "Value type must be 256-bit wide");
5672 SDValue V1 = Op.getOperand(0);
5673 SDValue V2 = Op.getOperand(1);
5674 unsigned NumElems = ResVT.getVectorNumElements();
5676 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5679 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5680 assert(Op.getNumOperands() == 2);
5682 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5683 // from two other 128-bit ones.
5684 return LowerAVXCONCAT_VECTORS(Op, DAG);
5687 // Try to lower a shuffle node into a simple blend instruction.
5689 LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
5690 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5691 SDValue V1 = SVOp->getOperand(0);
5692 SDValue V2 = SVOp->getOperand(1);
5693 DebugLoc dl = SVOp->getDebugLoc();
5694 MVT VT = SVOp->getValueType(0).getSimpleVT();
5695 MVT EltVT = VT.getVectorElementType();
5696 unsigned NumElems = VT.getVectorNumElements();
5698 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
5700 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
5703 // Check the mask for BLEND and build the value.
5704 unsigned MaskValue = 0;
5705 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
5706 unsigned NumLanes = (NumElems-1)/8 + 1;
5707 unsigned NumElemsInLane = NumElems / NumLanes;
5709 // Blend for v16i16 should be symetric for the both lanes.
5710 for (unsigned i = 0; i < NumElemsInLane; ++i) {
5712 int SndLaneEltIdx = (NumLanes == 2) ?
5713 SVOp->getMaskElt(i + NumElemsInLane) : -1;
5714 int EltIdx = SVOp->getMaskElt(i);
5716 if ((EltIdx < 0 || EltIdx == (int)i) &&
5717 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
5720 if (((unsigned)EltIdx == (i + NumElems)) &&
5721 (SndLaneEltIdx < 0 ||
5722 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
5723 MaskValue |= (1<<i);
5728 // Convert i32 vectors to floating point if it is not AVX2.
5729 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
5731 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
5732 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
5734 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
5735 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
5738 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
5739 DAG.getConstant(MaskValue, MVT::i32));
5740 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
5743 // v8i16 shuffles - Prefer shuffles in the following order:
5744 // 1. [all] pshuflw, pshufhw, optional move
5745 // 2. [ssse3] 1 x pshufb
5746 // 3. [ssse3] 2 x pshufb + 1 x por
5747 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5749 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
5750 SelectionDAG &DAG) {
5751 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5752 SDValue V1 = SVOp->getOperand(0);
5753 SDValue V2 = SVOp->getOperand(1);
5754 DebugLoc dl = SVOp->getDebugLoc();
5755 SmallVector<int, 8> MaskVals;
5757 // Determine if more than 1 of the words in each of the low and high quadwords
5758 // of the result come from the same quadword of one of the two inputs. Undef
5759 // mask values count as coming from any quadword, for better codegen.
5760 unsigned LoQuad[] = { 0, 0, 0, 0 };
5761 unsigned HiQuad[] = { 0, 0, 0, 0 };
5762 std::bitset<4> InputQuads;
5763 for (unsigned i = 0; i < 8; ++i) {
5764 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
5765 int EltIdx = SVOp->getMaskElt(i);
5766 MaskVals.push_back(EltIdx);
5775 InputQuads.set(EltIdx / 4);
5778 int BestLoQuad = -1;
5779 unsigned MaxQuad = 1;
5780 for (unsigned i = 0; i < 4; ++i) {
5781 if (LoQuad[i] > MaxQuad) {
5783 MaxQuad = LoQuad[i];
5787 int BestHiQuad = -1;
5789 for (unsigned i = 0; i < 4; ++i) {
5790 if (HiQuad[i] > MaxQuad) {
5792 MaxQuad = HiQuad[i];
5796 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5797 // of the two input vectors, shuffle them into one input vector so only a
5798 // single pshufb instruction is necessary. If There are more than 2 input
5799 // quads, disable the next transformation since it does not help SSSE3.
5800 bool V1Used = InputQuads[0] || InputQuads[1];
5801 bool V2Used = InputQuads[2] || InputQuads[3];
5802 if (Subtarget->hasSSSE3()) {
5803 if (InputQuads.count() == 2 && V1Used && V2Used) {
5804 BestLoQuad = InputQuads[0] ? 0 : 1;
5805 BestHiQuad = InputQuads[2] ? 2 : 3;
5807 if (InputQuads.count() > 2) {
5813 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5814 // the shuffle mask. If a quad is scored as -1, that means that it contains
5815 // words from all 4 input quadwords.
5817 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5819 BestLoQuad < 0 ? 0 : BestLoQuad,
5820 BestHiQuad < 0 ? 1 : BestHiQuad
5822 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5823 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5824 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5825 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5827 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5828 // source words for the shuffle, to aid later transformations.
5829 bool AllWordsInNewV = true;
5830 bool InOrder[2] = { true, true };
5831 for (unsigned i = 0; i != 8; ++i) {
5832 int idx = MaskVals[i];
5834 InOrder[i/4] = false;
5835 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5837 AllWordsInNewV = false;
5841 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5842 if (AllWordsInNewV) {
5843 for (int i = 0; i != 8; ++i) {
5844 int idx = MaskVals[i];
5847 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5848 if ((idx != i) && idx < 4)
5850 if ((idx != i) && idx > 3)
5859 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5860 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5861 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5862 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5863 unsigned TargetMask = 0;
5864 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5865 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5866 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5867 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
5868 getShufflePSHUFLWImmediate(SVOp);
5869 V1 = NewV.getOperand(0);
5870 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5874 // Promote splats to a larger type which usually leads to more efficient code.
5875 // FIXME: Is this true if pshufb is available?
5876 if (SVOp->isSplat())
5877 return PromoteSplat(SVOp, DAG);
5879 // If we have SSSE3, and all words of the result are from 1 input vector,
5880 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5881 // is present, fall back to case 4.
5882 if (Subtarget->hasSSSE3()) {
5883 SmallVector<SDValue,16> pshufbMask;
5885 // If we have elements from both input vectors, set the high bit of the
5886 // shuffle mask element to zero out elements that come from V2 in the V1
5887 // mask, and elements that come from V1 in the V2 mask, so that the two
5888 // results can be OR'd together.
5889 bool TwoInputs = V1Used && V2Used;
5890 for (unsigned i = 0; i != 8; ++i) {
5891 int EltIdx = MaskVals[i] * 2;
5892 int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
5893 int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
5894 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
5895 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
5897 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5898 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5899 DAG.getNode(ISD::BUILD_VECTOR, dl,
5900 MVT::v16i8, &pshufbMask[0], 16));
5902 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5904 // Calculate the shuffle mask for the second input, shuffle it, and
5905 // OR it with the first shuffled input.
5907 for (unsigned i = 0; i != 8; ++i) {
5908 int EltIdx = MaskVals[i] * 2;
5909 int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
5910 int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
5911 pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
5912 pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
5914 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5915 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5916 DAG.getNode(ISD::BUILD_VECTOR, dl,
5917 MVT::v16i8, &pshufbMask[0], 16));
5918 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5919 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5922 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5923 // and update MaskVals with new element order.
5924 std::bitset<8> InOrder;
5925 if (BestLoQuad >= 0) {
5926 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
5927 for (int i = 0; i != 4; ++i) {
5928 int idx = MaskVals[i];
5931 } else if ((idx / 4) == BestLoQuad) {
5936 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5939 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
5940 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5941 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5943 getShufflePSHUFLWImmediate(SVOp), DAG);
5947 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5948 // and update MaskVals with the new element order.
5949 if (BestHiQuad >= 0) {
5950 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
5951 for (unsigned i = 4; i != 8; ++i) {
5952 int idx = MaskVals[i];
5955 } else if ((idx / 4) == BestHiQuad) {
5956 MaskV[i] = (idx & 3) + 4;
5960 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5963 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3()) {
5964 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
5965 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5967 getShufflePSHUFHWImmediate(SVOp), DAG);
5971 // In case BestHi & BestLo were both -1, which means each quadword has a word
5972 // from each of the four input quadwords, calculate the InOrder bitvector now
5973 // before falling through to the insert/extract cleanup.
5974 if (BestLoQuad == -1 && BestHiQuad == -1) {
5976 for (int i = 0; i != 8; ++i)
5977 if (MaskVals[i] < 0 || MaskVals[i] == i)
5981 // The other elements are put in the right place using pextrw and pinsrw.
5982 for (unsigned i = 0; i != 8; ++i) {
5985 int EltIdx = MaskVals[i];
5988 SDValue ExtOp = (EltIdx < 8) ?
5989 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5990 DAG.getIntPtrConstant(EltIdx)) :
5991 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5992 DAG.getIntPtrConstant(EltIdx - 8));
5993 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5994 DAG.getIntPtrConstant(i));
5999 // v16i8 shuffles - Prefer shuffles in the following order:
6000 // 1. [ssse3] 1 x pshufb
6001 // 2. [ssse3] 2 x pshufb + 1 x por
6002 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
6004 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
6006 const X86TargetLowering &TLI) {
6007 SDValue V1 = SVOp->getOperand(0);
6008 SDValue V2 = SVOp->getOperand(1);
6009 DebugLoc dl = SVOp->getDebugLoc();
6010 ArrayRef<int> MaskVals = SVOp->getMask();
6012 // Promote splats to a larger type which usually leads to more efficient code.
6013 // FIXME: Is this true if pshufb is available?
6014 if (SVOp->isSplat())
6015 return PromoteSplat(SVOp, DAG);
6017 // If we have SSSE3, case 1 is generated when all result bytes come from
6018 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
6019 // present, fall back to case 3.
6021 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
6022 if (TLI.getSubtarget()->hasSSSE3()) {
6023 SmallVector<SDValue,16> pshufbMask;
6025 // If all result elements are from one input vector, then only translate
6026 // undef mask values to 0x80 (zero out result) in the pshufb mask.
6028 // Otherwise, we have elements from both input vectors, and must zero out
6029 // elements that come from V2 in the first mask, and V1 in the second mask
6030 // so that we can OR them together.
6031 for (unsigned i = 0; i != 16; ++i) {
6032 int EltIdx = MaskVals[i];
6033 if (EltIdx < 0 || EltIdx >= 16)
6035 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6037 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
6038 DAG.getNode(ISD::BUILD_VECTOR, dl,
6039 MVT::v16i8, &pshufbMask[0], 16));
6041 // As PSHUFB will zero elements with negative indices, it's safe to ignore
6042 // the 2nd operand if it's undefined or zero.
6043 if (V2.getOpcode() == ISD::UNDEF ||
6044 ISD::isBuildVectorAllZeros(V2.getNode()))
6047 // Calculate the shuffle mask for the second input, shuffle it, and
6048 // OR it with the first shuffled input.
6050 for (unsigned i = 0; i != 16; ++i) {
6051 int EltIdx = MaskVals[i];
6052 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
6053 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6055 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
6056 DAG.getNode(ISD::BUILD_VECTOR, dl,
6057 MVT::v16i8, &pshufbMask[0], 16));
6058 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
6061 // No SSSE3 - Calculate in place words and then fix all out of place words
6062 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
6063 // the 16 different words that comprise the two doublequadword input vectors.
6064 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
6065 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
6067 for (int i = 0; i != 8; ++i) {
6068 int Elt0 = MaskVals[i*2];
6069 int Elt1 = MaskVals[i*2+1];
6071 // This word of the result is all undef, skip it.
6072 if (Elt0 < 0 && Elt1 < 0)
6075 // This word of the result is already in the correct place, skip it.
6076 if ((Elt0 == i*2) && (Elt1 == i*2+1))
6079 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
6080 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
6083 // If Elt0 and Elt1 are defined, are consecutive, and can be load
6084 // using a single extract together, load it and store it.
6085 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
6086 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6087 DAG.getIntPtrConstant(Elt1 / 2));
6088 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6089 DAG.getIntPtrConstant(i));
6093 // If Elt1 is defined, extract it from the appropriate source. If the
6094 // source byte is not also odd, shift the extracted word left 8 bits
6095 // otherwise clear the bottom 8 bits if we need to do an or.
6097 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
6098 DAG.getIntPtrConstant(Elt1 / 2));
6099 if ((Elt1 & 1) == 0)
6100 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
6102 TLI.getShiftAmountTy(InsElt.getValueType())));
6104 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
6105 DAG.getConstant(0xFF00, MVT::i16));
6107 // If Elt0 is defined, extract it from the appropriate source. If the
6108 // source byte is not also even, shift the extracted word right 8 bits. If
6109 // Elt1 was also defined, OR the extracted values together before
6110 // inserting them in the result.
6112 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
6113 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
6114 if ((Elt0 & 1) != 0)
6115 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
6117 TLI.getShiftAmountTy(InsElt0.getValueType())));
6119 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
6120 DAG.getConstant(0x00FF, MVT::i16));
6121 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
6124 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
6125 DAG.getIntPtrConstant(i));
6127 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
6130 // v32i8 shuffles - Translate to VPSHUFB if possible.
6132 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
6133 const X86Subtarget *Subtarget,
6134 SelectionDAG &DAG) {
6135 MVT VT = SVOp->getValueType(0).getSimpleVT();
6136 SDValue V1 = SVOp->getOperand(0);
6137 SDValue V2 = SVOp->getOperand(1);
6138 DebugLoc dl = SVOp->getDebugLoc();
6139 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
6141 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6142 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
6143 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
6145 // VPSHUFB may be generated if
6146 // (1) one of input vector is undefined or zeroinitializer.
6147 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
6148 // And (2) the mask indexes don't cross the 128-bit lane.
6149 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
6150 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
6153 if (V1IsAllZero && !V2IsAllZero) {
6154 CommuteVectorShuffleMask(MaskVals, 32);
6157 SmallVector<SDValue, 32> pshufbMask;
6158 for (unsigned i = 0; i != 32; i++) {
6159 int EltIdx = MaskVals[i];
6160 if (EltIdx < 0 || EltIdx >= 32)
6163 if ((EltIdx >= 16 && i < 16) || (EltIdx < 16 && i >= 16))
6164 // Cross lane is not allowed.
6168 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
6170 return DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, V1,
6171 DAG.getNode(ISD::BUILD_VECTOR, dl,
6172 MVT::v32i8, &pshufbMask[0], 32));
6175 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
6176 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
6177 /// done when every pair / quad of shuffle mask elements point to elements in
6178 /// the right sequence. e.g.
6179 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
6181 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
6182 SelectionDAG &DAG) {
6183 MVT VT = SVOp->getValueType(0).getSimpleVT();
6184 DebugLoc dl = SVOp->getDebugLoc();
6185 unsigned NumElems = VT.getVectorNumElements();
6188 switch (VT.SimpleTy) {
6189 default: llvm_unreachable("Unexpected!");
6190 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
6191 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
6192 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
6193 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
6194 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
6195 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
6198 SmallVector<int, 8> MaskVec;
6199 for (unsigned i = 0; i != NumElems; i += Scale) {
6201 for (unsigned j = 0; j != Scale; ++j) {
6202 int EltIdx = SVOp->getMaskElt(i+j);
6206 StartIdx = (EltIdx / Scale);
6207 if (EltIdx != (int)(StartIdx*Scale + j))
6210 MaskVec.push_back(StartIdx);
6213 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
6214 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
6215 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
6218 /// getVZextMovL - Return a zero-extending vector move low node.
6220 static SDValue getVZextMovL(MVT VT, EVT OpVT,
6221 SDValue SrcOp, SelectionDAG &DAG,
6222 const X86Subtarget *Subtarget, DebugLoc dl) {
6223 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
6224 LoadSDNode *LD = NULL;
6225 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
6226 LD = dyn_cast<LoadSDNode>(SrcOp);
6228 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
6230 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
6231 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
6232 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
6233 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
6234 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
6236 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
6237 return DAG.getNode(ISD::BITCAST, dl, VT,
6238 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6239 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6247 return DAG.getNode(ISD::BITCAST, dl, VT,
6248 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
6249 DAG.getNode(ISD::BITCAST, dl,
6253 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
6254 /// which could not be matched by any known target speficic shuffle
6256 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6258 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
6259 if (NewOp.getNode())
6262 MVT VT = SVOp->getValueType(0).getSimpleVT();
6264 unsigned NumElems = VT.getVectorNumElements();
6265 unsigned NumLaneElems = NumElems / 2;
6267 DebugLoc dl = SVOp->getDebugLoc();
6268 MVT EltVT = VT.getVectorElementType();
6269 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
6272 SmallVector<int, 16> Mask;
6273 for (unsigned l = 0; l < 2; ++l) {
6274 // Build a shuffle mask for the output, discovering on the fly which
6275 // input vectors to use as shuffle operands (recorded in InputUsed).
6276 // If building a suitable shuffle vector proves too hard, then bail
6277 // out with UseBuildVector set.
6278 bool UseBuildVector = false;
6279 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
6280 unsigned LaneStart = l * NumLaneElems;
6281 for (unsigned i = 0; i != NumLaneElems; ++i) {
6282 // The mask element. This indexes into the input.
6283 int Idx = SVOp->getMaskElt(i+LaneStart);
6285 // the mask element does not index into any input vector.
6290 // The input vector this mask element indexes into.
6291 int Input = Idx / NumLaneElems;
6293 // Turn the index into an offset from the start of the input vector.
6294 Idx -= Input * NumLaneElems;
6296 // Find or create a shuffle vector operand to hold this input.
6298 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
6299 if (InputUsed[OpNo] == Input)
6300 // This input vector is already an operand.
6302 if (InputUsed[OpNo] < 0) {
6303 // Create a new operand for this input vector.
6304 InputUsed[OpNo] = Input;
6309 if (OpNo >= array_lengthof(InputUsed)) {
6310 // More than two input vectors used! Give up on trying to create a
6311 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
6312 UseBuildVector = true;
6316 // Add the mask index for the new shuffle vector.
6317 Mask.push_back(Idx + OpNo * NumLaneElems);
6320 if (UseBuildVector) {
6321 SmallVector<SDValue, 16> SVOps;
6322 for (unsigned i = 0; i != NumLaneElems; ++i) {
6323 // The mask element. This indexes into the input.
6324 int Idx = SVOp->getMaskElt(i+LaneStart);
6326 SVOps.push_back(DAG.getUNDEF(EltVT));
6330 // The input vector this mask element indexes into.
6331 int Input = Idx / NumElems;
6333 // Turn the index into an offset from the start of the input vector.
6334 Idx -= Input * NumElems;
6336 // Extract the vector element by hand.
6337 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
6338 SVOp->getOperand(Input),
6339 DAG.getIntPtrConstant(Idx)));
6342 // Construct the output using a BUILD_VECTOR.
6343 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, &SVOps[0],
6345 } else if (InputUsed[0] < 0) {
6346 // No input vectors were used! The result is undefined.
6347 Output[l] = DAG.getUNDEF(NVT);
6349 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
6350 (InputUsed[0] % 2) * NumLaneElems,
6352 // If only one input was used, use an undefined vector for the other.
6353 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
6354 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
6355 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
6356 // At least one input vector was used. Create a new shuffle vector.
6357 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
6363 // Concatenate the result back
6364 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
6367 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
6368 /// 4 elements, and match them with several different shuffle types.
6370 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
6371 SDValue V1 = SVOp->getOperand(0);
6372 SDValue V2 = SVOp->getOperand(1);
6373 DebugLoc dl = SVOp->getDebugLoc();
6374 MVT VT = SVOp->getValueType(0).getSimpleVT();
6376 assert(VT.is128BitVector() && "Unsupported vector size");
6378 std::pair<int, int> Locs[4];
6379 int Mask1[] = { -1, -1, -1, -1 };
6380 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
6384 for (unsigned i = 0; i != 4; ++i) {
6385 int Idx = PermMask[i];
6387 Locs[i] = std::make_pair(-1, -1);
6389 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
6391 Locs[i] = std::make_pair(0, NumLo);
6395 Locs[i] = std::make_pair(1, NumHi);
6397 Mask1[2+NumHi] = Idx;
6403 if (NumLo <= 2 && NumHi <= 2) {
6404 // If no more than two elements come from either vector. This can be
6405 // implemented with two shuffles. First shuffle gather the elements.
6406 // The second shuffle, which takes the first shuffle as both of its
6407 // vector operands, put the elements into the right order.
6408 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6410 int Mask2[] = { -1, -1, -1, -1 };
6412 for (unsigned i = 0; i != 4; ++i)
6413 if (Locs[i].first != -1) {
6414 unsigned Idx = (i < 2) ? 0 : 4;
6415 Idx += Locs[i].first * 2 + Locs[i].second;
6419 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6422 if (NumLo == 3 || NumHi == 3) {
6423 // Otherwise, we must have three elements from one vector, call it X, and
6424 // one element from the other, call it Y. First, use a shufps to build an
6425 // intermediate vector with the one element from Y and the element from X
6426 // that will be in the same half in the final destination (the indexes don't
6427 // matter). Then, use a shufps to build the final vector, taking the half
6428 // containing the element from Y from the intermediate, and the other half
6431 // Normalize it so the 3 elements come from V1.
6432 CommuteVectorShuffleMask(PermMask, 4);
6436 // Find the element from V2.
6438 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6439 int Val = PermMask[HiIndex];
6446 Mask1[0] = PermMask[HiIndex];
6448 Mask1[2] = PermMask[HiIndex^1];
6450 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6453 Mask1[0] = PermMask[0];
6454 Mask1[1] = PermMask[1];
6455 Mask1[2] = HiIndex & 1 ? 6 : 4;
6456 Mask1[3] = HiIndex & 1 ? 4 : 6;
6457 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6460 Mask1[0] = HiIndex & 1 ? 2 : 0;
6461 Mask1[1] = HiIndex & 1 ? 0 : 2;
6462 Mask1[2] = PermMask[2];
6463 Mask1[3] = PermMask[3];
6468 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6471 // Break it into (shuffle shuffle_hi, shuffle_lo).
6472 int LoMask[] = { -1, -1, -1, -1 };
6473 int HiMask[] = { -1, -1, -1, -1 };
6475 int *MaskPtr = LoMask;
6476 unsigned MaskIdx = 0;
6479 for (unsigned i = 0; i != 4; ++i) {
6486 int Idx = PermMask[i];
6488 Locs[i] = std::make_pair(-1, -1);
6489 } else if (Idx < 4) {
6490 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6491 MaskPtr[LoIdx] = Idx;
6494 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6495 MaskPtr[HiIdx] = Idx;
6500 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6501 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6502 int MaskOps[] = { -1, -1, -1, -1 };
6503 for (unsigned i = 0; i != 4; ++i)
6504 if (Locs[i].first != -1)
6505 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
6506 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6509 static bool MayFoldVectorLoad(SDValue V) {
6510 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6511 V = V.getOperand(0);
6513 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6514 V = V.getOperand(0);
6515 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
6516 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
6517 // BUILD_VECTOR (load), undef
6518 V = V.getOperand(0);
6520 return MayFoldLoad(V);
6524 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
6525 EVT VT = Op.getValueType();
6527 // Canonizalize to v2f64.
6528 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6529 return DAG.getNode(ISD::BITCAST, dl, VT,
6530 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6535 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
6537 SDValue V1 = Op.getOperand(0);
6538 SDValue V2 = Op.getOperand(1);
6539 EVT VT = Op.getValueType();
6541 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6543 if (HasSSE2 && VT == MVT::v2f64)
6544 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6546 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
6547 return DAG.getNode(ISD::BITCAST, dl, VT,
6548 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
6549 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
6550 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
6554 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
6555 SDValue V1 = Op.getOperand(0);
6556 SDValue V2 = Op.getOperand(1);
6557 EVT VT = Op.getValueType();
6559 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6560 "unsupported shuffle type");
6562 if (V2.getOpcode() == ISD::UNDEF)
6566 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6570 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6571 SDValue V1 = Op.getOperand(0);
6572 SDValue V2 = Op.getOperand(1);
6573 EVT VT = Op.getValueType();
6574 unsigned NumElems = VT.getVectorNumElements();
6576 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6577 // operand of these instructions is only memory, so check if there's a
6578 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6580 bool CanFoldLoad = false;
6582 // Trivial case, when V2 comes from a load.
6583 if (MayFoldVectorLoad(V2))
6586 // When V1 is a load, it can be folded later into a store in isel, example:
6587 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6589 // (MOVLPSmr addr:$src1, VR128:$src2)
6590 // So, recognize this potential and also use MOVLPS or MOVLPD
6591 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6594 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6596 if (HasSSE2 && NumElems == 2)
6597 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6600 // If we don't care about the second element, proceed to use movss.
6601 if (SVOp->getMaskElt(1) != -1)
6602 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6605 // movl and movlp will both match v2i64, but v2i64 is never matched by
6606 // movl earlier because we make it strict to avoid messing with the movlp load
6607 // folding logic (see the code above getMOVLP call). Match it here then,
6608 // this is horrible, but will stay like this until we move all shuffle
6609 // matching to x86 specific nodes. Note that for the 1st condition all
6610 // types are matched with movsd.
6612 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
6613 // as to remove this logic from here, as much as possible
6614 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
6615 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6616 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6619 assert(VT != MVT::v4i32 && "unsupported shuffle type");
6621 // Invert the operand order and use SHUFPS to match it.
6622 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
6623 getShuffleSHUFImmediate(SVOp), DAG);
6626 // Reduce a vector shuffle to zext.
6628 X86TargetLowering::LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
6629 // PMOVZX is only available from SSE41.
6630 if (!Subtarget->hasSSE41())
6633 EVT VT = Op.getValueType();
6635 // Only AVX2 support 256-bit vector integer extending.
6636 if (!Subtarget->hasInt256() && VT.is256BitVector())
6639 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6640 DebugLoc DL = Op.getDebugLoc();
6641 SDValue V1 = Op.getOperand(0);
6642 SDValue V2 = Op.getOperand(1);
6643 unsigned NumElems = VT.getVectorNumElements();
6645 // Extending is an unary operation and the element type of the source vector
6646 // won't be equal to or larger than i64.
6647 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
6648 VT.getVectorElementType() == MVT::i64)
6651 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
6652 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
6653 while ((1U << Shift) < NumElems) {
6654 if (SVOp->getMaskElt(1U << Shift) == 1)
6657 // The maximal ratio is 8, i.e. from i8 to i64.
6662 // Check the shuffle mask.
6663 unsigned Mask = (1U << Shift) - 1;
6664 for (unsigned i = 0; i != NumElems; ++i) {
6665 int EltIdx = SVOp->getMaskElt(i);
6666 if ((i & Mask) != 0 && EltIdx != -1)
6668 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
6672 LLVMContext *Context = DAG.getContext();
6673 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
6674 EVT NeVT = EVT::getIntegerVT(*Context, NBits);
6675 EVT NVT = EVT::getVectorVT(*Context, NeVT, NumElems >> Shift);
6677 if (!isTypeLegal(NVT))
6680 // Simplify the operand as it's prepared to be fed into shuffle.
6681 unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
6682 if (V1.getOpcode() == ISD::BITCAST &&
6683 V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
6684 V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6686 .getOperand(0).getValueType().getSizeInBits() == SignificantBits) {
6687 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
6688 SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
6689 ConstantSDNode *CIdx =
6690 dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
6691 // If it's foldable, i.e. normal load with single use, we will let code
6692 // selection to fold it. Otherwise, we will short the conversion sequence.
6693 if (CIdx && CIdx->getZExtValue() == 0 &&
6694 (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
6695 if (V.getValueSizeInBits() > V1.getValueSizeInBits()) {
6696 // The "ext_vec_elt" node is wider than the result node.
6697 // In this case we should extract subvector from V.
6698 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
6699 unsigned Ratio = V.getValueSizeInBits() / V1.getValueSizeInBits();
6700 EVT FullVT = V.getValueType();
6701 EVT SubVecVT = EVT::getVectorVT(*Context,
6702 FullVT.getVectorElementType(),
6703 FullVT.getVectorNumElements()/Ratio);
6704 V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
6705 DAG.getIntPtrConstant(0));
6707 V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V);
6711 return DAG.getNode(ISD::BITCAST, DL, VT,
6712 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
6716 X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
6717 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6718 MVT VT = Op.getValueType().getSimpleVT();
6719 DebugLoc dl = Op.getDebugLoc();
6720 SDValue V1 = Op.getOperand(0);
6721 SDValue V2 = Op.getOperand(1);
6723 if (isZeroShuffle(SVOp))
6724 return getZeroVector(VT, Subtarget, DAG, dl);
6726 // Handle splat operations
6727 if (SVOp->isSplat()) {
6728 // Use vbroadcast whenever the splat comes from a foldable load
6729 SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
6730 if (Broadcast.getNode())
6734 // Check integer expanding shuffles.
6735 SDValue NewOp = LowerVectorIntExtend(Op, DAG);
6736 if (NewOp.getNode())
6739 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6741 if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
6742 VT == MVT::v16i16 || VT == MVT::v32i8) {
6743 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
6744 if (NewOp.getNode())
6745 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6746 } else if ((VT == MVT::v4i32 ||
6747 (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6748 // FIXME: Figure out a cleaner way to do this.
6749 // Try to make use of movq to zero out the top part.
6750 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6751 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
6752 if (NewOp.getNode()) {
6753 MVT NewVT = NewOp.getValueType().getSimpleVT();
6754 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
6755 NewVT, true, false))
6756 return getVZextMovL(VT, NewVT, NewOp.getOperand(0),
6757 DAG, Subtarget, dl);
6759 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6760 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
6761 if (NewOp.getNode()) {
6762 MVT NewVT = NewOp.getValueType().getSimpleVT();
6763 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
6764 return getVZextMovL(VT, NewVT, NewOp.getOperand(1),
6765 DAG, Subtarget, dl);
6773 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6774 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6775 SDValue V1 = Op.getOperand(0);
6776 SDValue V2 = Op.getOperand(1);
6777 MVT VT = Op.getValueType().getSimpleVT();
6778 DebugLoc dl = Op.getDebugLoc();
6779 unsigned NumElems = VT.getVectorNumElements();
6780 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6781 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6782 bool V1IsSplat = false;
6783 bool V2IsSplat = false;
6784 bool HasSSE2 = Subtarget->hasSSE2();
6785 bool HasFp256 = Subtarget->hasFp256();
6786 bool HasInt256 = Subtarget->hasInt256();
6787 MachineFunction &MF = DAG.getMachineFunction();
6788 bool OptForSize = MF.getFunction()->getAttributes().
6789 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
6791 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
6793 if (V1IsUndef && V2IsUndef)
6794 return DAG.getUNDEF(VT);
6796 assert(!V1IsUndef && "Op 1 of shuffle should not be undef");
6798 // Vector shuffle lowering takes 3 steps:
6800 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6801 // narrowing and commutation of operands should be handled.
6802 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6804 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6805 // so the shuffle can be broken into other shuffles and the legalizer can
6806 // try the lowering again.
6808 // The general idea is that no vector_shuffle operation should be left to
6809 // be matched during isel, all of them must be converted to a target specific
6812 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6813 // narrowing and commutation of operands should be handled. The actual code
6814 // doesn't include all of those, work in progress...
6815 SDValue NewOp = NormalizeVectorShuffle(Op, DAG);
6816 if (NewOp.getNode())
6819 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
6821 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6822 // unpckh_undef). Only use pshufd if speed is more important than size.
6823 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
6824 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
6825 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
6826 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
6828 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
6829 V2IsUndef && MayFoldVectorLoad(V1))
6830 return getMOVDDup(Op, dl, V1, DAG);
6832 if (isMOVHLPS_v_undef_Mask(M, VT))
6833 return getMOVHighToLow(Op, dl, DAG);
6835 // Use to match splats
6836 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
6837 (VT == MVT::v2f64 || VT == MVT::v2i64))
6838 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
6840 if (isPSHUFDMask(M, VT)) {
6841 // The actual implementation will match the mask in the if above and then
6842 // during isel it can match several different instructions, not only pshufd
6843 // as its name says, sad but true, emulate the behavior for now...
6844 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6845 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6847 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
6849 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6850 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6852 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
6853 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
6856 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
6860 // Check if this can be converted into a logical shift.
6861 bool isLeft = false;
6864 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6865 if (isShift && ShVal.hasOneUse()) {
6866 // If the shifted value has multiple uses, it may be cheaper to use
6867 // v_set0 + movlhps or movhlps, etc.
6868 MVT EltVT = VT.getVectorElementType();
6869 ShAmt *= EltVT.getSizeInBits();
6870 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6873 if (isMOVLMask(M, VT)) {
6874 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6875 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6876 if (!isMOVLPMask(M, VT)) {
6877 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6878 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6880 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6881 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6885 // FIXME: fold these into legal mask.
6886 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
6887 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6889 if (isMOVHLPSMask(M, VT))
6890 return getMOVHighToLow(Op, dl, DAG);
6892 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
6893 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6895 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
6896 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6898 if (isMOVLPMask(M, VT))
6899 return getMOVLP(Op, dl, DAG, HasSSE2);
6901 if (ShouldXformToMOVHLPS(M, VT) ||
6902 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
6903 return CommuteVectorShuffle(SVOp, DAG);
6906 // No better options. Use a vshldq / vsrldq.
6907 MVT EltVT = VT.getVectorElementType();
6908 ShAmt *= EltVT.getSizeInBits();
6909 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6912 bool Commuted = false;
6913 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6914 // 1,1,1,1 -> v8i16 though.
6915 V1IsSplat = isSplatVector(V1.getNode());
6916 V2IsSplat = isSplatVector(V2.getNode());
6918 // Canonicalize the splat or undef, if present, to be on the RHS.
6919 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
6920 CommuteVectorShuffleMask(M, NumElems);
6922 std::swap(V1IsSplat, V2IsSplat);
6926 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
6927 // Shuffling low element of v1 into undef, just return v1.
6930 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6931 // the instruction selector will not match, so get a canonical MOVL with
6932 // swapped operands to undo the commute.
6933 return getMOVL(DAG, dl, VT, V2, V1);
6936 if (isUNPCKLMask(M, VT, HasInt256))
6937 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6939 if (isUNPCKHMask(M, VT, HasInt256))
6940 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6943 // Normalize mask so all entries that point to V2 points to its first
6944 // element then try to match unpck{h|l} again. If match, return a
6945 // new vector_shuffle with the corrected mask.p
6946 SmallVector<int, 8> NewMask(M.begin(), M.end());
6947 NormalizeMask(NewMask, NumElems);
6948 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
6949 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6950 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
6951 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6955 // Commute is back and try unpck* again.
6956 // FIXME: this seems wrong.
6957 CommuteVectorShuffleMask(M, NumElems);
6959 std::swap(V1IsSplat, V2IsSplat);
6962 if (isUNPCKLMask(M, VT, HasInt256))
6963 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
6965 if (isUNPCKHMask(M, VT, HasInt256))
6966 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
6969 // Normalize the node to match x86 shuffle ops if needed
6970 if (!V2IsUndef && (isSHUFPMask(M, VT, HasFp256, /* Commuted */ true)))
6971 return CommuteVectorShuffle(SVOp, DAG);
6973 // The checks below are all present in isShuffleMaskLegal, but they are
6974 // inlined here right now to enable us to directly emit target specific
6975 // nodes, and remove one by one until they don't return Op anymore.
6977 if (isPALIGNRMask(M, VT, Subtarget))
6978 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
6979 getShufflePALIGNRImmediate(SVOp),
6982 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6983 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6984 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6985 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
6988 if (isPSHUFHWMask(M, VT, HasInt256))
6989 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6990 getShufflePSHUFHWImmediate(SVOp),
6993 if (isPSHUFLWMask(M, VT, HasInt256))
6994 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6995 getShufflePSHUFLWImmediate(SVOp),
6998 if (isSHUFPMask(M, VT, HasFp256))
6999 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
7000 getShuffleSHUFImmediate(SVOp), DAG);
7002 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
7003 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
7004 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
7005 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
7007 //===--------------------------------------------------------------------===//
7008 // Generate target specific nodes for 128 or 256-bit shuffles only
7009 // supported in the AVX instruction set.
7012 // Handle VMOVDDUPY permutations
7013 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
7014 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
7016 // Handle VPERMILPS/D* permutations
7017 if (isVPERMILPMask(M, VT, HasFp256)) {
7018 if (HasInt256 && VT == MVT::v8i32)
7019 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
7020 getShuffleSHUFImmediate(SVOp), DAG);
7021 return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
7022 getShuffleSHUFImmediate(SVOp), DAG);
7025 // Handle VPERM2F128/VPERM2I128 permutations
7026 if (isVPERM2X128Mask(M, VT, HasFp256))
7027 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
7028 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
7030 SDValue BlendOp = LowerVECTOR_SHUFFLEtoBlend(SVOp, Subtarget, DAG);
7031 if (BlendOp.getNode())
7034 if (V2IsUndef && HasInt256 && (VT == MVT::v8i32 || VT == MVT::v8f32)) {
7035 SmallVector<SDValue, 8> permclMask;
7036 for (unsigned i = 0; i != 8; ++i) {
7037 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MVT::i32));
7039 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32,
7041 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
7042 return DAG.getNode(X86ISD::VPERMV, dl, VT,
7043 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
7046 if (V2IsUndef && HasInt256 && (VT == MVT::v4i64 || VT == MVT::v4f64))
7047 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1,
7048 getShuffleCLImmediate(SVOp), DAG);
7050 //===--------------------------------------------------------------------===//
7051 // Since no target specific shuffle was selected for this generic one,
7052 // lower it into other known shuffles. FIXME: this isn't true yet, but
7053 // this is the plan.
7056 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
7057 if (VT == MVT::v8i16) {
7058 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
7059 if (NewOp.getNode())
7063 if (VT == MVT::v16i8) {
7064 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
7065 if (NewOp.getNode())
7069 if (VT == MVT::v32i8) {
7070 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
7071 if (NewOp.getNode())
7075 // Handle all 128-bit wide vectors with 4 elements, and match them with
7076 // several different shuffle types.
7077 if (NumElems == 4 && VT.is128BitVector())
7078 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
7080 // Handle general 256-bit shuffles
7081 if (VT.is256BitVector())
7082 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
7087 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7088 MVT VT = Op.getValueType().getSimpleVT();
7089 DebugLoc dl = Op.getDebugLoc();
7091 if (!Op.getOperand(0).getValueType().getSimpleVT().is128BitVector())
7094 if (VT.getSizeInBits() == 8) {
7095 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
7096 Op.getOperand(0), Op.getOperand(1));
7097 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7098 DAG.getValueType(VT));
7099 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7102 if (VT.getSizeInBits() == 16) {
7103 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7104 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
7106 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7107 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7108 DAG.getNode(ISD::BITCAST, dl,
7112 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
7113 Op.getOperand(0), Op.getOperand(1));
7114 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
7115 DAG.getValueType(VT));
7116 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7119 if (VT == MVT::f32) {
7120 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
7121 // the result back to FR32 register. It's only worth matching if the
7122 // result has a single use which is a store or a bitcast to i32. And in
7123 // the case of a store, it's not worth it if the index is a constant 0,
7124 // because a MOVSSmr can be used instead, which is smaller and faster.
7125 if (!Op.hasOneUse())
7127 SDNode *User = *Op.getNode()->use_begin();
7128 if ((User->getOpcode() != ISD::STORE ||
7129 (isa<ConstantSDNode>(Op.getOperand(1)) &&
7130 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
7131 (User->getOpcode() != ISD::BITCAST ||
7132 User->getValueType(0) != MVT::i32))
7134 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7135 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
7138 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
7141 if (VT == MVT::i32 || VT == MVT::i64) {
7142 // ExtractPS/pextrq works with constant index.
7143 if (isa<ConstantSDNode>(Op.getOperand(1)))
7150 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
7151 SelectionDAG &DAG) const {
7152 if (!isa<ConstantSDNode>(Op.getOperand(1)))
7155 SDValue Vec = Op.getOperand(0);
7156 MVT VecVT = Vec.getValueType().getSimpleVT();
7158 // If this is a 256-bit vector result, first extract the 128-bit vector and
7159 // then extract the element from the 128-bit vector.
7160 if (VecVT.is256BitVector()) {
7161 DebugLoc dl = Op.getNode()->getDebugLoc();
7162 unsigned NumElems = VecVT.getVectorNumElements();
7163 SDValue Idx = Op.getOperand(1);
7164 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7166 // Get the 128-bit vector.
7167 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
7169 if (IdxVal >= NumElems/2)
7170 IdxVal -= NumElems/2;
7171 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
7172 DAG.getConstant(IdxVal, MVT::i32));
7175 assert(VecVT.is128BitVector() && "Unexpected vector length");
7177 if (Subtarget->hasSSE41()) {
7178 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
7183 MVT VT = Op.getValueType().getSimpleVT();
7184 DebugLoc dl = Op.getDebugLoc();
7185 // TODO: handle v16i8.
7186 if (VT.getSizeInBits() == 16) {
7187 SDValue Vec = Op.getOperand(0);
7188 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7190 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
7191 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
7192 DAG.getNode(ISD::BITCAST, dl,
7195 // Transform it so it match pextrw which produces a 32-bit result.
7196 MVT EltVT = MVT::i32;
7197 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
7198 Op.getOperand(0), Op.getOperand(1));
7199 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
7200 DAG.getValueType(VT));
7201 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
7204 if (VT.getSizeInBits() == 32) {
7205 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7209 // SHUFPS the element to the lowest double word, then movss.
7210 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
7211 MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
7212 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7213 DAG.getUNDEF(VVT), Mask);
7214 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7215 DAG.getIntPtrConstant(0));
7218 if (VT.getSizeInBits() == 64) {
7219 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
7220 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
7221 // to match extract_elt for f64.
7222 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7226 // UNPCKHPD the element to the lowest double word, then movsd.
7227 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
7228 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
7229 int Mask[2] = { 1, -1 };
7230 MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
7231 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
7232 DAG.getUNDEF(VVT), Mask);
7233 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
7234 DAG.getIntPtrConstant(0));
7240 static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
7241 MVT VT = Op.getValueType().getSimpleVT();
7242 MVT EltVT = VT.getVectorElementType();
7243 DebugLoc dl = Op.getDebugLoc();
7245 SDValue N0 = Op.getOperand(0);
7246 SDValue N1 = Op.getOperand(1);
7247 SDValue N2 = Op.getOperand(2);
7249 if (!VT.is128BitVector())
7252 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
7253 isa<ConstantSDNode>(N2)) {
7255 if (VT == MVT::v8i16)
7256 Opc = X86ISD::PINSRW;
7257 else if (VT == MVT::v16i8)
7258 Opc = X86ISD::PINSRB;
7260 Opc = X86ISD::PINSRB;
7262 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
7264 if (N1.getValueType() != MVT::i32)
7265 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7266 if (N2.getValueType() != MVT::i32)
7267 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7268 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
7271 if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
7272 // Bits [7:6] of the constant are the source select. This will always be
7273 // zero here. The DAG Combiner may combine an extract_elt index into these
7274 // bits. For example (insert (extract, 3), 2) could be matched by putting
7275 // the '3' into bits [7:6] of X86ISD::INSERTPS.
7276 // Bits [5:4] of the constant are the destination select. This is the
7277 // value of the incoming immediate.
7278 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
7279 // combine either bitwise AND or insert of float 0.0 to set these bits.
7280 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
7281 // Create this as a scalar to vector..
7282 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
7283 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
7286 if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
7287 // PINSR* works with constant index.
7294 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
7295 MVT VT = Op.getValueType().getSimpleVT();
7296 MVT EltVT = VT.getVectorElementType();
7298 DebugLoc dl = Op.getDebugLoc();
7299 SDValue N0 = Op.getOperand(0);
7300 SDValue N1 = Op.getOperand(1);
7301 SDValue N2 = Op.getOperand(2);
7303 // If this is a 256-bit vector result, first extract the 128-bit vector,
7304 // insert the element into the extracted half and then place it back.
7305 if (VT.is256BitVector()) {
7306 if (!isa<ConstantSDNode>(N2))
7309 // Get the desired 128-bit vector half.
7310 unsigned NumElems = VT.getVectorNumElements();
7311 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7312 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
7314 // Insert the element into the desired half.
7315 bool Upper = IdxVal >= NumElems/2;
7316 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
7317 DAG.getConstant(Upper ? IdxVal-NumElems/2 : IdxVal, MVT::i32));
7319 // Insert the changed part back to the 256-bit vector
7320 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
7323 if (Subtarget->hasSSE41())
7324 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7326 if (EltVT == MVT::i8)
7329 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7330 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7331 // as its second argument.
7332 if (N1.getValueType() != MVT::i32)
7333 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7334 if (N2.getValueType() != MVT::i32)
7335 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7336 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7341 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
7342 LLVMContext *Context = DAG.getContext();
7343 DebugLoc dl = Op.getDebugLoc();
7344 MVT OpVT = Op.getValueType().getSimpleVT();
7346 // If this is a 256-bit vector result, first insert into a 128-bit
7347 // vector and then insert into the 256-bit vector.
7348 if (!OpVT.is128BitVector()) {
7349 // Insert into a 128-bit vector.
7350 EVT VT128 = EVT::getVectorVT(*Context,
7351 OpVT.getVectorElementType(),
7352 OpVT.getVectorNumElements() / 2);
7354 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7356 // Insert the 128-bit vector.
7357 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
7360 if (OpVT == MVT::v1i64 &&
7361 Op.getOperand(0).getValueType() == MVT::i64)
7362 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7364 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7365 assert(OpVT.is128BitVector() && "Expected an SSE type!");
7366 return DAG.getNode(ISD::BITCAST, dl, OpVT,
7367 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7370 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7371 // a simple subregister reference or explicit instructions to grab
7372 // upper bits of a vector.
7373 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7374 SelectionDAG &DAG) {
7375 if (Subtarget->hasFp256()) {
7376 DebugLoc dl = Op.getNode()->getDebugLoc();
7377 SDValue Vec = Op.getNode()->getOperand(0);
7378 SDValue Idx = Op.getNode()->getOperand(1);
7380 if (Op.getNode()->getValueType(0).is128BitVector() &&
7381 Vec.getNode()->getValueType(0).is256BitVector() &&
7382 isa<ConstantSDNode>(Idx)) {
7383 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7384 return Extract128BitVector(Vec, IdxVal, DAG, dl);
7390 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7391 // simple superregister reference or explicit instructions to insert
7392 // the upper bits of a vector.
7393 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
7394 SelectionDAG &DAG) {
7395 if (Subtarget->hasFp256()) {
7396 DebugLoc dl = Op.getNode()->getDebugLoc();
7397 SDValue Vec = Op.getNode()->getOperand(0);
7398 SDValue SubVec = Op.getNode()->getOperand(1);
7399 SDValue Idx = Op.getNode()->getOperand(2);
7401 if (Op.getNode()->getValueType(0).is256BitVector() &&
7402 SubVec.getNode()->getValueType(0).is128BitVector() &&
7403 isa<ConstantSDNode>(Idx)) {
7404 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
7405 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
7411 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7412 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7413 // one of the above mentioned nodes. It has to be wrapped because otherwise
7414 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7415 // be used to form addressing mode. These wrapped nodes will be selected
7418 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7419 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7421 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7423 unsigned char OpFlag = 0;
7424 unsigned WrapperKind = X86ISD::Wrapper;
7425 CodeModel::Model M = getTargetMachine().getCodeModel();
7427 if (Subtarget->isPICStyleRIPRel() &&
7428 (M == CodeModel::Small || M == CodeModel::Kernel))
7429 WrapperKind = X86ISD::WrapperRIP;
7430 else if (Subtarget->isPICStyleGOT())
7431 OpFlag = X86II::MO_GOTOFF;
7432 else if (Subtarget->isPICStyleStubPIC())
7433 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7435 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7437 CP->getOffset(), OpFlag);
7438 DebugLoc DL = CP->getDebugLoc();
7439 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7440 // With PIC, the address is actually $g + Offset.
7442 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7443 DAG.getNode(X86ISD::GlobalBaseReg,
7444 DebugLoc(), getPointerTy()),
7451 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7452 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7454 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7456 unsigned char OpFlag = 0;
7457 unsigned WrapperKind = X86ISD::Wrapper;
7458 CodeModel::Model M = getTargetMachine().getCodeModel();
7460 if (Subtarget->isPICStyleRIPRel() &&
7461 (M == CodeModel::Small || M == CodeModel::Kernel))
7462 WrapperKind = X86ISD::WrapperRIP;
7463 else if (Subtarget->isPICStyleGOT())
7464 OpFlag = X86II::MO_GOTOFF;
7465 else if (Subtarget->isPICStyleStubPIC())
7466 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7468 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7470 DebugLoc DL = JT->getDebugLoc();
7471 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7473 // With PIC, the address is actually $g + Offset.
7475 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7476 DAG.getNode(X86ISD::GlobalBaseReg,
7477 DebugLoc(), getPointerTy()),
7484 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7485 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7487 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7489 unsigned char OpFlag = 0;
7490 unsigned WrapperKind = X86ISD::Wrapper;
7491 CodeModel::Model M = getTargetMachine().getCodeModel();
7493 if (Subtarget->isPICStyleRIPRel() &&
7494 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7495 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7496 OpFlag = X86II::MO_GOTPCREL;
7497 WrapperKind = X86ISD::WrapperRIP;
7498 } else if (Subtarget->isPICStyleGOT()) {
7499 OpFlag = X86II::MO_GOT;
7500 } else if (Subtarget->isPICStyleStubPIC()) {
7501 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7502 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7503 OpFlag = X86II::MO_DARWIN_NONLAZY;
7506 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7508 DebugLoc DL = Op.getDebugLoc();
7509 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7511 // With PIC, the address is actually $g + Offset.
7512 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7513 !Subtarget->is64Bit()) {
7514 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7515 DAG.getNode(X86ISD::GlobalBaseReg,
7516 DebugLoc(), getPointerTy()),
7520 // For symbols that require a load from a stub to get the address, emit the
7522 if (isGlobalStubReference(OpFlag))
7523 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7524 MachinePointerInfo::getGOT(), false, false, false, 0);
7530 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7531 // Create the TargetBlockAddressAddress node.
7532 unsigned char OpFlags =
7533 Subtarget->ClassifyBlockAddressReference();
7534 CodeModel::Model M = getTargetMachine().getCodeModel();
7535 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
7536 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
7537 DebugLoc dl = Op.getDebugLoc();
7538 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
7541 if (Subtarget->isPICStyleRIPRel() &&
7542 (M == CodeModel::Small || M == CodeModel::Kernel))
7543 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7545 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7547 // With PIC, the address is actually $g + Offset.
7548 if (isGlobalRelativeToPICBase(OpFlags)) {
7549 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7550 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7558 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
7559 int64_t Offset, SelectionDAG &DAG) const {
7560 // Create the TargetGlobalAddress node, folding in the constant
7561 // offset if it is legal.
7562 unsigned char OpFlags =
7563 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
7564 CodeModel::Model M = getTargetMachine().getCodeModel();
7566 if (OpFlags == X86II::MO_NO_FLAG &&
7567 X86::isOffsetSuitableForCodeModel(Offset, M)) {
7568 // A direct static reference to a global.
7569 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
7572 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
7575 if (Subtarget->isPICStyleRIPRel() &&
7576 (M == CodeModel::Small || M == CodeModel::Kernel))
7577 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7579 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7581 // With PIC, the address is actually $g + Offset.
7582 if (isGlobalRelativeToPICBase(OpFlags)) {
7583 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7584 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7588 // For globals that require a load from a stub to get the address, emit the
7590 if (isGlobalStubReference(OpFlags))
7591 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7592 MachinePointerInfo::getGOT(), false, false, false, 0);
7594 // If there was a non-zero offset that we didn't fold, create an explicit
7597 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7598 DAG.getConstant(Offset, getPointerTy()));
7604 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
7605 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
7606 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
7607 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
7611 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
7612 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
7613 unsigned char OperandFlags, bool LocalDynamic = false) {
7614 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7615 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7616 DebugLoc dl = GA->getDebugLoc();
7617 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7618 GA->getValueType(0),
7622 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
7626 SDValue Ops[] = { Chain, TGA, *InFlag };
7627 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 3);
7629 SDValue Ops[] = { Chain, TGA };
7630 Chain = DAG.getNode(CallType, dl, NodeTys, Ops, 2);
7633 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
7634 MFI->setAdjustsStack(true);
7636 SDValue Flag = Chain.getValue(1);
7637 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
7640 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
7642 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7645 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
7646 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7647 DAG.getNode(X86ISD::GlobalBaseReg,
7648 DebugLoc(), PtrVT), InFlag);
7649 InFlag = Chain.getValue(1);
7651 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
7654 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
7656 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7658 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
7659 X86::RAX, X86II::MO_TLSGD);
7662 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
7666 DebugLoc dl = GA->getDebugLoc();
7668 // Get the start address of the TLS block for this module.
7669 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
7670 .getInfo<X86MachineFunctionInfo>();
7671 MFI->incNumLocalDynamicTLSAccesses();
7675 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
7676 X86II::MO_TLSLD, /*LocalDynamic=*/true);
7679 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7680 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT), InFlag);
7681 InFlag = Chain.getValue(1);
7682 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
7683 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
7686 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
7690 unsigned char OperandFlags = X86II::MO_DTPOFF;
7691 unsigned WrapperKind = X86ISD::Wrapper;
7692 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7693 GA->getValueType(0),
7694 GA->getOffset(), OperandFlags);
7695 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7697 // Add x@dtpoff with the base.
7698 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
7701 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
7702 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7703 const EVT PtrVT, TLSModel::Model model,
7704 bool is64Bit, bool isPIC) {
7705 DebugLoc dl = GA->getDebugLoc();
7707 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
7708 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
7709 is64Bit ? 257 : 256));
7711 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
7712 DAG.getIntPtrConstant(0),
7713 MachinePointerInfo(Ptr),
7714 false, false, false, 0);
7716 unsigned char OperandFlags = 0;
7717 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
7719 unsigned WrapperKind = X86ISD::Wrapper;
7720 if (model == TLSModel::LocalExec) {
7721 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
7722 } else if (model == TLSModel::InitialExec) {
7724 OperandFlags = X86II::MO_GOTTPOFF;
7725 WrapperKind = X86ISD::WrapperRIP;
7727 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
7730 llvm_unreachable("Unexpected model");
7733 // emit "addl x@ntpoff,%eax" (local exec)
7734 // or "addl x@indntpoff,%eax" (initial exec)
7735 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
7736 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7737 GA->getValueType(0),
7738 GA->getOffset(), OperandFlags);
7739 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7741 if (model == TLSModel::InitialExec) {
7742 if (isPIC && !is64Bit) {
7743 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
7744 DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), PtrVT),
7748 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
7749 MachinePointerInfo::getGOT(), false, false, false,
7753 // The address of the thread local variable is the add of the thread
7754 // pointer with the offset of the variable.
7755 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
7759 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
7761 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
7762 const GlobalValue *GV = GA->getGlobal();
7764 if (Subtarget->isTargetELF()) {
7765 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
7768 case TLSModel::GeneralDynamic:
7769 if (Subtarget->is64Bit())
7770 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
7771 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
7772 case TLSModel::LocalDynamic:
7773 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
7774 Subtarget->is64Bit());
7775 case TLSModel::InitialExec:
7776 case TLSModel::LocalExec:
7777 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
7778 Subtarget->is64Bit(),
7779 getTargetMachine().getRelocationModel() == Reloc::PIC_);
7781 llvm_unreachable("Unknown TLS model.");
7784 if (Subtarget->isTargetDarwin()) {
7785 // Darwin only has one model of TLS. Lower to that.
7786 unsigned char OpFlag = 0;
7787 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
7788 X86ISD::WrapperRIP : X86ISD::Wrapper;
7790 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7792 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
7793 !Subtarget->is64Bit();
7795 OpFlag = X86II::MO_TLVP_PIC_BASE;
7797 OpFlag = X86II::MO_TLVP;
7798 DebugLoc DL = Op.getDebugLoc();
7799 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7800 GA->getValueType(0),
7801 GA->getOffset(), OpFlag);
7802 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7804 // With PIC32, the address is actually $g + Offset.
7806 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7807 DAG.getNode(X86ISD::GlobalBaseReg,
7808 DebugLoc(), getPointerTy()),
7811 // Lowering the machine isd will make sure everything is in the right
7813 SDValue Chain = DAG.getEntryNode();
7814 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7815 SDValue Args[] = { Chain, Offset };
7816 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7818 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7819 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7820 MFI->setAdjustsStack(true);
7822 // And our return value (tls address) is in the standard call return value
7824 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7825 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
7829 if (Subtarget->isTargetWindows() || Subtarget->isTargetMingw()) {
7830 // Just use the implicit TLS architecture
7831 // Need to generate someting similar to:
7832 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
7834 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
7835 // mov rcx, qword [rdx+rcx*8]
7836 // mov eax, .tls$:tlsvar
7837 // [rax+rcx] contains the address
7838 // Windows 64bit: gs:0x58
7839 // Windows 32bit: fs:__tls_array
7841 // If GV is an alias then use the aliasee for determining
7842 // thread-localness.
7843 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
7844 GV = GA->resolveAliasedGlobal(false);
7845 DebugLoc dl = GA->getDebugLoc();
7846 SDValue Chain = DAG.getEntryNode();
7848 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
7849 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
7850 // use its literal value of 0x2C.
7851 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
7852 ? Type::getInt8PtrTy(*DAG.getContext(),
7854 : Type::getInt32PtrTy(*DAG.getContext(),
7857 SDValue TlsArray = Subtarget->is64Bit() ? DAG.getIntPtrConstant(0x58) :
7858 (Subtarget->isTargetMingw() ? DAG.getIntPtrConstant(0x2C) :
7859 DAG.getExternalSymbol("_tls_array", getPointerTy()));
7861 SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
7862 MachinePointerInfo(Ptr),
7863 false, false, false, 0);
7865 // Load the _tls_index variable
7866 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
7867 if (Subtarget->is64Bit())
7868 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
7869 IDX, MachinePointerInfo(), MVT::i32,
7872 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
7873 false, false, false, 0);
7875 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
7877 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
7879 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
7880 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
7881 false, false, false, 0);
7883 // Get the offset of start of .tls section
7884 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7885 GA->getValueType(0),
7886 GA->getOffset(), X86II::MO_SECREL);
7887 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
7889 // The address of the thread local variable is the add of the thread
7890 // pointer with the offset of the variable.
7891 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
7894 llvm_unreachable("TLS not implemented for this target.");
7897 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
7898 /// and take a 2 x i32 value to shift plus a shift amount.
7899 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
7900 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7901 EVT VT = Op.getValueType();
7902 unsigned VTBits = VT.getSizeInBits();
7903 DebugLoc dl = Op.getDebugLoc();
7904 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7905 SDValue ShOpLo = Op.getOperand(0);
7906 SDValue ShOpHi = Op.getOperand(1);
7907 SDValue ShAmt = Op.getOperand(2);
7908 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7909 DAG.getConstant(VTBits - 1, MVT::i8))
7910 : DAG.getConstant(0, VT);
7913 if (Op.getOpcode() == ISD::SHL_PARTS) {
7914 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7915 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7917 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7918 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7921 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7922 DAG.getConstant(VTBits, MVT::i8));
7923 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7924 AndNode, DAG.getConstant(0, MVT::i8));
7927 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7928 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7929 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7931 if (Op.getOpcode() == ISD::SHL_PARTS) {
7932 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7933 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7935 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7936 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7939 SDValue Ops[2] = { Lo, Hi };
7940 return DAG.getMergeValues(Ops, 2, dl);
7943 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7944 SelectionDAG &DAG) const {
7945 EVT SrcVT = Op.getOperand(0).getValueType();
7947 if (SrcVT.isVector())
7950 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7951 "Unknown SINT_TO_FP to lower!");
7953 // These are really Legal; return the operand so the caller accepts it as
7955 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7957 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7958 Subtarget->is64Bit()) {
7962 DebugLoc dl = Op.getDebugLoc();
7963 unsigned Size = SrcVT.getSizeInBits()/8;
7964 MachineFunction &MF = DAG.getMachineFunction();
7965 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7966 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7967 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7969 MachinePointerInfo::getFixedStack(SSFI),
7971 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7974 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7976 SelectionDAG &DAG) const {
7978 DebugLoc DL = Op.getDebugLoc();
7980 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7982 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7984 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7986 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7988 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7989 MachineMemOperand *MMO;
7991 int SSFI = FI->getIndex();
7993 DAG.getMachineFunction()
7994 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7995 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7997 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7998 StackSlot = StackSlot.getOperand(1);
8000 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
8001 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
8003 Tys, Ops, array_lengthof(Ops),
8007 Chain = Result.getValue(1);
8008 SDValue InFlag = Result.getValue(2);
8010 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
8011 // shouldn't be necessary except that RFP cannot be live across
8012 // multiple blocks. When stackifier is fixed, they can be uncoupled.
8013 MachineFunction &MF = DAG.getMachineFunction();
8014 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
8015 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
8016 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8017 Tys = DAG.getVTList(MVT::Other);
8019 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
8021 MachineMemOperand *MMO =
8022 DAG.getMachineFunction()
8023 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8024 MachineMemOperand::MOStore, SSFISize, SSFISize);
8026 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
8027 Ops, array_lengthof(Ops),
8028 Op.getValueType(), MMO);
8029 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
8030 MachinePointerInfo::getFixedStack(SSFI),
8031 false, false, false, 0);
8037 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
8038 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
8039 SelectionDAG &DAG) const {
8040 // This algorithm is not obvious. Here it is what we're trying to output:
8043 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
8044 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
8048 pshufd $0x4e, %xmm0, %xmm1
8053 DebugLoc dl = Op.getDebugLoc();
8054 LLVMContext *Context = DAG.getContext();
8056 // Build some magic constants.
8057 const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
8058 Constant *C0 = ConstantDataVector::get(*Context, CV0);
8059 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
8061 SmallVector<Constant*,2> CV1;
8063 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8064 APInt(64, 0x4330000000000000ULL))));
8066 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8067 APInt(64, 0x4530000000000000ULL))));
8068 Constant *C1 = ConstantVector::get(CV1);
8069 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
8071 // Load the 64-bit value into an XMM register.
8072 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
8074 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
8075 MachinePointerInfo::getConstantPool(),
8076 false, false, false, 16);
8077 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
8078 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
8081 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
8082 MachinePointerInfo::getConstantPool(),
8083 false, false, false, 16);
8084 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
8085 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
8088 if (Subtarget->hasSSE3()) {
8089 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
8090 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
8092 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
8093 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
8095 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
8096 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
8100 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
8101 DAG.getIntPtrConstant(0));
8104 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
8105 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
8106 SelectionDAG &DAG) const {
8107 DebugLoc dl = Op.getDebugLoc();
8108 // FP constant to bias correct the final result.
8109 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
8112 // Load the 32-bit value into an XMM register.
8113 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
8116 // Zero out the upper parts of the register.
8117 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
8119 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8120 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
8121 DAG.getIntPtrConstant(0));
8123 // Or the load with the bias.
8124 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
8125 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8126 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8128 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8129 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
8130 MVT::v2f64, Bias)));
8131 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8132 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
8133 DAG.getIntPtrConstant(0));
8135 // Subtract the bias.
8136 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
8138 // Handle final rounding.
8139 EVT DestVT = Op.getValueType();
8141 if (DestVT.bitsLT(MVT::f64))
8142 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
8143 DAG.getIntPtrConstant(0));
8144 if (DestVT.bitsGT(MVT::f64))
8145 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
8147 // Handle final rounding.
8151 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
8152 SelectionDAG &DAG) const {
8153 SDValue N0 = Op.getOperand(0);
8154 EVT SVT = N0.getValueType();
8155 DebugLoc dl = Op.getDebugLoc();
8157 assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
8158 SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
8159 "Custom UINT_TO_FP is not supported!");
8161 EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
8162 SVT.getVectorNumElements());
8163 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
8164 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
8167 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
8168 SelectionDAG &DAG) const {
8169 SDValue N0 = Op.getOperand(0);
8170 DebugLoc dl = Op.getDebugLoc();
8172 if (Op.getValueType().isVector())
8173 return lowerUINT_TO_FP_vec(Op, DAG);
8175 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
8176 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
8177 // the optimization here.
8178 if (DAG.SignBitIsZero(N0))
8179 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
8181 EVT SrcVT = N0.getValueType();
8182 EVT DstVT = Op.getValueType();
8183 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
8184 return LowerUINT_TO_FP_i64(Op, DAG);
8185 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
8186 return LowerUINT_TO_FP_i32(Op, DAG);
8187 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
8190 // Make a 64-bit buffer, and use it to build an FILD.
8191 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
8192 if (SrcVT == MVT::i32) {
8193 SDValue WordOff = DAG.getConstant(4, getPointerTy());
8194 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
8195 getPointerTy(), StackSlot, WordOff);
8196 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8197 StackSlot, MachinePointerInfo(),
8199 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
8200 OffsetSlot, MachinePointerInfo(),
8202 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
8206 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
8207 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
8208 StackSlot, MachinePointerInfo(),
8210 // For i64 source, we need to add the appropriate power of 2 if the input
8211 // was negative. This is the same as the optimization in
8212 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
8213 // we must be careful to do the computation in x87 extended precision, not
8214 // in SSE. (The generic code can't know it's OK to do this, or how to.)
8215 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
8216 MachineMemOperand *MMO =
8217 DAG.getMachineFunction()
8218 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8219 MachineMemOperand::MOLoad, 8, 8);
8221 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
8222 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
8223 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
8226 APInt FF(32, 0x5F800000ULL);
8228 // Check whether the sign bit is set.
8229 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
8230 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
8233 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
8234 SDValue FudgePtr = DAG.getConstantPool(
8235 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
8238 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
8239 SDValue Zero = DAG.getIntPtrConstant(0);
8240 SDValue Four = DAG.getIntPtrConstant(4);
8241 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
8243 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
8245 // Load the value out, extending it from f32 to f80.
8246 // FIXME: Avoid the extend by constructing the right constant pool?
8247 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
8248 FudgePtr, MachinePointerInfo::getConstantPool(),
8249 MVT::f32, false, false, 4);
8250 // Extend everything to 80 bits to force it to be done on x87.
8251 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
8252 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
8255 std::pair<SDValue,SDValue>
8256 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
8257 bool IsSigned, bool IsReplace) const {
8258 DebugLoc DL = Op.getDebugLoc();
8260 EVT DstTy = Op.getValueType();
8262 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
8263 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
8267 assert(DstTy.getSimpleVT() <= MVT::i64 &&
8268 DstTy.getSimpleVT() >= MVT::i16 &&
8269 "Unknown FP_TO_INT to lower!");
8271 // These are really Legal.
8272 if (DstTy == MVT::i32 &&
8273 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8274 return std::make_pair(SDValue(), SDValue());
8275 if (Subtarget->is64Bit() &&
8276 DstTy == MVT::i64 &&
8277 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
8278 return std::make_pair(SDValue(), SDValue());
8280 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
8281 // stack slot, or into the FTOL runtime function.
8282 MachineFunction &MF = DAG.getMachineFunction();
8283 unsigned MemSize = DstTy.getSizeInBits()/8;
8284 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8285 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8288 if (!IsSigned && isIntegerTypeFTOL(DstTy))
8289 Opc = X86ISD::WIN_FTOL;
8291 switch (DstTy.getSimpleVT().SimpleTy) {
8292 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
8293 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
8294 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
8295 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
8298 SDValue Chain = DAG.getEntryNode();
8299 SDValue Value = Op.getOperand(0);
8300 EVT TheVT = Op.getOperand(0).getValueType();
8301 // FIXME This causes a redundant load/store if the SSE-class value is already
8302 // in memory, such as if it is on the callstack.
8303 if (isScalarFPTypeInSSEReg(TheVT)) {
8304 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
8305 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
8306 MachinePointerInfo::getFixedStack(SSFI),
8308 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
8310 Chain, StackSlot, DAG.getValueType(TheVT)
8313 MachineMemOperand *MMO =
8314 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8315 MachineMemOperand::MOLoad, MemSize, MemSize);
8316 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
8318 Chain = Value.getValue(1);
8319 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
8320 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8323 MachineMemOperand *MMO =
8324 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8325 MachineMemOperand::MOStore, MemSize, MemSize);
8327 if (Opc != X86ISD::WIN_FTOL) {
8328 // Build the FP_TO_INT*_IN_MEM
8329 SDValue Ops[] = { Chain, Value, StackSlot };
8330 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
8331 Ops, 3, DstTy, MMO);
8332 return std::make_pair(FIST, StackSlot);
8334 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
8335 DAG.getVTList(MVT::Other, MVT::Glue),
8337 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
8338 MVT::i32, ftol.getValue(1));
8339 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
8340 MVT::i32, eax.getValue(2));
8341 SDValue Ops[] = { eax, edx };
8342 SDValue pair = IsReplace
8343 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops, 2)
8344 : DAG.getMergeValues(Ops, 2, DL);
8345 return std::make_pair(pair, SDValue());
8349 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
8350 const X86Subtarget *Subtarget) {
8351 MVT VT = Op->getValueType(0).getSimpleVT();
8352 SDValue In = Op->getOperand(0);
8353 MVT InVT = In.getValueType().getSimpleVT();
8354 DebugLoc dl = Op->getDebugLoc();
8356 // Optimize vectors in AVX mode:
8359 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
8360 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
8361 // Concat upper and lower parts.
8364 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
8365 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
8366 // Concat upper and lower parts.
8369 if (((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
8370 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
8373 if (Subtarget->hasInt256())
8374 return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
8376 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
8377 SDValue Undef = DAG.getUNDEF(InVT);
8378 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
8379 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8380 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
8382 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
8383 VT.getVectorNumElements()/2);
8385 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
8386 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
8388 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
8391 SDValue X86TargetLowering::LowerANY_EXTEND(SDValue Op,
8392 SelectionDAG &DAG) const {
8393 if (Subtarget->hasFp256()) {
8394 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8401 SDValue X86TargetLowering::LowerZERO_EXTEND(SDValue Op,
8402 SelectionDAG &DAG) const {
8403 DebugLoc DL = Op.getDebugLoc();
8404 MVT VT = Op.getValueType().getSimpleVT();
8405 SDValue In = Op.getOperand(0);
8406 MVT SVT = In.getValueType().getSimpleVT();
8408 if (Subtarget->hasFp256()) {
8409 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
8414 if (!VT.is256BitVector() || !SVT.is128BitVector() ||
8415 VT.getVectorNumElements() != SVT.getVectorNumElements())
8418 assert(Subtarget->hasFp256() && "256-bit vector is observed without AVX!");
8420 // AVX2 has better support of integer extending.
8421 if (Subtarget->hasInt256())
8422 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
8424 SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
8425 static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
8426 SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
8427 DAG.getVectorShuffle(MVT::v8i16, DL, In,
8428 DAG.getUNDEF(MVT::v8i16),
8431 return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
8434 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
8435 DebugLoc DL = Op.getDebugLoc();
8436 MVT VT = Op.getValueType().getSimpleVT();
8437 SDValue In = Op.getOperand(0);
8438 MVT SVT = In.getValueType().getSimpleVT();
8440 if ((VT == MVT::v4i32) && (SVT == MVT::v4i64)) {
8441 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
8442 if (Subtarget->hasInt256()) {
8443 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
8444 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
8445 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
8447 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
8448 DAG.getIntPtrConstant(0));
8451 // On AVX, v4i64 -> v4i32 becomes a sequence that uses PSHUFD and MOVLHPS.
8452 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8453 DAG.getIntPtrConstant(0));
8454 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8455 DAG.getIntPtrConstant(2));
8457 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
8458 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
8461 static const int ShufMask1[] = {0, 2, 0, 0};
8462 SDValue Undef = DAG.getUNDEF(VT);
8463 OpLo = DAG.getVectorShuffle(VT, DL, OpLo, Undef, ShufMask1);
8464 OpHi = DAG.getVectorShuffle(VT, DL, OpHi, Undef, ShufMask1);
8466 // The MOVLHPS mask:
8467 static const int ShufMask2[] = {0, 1, 4, 5};
8468 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask2);
8471 if ((VT == MVT::v8i16) && (SVT == MVT::v8i32)) {
8472 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
8473 if (Subtarget->hasInt256()) {
8474 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
8476 SmallVector<SDValue,32> pshufbMask;
8477 for (unsigned i = 0; i < 2; ++i) {
8478 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
8479 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
8480 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
8481 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
8482 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
8483 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
8484 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
8485 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
8486 for (unsigned j = 0; j < 8; ++j)
8487 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
8489 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8,
8490 &pshufbMask[0], 32);
8491 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
8492 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
8494 static const int ShufMask[] = {0, 2, -1, -1};
8495 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
8497 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
8498 DAG.getIntPtrConstant(0));
8499 return DAG.getNode(ISD::BITCAST, DL, VT, In);
8502 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
8503 DAG.getIntPtrConstant(0));
8505 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
8506 DAG.getIntPtrConstant(4));
8508 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
8509 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
8512 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
8513 -1, -1, -1, -1, -1, -1, -1, -1};
8515 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
8516 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
8517 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
8519 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
8520 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
8522 // The MOVLHPS Mask:
8523 static const int ShufMask2[] = {0, 1, 4, 5};
8524 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
8525 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
8528 // Handle truncation of V256 to V128 using shuffles.
8529 if (!VT.is128BitVector() || !SVT.is256BitVector())
8532 assert(VT.getVectorNumElements() != SVT.getVectorNumElements() &&
8534 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
8536 unsigned NumElems = VT.getVectorNumElements();
8537 EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
8540 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
8541 // Prepare truncation shuffle mask
8542 for (unsigned i = 0; i != NumElems; ++i)
8544 SDValue V = DAG.getVectorShuffle(NVT, DL,
8545 DAG.getNode(ISD::BITCAST, DL, NVT, In),
8546 DAG.getUNDEF(NVT), &MaskVec[0]);
8547 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
8548 DAG.getIntPtrConstant(0));
8551 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
8552 SelectionDAG &DAG) const {
8553 MVT VT = Op.getValueType().getSimpleVT();
8554 if (VT.isVector()) {
8555 if (VT == MVT::v8i16)
8556 return DAG.getNode(ISD::TRUNCATE, Op.getDebugLoc(), VT,
8557 DAG.getNode(ISD::FP_TO_SINT, Op.getDebugLoc(),
8558 MVT::v8i32, Op.getOperand(0)));
8562 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8563 /*IsSigned=*/ true, /*IsReplace=*/ false);
8564 SDValue FIST = Vals.first, StackSlot = Vals.second;
8565 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
8566 if (FIST.getNode() == 0) return Op;
8568 if (StackSlot.getNode())
8570 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
8571 FIST, StackSlot, MachinePointerInfo(),
8572 false, false, false, 0);
8574 // The node is the result.
8578 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
8579 SelectionDAG &DAG) const {
8580 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
8581 /*IsSigned=*/ false, /*IsReplace=*/ false);
8582 SDValue FIST = Vals.first, StackSlot = Vals.second;
8583 assert(FIST.getNode() && "Unexpected failure");
8585 if (StackSlot.getNode())
8587 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
8588 FIST, StackSlot, MachinePointerInfo(),
8589 false, false, false, 0);
8591 // The node is the result.
8595 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
8596 DebugLoc DL = Op.getDebugLoc();
8597 MVT VT = Op.getValueType().getSimpleVT();
8598 SDValue In = Op.getOperand(0);
8599 MVT SVT = In.getValueType().getSimpleVT();
8601 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
8603 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
8604 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
8605 In, DAG.getUNDEF(SVT)));
8608 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
8609 LLVMContext *Context = DAG.getContext();
8610 DebugLoc dl = Op.getDebugLoc();
8611 MVT VT = Op.getValueType().getSimpleVT();
8613 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
8614 if (VT.isVector()) {
8615 EltVT = VT.getVectorElementType();
8616 NumElts = VT.getVectorNumElements();
8619 if (EltVT == MVT::f64)
8620 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8621 APInt(64, ~(1ULL << 63))));
8623 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
8624 APInt(32, ~(1U << 31))));
8625 C = ConstantVector::getSplat(NumElts, C);
8626 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
8627 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8628 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8629 MachinePointerInfo::getConstantPool(),
8630 false, false, false, Alignment);
8631 if (VT.isVector()) {
8632 MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8633 return DAG.getNode(ISD::BITCAST, dl, VT,
8634 DAG.getNode(ISD::AND, dl, ANDVT,
8635 DAG.getNode(ISD::BITCAST, dl, ANDVT,
8637 DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
8639 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
8642 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
8643 LLVMContext *Context = DAG.getContext();
8644 DebugLoc dl = Op.getDebugLoc();
8645 MVT VT = Op.getValueType().getSimpleVT();
8647 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
8648 if (VT.isVector()) {
8649 EltVT = VT.getVectorElementType();
8650 NumElts = VT.getVectorNumElements();
8653 if (EltVT == MVT::f64)
8654 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
8655 APInt(64, 1ULL << 63)));
8657 C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
8658 APInt(32, 1U << 31)));
8659 C = ConstantVector::getSplat(NumElts, C);
8660 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
8661 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8662 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8663 MachinePointerInfo::getConstantPool(),
8664 false, false, false, Alignment);
8665 if (VT.isVector()) {
8666 MVT XORVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8667 return DAG.getNode(ISD::BITCAST, dl, VT,
8668 DAG.getNode(ISD::XOR, dl, XORVT,
8669 DAG.getNode(ISD::BITCAST, dl, XORVT,
8671 DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
8674 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
8677 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
8678 LLVMContext *Context = DAG.getContext();
8679 SDValue Op0 = Op.getOperand(0);
8680 SDValue Op1 = Op.getOperand(1);
8681 DebugLoc dl = Op.getDebugLoc();
8682 MVT VT = Op.getValueType().getSimpleVT();
8683 MVT SrcVT = Op1.getValueType().getSimpleVT();
8685 // If second operand is smaller, extend it first.
8686 if (SrcVT.bitsLT(VT)) {
8687 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
8690 // And if it is bigger, shrink it first.
8691 if (SrcVT.bitsGT(VT)) {
8692 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
8696 // At this point the operands and the result should have the same
8697 // type, and that won't be f80 since that is not custom lowered.
8699 // First get the sign bit of second operand.
8700 SmallVector<Constant*,4> CV;
8701 if (SrcVT == MVT::f64) {
8702 const fltSemantics &Sem = APFloat::IEEEdouble;
8703 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
8704 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
8706 const fltSemantics &Sem = APFloat::IEEEsingle;
8707 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
8708 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8709 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8710 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8712 Constant *C = ConstantVector::get(CV);
8713 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8714 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
8715 MachinePointerInfo::getConstantPool(),
8716 false, false, false, 16);
8717 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
8719 // Shift sign bit right or left if the two operands have different types.
8720 if (SrcVT.bitsGT(VT)) {
8721 // Op0 is MVT::f32, Op1 is MVT::f64.
8722 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
8723 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
8724 DAG.getConstant(32, MVT::i32));
8725 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
8726 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
8727 DAG.getIntPtrConstant(0));
8730 // Clear first operand sign bit.
8732 if (VT == MVT::f64) {
8733 const fltSemantics &Sem = APFloat::IEEEdouble;
8734 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
8735 APInt(64, ~(1ULL << 63)))));
8736 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
8738 const fltSemantics &Sem = APFloat::IEEEsingle;
8739 CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
8740 APInt(32, ~(1U << 31)))));
8741 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8742 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8743 CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
8745 C = ConstantVector::get(CV);
8746 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8747 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8748 MachinePointerInfo::getConstantPool(),
8749 false, false, false, 16);
8750 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
8752 // Or the value with the sign bit.
8753 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
8756 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
8757 SDValue N0 = Op.getOperand(0);
8758 DebugLoc dl = Op.getDebugLoc();
8759 MVT VT = Op.getValueType().getSimpleVT();
8761 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
8762 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
8763 DAG.getConstant(1, VT));
8764 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
8767 // LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
8769 SDValue X86TargetLowering::LowerVectorAllZeroTest(SDValue Op,
8770 SelectionDAG &DAG) const {
8771 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
8773 if (!Subtarget->hasSSE41())
8776 if (!Op->hasOneUse())
8779 SDNode *N = Op.getNode();
8780 DebugLoc DL = N->getDebugLoc();
8782 SmallVector<SDValue, 8> Opnds;
8783 DenseMap<SDValue, unsigned> VecInMap;
8784 EVT VT = MVT::Other;
8786 // Recognize a special case where a vector is casted into wide integer to
8788 Opnds.push_back(N->getOperand(0));
8789 Opnds.push_back(N->getOperand(1));
8791 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
8792 SmallVector<SDValue, 8>::const_iterator I = Opnds.begin() + Slot;
8793 // BFS traverse all OR'd operands.
8794 if (I->getOpcode() == ISD::OR) {
8795 Opnds.push_back(I->getOperand(0));
8796 Opnds.push_back(I->getOperand(1));
8797 // Re-evaluate the number of nodes to be traversed.
8798 e += 2; // 2 more nodes (LHS and RHS) are pushed.
8802 // Quit if a non-EXTRACT_VECTOR_ELT
8803 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8806 // Quit if without a constant index.
8807 SDValue Idx = I->getOperand(1);
8808 if (!isa<ConstantSDNode>(Idx))
8811 SDValue ExtractedFromVec = I->getOperand(0);
8812 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
8813 if (M == VecInMap.end()) {
8814 VT = ExtractedFromVec.getValueType();
8815 // Quit if not 128/256-bit vector.
8816 if (!VT.is128BitVector() && !VT.is256BitVector())
8818 // Quit if not the same type.
8819 if (VecInMap.begin() != VecInMap.end() &&
8820 VT != VecInMap.begin()->first.getValueType())
8822 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
8824 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
8827 assert((VT.is128BitVector() || VT.is256BitVector()) &&
8828 "Not extracted from 128-/256-bit vector.");
8830 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
8831 SmallVector<SDValue, 8> VecIns;
8833 for (DenseMap<SDValue, unsigned>::const_iterator
8834 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
8835 // Quit if not all elements are used.
8836 if (I->second != FullMask)
8838 VecIns.push_back(I->first);
8841 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
8843 // Cast all vectors into TestVT for PTEST.
8844 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
8845 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
8847 // If more than one full vectors are evaluated, OR them first before PTEST.
8848 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
8849 // Each iteration will OR 2 nodes and append the result until there is only
8850 // 1 node left, i.e. the final OR'd value of all vectors.
8851 SDValue LHS = VecIns[Slot];
8852 SDValue RHS = VecIns[Slot + 1];
8853 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
8856 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
8857 VecIns.back(), VecIns.back());
8860 /// Emit nodes that will be selected as "test Op0,Op0", or something
8862 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
8863 SelectionDAG &DAG) const {
8864 DebugLoc dl = Op.getDebugLoc();
8866 // CF and OF aren't always set the way we want. Determine which
8867 // of these we need.
8868 bool NeedCF = false;
8869 bool NeedOF = false;
8872 case X86::COND_A: case X86::COND_AE:
8873 case X86::COND_B: case X86::COND_BE:
8876 case X86::COND_G: case X86::COND_GE:
8877 case X86::COND_L: case X86::COND_LE:
8878 case X86::COND_O: case X86::COND_NO:
8883 // See if we can use the EFLAGS value from the operand instead of
8884 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
8885 // we prove that the arithmetic won't overflow, we can't use OF or CF.
8886 if (Op.getResNo() != 0 || NeedOF || NeedCF)
8887 // Emit a CMP with 0, which is the TEST pattern.
8888 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8889 DAG.getConstant(0, Op.getValueType()));
8891 unsigned Opcode = 0;
8892 unsigned NumOperands = 0;
8894 // Truncate operations may prevent the merge of the SETCC instruction
8895 // and the arithmetic intruction before it. Attempt to truncate the operands
8896 // of the arithmetic instruction and use a reduced bit-width instruction.
8897 bool NeedTruncation = false;
8898 SDValue ArithOp = Op;
8899 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
8900 SDValue Arith = Op->getOperand(0);
8901 // Both the trunc and the arithmetic op need to have one user each.
8902 if (Arith->hasOneUse())
8903 switch (Arith.getOpcode()) {
8910 NeedTruncation = true;
8916 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
8917 // which may be the result of a CAST. We use the variable 'Op', which is the
8918 // non-casted variable when we check for possible users.
8919 switch (ArithOp.getOpcode()) {
8921 // Due to an isel shortcoming, be conservative if this add is likely to be
8922 // selected as part of a load-modify-store instruction. When the root node
8923 // in a match is a store, isel doesn't know how to remap non-chain non-flag
8924 // uses of other nodes in the match, such as the ADD in this case. This
8925 // leads to the ADD being left around and reselected, with the result being
8926 // two adds in the output. Alas, even if none our users are stores, that
8927 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
8928 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
8929 // climbing the DAG back to the root, and it doesn't seem to be worth the
8931 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8932 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8933 if (UI->getOpcode() != ISD::CopyToReg &&
8934 UI->getOpcode() != ISD::SETCC &&
8935 UI->getOpcode() != ISD::STORE)
8938 if (ConstantSDNode *C =
8939 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
8940 // An add of one will be selected as an INC.
8941 if (C->getAPIntValue() == 1) {
8942 Opcode = X86ISD::INC;
8947 // An add of negative one (subtract of one) will be selected as a DEC.
8948 if (C->getAPIntValue().isAllOnesValue()) {
8949 Opcode = X86ISD::DEC;
8955 // Otherwise use a regular EFLAGS-setting add.
8956 Opcode = X86ISD::ADD;
8960 // If the primary and result isn't used, don't bother using X86ISD::AND,
8961 // because a TEST instruction will be better.
8962 bool NonFlagUse = false;
8963 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8964 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
8966 unsigned UOpNo = UI.getOperandNo();
8967 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
8968 // Look pass truncate.
8969 UOpNo = User->use_begin().getOperandNo();
8970 User = *User->use_begin();
8973 if (User->getOpcode() != ISD::BRCOND &&
8974 User->getOpcode() != ISD::SETCC &&
8975 !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) {
8988 // Due to the ISEL shortcoming noted above, be conservative if this op is
8989 // likely to be selected as part of a load-modify-store instruction.
8990 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8991 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8992 if (UI->getOpcode() == ISD::STORE)
8995 // Otherwise use a regular EFLAGS-setting instruction.
8996 switch (ArithOp.getOpcode()) {
8997 default: llvm_unreachable("unexpected operator!");
8998 case ISD::SUB: Opcode = X86ISD::SUB; break;
8999 case ISD::XOR: Opcode = X86ISD::XOR; break;
9000 case ISD::AND: Opcode = X86ISD::AND; break;
9002 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
9003 SDValue EFLAGS = LowerVectorAllZeroTest(Op, DAG);
9004 if (EFLAGS.getNode())
9007 Opcode = X86ISD::OR;
9021 return SDValue(Op.getNode(), 1);
9027 // If we found that truncation is beneficial, perform the truncation and
9029 if (NeedTruncation) {
9030 EVT VT = Op.getValueType();
9031 SDValue WideVal = Op->getOperand(0);
9032 EVT WideVT = WideVal.getValueType();
9033 unsigned ConvertedOp = 0;
9034 // Use a target machine opcode to prevent further DAGCombine
9035 // optimizations that may separate the arithmetic operations
9036 // from the setcc node.
9037 switch (WideVal.getOpcode()) {
9039 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
9040 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
9041 case ISD::AND: ConvertedOp = X86ISD::AND; break;
9042 case ISD::OR: ConvertedOp = X86ISD::OR; break;
9043 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
9047 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9048 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
9049 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
9050 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
9051 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
9057 // Emit a CMP with 0, which is the TEST pattern.
9058 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
9059 DAG.getConstant(0, Op.getValueType()));
9061 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
9062 SmallVector<SDValue, 4> Ops;
9063 for (unsigned i = 0; i != NumOperands; ++i)
9064 Ops.push_back(Op.getOperand(i));
9066 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
9067 DAG.ReplaceAllUsesWith(Op, New);
9068 return SDValue(New.getNode(), 1);
9071 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
9073 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
9074 SelectionDAG &DAG) const {
9075 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
9076 if (C->getAPIntValue() == 0)
9077 return EmitTest(Op0, X86CC, DAG);
9079 DebugLoc dl = Op0.getDebugLoc();
9080 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
9081 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
9082 // Use SUB instead of CMP to enable CSE between SUB and CMP.
9083 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
9084 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
9086 return SDValue(Sub.getNode(), 1);
9088 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
9091 /// Convert a comparison if required by the subtarget.
9092 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
9093 SelectionDAG &DAG) const {
9094 // If the subtarget does not support the FUCOMI instruction, floating-point
9095 // comparisons have to be converted.
9096 if (Subtarget->hasCMov() ||
9097 Cmp.getOpcode() != X86ISD::CMP ||
9098 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
9099 !Cmp.getOperand(1).getValueType().isFloatingPoint())
9102 // The instruction selector will select an FUCOM instruction instead of
9103 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
9104 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
9105 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
9106 DebugLoc dl = Cmp.getDebugLoc();
9107 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
9108 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
9109 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
9110 DAG.getConstant(8, MVT::i8));
9111 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
9112 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
9115 static bool isAllOnes(SDValue V) {
9116 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9117 return C && C->isAllOnesValue();
9120 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
9121 /// if it's possible.
9122 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
9123 DebugLoc dl, SelectionDAG &DAG) const {
9124 SDValue Op0 = And.getOperand(0);
9125 SDValue Op1 = And.getOperand(1);
9126 if (Op0.getOpcode() == ISD::TRUNCATE)
9127 Op0 = Op0.getOperand(0);
9128 if (Op1.getOpcode() == ISD::TRUNCATE)
9129 Op1 = Op1.getOperand(0);
9132 if (Op1.getOpcode() == ISD::SHL)
9133 std::swap(Op0, Op1);
9134 if (Op0.getOpcode() == ISD::SHL) {
9135 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
9136 if (And00C->getZExtValue() == 1) {
9137 // If we looked past a truncate, check that it's only truncating away
9139 unsigned BitWidth = Op0.getValueSizeInBits();
9140 unsigned AndBitWidth = And.getValueSizeInBits();
9141 if (BitWidth > AndBitWidth) {
9143 DAG.ComputeMaskedBits(Op0, Zeros, Ones);
9144 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
9148 RHS = Op0.getOperand(1);
9150 } else if (Op1.getOpcode() == ISD::Constant) {
9151 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
9152 uint64_t AndRHSVal = AndRHS->getZExtValue();
9153 SDValue AndLHS = Op0;
9155 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
9156 LHS = AndLHS.getOperand(0);
9157 RHS = AndLHS.getOperand(1);
9160 // Use BT if the immediate can't be encoded in a TEST instruction.
9161 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
9163 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
9167 if (LHS.getNode()) {
9168 // If the LHS is of the form (x ^ -1) then replace the LHS with x and flip
9169 // the condition code later.
9170 bool Invert = false;
9171 if (LHS.getOpcode() == ISD::XOR && isAllOnes(LHS.getOperand(1))) {
9173 LHS = LHS.getOperand(0);
9176 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
9177 // instruction. Since the shift amount is in-range-or-undefined, we know
9178 // that doing a bittest on the i32 value is ok. We extend to i32 because
9179 // the encoding for the i16 version is larger than the i32 version.
9180 // Also promote i16 to i32 for performance / code size reason.
9181 if (LHS.getValueType() == MVT::i8 ||
9182 LHS.getValueType() == MVT::i16)
9183 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
9185 // If the operand types disagree, extend the shift amount to match. Since
9186 // BT ignores high bits (like shifts) we can use anyextend.
9187 if (LHS.getValueType() != RHS.getValueType())
9188 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
9190 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
9191 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
9192 // Flip the condition if the LHS was a not instruction
9194 Cond = X86::GetOppositeBranchCondition(Cond);
9195 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9196 DAG.getConstant(Cond, MVT::i8), BT);
9202 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
9203 // ones, and then concatenate the result back.
9204 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
9205 MVT VT = Op.getValueType().getSimpleVT();
9207 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
9208 "Unsupported value type for operation");
9210 unsigned NumElems = VT.getVectorNumElements();
9211 DebugLoc dl = Op.getDebugLoc();
9212 SDValue CC = Op.getOperand(2);
9214 // Extract the LHS vectors
9215 SDValue LHS = Op.getOperand(0);
9216 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
9217 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
9219 // Extract the RHS vectors
9220 SDValue RHS = Op.getOperand(1);
9221 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
9222 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
9224 // Issue the operation on the smaller types and concatenate the result back
9225 MVT EltVT = VT.getVectorElementType();
9226 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9227 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9228 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
9229 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
9232 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
9233 SelectionDAG &DAG) {
9235 SDValue Op0 = Op.getOperand(0);
9236 SDValue Op1 = Op.getOperand(1);
9237 SDValue CC = Op.getOperand(2);
9238 MVT VT = Op.getValueType().getSimpleVT();
9239 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
9240 bool isFP = Op.getOperand(1).getValueType().getSimpleVT().isFloatingPoint();
9241 DebugLoc dl = Op.getDebugLoc();
9245 MVT EltVT = Op0.getValueType().getVectorElementType().getSimpleVT();
9246 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
9252 // SSE Condition code mapping:
9261 switch (SetCCOpcode) {
9262 default: llvm_unreachable("Unexpected SETCC condition");
9264 case ISD::SETEQ: SSECC = 0; break;
9266 case ISD::SETGT: Swap = true; // Fallthrough
9268 case ISD::SETOLT: SSECC = 1; break;
9270 case ISD::SETGE: Swap = true; // Fallthrough
9272 case ISD::SETOLE: SSECC = 2; break;
9273 case ISD::SETUO: SSECC = 3; break;
9275 case ISD::SETNE: SSECC = 4; break;
9276 case ISD::SETULE: Swap = true; // Fallthrough
9277 case ISD::SETUGE: SSECC = 5; break;
9278 case ISD::SETULT: Swap = true; // Fallthrough
9279 case ISD::SETUGT: SSECC = 6; break;
9280 case ISD::SETO: SSECC = 7; break;
9282 case ISD::SETONE: SSECC = 8; break;
9285 std::swap(Op0, Op1);
9287 // In the two special cases we can't handle, emit two comparisons.
9290 unsigned CombineOpc;
9291 if (SetCCOpcode == ISD::SETUEQ) {
9292 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
9294 assert(SetCCOpcode == ISD::SETONE);
9295 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
9298 SDValue Cmp0 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9299 DAG.getConstant(CC0, MVT::i8));
9300 SDValue Cmp1 = DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9301 DAG.getConstant(CC1, MVT::i8));
9302 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
9304 // Handle all other FP comparisons here.
9305 return DAG.getNode(X86ISD::CMPP, dl, VT, Op0, Op1,
9306 DAG.getConstant(SSECC, MVT::i8));
9309 // Break 256-bit integer vector compare into smaller ones.
9310 if (VT.is256BitVector() && !Subtarget->hasInt256())
9311 return Lower256IntVSETCC(Op, DAG);
9313 // We are handling one of the integer comparisons here. Since SSE only has
9314 // GT and EQ comparisons for integer, swapping operands and multiple
9315 // operations may be required for some comparisons.
9317 bool Swap = false, Invert = false, FlipSigns = false;
9319 switch (SetCCOpcode) {
9320 default: llvm_unreachable("Unexpected SETCC condition");
9321 case ISD::SETNE: Invert = true;
9322 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
9323 case ISD::SETLT: Swap = true;
9324 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
9325 case ISD::SETGE: Swap = true;
9326 case ISD::SETLE: Opc = X86ISD::PCMPGT; Invert = true; break;
9327 case ISD::SETULT: Swap = true;
9328 case ISD::SETUGT: Opc = X86ISD::PCMPGT; FlipSigns = true; break;
9329 case ISD::SETUGE: Swap = true;
9330 case ISD::SETULE: Opc = X86ISD::PCMPGT; FlipSigns = true; Invert = true; break;
9333 std::swap(Op0, Op1);
9335 // Check that the operation in question is available (most are plain SSE2,
9336 // but PCMPGTQ and PCMPEQQ have different requirements).
9337 if (VT == MVT::v2i64) {
9338 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42())
9340 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
9341 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
9342 // pcmpeqd + pshufd + pand.
9343 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
9345 // First cast everything to the right type,
9346 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
9347 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
9350 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
9352 // Make sure the lower and upper halves are both all-ones.
9353 const int Mask[] = { 1, 0, 3, 2 };
9354 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
9355 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
9358 Result = DAG.getNOT(dl, Result, MVT::v4i32);
9360 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
9364 // Since SSE has no unsigned integer comparisons, we need to flip the sign
9365 // bits of the inputs before performing those operations.
9367 EVT EltVT = VT.getVectorElementType();
9368 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
9370 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
9371 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
9373 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
9374 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
9377 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
9379 // If the logical-not of the result is required, perform that now.
9381 Result = DAG.getNOT(dl, Result, VT);
9386 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
9388 MVT VT = Op.getValueType().getSimpleVT();
9390 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
9392 assert(VT == MVT::i8 && "SetCC type must be 8-bit integer");
9393 SDValue Op0 = Op.getOperand(0);
9394 SDValue Op1 = Op.getOperand(1);
9395 DebugLoc dl = Op.getDebugLoc();
9396 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
9398 // Optimize to BT if possible.
9399 // Lower (X & (1 << N)) == 0 to BT(X, N).
9400 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
9401 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
9402 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
9403 Op1.getOpcode() == ISD::Constant &&
9404 cast<ConstantSDNode>(Op1)->isNullValue() &&
9405 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
9406 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
9407 if (NewSetCC.getNode())
9411 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
9413 if (Op1.getOpcode() == ISD::Constant &&
9414 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
9415 cast<ConstantSDNode>(Op1)->isNullValue()) &&
9416 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
9418 // If the input is a setcc, then reuse the input setcc or use a new one with
9419 // the inverted condition.
9420 if (Op0.getOpcode() == X86ISD::SETCC) {
9421 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
9422 bool Invert = (CC == ISD::SETNE) ^
9423 cast<ConstantSDNode>(Op1)->isNullValue();
9424 if (!Invert) return Op0;
9426 CCode = X86::GetOppositeBranchCondition(CCode);
9427 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9428 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
9432 bool isFP = Op1.getValueType().getSimpleVT().isFloatingPoint();
9433 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
9434 if (X86CC == X86::COND_INVALID)
9437 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
9438 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
9439 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9440 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
9443 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
9444 static bool isX86LogicalCmp(SDValue Op) {
9445 unsigned Opc = Op.getNode()->getOpcode();
9446 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
9447 Opc == X86ISD::SAHF)
9449 if (Op.getResNo() == 1 &&
9450 (Opc == X86ISD::ADD ||
9451 Opc == X86ISD::SUB ||
9452 Opc == X86ISD::ADC ||
9453 Opc == X86ISD::SBB ||
9454 Opc == X86ISD::SMUL ||
9455 Opc == X86ISD::UMUL ||
9456 Opc == X86ISD::INC ||
9457 Opc == X86ISD::DEC ||
9458 Opc == X86ISD::OR ||
9459 Opc == X86ISD::XOR ||
9460 Opc == X86ISD::AND))
9463 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
9469 static bool isZero(SDValue V) {
9470 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
9471 return C && C->isNullValue();
9474 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
9475 if (V.getOpcode() != ISD::TRUNCATE)
9478 SDValue VOp0 = V.getOperand(0);
9479 unsigned InBits = VOp0.getValueSizeInBits();
9480 unsigned Bits = V.getValueSizeInBits();
9481 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
9484 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
9485 bool addTest = true;
9486 SDValue Cond = Op.getOperand(0);
9487 SDValue Op1 = Op.getOperand(1);
9488 SDValue Op2 = Op.getOperand(2);
9489 DebugLoc DL = Op.getDebugLoc();
9492 if (Cond.getOpcode() == ISD::SETCC) {
9493 SDValue NewCond = LowerSETCC(Cond, DAG);
9494 if (NewCond.getNode())
9498 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
9499 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
9500 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
9501 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
9502 if (Cond.getOpcode() == X86ISD::SETCC &&
9503 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
9504 isZero(Cond.getOperand(1).getOperand(1))) {
9505 SDValue Cmp = Cond.getOperand(1);
9507 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
9509 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
9510 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
9511 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
9513 SDValue CmpOp0 = Cmp.getOperand(0);
9514 // Apply further optimizations for special cases
9515 // (select (x != 0), -1, 0) -> neg & sbb
9516 // (select (x == 0), 0, -1) -> neg & sbb
9517 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
9518 if (YC->isNullValue() &&
9519 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
9520 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
9521 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
9522 DAG.getConstant(0, CmpOp0.getValueType()),
9524 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9525 DAG.getConstant(X86::COND_B, MVT::i8),
9526 SDValue(Neg.getNode(), 1));
9530 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
9531 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
9532 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9534 SDValue Res = // Res = 0 or -1.
9535 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9536 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
9538 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
9539 Res = DAG.getNOT(DL, Res, Res.getValueType());
9541 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
9542 if (N2C == 0 || !N2C->isNullValue())
9543 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
9548 // Look past (and (setcc_carry (cmp ...)), 1).
9549 if (Cond.getOpcode() == ISD::AND &&
9550 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
9551 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
9552 if (C && C->getAPIntValue() == 1)
9553 Cond = Cond.getOperand(0);
9556 // If condition flag is set by a X86ISD::CMP, then use it as the condition
9557 // setting operand in place of the X86ISD::SETCC.
9558 unsigned CondOpcode = Cond.getOpcode();
9559 if (CondOpcode == X86ISD::SETCC ||
9560 CondOpcode == X86ISD::SETCC_CARRY) {
9561 CC = Cond.getOperand(0);
9563 SDValue Cmp = Cond.getOperand(1);
9564 unsigned Opc = Cmp.getOpcode();
9565 MVT VT = Op.getValueType().getSimpleVT();
9567 bool IllegalFPCMov = false;
9568 if (VT.isFloatingPoint() && !VT.isVector() &&
9569 !isScalarFPTypeInSSEReg(VT)) // FPStack?
9570 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
9572 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
9573 Opc == X86ISD::BT) { // FIXME
9577 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
9578 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
9579 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
9580 Cond.getOperand(0).getValueType() != MVT::i8)) {
9581 SDValue LHS = Cond.getOperand(0);
9582 SDValue RHS = Cond.getOperand(1);
9586 switch (CondOpcode) {
9587 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
9588 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
9589 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
9590 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
9591 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
9592 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
9593 default: llvm_unreachable("unexpected overflowing operator");
9595 if (CondOpcode == ISD::UMULO)
9596 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
9599 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
9601 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
9603 if (CondOpcode == ISD::UMULO)
9604 Cond = X86Op.getValue(2);
9606 Cond = X86Op.getValue(1);
9608 CC = DAG.getConstant(X86Cond, MVT::i8);
9613 // Look pass the truncate if the high bits are known zero.
9614 if (isTruncWithZeroHighBitsInput(Cond, DAG))
9615 Cond = Cond.getOperand(0);
9617 // We know the result of AND is compared against zero. Try to match
9619 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
9620 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
9621 if (NewSetCC.getNode()) {
9622 CC = NewSetCC.getOperand(0);
9623 Cond = NewSetCC.getOperand(1);
9630 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9631 Cond = EmitTest(Cond, X86::COND_NE, DAG);
9634 // a < b ? -1 : 0 -> RES = ~setcc_carry
9635 // a < b ? 0 : -1 -> RES = setcc_carry
9636 // a >= b ? -1 : 0 -> RES = setcc_carry
9637 // a >= b ? 0 : -1 -> RES = ~setcc_carry
9638 if (Cond.getOpcode() == X86ISD::SUB) {
9639 Cond = ConvertCmpIfNecessary(Cond, DAG);
9640 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
9642 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
9643 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
9644 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
9645 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
9646 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
9647 return DAG.getNOT(DL, Res, Res.getValueType());
9652 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
9653 // widen the cmov and push the truncate through. This avoids introducing a new
9654 // branch during isel and doesn't add any extensions.
9655 if (Op.getValueType() == MVT::i8 &&
9656 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
9657 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
9658 if (T1.getValueType() == T2.getValueType() &&
9659 // Blacklist CopyFromReg to avoid partial register stalls.
9660 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
9661 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
9662 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
9663 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
9667 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
9668 // condition is true.
9669 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
9670 SDValue Ops[] = { Op2, Op1, CC, Cond };
9671 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
9674 SDValue X86TargetLowering::LowerSIGN_EXTEND(SDValue Op,
9675 SelectionDAG &DAG) const {
9676 MVT VT = Op->getValueType(0).getSimpleVT();
9677 SDValue In = Op->getOperand(0);
9678 MVT InVT = In.getValueType().getSimpleVT();
9679 DebugLoc dl = Op->getDebugLoc();
9681 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
9682 (VT != MVT::v8i32 || InVT != MVT::v8i16))
9685 if (Subtarget->hasInt256())
9686 return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
9688 // Optimize vectors in AVX mode
9689 // Sign extend v8i16 to v8i32 and
9692 // Divide input vector into two parts
9693 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
9694 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
9695 // concat the vectors to original VT
9697 unsigned NumElems = InVT.getVectorNumElements();
9698 SDValue Undef = DAG.getUNDEF(InVT);
9700 SmallVector<int,8> ShufMask1(NumElems, -1);
9701 for (unsigned i = 0; i != NumElems/2; ++i)
9704 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
9706 SmallVector<int,8> ShufMask2(NumElems, -1);
9707 for (unsigned i = 0; i != NumElems/2; ++i)
9708 ShufMask2[i] = i + NumElems/2;
9710 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
9712 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
9713 VT.getVectorNumElements()/2);
9715 OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
9716 OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
9718 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
9721 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
9722 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
9723 // from the AND / OR.
9724 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
9725 Opc = Op.getOpcode();
9726 if (Opc != ISD::OR && Opc != ISD::AND)
9728 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
9729 Op.getOperand(0).hasOneUse() &&
9730 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
9731 Op.getOperand(1).hasOneUse());
9734 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
9735 // 1 and that the SETCC node has a single use.
9736 static bool isXor1OfSetCC(SDValue Op) {
9737 if (Op.getOpcode() != ISD::XOR)
9739 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
9740 if (N1C && N1C->getAPIntValue() == 1) {
9741 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
9742 Op.getOperand(0).hasOneUse();
9747 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
9748 bool addTest = true;
9749 SDValue Chain = Op.getOperand(0);
9750 SDValue Cond = Op.getOperand(1);
9751 SDValue Dest = Op.getOperand(2);
9752 DebugLoc dl = Op.getDebugLoc();
9754 bool Inverted = false;
9756 if (Cond.getOpcode() == ISD::SETCC) {
9757 // Check for setcc([su]{add,sub,mul}o == 0).
9758 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
9759 isa<ConstantSDNode>(Cond.getOperand(1)) &&
9760 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
9761 Cond.getOperand(0).getResNo() == 1 &&
9762 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
9763 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
9764 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
9765 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
9766 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
9767 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
9769 Cond = Cond.getOperand(0);
9771 SDValue NewCond = LowerSETCC(Cond, DAG);
9772 if (NewCond.getNode())
9777 // FIXME: LowerXALUO doesn't handle these!!
9778 else if (Cond.getOpcode() == X86ISD::ADD ||
9779 Cond.getOpcode() == X86ISD::SUB ||
9780 Cond.getOpcode() == X86ISD::SMUL ||
9781 Cond.getOpcode() == X86ISD::UMUL)
9782 Cond = LowerXALUO(Cond, DAG);
9785 // Look pass (and (setcc_carry (cmp ...)), 1).
9786 if (Cond.getOpcode() == ISD::AND &&
9787 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
9788 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
9789 if (C && C->getAPIntValue() == 1)
9790 Cond = Cond.getOperand(0);
9793 // If condition flag is set by a X86ISD::CMP, then use it as the condition
9794 // setting operand in place of the X86ISD::SETCC.
9795 unsigned CondOpcode = Cond.getOpcode();
9796 if (CondOpcode == X86ISD::SETCC ||
9797 CondOpcode == X86ISD::SETCC_CARRY) {
9798 CC = Cond.getOperand(0);
9800 SDValue Cmp = Cond.getOperand(1);
9801 unsigned Opc = Cmp.getOpcode();
9802 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
9803 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
9807 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
9811 // These can only come from an arithmetic instruction with overflow,
9812 // e.g. SADDO, UADDO.
9813 Cond = Cond.getNode()->getOperand(1);
9819 CondOpcode = Cond.getOpcode();
9820 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
9821 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
9822 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
9823 Cond.getOperand(0).getValueType() != MVT::i8)) {
9824 SDValue LHS = Cond.getOperand(0);
9825 SDValue RHS = Cond.getOperand(1);
9829 switch (CondOpcode) {
9830 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
9831 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
9832 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
9833 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
9834 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
9835 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
9836 default: llvm_unreachable("unexpected overflowing operator");
9839 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
9840 if (CondOpcode == ISD::UMULO)
9841 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
9844 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
9846 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
9848 if (CondOpcode == ISD::UMULO)
9849 Cond = X86Op.getValue(2);
9851 Cond = X86Op.getValue(1);
9853 CC = DAG.getConstant(X86Cond, MVT::i8);
9857 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
9858 SDValue Cmp = Cond.getOperand(0).getOperand(1);
9859 if (CondOpc == ISD::OR) {
9860 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
9861 // two branches instead of an explicit OR instruction with a
9863 if (Cmp == Cond.getOperand(1).getOperand(1) &&
9864 isX86LogicalCmp(Cmp)) {
9865 CC = Cond.getOperand(0).getOperand(0);
9866 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9867 Chain, Dest, CC, Cmp);
9868 CC = Cond.getOperand(1).getOperand(0);
9872 } else { // ISD::AND
9873 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
9874 // two branches instead of an explicit AND instruction with a
9875 // separate test. However, we only do this if this block doesn't
9876 // have a fall-through edge, because this requires an explicit
9877 // jmp when the condition is false.
9878 if (Cmp == Cond.getOperand(1).getOperand(1) &&
9879 isX86LogicalCmp(Cmp) &&
9880 Op.getNode()->hasOneUse()) {
9881 X86::CondCode CCode =
9882 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
9883 CCode = X86::GetOppositeBranchCondition(CCode);
9884 CC = DAG.getConstant(CCode, MVT::i8);
9885 SDNode *User = *Op.getNode()->use_begin();
9886 // Look for an unconditional branch following this conditional branch.
9887 // We need this because we need to reverse the successors in order
9888 // to implement FCMP_OEQ.
9889 if (User->getOpcode() == ISD::BR) {
9890 SDValue FalseBB = User->getOperand(1);
9892 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9893 assert(NewBR == User);
9897 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9898 Chain, Dest, CC, Cmp);
9899 X86::CondCode CCode =
9900 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
9901 CCode = X86::GetOppositeBranchCondition(CCode);
9902 CC = DAG.getConstant(CCode, MVT::i8);
9908 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
9909 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
9910 // It should be transformed during dag combiner except when the condition
9911 // is set by a arithmetics with overflow node.
9912 X86::CondCode CCode =
9913 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
9914 CCode = X86::GetOppositeBranchCondition(CCode);
9915 CC = DAG.getConstant(CCode, MVT::i8);
9916 Cond = Cond.getOperand(0).getOperand(1);
9918 } else if (Cond.getOpcode() == ISD::SETCC &&
9919 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
9920 // For FCMP_OEQ, we can emit
9921 // two branches instead of an explicit AND instruction with a
9922 // separate test. However, we only do this if this block doesn't
9923 // have a fall-through edge, because this requires an explicit
9924 // jmp when the condition is false.
9925 if (Op.getNode()->hasOneUse()) {
9926 SDNode *User = *Op.getNode()->use_begin();
9927 // Look for an unconditional branch following this conditional branch.
9928 // We need this because we need to reverse the successors in order
9929 // to implement FCMP_OEQ.
9930 if (User->getOpcode() == ISD::BR) {
9931 SDValue FalseBB = User->getOperand(1);
9933 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9934 assert(NewBR == User);
9938 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
9939 Cond.getOperand(0), Cond.getOperand(1));
9940 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9941 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9942 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9943 Chain, Dest, CC, Cmp);
9944 CC = DAG.getConstant(X86::COND_P, MVT::i8);
9949 } else if (Cond.getOpcode() == ISD::SETCC &&
9950 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
9951 // For FCMP_UNE, we can emit
9952 // two branches instead of an explicit AND instruction with a
9953 // separate test. However, we only do this if this block doesn't
9954 // have a fall-through edge, because this requires an explicit
9955 // jmp when the condition is false.
9956 if (Op.getNode()->hasOneUse()) {
9957 SDNode *User = *Op.getNode()->use_begin();
9958 // Look for an unconditional branch following this conditional branch.
9959 // We need this because we need to reverse the successors in order
9960 // to implement FCMP_UNE.
9961 if (User->getOpcode() == ISD::BR) {
9962 SDValue FalseBB = User->getOperand(1);
9964 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
9965 assert(NewBR == User);
9968 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
9969 Cond.getOperand(0), Cond.getOperand(1));
9970 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
9971 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
9972 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
9973 Chain, Dest, CC, Cmp);
9974 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
9984 // Look pass the truncate if the high bits are known zero.
9985 if (isTruncWithZeroHighBitsInput(Cond, DAG))
9986 Cond = Cond.getOperand(0);
9988 // We know the result of AND is compared against zero. Try to match
9990 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
9991 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
9992 if (NewSetCC.getNode()) {
9993 CC = NewSetCC.getOperand(0);
9994 Cond = NewSetCC.getOperand(1);
10001 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
10002 Cond = EmitTest(Cond, X86::COND_NE, DAG);
10004 Cond = ConvertCmpIfNecessary(Cond, DAG);
10005 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
10006 Chain, Dest, CC, Cond);
10009 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
10010 // Calls to _alloca is needed to probe the stack when allocating more than 4k
10011 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
10012 // that the guard pages used by the OS virtual memory manager are allocated in
10013 // correct sequence.
10015 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
10016 SelectionDAG &DAG) const {
10017 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
10018 getTargetMachine().Options.EnableSegmentedStacks) &&
10019 "This should be used only on Windows targets or when segmented stacks "
10021 assert(!Subtarget->isTargetEnvMacho() && "Not implemented");
10022 DebugLoc dl = Op.getDebugLoc();
10025 SDValue Chain = Op.getOperand(0);
10026 SDValue Size = Op.getOperand(1);
10027 // FIXME: Ensure alignment here
10029 bool Is64Bit = Subtarget->is64Bit();
10030 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
10032 if (getTargetMachine().Options.EnableSegmentedStacks) {
10033 MachineFunction &MF = DAG.getMachineFunction();
10034 MachineRegisterInfo &MRI = MF.getRegInfo();
10037 // The 64 bit implementation of segmented stacks needs to clobber both r10
10038 // r11. This makes it impossible to use it along with nested parameters.
10039 const Function *F = MF.getFunction();
10041 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
10043 if (I->hasNestAttr())
10044 report_fatal_error("Cannot use segmented stacks with functions that "
10045 "have nested arguments.");
10048 const TargetRegisterClass *AddrRegClass =
10049 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
10050 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
10051 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
10052 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
10053 DAG.getRegister(Vreg, SPTy));
10054 SDValue Ops1[2] = { Value, Chain };
10055 return DAG.getMergeValues(Ops1, 2, dl);
10058 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
10060 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
10061 Flag = Chain.getValue(1);
10062 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10064 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
10065 Flag = Chain.getValue(1);
10067 Chain = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
10070 SDValue Ops1[2] = { Chain.getValue(0), Chain };
10071 return DAG.getMergeValues(Ops1, 2, dl);
10075 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
10076 MachineFunction &MF = DAG.getMachineFunction();
10077 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
10079 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10080 DebugLoc DL = Op.getDebugLoc();
10082 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
10083 // vastart just stores the address of the VarArgsFrameIndex slot into the
10084 // memory location argument.
10085 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10087 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
10088 MachinePointerInfo(SV), false, false, 0);
10092 // gp_offset (0 - 6 * 8)
10093 // fp_offset (48 - 48 + 8 * 16)
10094 // overflow_arg_area (point to parameters coming in memory).
10096 SmallVector<SDValue, 8> MemOps;
10097 SDValue FIN = Op.getOperand(1);
10099 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
10100 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
10102 FIN, MachinePointerInfo(SV), false, false, 0);
10103 MemOps.push_back(Store);
10106 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10107 FIN, DAG.getIntPtrConstant(4));
10108 Store = DAG.getStore(Op.getOperand(0), DL,
10109 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
10111 FIN, MachinePointerInfo(SV, 4), false, false, 0);
10112 MemOps.push_back(Store);
10114 // Store ptr to overflow_arg_area
10115 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10116 FIN, DAG.getIntPtrConstant(4));
10117 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
10119 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
10120 MachinePointerInfo(SV, 8),
10122 MemOps.push_back(Store);
10124 // Store ptr to reg_save_area.
10125 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10126 FIN, DAG.getIntPtrConstant(8));
10127 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
10129 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
10130 MachinePointerInfo(SV, 16), false, false, 0);
10131 MemOps.push_back(Store);
10132 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
10133 &MemOps[0], MemOps.size());
10136 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
10137 assert(Subtarget->is64Bit() &&
10138 "LowerVAARG only handles 64-bit va_arg!");
10139 assert((Subtarget->isTargetLinux() ||
10140 Subtarget->isTargetDarwin()) &&
10141 "Unhandled target in LowerVAARG");
10142 assert(Op.getNode()->getNumOperands() == 4);
10143 SDValue Chain = Op.getOperand(0);
10144 SDValue SrcPtr = Op.getOperand(1);
10145 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
10146 unsigned Align = Op.getConstantOperandVal(3);
10147 DebugLoc dl = Op.getDebugLoc();
10149 EVT ArgVT = Op.getNode()->getValueType(0);
10150 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10151 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
10154 // Decide which area this value should be read from.
10155 // TODO: Implement the AMD64 ABI in its entirety. This simple
10156 // selection mechanism works only for the basic types.
10157 if (ArgVT == MVT::f80) {
10158 llvm_unreachable("va_arg for f80 not yet implemented");
10159 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
10160 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
10161 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
10162 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
10164 llvm_unreachable("Unhandled argument type in LowerVAARG");
10167 if (ArgMode == 2) {
10168 // Sanity Check: Make sure using fp_offset makes sense.
10169 assert(!getTargetMachine().Options.UseSoftFloat &&
10170 !(DAG.getMachineFunction()
10171 .getFunction()->getAttributes()
10172 .hasAttribute(AttributeSet::FunctionIndex,
10173 Attribute::NoImplicitFloat)) &&
10174 Subtarget->hasSSE1());
10177 // Insert VAARG_64 node into the DAG
10178 // VAARG_64 returns two values: Variable Argument Address, Chain
10179 SmallVector<SDValue, 11> InstOps;
10180 InstOps.push_back(Chain);
10181 InstOps.push_back(SrcPtr);
10182 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
10183 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
10184 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
10185 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
10186 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
10187 VTs, &InstOps[0], InstOps.size(),
10189 MachinePointerInfo(SV),
10191 /*Volatile=*/false,
10193 /*WriteMem=*/true);
10194 Chain = VAARG.getValue(1);
10196 // Load the next argument and return it
10197 return DAG.getLoad(ArgVT, dl,
10200 MachinePointerInfo(),
10201 false, false, false, 0);
10204 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
10205 SelectionDAG &DAG) {
10206 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
10207 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
10208 SDValue Chain = Op.getOperand(0);
10209 SDValue DstPtr = Op.getOperand(1);
10210 SDValue SrcPtr = Op.getOperand(2);
10211 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
10212 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
10213 DebugLoc DL = Op.getDebugLoc();
10215 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
10216 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
10218 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
10221 // getTargetVShiftNode - Handle vector element shifts where the shift amount
10222 // may or may not be a constant. Takes immediate version of shift as input.
10223 static SDValue getTargetVShiftNode(unsigned Opc, DebugLoc dl, EVT VT,
10224 SDValue SrcOp, SDValue ShAmt,
10225 SelectionDAG &DAG) {
10226 assert(ShAmt.getValueType() == MVT::i32 && "ShAmt is not i32");
10228 if (isa<ConstantSDNode>(ShAmt)) {
10229 // Constant may be a TargetConstant. Use a regular constant.
10230 uint32_t ShiftAmt = cast<ConstantSDNode>(ShAmt)->getZExtValue();
10232 default: llvm_unreachable("Unknown target vector shift node");
10233 case X86ISD::VSHLI:
10234 case X86ISD::VSRLI:
10235 case X86ISD::VSRAI:
10236 return DAG.getNode(Opc, dl, VT, SrcOp,
10237 DAG.getConstant(ShiftAmt, MVT::i32));
10241 // Change opcode to non-immediate version
10243 default: llvm_unreachable("Unknown target vector shift node");
10244 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
10245 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
10246 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
10249 // Need to build a vector containing shift amount
10250 // Shift amount is 32-bits, but SSE instructions read 64-bit, so fill with 0
10253 ShOps[1] = DAG.getConstant(0, MVT::i32);
10254 ShOps[2] = ShOps[3] = DAG.getUNDEF(MVT::i32);
10255 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, &ShOps[0], 4);
10257 // The return type has to be a 128-bit type with the same element
10258 // type as the input type.
10259 MVT EltVT = VT.getVectorElementType().getSimpleVT();
10260 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
10262 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
10263 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
10266 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
10267 DebugLoc dl = Op.getDebugLoc();
10268 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10270 default: return SDValue(); // Don't custom lower most intrinsics.
10271 // Comparison intrinsics.
10272 case Intrinsic::x86_sse_comieq_ss:
10273 case Intrinsic::x86_sse_comilt_ss:
10274 case Intrinsic::x86_sse_comile_ss:
10275 case Intrinsic::x86_sse_comigt_ss:
10276 case Intrinsic::x86_sse_comige_ss:
10277 case Intrinsic::x86_sse_comineq_ss:
10278 case Intrinsic::x86_sse_ucomieq_ss:
10279 case Intrinsic::x86_sse_ucomilt_ss:
10280 case Intrinsic::x86_sse_ucomile_ss:
10281 case Intrinsic::x86_sse_ucomigt_ss:
10282 case Intrinsic::x86_sse_ucomige_ss:
10283 case Intrinsic::x86_sse_ucomineq_ss:
10284 case Intrinsic::x86_sse2_comieq_sd:
10285 case Intrinsic::x86_sse2_comilt_sd:
10286 case Intrinsic::x86_sse2_comile_sd:
10287 case Intrinsic::x86_sse2_comigt_sd:
10288 case Intrinsic::x86_sse2_comige_sd:
10289 case Intrinsic::x86_sse2_comineq_sd:
10290 case Intrinsic::x86_sse2_ucomieq_sd:
10291 case Intrinsic::x86_sse2_ucomilt_sd:
10292 case Intrinsic::x86_sse2_ucomile_sd:
10293 case Intrinsic::x86_sse2_ucomigt_sd:
10294 case Intrinsic::x86_sse2_ucomige_sd:
10295 case Intrinsic::x86_sse2_ucomineq_sd: {
10299 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10300 case Intrinsic::x86_sse_comieq_ss:
10301 case Intrinsic::x86_sse2_comieq_sd:
10302 Opc = X86ISD::COMI;
10305 case Intrinsic::x86_sse_comilt_ss:
10306 case Intrinsic::x86_sse2_comilt_sd:
10307 Opc = X86ISD::COMI;
10310 case Intrinsic::x86_sse_comile_ss:
10311 case Intrinsic::x86_sse2_comile_sd:
10312 Opc = X86ISD::COMI;
10315 case Intrinsic::x86_sse_comigt_ss:
10316 case Intrinsic::x86_sse2_comigt_sd:
10317 Opc = X86ISD::COMI;
10320 case Intrinsic::x86_sse_comige_ss:
10321 case Intrinsic::x86_sse2_comige_sd:
10322 Opc = X86ISD::COMI;
10325 case Intrinsic::x86_sse_comineq_ss:
10326 case Intrinsic::x86_sse2_comineq_sd:
10327 Opc = X86ISD::COMI;
10330 case Intrinsic::x86_sse_ucomieq_ss:
10331 case Intrinsic::x86_sse2_ucomieq_sd:
10332 Opc = X86ISD::UCOMI;
10335 case Intrinsic::x86_sse_ucomilt_ss:
10336 case Intrinsic::x86_sse2_ucomilt_sd:
10337 Opc = X86ISD::UCOMI;
10340 case Intrinsic::x86_sse_ucomile_ss:
10341 case Intrinsic::x86_sse2_ucomile_sd:
10342 Opc = X86ISD::UCOMI;
10345 case Intrinsic::x86_sse_ucomigt_ss:
10346 case Intrinsic::x86_sse2_ucomigt_sd:
10347 Opc = X86ISD::UCOMI;
10350 case Intrinsic::x86_sse_ucomige_ss:
10351 case Intrinsic::x86_sse2_ucomige_sd:
10352 Opc = X86ISD::UCOMI;
10355 case Intrinsic::x86_sse_ucomineq_ss:
10356 case Intrinsic::x86_sse2_ucomineq_sd:
10357 Opc = X86ISD::UCOMI;
10362 SDValue LHS = Op.getOperand(1);
10363 SDValue RHS = Op.getOperand(2);
10364 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
10365 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
10366 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
10367 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10368 DAG.getConstant(X86CC, MVT::i8), Cond);
10369 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10372 // Arithmetic intrinsics.
10373 case Intrinsic::x86_sse2_pmulu_dq:
10374 case Intrinsic::x86_avx2_pmulu_dq:
10375 return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
10376 Op.getOperand(1), Op.getOperand(2));
10378 // SSE2/AVX2 sub with unsigned saturation intrinsics
10379 case Intrinsic::x86_sse2_psubus_b:
10380 case Intrinsic::x86_sse2_psubus_w:
10381 case Intrinsic::x86_avx2_psubus_b:
10382 case Intrinsic::x86_avx2_psubus_w:
10383 return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
10384 Op.getOperand(1), Op.getOperand(2));
10386 // SSE3/AVX horizontal add/sub intrinsics
10387 case Intrinsic::x86_sse3_hadd_ps:
10388 case Intrinsic::x86_sse3_hadd_pd:
10389 case Intrinsic::x86_avx_hadd_ps_256:
10390 case Intrinsic::x86_avx_hadd_pd_256:
10391 case Intrinsic::x86_sse3_hsub_ps:
10392 case Intrinsic::x86_sse3_hsub_pd:
10393 case Intrinsic::x86_avx_hsub_ps_256:
10394 case Intrinsic::x86_avx_hsub_pd_256:
10395 case Intrinsic::x86_ssse3_phadd_w_128:
10396 case Intrinsic::x86_ssse3_phadd_d_128:
10397 case Intrinsic::x86_avx2_phadd_w:
10398 case Intrinsic::x86_avx2_phadd_d:
10399 case Intrinsic::x86_ssse3_phsub_w_128:
10400 case Intrinsic::x86_ssse3_phsub_d_128:
10401 case Intrinsic::x86_avx2_phsub_w:
10402 case Intrinsic::x86_avx2_phsub_d: {
10405 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10406 case Intrinsic::x86_sse3_hadd_ps:
10407 case Intrinsic::x86_sse3_hadd_pd:
10408 case Intrinsic::x86_avx_hadd_ps_256:
10409 case Intrinsic::x86_avx_hadd_pd_256:
10410 Opcode = X86ISD::FHADD;
10412 case Intrinsic::x86_sse3_hsub_ps:
10413 case Intrinsic::x86_sse3_hsub_pd:
10414 case Intrinsic::x86_avx_hsub_ps_256:
10415 case Intrinsic::x86_avx_hsub_pd_256:
10416 Opcode = X86ISD::FHSUB;
10418 case Intrinsic::x86_ssse3_phadd_w_128:
10419 case Intrinsic::x86_ssse3_phadd_d_128:
10420 case Intrinsic::x86_avx2_phadd_w:
10421 case Intrinsic::x86_avx2_phadd_d:
10422 Opcode = X86ISD::HADD;
10424 case Intrinsic::x86_ssse3_phsub_w_128:
10425 case Intrinsic::x86_ssse3_phsub_d_128:
10426 case Intrinsic::x86_avx2_phsub_w:
10427 case Intrinsic::x86_avx2_phsub_d:
10428 Opcode = X86ISD::HSUB;
10431 return DAG.getNode(Opcode, dl, Op.getValueType(),
10432 Op.getOperand(1), Op.getOperand(2));
10435 // SSE2/SSE41/AVX2 integer max/min intrinsics.
10436 case Intrinsic::x86_sse2_pmaxu_b:
10437 case Intrinsic::x86_sse41_pmaxuw:
10438 case Intrinsic::x86_sse41_pmaxud:
10439 case Intrinsic::x86_avx2_pmaxu_b:
10440 case Intrinsic::x86_avx2_pmaxu_w:
10441 case Intrinsic::x86_avx2_pmaxu_d:
10442 case Intrinsic::x86_sse2_pminu_b:
10443 case Intrinsic::x86_sse41_pminuw:
10444 case Intrinsic::x86_sse41_pminud:
10445 case Intrinsic::x86_avx2_pminu_b:
10446 case Intrinsic::x86_avx2_pminu_w:
10447 case Intrinsic::x86_avx2_pminu_d:
10448 case Intrinsic::x86_sse41_pmaxsb:
10449 case Intrinsic::x86_sse2_pmaxs_w:
10450 case Intrinsic::x86_sse41_pmaxsd:
10451 case Intrinsic::x86_avx2_pmaxs_b:
10452 case Intrinsic::x86_avx2_pmaxs_w:
10453 case Intrinsic::x86_avx2_pmaxs_d:
10454 case Intrinsic::x86_sse41_pminsb:
10455 case Intrinsic::x86_sse2_pmins_w:
10456 case Intrinsic::x86_sse41_pminsd:
10457 case Intrinsic::x86_avx2_pmins_b:
10458 case Intrinsic::x86_avx2_pmins_w:
10459 case Intrinsic::x86_avx2_pmins_d: {
10462 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10463 case Intrinsic::x86_sse2_pmaxu_b:
10464 case Intrinsic::x86_sse41_pmaxuw:
10465 case Intrinsic::x86_sse41_pmaxud:
10466 case Intrinsic::x86_avx2_pmaxu_b:
10467 case Intrinsic::x86_avx2_pmaxu_w:
10468 case Intrinsic::x86_avx2_pmaxu_d:
10469 Opcode = X86ISD::UMAX;
10471 case Intrinsic::x86_sse2_pminu_b:
10472 case Intrinsic::x86_sse41_pminuw:
10473 case Intrinsic::x86_sse41_pminud:
10474 case Intrinsic::x86_avx2_pminu_b:
10475 case Intrinsic::x86_avx2_pminu_w:
10476 case Intrinsic::x86_avx2_pminu_d:
10477 Opcode = X86ISD::UMIN;
10479 case Intrinsic::x86_sse41_pmaxsb:
10480 case Intrinsic::x86_sse2_pmaxs_w:
10481 case Intrinsic::x86_sse41_pmaxsd:
10482 case Intrinsic::x86_avx2_pmaxs_b:
10483 case Intrinsic::x86_avx2_pmaxs_w:
10484 case Intrinsic::x86_avx2_pmaxs_d:
10485 Opcode = X86ISD::SMAX;
10487 case Intrinsic::x86_sse41_pminsb:
10488 case Intrinsic::x86_sse2_pmins_w:
10489 case Intrinsic::x86_sse41_pminsd:
10490 case Intrinsic::x86_avx2_pmins_b:
10491 case Intrinsic::x86_avx2_pmins_w:
10492 case Intrinsic::x86_avx2_pmins_d:
10493 Opcode = X86ISD::SMIN;
10496 return DAG.getNode(Opcode, dl, Op.getValueType(),
10497 Op.getOperand(1), Op.getOperand(2));
10500 // SSE/SSE2/AVX floating point max/min intrinsics.
10501 case Intrinsic::x86_sse_max_ps:
10502 case Intrinsic::x86_sse2_max_pd:
10503 case Intrinsic::x86_avx_max_ps_256:
10504 case Intrinsic::x86_avx_max_pd_256:
10505 case Intrinsic::x86_sse_min_ps:
10506 case Intrinsic::x86_sse2_min_pd:
10507 case Intrinsic::x86_avx_min_ps_256:
10508 case Intrinsic::x86_avx_min_pd_256: {
10511 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10512 case Intrinsic::x86_sse_max_ps:
10513 case Intrinsic::x86_sse2_max_pd:
10514 case Intrinsic::x86_avx_max_ps_256:
10515 case Intrinsic::x86_avx_max_pd_256:
10516 Opcode = X86ISD::FMAX;
10518 case Intrinsic::x86_sse_min_ps:
10519 case Intrinsic::x86_sse2_min_pd:
10520 case Intrinsic::x86_avx_min_ps_256:
10521 case Intrinsic::x86_avx_min_pd_256:
10522 Opcode = X86ISD::FMIN;
10525 return DAG.getNode(Opcode, dl, Op.getValueType(),
10526 Op.getOperand(1), Op.getOperand(2));
10529 // AVX2 variable shift intrinsics
10530 case Intrinsic::x86_avx2_psllv_d:
10531 case Intrinsic::x86_avx2_psllv_q:
10532 case Intrinsic::x86_avx2_psllv_d_256:
10533 case Intrinsic::x86_avx2_psllv_q_256:
10534 case Intrinsic::x86_avx2_psrlv_d:
10535 case Intrinsic::x86_avx2_psrlv_q:
10536 case Intrinsic::x86_avx2_psrlv_d_256:
10537 case Intrinsic::x86_avx2_psrlv_q_256:
10538 case Intrinsic::x86_avx2_psrav_d:
10539 case Intrinsic::x86_avx2_psrav_d_256: {
10542 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10543 case Intrinsic::x86_avx2_psllv_d:
10544 case Intrinsic::x86_avx2_psllv_q:
10545 case Intrinsic::x86_avx2_psllv_d_256:
10546 case Intrinsic::x86_avx2_psllv_q_256:
10549 case Intrinsic::x86_avx2_psrlv_d:
10550 case Intrinsic::x86_avx2_psrlv_q:
10551 case Intrinsic::x86_avx2_psrlv_d_256:
10552 case Intrinsic::x86_avx2_psrlv_q_256:
10555 case Intrinsic::x86_avx2_psrav_d:
10556 case Intrinsic::x86_avx2_psrav_d_256:
10560 return DAG.getNode(Opcode, dl, Op.getValueType(),
10561 Op.getOperand(1), Op.getOperand(2));
10564 case Intrinsic::x86_ssse3_pshuf_b_128:
10565 case Intrinsic::x86_avx2_pshuf_b:
10566 return DAG.getNode(X86ISD::PSHUFB, dl, Op.getValueType(),
10567 Op.getOperand(1), Op.getOperand(2));
10569 case Intrinsic::x86_ssse3_psign_b_128:
10570 case Intrinsic::x86_ssse3_psign_w_128:
10571 case Intrinsic::x86_ssse3_psign_d_128:
10572 case Intrinsic::x86_avx2_psign_b:
10573 case Intrinsic::x86_avx2_psign_w:
10574 case Intrinsic::x86_avx2_psign_d:
10575 return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
10576 Op.getOperand(1), Op.getOperand(2));
10578 case Intrinsic::x86_sse41_insertps:
10579 return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
10580 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
10582 case Intrinsic::x86_avx_vperm2f128_ps_256:
10583 case Intrinsic::x86_avx_vperm2f128_pd_256:
10584 case Intrinsic::x86_avx_vperm2f128_si_256:
10585 case Intrinsic::x86_avx2_vperm2i128:
10586 return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
10587 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
10589 case Intrinsic::x86_avx2_permd:
10590 case Intrinsic::x86_avx2_permps:
10591 // Operands intentionally swapped. Mask is last operand to intrinsic,
10592 // but second operand for node/intruction.
10593 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
10594 Op.getOperand(2), Op.getOperand(1));
10596 case Intrinsic::x86_sse_sqrt_ps:
10597 case Intrinsic::x86_sse2_sqrt_pd:
10598 case Intrinsic::x86_avx_sqrt_ps_256:
10599 case Intrinsic::x86_avx_sqrt_pd_256:
10600 return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
10602 // ptest and testp intrinsics. The intrinsic these come from are designed to
10603 // return an integer value, not just an instruction so lower it to the ptest
10604 // or testp pattern and a setcc for the result.
10605 case Intrinsic::x86_sse41_ptestz:
10606 case Intrinsic::x86_sse41_ptestc:
10607 case Intrinsic::x86_sse41_ptestnzc:
10608 case Intrinsic::x86_avx_ptestz_256:
10609 case Intrinsic::x86_avx_ptestc_256:
10610 case Intrinsic::x86_avx_ptestnzc_256:
10611 case Intrinsic::x86_avx_vtestz_ps:
10612 case Intrinsic::x86_avx_vtestc_ps:
10613 case Intrinsic::x86_avx_vtestnzc_ps:
10614 case Intrinsic::x86_avx_vtestz_pd:
10615 case Intrinsic::x86_avx_vtestc_pd:
10616 case Intrinsic::x86_avx_vtestnzc_pd:
10617 case Intrinsic::x86_avx_vtestz_ps_256:
10618 case Intrinsic::x86_avx_vtestc_ps_256:
10619 case Intrinsic::x86_avx_vtestnzc_ps_256:
10620 case Intrinsic::x86_avx_vtestz_pd_256:
10621 case Intrinsic::x86_avx_vtestc_pd_256:
10622 case Intrinsic::x86_avx_vtestnzc_pd_256: {
10623 bool IsTestPacked = false;
10626 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
10627 case Intrinsic::x86_avx_vtestz_ps:
10628 case Intrinsic::x86_avx_vtestz_pd:
10629 case Intrinsic::x86_avx_vtestz_ps_256:
10630 case Intrinsic::x86_avx_vtestz_pd_256:
10631 IsTestPacked = true; // Fallthrough
10632 case Intrinsic::x86_sse41_ptestz:
10633 case Intrinsic::x86_avx_ptestz_256:
10635 X86CC = X86::COND_E;
10637 case Intrinsic::x86_avx_vtestc_ps:
10638 case Intrinsic::x86_avx_vtestc_pd:
10639 case Intrinsic::x86_avx_vtestc_ps_256:
10640 case Intrinsic::x86_avx_vtestc_pd_256:
10641 IsTestPacked = true; // Fallthrough
10642 case Intrinsic::x86_sse41_ptestc:
10643 case Intrinsic::x86_avx_ptestc_256:
10645 X86CC = X86::COND_B;
10647 case Intrinsic::x86_avx_vtestnzc_ps:
10648 case Intrinsic::x86_avx_vtestnzc_pd:
10649 case Intrinsic::x86_avx_vtestnzc_ps_256:
10650 case Intrinsic::x86_avx_vtestnzc_pd_256:
10651 IsTestPacked = true; // Fallthrough
10652 case Intrinsic::x86_sse41_ptestnzc:
10653 case Intrinsic::x86_avx_ptestnzc_256:
10655 X86CC = X86::COND_A;
10659 SDValue LHS = Op.getOperand(1);
10660 SDValue RHS = Op.getOperand(2);
10661 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
10662 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
10663 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
10664 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
10665 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10668 // SSE/AVX shift intrinsics
10669 case Intrinsic::x86_sse2_psll_w:
10670 case Intrinsic::x86_sse2_psll_d:
10671 case Intrinsic::x86_sse2_psll_q:
10672 case Intrinsic::x86_avx2_psll_w:
10673 case Intrinsic::x86_avx2_psll_d:
10674 case Intrinsic::x86_avx2_psll_q:
10675 case Intrinsic::x86_sse2_psrl_w:
10676 case Intrinsic::x86_sse2_psrl_d:
10677 case Intrinsic::x86_sse2_psrl_q:
10678 case Intrinsic::x86_avx2_psrl_w:
10679 case Intrinsic::x86_avx2_psrl_d:
10680 case Intrinsic::x86_avx2_psrl_q:
10681 case Intrinsic::x86_sse2_psra_w:
10682 case Intrinsic::x86_sse2_psra_d:
10683 case Intrinsic::x86_avx2_psra_w:
10684 case Intrinsic::x86_avx2_psra_d: {
10687 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10688 case Intrinsic::x86_sse2_psll_w:
10689 case Intrinsic::x86_sse2_psll_d:
10690 case Intrinsic::x86_sse2_psll_q:
10691 case Intrinsic::x86_avx2_psll_w:
10692 case Intrinsic::x86_avx2_psll_d:
10693 case Intrinsic::x86_avx2_psll_q:
10694 Opcode = X86ISD::VSHL;
10696 case Intrinsic::x86_sse2_psrl_w:
10697 case Intrinsic::x86_sse2_psrl_d:
10698 case Intrinsic::x86_sse2_psrl_q:
10699 case Intrinsic::x86_avx2_psrl_w:
10700 case Intrinsic::x86_avx2_psrl_d:
10701 case Intrinsic::x86_avx2_psrl_q:
10702 Opcode = X86ISD::VSRL;
10704 case Intrinsic::x86_sse2_psra_w:
10705 case Intrinsic::x86_sse2_psra_d:
10706 case Intrinsic::x86_avx2_psra_w:
10707 case Intrinsic::x86_avx2_psra_d:
10708 Opcode = X86ISD::VSRA;
10711 return DAG.getNode(Opcode, dl, Op.getValueType(),
10712 Op.getOperand(1), Op.getOperand(2));
10715 // SSE/AVX immediate shift intrinsics
10716 case Intrinsic::x86_sse2_pslli_w:
10717 case Intrinsic::x86_sse2_pslli_d:
10718 case Intrinsic::x86_sse2_pslli_q:
10719 case Intrinsic::x86_avx2_pslli_w:
10720 case Intrinsic::x86_avx2_pslli_d:
10721 case Intrinsic::x86_avx2_pslli_q:
10722 case Intrinsic::x86_sse2_psrli_w:
10723 case Intrinsic::x86_sse2_psrli_d:
10724 case Intrinsic::x86_sse2_psrli_q:
10725 case Intrinsic::x86_avx2_psrli_w:
10726 case Intrinsic::x86_avx2_psrli_d:
10727 case Intrinsic::x86_avx2_psrli_q:
10728 case Intrinsic::x86_sse2_psrai_w:
10729 case Intrinsic::x86_sse2_psrai_d:
10730 case Intrinsic::x86_avx2_psrai_w:
10731 case Intrinsic::x86_avx2_psrai_d: {
10734 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10735 case Intrinsic::x86_sse2_pslli_w:
10736 case Intrinsic::x86_sse2_pslli_d:
10737 case Intrinsic::x86_sse2_pslli_q:
10738 case Intrinsic::x86_avx2_pslli_w:
10739 case Intrinsic::x86_avx2_pslli_d:
10740 case Intrinsic::x86_avx2_pslli_q:
10741 Opcode = X86ISD::VSHLI;
10743 case Intrinsic::x86_sse2_psrli_w:
10744 case Intrinsic::x86_sse2_psrli_d:
10745 case Intrinsic::x86_sse2_psrli_q:
10746 case Intrinsic::x86_avx2_psrli_w:
10747 case Intrinsic::x86_avx2_psrli_d:
10748 case Intrinsic::x86_avx2_psrli_q:
10749 Opcode = X86ISD::VSRLI;
10751 case Intrinsic::x86_sse2_psrai_w:
10752 case Intrinsic::x86_sse2_psrai_d:
10753 case Intrinsic::x86_avx2_psrai_w:
10754 case Intrinsic::x86_avx2_psrai_d:
10755 Opcode = X86ISD::VSRAI;
10758 return getTargetVShiftNode(Opcode, dl, Op.getValueType(),
10759 Op.getOperand(1), Op.getOperand(2), DAG);
10762 case Intrinsic::x86_sse42_pcmpistria128:
10763 case Intrinsic::x86_sse42_pcmpestria128:
10764 case Intrinsic::x86_sse42_pcmpistric128:
10765 case Intrinsic::x86_sse42_pcmpestric128:
10766 case Intrinsic::x86_sse42_pcmpistrio128:
10767 case Intrinsic::x86_sse42_pcmpestrio128:
10768 case Intrinsic::x86_sse42_pcmpistris128:
10769 case Intrinsic::x86_sse42_pcmpestris128:
10770 case Intrinsic::x86_sse42_pcmpistriz128:
10771 case Intrinsic::x86_sse42_pcmpestriz128: {
10775 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10776 case Intrinsic::x86_sse42_pcmpistria128:
10777 Opcode = X86ISD::PCMPISTRI;
10778 X86CC = X86::COND_A;
10780 case Intrinsic::x86_sse42_pcmpestria128:
10781 Opcode = X86ISD::PCMPESTRI;
10782 X86CC = X86::COND_A;
10784 case Intrinsic::x86_sse42_pcmpistric128:
10785 Opcode = X86ISD::PCMPISTRI;
10786 X86CC = X86::COND_B;
10788 case Intrinsic::x86_sse42_pcmpestric128:
10789 Opcode = X86ISD::PCMPESTRI;
10790 X86CC = X86::COND_B;
10792 case Intrinsic::x86_sse42_pcmpistrio128:
10793 Opcode = X86ISD::PCMPISTRI;
10794 X86CC = X86::COND_O;
10796 case Intrinsic::x86_sse42_pcmpestrio128:
10797 Opcode = X86ISD::PCMPESTRI;
10798 X86CC = X86::COND_O;
10800 case Intrinsic::x86_sse42_pcmpistris128:
10801 Opcode = X86ISD::PCMPISTRI;
10802 X86CC = X86::COND_S;
10804 case Intrinsic::x86_sse42_pcmpestris128:
10805 Opcode = X86ISD::PCMPESTRI;
10806 X86CC = X86::COND_S;
10808 case Intrinsic::x86_sse42_pcmpistriz128:
10809 Opcode = X86ISD::PCMPISTRI;
10810 X86CC = X86::COND_E;
10812 case Intrinsic::x86_sse42_pcmpestriz128:
10813 Opcode = X86ISD::PCMPESTRI;
10814 X86CC = X86::COND_E;
10817 SmallVector<SDValue, 5> NewOps;
10818 NewOps.append(Op->op_begin()+1, Op->op_end());
10819 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
10820 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
10821 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10822 DAG.getConstant(X86CC, MVT::i8),
10823 SDValue(PCMP.getNode(), 1));
10824 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
10827 case Intrinsic::x86_sse42_pcmpistri128:
10828 case Intrinsic::x86_sse42_pcmpestri128: {
10830 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
10831 Opcode = X86ISD::PCMPISTRI;
10833 Opcode = X86ISD::PCMPESTRI;
10835 SmallVector<SDValue, 5> NewOps;
10836 NewOps.append(Op->op_begin()+1, Op->op_end());
10837 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
10838 return DAG.getNode(Opcode, dl, VTs, NewOps.data(), NewOps.size());
10840 case Intrinsic::x86_fma_vfmadd_ps:
10841 case Intrinsic::x86_fma_vfmadd_pd:
10842 case Intrinsic::x86_fma_vfmsub_ps:
10843 case Intrinsic::x86_fma_vfmsub_pd:
10844 case Intrinsic::x86_fma_vfnmadd_ps:
10845 case Intrinsic::x86_fma_vfnmadd_pd:
10846 case Intrinsic::x86_fma_vfnmsub_ps:
10847 case Intrinsic::x86_fma_vfnmsub_pd:
10848 case Intrinsic::x86_fma_vfmaddsub_ps:
10849 case Intrinsic::x86_fma_vfmaddsub_pd:
10850 case Intrinsic::x86_fma_vfmsubadd_ps:
10851 case Intrinsic::x86_fma_vfmsubadd_pd:
10852 case Intrinsic::x86_fma_vfmadd_ps_256:
10853 case Intrinsic::x86_fma_vfmadd_pd_256:
10854 case Intrinsic::x86_fma_vfmsub_ps_256:
10855 case Intrinsic::x86_fma_vfmsub_pd_256:
10856 case Intrinsic::x86_fma_vfnmadd_ps_256:
10857 case Intrinsic::x86_fma_vfnmadd_pd_256:
10858 case Intrinsic::x86_fma_vfnmsub_ps_256:
10859 case Intrinsic::x86_fma_vfnmsub_pd_256:
10860 case Intrinsic::x86_fma_vfmaddsub_ps_256:
10861 case Intrinsic::x86_fma_vfmaddsub_pd_256:
10862 case Intrinsic::x86_fma_vfmsubadd_ps_256:
10863 case Intrinsic::x86_fma_vfmsubadd_pd_256: {
10866 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
10867 case Intrinsic::x86_fma_vfmadd_ps:
10868 case Intrinsic::x86_fma_vfmadd_pd:
10869 case Intrinsic::x86_fma_vfmadd_ps_256:
10870 case Intrinsic::x86_fma_vfmadd_pd_256:
10871 Opc = X86ISD::FMADD;
10873 case Intrinsic::x86_fma_vfmsub_ps:
10874 case Intrinsic::x86_fma_vfmsub_pd:
10875 case Intrinsic::x86_fma_vfmsub_ps_256:
10876 case Intrinsic::x86_fma_vfmsub_pd_256:
10877 Opc = X86ISD::FMSUB;
10879 case Intrinsic::x86_fma_vfnmadd_ps:
10880 case Intrinsic::x86_fma_vfnmadd_pd:
10881 case Intrinsic::x86_fma_vfnmadd_ps_256:
10882 case Intrinsic::x86_fma_vfnmadd_pd_256:
10883 Opc = X86ISD::FNMADD;
10885 case Intrinsic::x86_fma_vfnmsub_ps:
10886 case Intrinsic::x86_fma_vfnmsub_pd:
10887 case Intrinsic::x86_fma_vfnmsub_ps_256:
10888 case Intrinsic::x86_fma_vfnmsub_pd_256:
10889 Opc = X86ISD::FNMSUB;
10891 case Intrinsic::x86_fma_vfmaddsub_ps:
10892 case Intrinsic::x86_fma_vfmaddsub_pd:
10893 case Intrinsic::x86_fma_vfmaddsub_ps_256:
10894 case Intrinsic::x86_fma_vfmaddsub_pd_256:
10895 Opc = X86ISD::FMADDSUB;
10897 case Intrinsic::x86_fma_vfmsubadd_ps:
10898 case Intrinsic::x86_fma_vfmsubadd_pd:
10899 case Intrinsic::x86_fma_vfmsubadd_ps_256:
10900 case Intrinsic::x86_fma_vfmsubadd_pd_256:
10901 Opc = X86ISD::FMSUBADD;
10905 return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
10906 Op.getOperand(2), Op.getOperand(3));
10911 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
10912 DebugLoc dl = Op.getDebugLoc();
10913 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10915 default: return SDValue(); // Don't custom lower most intrinsics.
10917 // RDRAND intrinsics.
10918 case Intrinsic::x86_rdrand_16:
10919 case Intrinsic::x86_rdrand_32:
10920 case Intrinsic::x86_rdrand_64: {
10921 // Emit the node with the right value type.
10922 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
10923 SDValue Result = DAG.getNode(X86ISD::RDRAND, dl, VTs, Op.getOperand(0));
10925 // If the value returned by RDRAND was valid (CF=1), return 1. Otherwise
10926 // return the value from Rand, which is always 0, casted to i32.
10927 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
10928 DAG.getConstant(1, Op->getValueType(1)),
10929 DAG.getConstant(X86::COND_B, MVT::i32),
10930 SDValue(Result.getNode(), 1) };
10931 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
10932 DAG.getVTList(Op->getValueType(1), MVT::Glue),
10935 // Return { result, isValid, chain }.
10936 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
10937 SDValue(Result.getNode(), 2));
10940 // XTEST intrinsics.
10941 case Intrinsic::x86_xtest: {
10942 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
10943 SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
10944 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
10945 DAG.getConstant(X86::COND_NE, MVT::i8),
10947 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
10948 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
10949 Ret, SDValue(InTrans.getNode(), 1));
10954 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
10955 SelectionDAG &DAG) const {
10956 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10957 MFI->setReturnAddressIsTaken(true);
10959 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10960 DebugLoc dl = Op.getDebugLoc();
10961 EVT PtrVT = getPointerTy();
10964 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
10966 DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
10967 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
10968 DAG.getNode(ISD::ADD, dl, PtrVT,
10969 FrameAddr, Offset),
10970 MachinePointerInfo(), false, false, false, 0);
10973 // Just load the return address.
10974 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
10975 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
10976 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
10979 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
10980 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10981 MFI->setFrameAddressIsTaken(true);
10983 EVT VT = Op.getValueType();
10984 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
10985 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
10986 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
10987 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
10989 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
10990 MachinePointerInfo(),
10991 false, false, false, 0);
10995 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
10996 SelectionDAG &DAG) const {
10997 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
11000 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
11001 SDValue Chain = Op.getOperand(0);
11002 SDValue Offset = Op.getOperand(1);
11003 SDValue Handler = Op.getOperand(2);
11004 DebugLoc dl = Op.getDebugLoc();
11006 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
11007 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
11009 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
11011 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
11012 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
11013 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
11014 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
11016 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
11018 return DAG.getNode(X86ISD::EH_RETURN, dl,
11020 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
11023 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
11024 SelectionDAG &DAG) const {
11025 DebugLoc DL = Op.getDebugLoc();
11026 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
11027 DAG.getVTList(MVT::i32, MVT::Other),
11028 Op.getOperand(0), Op.getOperand(1));
11031 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
11032 SelectionDAG &DAG) const {
11033 DebugLoc DL = Op.getDebugLoc();
11034 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
11035 Op.getOperand(0), Op.getOperand(1));
11038 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
11039 return Op.getOperand(0);
11042 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
11043 SelectionDAG &DAG) const {
11044 SDValue Root = Op.getOperand(0);
11045 SDValue Trmp = Op.getOperand(1); // trampoline
11046 SDValue FPtr = Op.getOperand(2); // nested function
11047 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
11048 DebugLoc dl = Op.getDebugLoc();
11050 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
11051 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
11053 if (Subtarget->is64Bit()) {
11054 SDValue OutChains[6];
11056 // Large code-model.
11057 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
11058 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
11060 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
11061 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
11063 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
11065 // Load the pointer to the nested function into R11.
11066 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
11067 SDValue Addr = Trmp;
11068 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11069 Addr, MachinePointerInfo(TrmpAddr),
11072 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11073 DAG.getConstant(2, MVT::i64));
11074 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
11075 MachinePointerInfo(TrmpAddr, 2),
11078 // Load the 'nest' parameter value into R10.
11079 // R10 is specified in X86CallingConv.td
11080 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
11081 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11082 DAG.getConstant(10, MVT::i64));
11083 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11084 Addr, MachinePointerInfo(TrmpAddr, 10),
11087 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11088 DAG.getConstant(12, MVT::i64));
11089 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
11090 MachinePointerInfo(TrmpAddr, 12),
11093 // Jump to the nested function.
11094 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
11095 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11096 DAG.getConstant(20, MVT::i64));
11097 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
11098 Addr, MachinePointerInfo(TrmpAddr, 20),
11101 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
11102 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
11103 DAG.getConstant(22, MVT::i64));
11104 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
11105 MachinePointerInfo(TrmpAddr, 22),
11108 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6);
11110 const Function *Func =
11111 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
11112 CallingConv::ID CC = Func->getCallingConv();
11117 llvm_unreachable("Unsupported calling convention");
11118 case CallingConv::C:
11119 case CallingConv::X86_StdCall: {
11120 // Pass 'nest' parameter in ECX.
11121 // Must be kept in sync with X86CallingConv.td
11122 NestReg = X86::ECX;
11124 // Check that ECX wasn't needed by an 'inreg' parameter.
11125 FunctionType *FTy = Func->getFunctionType();
11126 const AttributeSet &Attrs = Func->getAttributes();
11128 if (!Attrs.isEmpty() && !Func->isVarArg()) {
11129 unsigned InRegCount = 0;
11132 for (FunctionType::param_iterator I = FTy->param_begin(),
11133 E = FTy->param_end(); I != E; ++I, ++Idx)
11134 if (Attrs.hasAttribute(Idx, Attribute::InReg))
11135 // FIXME: should only count parameters that are lowered to integers.
11136 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
11138 if (InRegCount > 2) {
11139 report_fatal_error("Nest register in use - reduce number of inreg"
11145 case CallingConv::X86_FastCall:
11146 case CallingConv::X86_ThisCall:
11147 case CallingConv::Fast:
11148 // Pass 'nest' parameter in EAX.
11149 // Must be kept in sync with X86CallingConv.td
11150 NestReg = X86::EAX;
11154 SDValue OutChains[4];
11155 SDValue Addr, Disp;
11157 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11158 DAG.getConstant(10, MVT::i32));
11159 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
11161 // This is storing the opcode for MOV32ri.
11162 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
11163 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
11164 OutChains[0] = DAG.getStore(Root, dl,
11165 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
11166 Trmp, MachinePointerInfo(TrmpAddr),
11169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11170 DAG.getConstant(1, MVT::i32));
11171 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
11172 MachinePointerInfo(TrmpAddr, 1),
11175 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
11176 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11177 DAG.getConstant(5, MVT::i32));
11178 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
11179 MachinePointerInfo(TrmpAddr, 5),
11182 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
11183 DAG.getConstant(6, MVT::i32));
11184 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
11185 MachinePointerInfo(TrmpAddr, 6),
11188 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4);
11192 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
11193 SelectionDAG &DAG) const {
11195 The rounding mode is in bits 11:10 of FPSR, and has the following
11197 00 Round to nearest
11202 FLT_ROUNDS, on the other hand, expects the following:
11209 To perform the conversion, we do:
11210 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
11213 MachineFunction &MF = DAG.getMachineFunction();
11214 const TargetMachine &TM = MF.getTarget();
11215 const TargetFrameLowering &TFI = *TM.getFrameLowering();
11216 unsigned StackAlignment = TFI.getStackAlignment();
11217 EVT VT = Op.getValueType();
11218 DebugLoc DL = Op.getDebugLoc();
11220 // Save FP Control Word to stack slot
11221 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
11222 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11224 MachineMemOperand *MMO =
11225 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11226 MachineMemOperand::MOStore, 2, 2);
11228 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
11229 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
11230 DAG.getVTList(MVT::Other),
11231 Ops, 2, MVT::i16, MMO);
11233 // Load FP Control Word from stack slot
11234 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
11235 MachinePointerInfo(), false, false, false, 0);
11237 // Transform as necessary
11239 DAG.getNode(ISD::SRL, DL, MVT::i16,
11240 DAG.getNode(ISD::AND, DL, MVT::i16,
11241 CWD, DAG.getConstant(0x800, MVT::i16)),
11242 DAG.getConstant(11, MVT::i8));
11244 DAG.getNode(ISD::SRL, DL, MVT::i16,
11245 DAG.getNode(ISD::AND, DL, MVT::i16,
11246 CWD, DAG.getConstant(0x400, MVT::i16)),
11247 DAG.getConstant(9, MVT::i8));
11250 DAG.getNode(ISD::AND, DL, MVT::i16,
11251 DAG.getNode(ISD::ADD, DL, MVT::i16,
11252 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
11253 DAG.getConstant(1, MVT::i16)),
11254 DAG.getConstant(3, MVT::i16));
11256 return DAG.getNode((VT.getSizeInBits() < 16 ?
11257 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
11260 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
11261 EVT VT = Op.getValueType();
11263 unsigned NumBits = VT.getSizeInBits();
11264 DebugLoc dl = Op.getDebugLoc();
11266 Op = Op.getOperand(0);
11267 if (VT == MVT::i8) {
11268 // Zero extend to i32 since there is not an i8 bsr.
11270 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
11273 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
11274 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
11275 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
11277 // If src is zero (i.e. bsr sets ZF), returns NumBits.
11280 DAG.getConstant(NumBits+NumBits-1, OpVT),
11281 DAG.getConstant(X86::COND_E, MVT::i8),
11284 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
11286 // Finally xor with NumBits-1.
11287 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
11290 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
11294 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
11295 EVT VT = Op.getValueType();
11297 unsigned NumBits = VT.getSizeInBits();
11298 DebugLoc dl = Op.getDebugLoc();
11300 Op = Op.getOperand(0);
11301 if (VT == MVT::i8) {
11302 // Zero extend to i32 since there is not an i8 bsr.
11304 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
11307 // Issue a bsr (scan bits in reverse).
11308 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
11309 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
11311 // And xor with NumBits-1.
11312 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
11315 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
11319 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
11320 EVT VT = Op.getValueType();
11321 unsigned NumBits = VT.getSizeInBits();
11322 DebugLoc dl = Op.getDebugLoc();
11323 Op = Op.getOperand(0);
11325 // Issue a bsf (scan bits forward) which also sets EFLAGS.
11326 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
11327 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
11329 // If src is zero (i.e. bsf sets ZF), returns NumBits.
11332 DAG.getConstant(NumBits, VT),
11333 DAG.getConstant(X86::COND_E, MVT::i8),
11336 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops, array_lengthof(Ops));
11339 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
11340 // ones, and then concatenate the result back.
11341 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
11342 EVT VT = Op.getValueType();
11344 assert(VT.is256BitVector() && VT.isInteger() &&
11345 "Unsupported value type for operation");
11347 unsigned NumElems = VT.getVectorNumElements();
11348 DebugLoc dl = Op.getDebugLoc();
11350 // Extract the LHS vectors
11351 SDValue LHS = Op.getOperand(0);
11352 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
11353 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
11355 // Extract the RHS vectors
11356 SDValue RHS = Op.getOperand(1);
11357 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
11358 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
11360 MVT EltVT = VT.getVectorElementType().getSimpleVT();
11361 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
11363 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
11364 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
11365 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
11368 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
11369 assert(Op.getValueType().is256BitVector() &&
11370 Op.getValueType().isInteger() &&
11371 "Only handle AVX 256-bit vector integer operation");
11372 return Lower256IntArith(Op, DAG);
11375 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
11376 assert(Op.getValueType().is256BitVector() &&
11377 Op.getValueType().isInteger() &&
11378 "Only handle AVX 256-bit vector integer operation");
11379 return Lower256IntArith(Op, DAG);
11382 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
11383 SelectionDAG &DAG) {
11384 DebugLoc dl = Op.getDebugLoc();
11385 EVT VT = Op.getValueType();
11387 // Decompose 256-bit ops into smaller 128-bit ops.
11388 if (VT.is256BitVector() && !Subtarget->hasInt256())
11389 return Lower256IntArith(Op, DAG);
11391 SDValue A = Op.getOperand(0);
11392 SDValue B = Op.getOperand(1);
11394 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
11395 if (VT == MVT::v4i32) {
11396 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
11397 "Should not custom lower when pmuldq is available!");
11399 // Extract the odd parts.
11400 const int UnpackMask[] = { 1, -1, 3, -1 };
11401 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
11402 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
11404 // Multiply the even parts.
11405 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
11406 // Now multiply odd parts.
11407 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
11409 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
11410 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
11412 // Merge the two vectors back together with a shuffle. This expands into 2
11414 const int ShufMask[] = { 0, 4, 2, 6 };
11415 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
11418 assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
11419 "Only know how to lower V2I64/V4I64 multiply");
11421 // Ahi = psrlqi(a, 32);
11422 // Bhi = psrlqi(b, 32);
11424 // AloBlo = pmuludq(a, b);
11425 // AloBhi = pmuludq(a, Bhi);
11426 // AhiBlo = pmuludq(Ahi, b);
11428 // AloBhi = psllqi(AloBhi, 32);
11429 // AhiBlo = psllqi(AhiBlo, 32);
11430 // return AloBlo + AloBhi + AhiBlo;
11432 SDValue ShAmt = DAG.getConstant(32, MVT::i32);
11434 SDValue Ahi = DAG.getNode(X86ISD::VSRLI, dl, VT, A, ShAmt);
11435 SDValue Bhi = DAG.getNode(X86ISD::VSRLI, dl, VT, B, ShAmt);
11437 // Bit cast to 32-bit vectors for MULUDQ
11438 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 : MVT::v8i32;
11439 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
11440 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
11441 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
11442 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
11444 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
11445 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
11446 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
11448 AloBhi = DAG.getNode(X86ISD::VSHLI, dl, VT, AloBhi, ShAmt);
11449 AhiBlo = DAG.getNode(X86ISD::VSHLI, dl, VT, AhiBlo, ShAmt);
11451 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
11452 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
11455 SDValue X86TargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
11456 EVT VT = Op.getValueType();
11457 EVT EltTy = VT.getVectorElementType();
11458 unsigned NumElts = VT.getVectorNumElements();
11459 SDValue N0 = Op.getOperand(0);
11460 DebugLoc dl = Op.getDebugLoc();
11462 // Lower sdiv X, pow2-const.
11463 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(Op.getOperand(1));
11467 APInt SplatValue, SplatUndef;
11468 unsigned MinSplatBits;
11470 if (!C->isConstantSplat(SplatValue, SplatUndef, MinSplatBits, HasAnyUndefs))
11473 if ((SplatValue != 0) &&
11474 (SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
11475 unsigned lg2 = SplatValue.countTrailingZeros();
11476 // Splat the sign bit.
11477 SDValue Sz = DAG.getConstant(EltTy.getSizeInBits()-1, MVT::i32);
11478 SDValue SGN = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, N0, Sz, DAG);
11479 // Add (N0 < 0) ? abs2 - 1 : 0;
11480 SDValue Amt = DAG.getConstant(EltTy.getSizeInBits() - lg2, MVT::i32);
11481 SDValue SRL = getTargetVShiftNode(X86ISD::VSRLI, dl, VT, SGN, Amt, DAG);
11482 SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
11483 SDValue Lg2Amt = DAG.getConstant(lg2, MVT::i32);
11484 SDValue SRA = getTargetVShiftNode(X86ISD::VSRAI, dl, VT, ADD, Lg2Amt, DAG);
11486 // If we're dividing by a positive value, we're done. Otherwise, we must
11487 // negate the result.
11488 if (SplatValue.isNonNegative())
11491 SmallVector<SDValue, 16> V(NumElts, DAG.getConstant(0, EltTy));
11492 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], NumElts);
11493 return DAG.getNode(ISD::SUB, dl, VT, Zero, SRA);
11498 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
11499 const X86Subtarget *Subtarget) {
11500 EVT VT = Op.getValueType();
11501 DebugLoc dl = Op.getDebugLoc();
11502 SDValue R = Op.getOperand(0);
11503 SDValue Amt = Op.getOperand(1);
11505 // Optimize shl/srl/sra with constant shift amount.
11506 if (isSplatVector(Amt.getNode())) {
11507 SDValue SclrAmt = Amt->getOperand(0);
11508 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
11509 uint64_t ShiftAmt = C->getZExtValue();
11511 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
11512 (Subtarget->hasInt256() &&
11513 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16))) {
11514 if (Op.getOpcode() == ISD::SHL)
11515 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
11516 DAG.getConstant(ShiftAmt, MVT::i32));
11517 if (Op.getOpcode() == ISD::SRL)
11518 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
11519 DAG.getConstant(ShiftAmt, MVT::i32));
11520 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
11521 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
11522 DAG.getConstant(ShiftAmt, MVT::i32));
11525 if (VT == MVT::v16i8) {
11526 if (Op.getOpcode() == ISD::SHL) {
11527 // Make a large shift.
11528 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, R,
11529 DAG.getConstant(ShiftAmt, MVT::i32));
11530 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
11531 // Zero out the rightmost bits.
11532 SmallVector<SDValue, 16> V(16,
11533 DAG.getConstant(uint8_t(-1U << ShiftAmt),
11535 return DAG.getNode(ISD::AND, dl, VT, SHL,
11536 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
11538 if (Op.getOpcode() == ISD::SRL) {
11539 // Make a large shift.
11540 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v8i16, R,
11541 DAG.getConstant(ShiftAmt, MVT::i32));
11542 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
11543 // Zero out the leftmost bits.
11544 SmallVector<SDValue, 16> V(16,
11545 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
11547 return DAG.getNode(ISD::AND, dl, VT, SRL,
11548 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16));
11550 if (Op.getOpcode() == ISD::SRA) {
11551 if (ShiftAmt == 7) {
11552 // R s>> 7 === R s< 0
11553 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
11554 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
11557 // R s>> a === ((R u>> a) ^ m) - m
11558 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
11559 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
11561 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16);
11562 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
11563 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
11566 llvm_unreachable("Unknown shift opcode.");
11569 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
11570 if (Op.getOpcode() == ISD::SHL) {
11571 // Make a large shift.
11572 SDValue SHL = DAG.getNode(X86ISD::VSHLI, dl, MVT::v16i16, R,
11573 DAG.getConstant(ShiftAmt, MVT::i32));
11574 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
11575 // Zero out the rightmost bits.
11576 SmallVector<SDValue, 32> V(32,
11577 DAG.getConstant(uint8_t(-1U << ShiftAmt),
11579 return DAG.getNode(ISD::AND, dl, VT, SHL,
11580 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
11582 if (Op.getOpcode() == ISD::SRL) {
11583 // Make a large shift.
11584 SDValue SRL = DAG.getNode(X86ISD::VSRLI, dl, MVT::v16i16, R,
11585 DAG.getConstant(ShiftAmt, MVT::i32));
11586 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
11587 // Zero out the leftmost bits.
11588 SmallVector<SDValue, 32> V(32,
11589 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
11591 return DAG.getNode(ISD::AND, dl, VT, SRL,
11592 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32));
11594 if (Op.getOpcode() == ISD::SRA) {
11595 if (ShiftAmt == 7) {
11596 // R s>> 7 === R s< 0
11597 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
11598 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
11601 // R s>> a === ((R u>> a) ^ m) - m
11602 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
11603 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
11605 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 32);
11606 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
11607 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
11610 llvm_unreachable("Unknown shift opcode.");
11615 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
11616 if (!Subtarget->is64Bit() &&
11617 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
11618 Amt.getOpcode() == ISD::BITCAST &&
11619 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
11620 Amt = Amt.getOperand(0);
11621 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
11622 VT.getVectorNumElements();
11623 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
11624 uint64_t ShiftAmt = 0;
11625 for (unsigned i = 0; i != Ratio; ++i) {
11626 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
11630 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
11632 // Check remaining shift amounts.
11633 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
11634 uint64_t ShAmt = 0;
11635 for (unsigned j = 0; j != Ratio; ++j) {
11636 ConstantSDNode *C =
11637 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
11641 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
11643 if (ShAmt != ShiftAmt)
11646 switch (Op.getOpcode()) {
11648 llvm_unreachable("Unknown shift opcode!");
11650 return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
11651 DAG.getConstant(ShiftAmt, MVT::i32));
11653 return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
11654 DAG.getConstant(ShiftAmt, MVT::i32));
11656 return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
11657 DAG.getConstant(ShiftAmt, MVT::i32));
11664 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
11665 const X86Subtarget* Subtarget) {
11666 EVT VT = Op.getValueType();
11667 DebugLoc dl = Op.getDebugLoc();
11668 SDValue R = Op.getOperand(0);
11669 SDValue Amt = Op.getOperand(1);
11671 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
11672 VT == MVT::v4i32 || VT == MVT::v8i16 ||
11673 (Subtarget->hasInt256() &&
11674 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
11675 VT == MVT::v8i32 || VT == MVT::v16i16))) {
11677 EVT EltVT = VT.getVectorElementType();
11679 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
11680 unsigned NumElts = VT.getVectorNumElements();
11682 for (i = 0; i != NumElts; ++i) {
11683 if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
11687 for (j = i; j != NumElts; ++j) {
11688 SDValue Arg = Amt.getOperand(j);
11689 if (Arg.getOpcode() == ISD::UNDEF) continue;
11690 if (Arg != Amt.getOperand(i))
11693 if (i != NumElts && j == NumElts)
11694 BaseShAmt = Amt.getOperand(i);
11696 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
11697 Amt = Amt.getOperand(0);
11698 if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
11699 cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
11700 SDValue InVec = Amt.getOperand(0);
11701 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11702 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11704 for (; i != NumElts; ++i) {
11705 SDValue Arg = InVec.getOperand(i);
11706 if (Arg.getOpcode() == ISD::UNDEF) continue;
11710 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11711 if (ConstantSDNode *C =
11712 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11713 unsigned SplatIdx =
11714 cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
11715 if (C->getZExtValue() == SplatIdx)
11716 BaseShAmt = InVec.getOperand(1);
11719 if (BaseShAmt.getNode() == 0)
11720 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
11721 DAG.getIntPtrConstant(0));
11725 if (BaseShAmt.getNode()) {
11726 if (EltVT.bitsGT(MVT::i32))
11727 BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
11728 else if (EltVT.bitsLT(MVT::i32))
11729 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
11731 switch (Op.getOpcode()) {
11733 llvm_unreachable("Unknown shift opcode!");
11735 switch (VT.getSimpleVT().SimpleTy) {
11736 default: return SDValue();
11743 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
11746 switch (VT.getSimpleVT().SimpleTy) {
11747 default: return SDValue();
11752 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
11755 switch (VT.getSimpleVT().SimpleTy) {
11756 default: return SDValue();
11763 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
11769 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
11770 if (!Subtarget->is64Bit() &&
11771 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
11772 Amt.getOpcode() == ISD::BITCAST &&
11773 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
11774 Amt = Amt.getOperand(0);
11775 unsigned Ratio = Amt.getValueType().getVectorNumElements() /
11776 VT.getVectorNumElements();
11777 std::vector<SDValue> Vals(Ratio);
11778 for (unsigned i = 0; i != Ratio; ++i)
11779 Vals[i] = Amt.getOperand(i);
11780 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
11781 for (unsigned j = 0; j != Ratio; ++j)
11782 if (Vals[j] != Amt.getOperand(i + j))
11785 switch (Op.getOpcode()) {
11787 llvm_unreachable("Unknown shift opcode!");
11789 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
11791 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
11793 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
11800 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
11802 EVT VT = Op.getValueType();
11803 DebugLoc dl = Op.getDebugLoc();
11804 SDValue R = Op.getOperand(0);
11805 SDValue Amt = Op.getOperand(1);
11808 if (!Subtarget->hasSSE2())
11811 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
11815 V = LowerScalarVariableShift(Op, DAG, Subtarget);
11819 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
11820 if (Subtarget->hasInt256()) {
11821 if (Op.getOpcode() == ISD::SRL &&
11822 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
11823 VT == MVT::v4i64 || VT == MVT::v8i32))
11825 if (Op.getOpcode() == ISD::SHL &&
11826 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
11827 VT == MVT::v4i64 || VT == MVT::v8i32))
11829 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
11833 // Lower SHL with variable shift amount.
11834 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
11835 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
11837 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
11838 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
11839 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
11840 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
11842 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
11843 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
11846 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
11847 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
11849 // Turn 'a' into a mask suitable for VSELECT
11850 SDValue VSelM = DAG.getConstant(0x80, VT);
11851 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11852 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11854 SDValue CM1 = DAG.getConstant(0x0f, VT);
11855 SDValue CM2 = DAG.getConstant(0x3f, VT);
11857 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
11858 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
11859 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
11860 DAG.getConstant(4, MVT::i32), DAG);
11861 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
11862 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
11865 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
11866 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11867 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11869 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
11870 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
11871 M = getTargetVShiftNode(X86ISD::VSHLI, dl, MVT::v8i16, M,
11872 DAG.getConstant(2, MVT::i32), DAG);
11873 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
11874 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
11877 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
11878 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
11879 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
11881 // return VSELECT(r, r+r, a);
11882 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
11883 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
11887 // Decompose 256-bit shifts into smaller 128-bit shifts.
11888 if (VT.is256BitVector()) {
11889 unsigned NumElems = VT.getVectorNumElements();
11890 MVT EltVT = VT.getVectorElementType().getSimpleVT();
11891 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
11893 // Extract the two vectors
11894 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
11895 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
11897 // Recreate the shift amount vectors
11898 SDValue Amt1, Amt2;
11899 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
11900 // Constant shift amount
11901 SmallVector<SDValue, 4> Amt1Csts;
11902 SmallVector<SDValue, 4> Amt2Csts;
11903 for (unsigned i = 0; i != NumElems/2; ++i)
11904 Amt1Csts.push_back(Amt->getOperand(i));
11905 for (unsigned i = NumElems/2; i != NumElems; ++i)
11906 Amt2Csts.push_back(Amt->getOperand(i));
11908 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
11909 &Amt1Csts[0], NumElems/2);
11910 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
11911 &Amt2Csts[0], NumElems/2);
11913 // Variable shift amount
11914 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
11915 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
11918 // Issue new vector shifts for the smaller types
11919 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
11920 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
11922 // Concatenate the result back
11923 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
11929 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
11930 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
11931 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
11932 // looks for this combo and may remove the "setcc" instruction if the "setcc"
11933 // has only one use.
11934 SDNode *N = Op.getNode();
11935 SDValue LHS = N->getOperand(0);
11936 SDValue RHS = N->getOperand(1);
11937 unsigned BaseOp = 0;
11939 DebugLoc DL = Op.getDebugLoc();
11940 switch (Op.getOpcode()) {
11941 default: llvm_unreachable("Unknown ovf instruction!");
11943 // A subtract of one will be selected as a INC. Note that INC doesn't
11944 // set CF, so we can't do this for UADDO.
11945 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
11947 BaseOp = X86ISD::INC;
11948 Cond = X86::COND_O;
11951 BaseOp = X86ISD::ADD;
11952 Cond = X86::COND_O;
11955 BaseOp = X86ISD::ADD;
11956 Cond = X86::COND_B;
11959 // A subtract of one will be selected as a DEC. Note that DEC doesn't
11960 // set CF, so we can't do this for USUBO.
11961 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
11963 BaseOp = X86ISD::DEC;
11964 Cond = X86::COND_O;
11967 BaseOp = X86ISD::SUB;
11968 Cond = X86::COND_O;
11971 BaseOp = X86ISD::SUB;
11972 Cond = X86::COND_B;
11975 BaseOp = X86ISD::SMUL;
11976 Cond = X86::COND_O;
11978 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
11979 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
11981 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
11984 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11985 DAG.getConstant(X86::COND_O, MVT::i32),
11986 SDValue(Sum.getNode(), 2));
11988 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
11992 // Also sets EFLAGS.
11993 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
11994 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
11997 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
11998 DAG.getConstant(Cond, MVT::i32),
11999 SDValue(Sum.getNode(), 1));
12001 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
12004 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
12005 SelectionDAG &DAG) const {
12006 DebugLoc dl = Op.getDebugLoc();
12007 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
12008 EVT VT = Op.getValueType();
12010 if (!Subtarget->hasSSE2() || !VT.isVector())
12013 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
12014 ExtraVT.getScalarType().getSizeInBits();
12015 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
12017 switch (VT.getSimpleVT().SimpleTy) {
12018 default: return SDValue();
12021 if (!Subtarget->hasFp256())
12023 if (!Subtarget->hasInt256()) {
12024 // needs to be split
12025 unsigned NumElems = VT.getVectorNumElements();
12027 // Extract the LHS vectors
12028 SDValue LHS = Op.getOperand(0);
12029 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12030 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12032 MVT EltVT = VT.getVectorElementType().getSimpleVT();
12033 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12035 EVT ExtraEltVT = ExtraVT.getVectorElementType();
12036 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
12037 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
12039 SDValue Extra = DAG.getValueType(ExtraVT);
12041 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
12042 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
12044 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
12049 // (sext (vzext x)) -> (vsext x)
12050 SDValue Op0 = Op.getOperand(0);
12051 SDValue Op00 = Op0.getOperand(0);
12053 // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
12054 if (Op0.getOpcode() == ISD::BITCAST &&
12055 Op00.getOpcode() == ISD::VECTOR_SHUFFLE)
12056 Tmp1 = LowerVectorIntExtend(Op00, DAG);
12057 if (Tmp1.getNode()) {
12058 SDValue Tmp1Op0 = Tmp1.getOperand(0);
12059 assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
12060 "This optimization is invalid without a VZEXT.");
12061 return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
12064 // If the above didn't work, then just use Shift-Left + Shift-Right.
12065 Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT, Op0, ShAmt, DAG);
12066 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, Tmp1, ShAmt, DAG);
12071 static SDValue LowerMEMBARRIER(SDValue Op, const X86Subtarget *Subtarget,
12072 SelectionDAG &DAG) {
12073 DebugLoc dl = Op.getDebugLoc();
12075 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
12076 // There isn't any reason to disable it if the target processor supports it.
12077 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
12078 SDValue Chain = Op.getOperand(0);
12079 SDValue Zero = DAG.getConstant(0, MVT::i32);
12081 DAG.getRegister(X86::ESP, MVT::i32), // Base
12082 DAG.getTargetConstant(1, MVT::i8), // Scale
12083 DAG.getRegister(0, MVT::i32), // Index
12084 DAG.getTargetConstant(0, MVT::i32), // Disp
12085 DAG.getRegister(0, MVT::i32), // Segment.
12090 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
12091 array_lengthof(Ops));
12092 return SDValue(Res, 0);
12095 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
12097 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
12099 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12100 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
12101 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
12102 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
12104 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
12105 if (!Op1 && !Op2 && !Op3 && Op4)
12106 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
12108 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
12109 if (Op1 && !Op2 && !Op3 && !Op4)
12110 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
12112 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
12114 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
12117 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
12118 SelectionDAG &DAG) {
12119 DebugLoc dl = Op.getDebugLoc();
12120 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
12121 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
12122 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
12123 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
12125 // The only fence that needs an instruction is a sequentially-consistent
12126 // cross-thread fence.
12127 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
12128 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
12129 // no-sse2). There isn't any reason to disable it if the target processor
12131 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
12132 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
12134 SDValue Chain = Op.getOperand(0);
12135 SDValue Zero = DAG.getConstant(0, MVT::i32);
12137 DAG.getRegister(X86::ESP, MVT::i32), // Base
12138 DAG.getTargetConstant(1, MVT::i8), // Scale
12139 DAG.getRegister(0, MVT::i32), // Index
12140 DAG.getTargetConstant(0, MVT::i32), // Disp
12141 DAG.getRegister(0, MVT::i32), // Segment.
12146 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
12147 array_lengthof(Ops));
12148 return SDValue(Res, 0);
12151 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
12152 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
12155 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
12156 SelectionDAG &DAG) {
12157 EVT T = Op.getValueType();
12158 DebugLoc DL = Op.getDebugLoc();
12161 switch(T.getSimpleVT().SimpleTy) {
12162 default: llvm_unreachable("Invalid value type!");
12163 case MVT::i8: Reg = X86::AL; size = 1; break;
12164 case MVT::i16: Reg = X86::AX; size = 2; break;
12165 case MVT::i32: Reg = X86::EAX; size = 4; break;
12167 assert(Subtarget->is64Bit() && "Node not type legal!");
12168 Reg = X86::RAX; size = 8;
12171 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
12172 Op.getOperand(2), SDValue());
12173 SDValue Ops[] = { cpIn.getValue(0),
12176 DAG.getTargetConstant(size, MVT::i8),
12177 cpIn.getValue(1) };
12178 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12179 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
12180 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
12183 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
12187 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
12188 SelectionDAG &DAG) {
12189 assert(Subtarget->is64Bit() && "Result not type legalized?");
12190 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12191 SDValue TheChain = Op.getOperand(0);
12192 DebugLoc dl = Op.getDebugLoc();
12193 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
12194 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
12195 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
12197 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
12198 DAG.getConstant(32, MVT::i8));
12200 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
12203 return DAG.getMergeValues(Ops, 2, dl);
12206 SDValue X86TargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
12207 EVT SrcVT = Op.getOperand(0).getValueType();
12208 EVT DstVT = Op.getValueType();
12209 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
12210 Subtarget->hasMMX() && "Unexpected custom BITCAST");
12211 assert((DstVT == MVT::i64 ||
12212 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
12213 "Unexpected custom BITCAST");
12214 // i64 <=> MMX conversions are Legal.
12215 if (SrcVT==MVT::i64 && DstVT.isVector())
12217 if (DstVT==MVT::i64 && SrcVT.isVector())
12219 // MMX <=> MMX conversions are Legal.
12220 if (SrcVT.isVector() && DstVT.isVector())
12222 // All other conversions need to be expanded.
12226 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
12227 SDNode *Node = Op.getNode();
12228 DebugLoc dl = Node->getDebugLoc();
12229 EVT T = Node->getValueType(0);
12230 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
12231 DAG.getConstant(0, T), Node->getOperand(2));
12232 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
12233 cast<AtomicSDNode>(Node)->getMemoryVT(),
12234 Node->getOperand(0),
12235 Node->getOperand(1), negOp,
12236 cast<AtomicSDNode>(Node)->getSrcValue(),
12237 cast<AtomicSDNode>(Node)->getAlignment(),
12238 cast<AtomicSDNode>(Node)->getOrdering(),
12239 cast<AtomicSDNode>(Node)->getSynchScope());
12242 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
12243 SDNode *Node = Op.getNode();
12244 DebugLoc dl = Node->getDebugLoc();
12245 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
12247 // Convert seq_cst store -> xchg
12248 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
12249 // FIXME: On 32-bit, store -> fist or movq would be more efficient
12250 // (The only way to get a 16-byte store is cmpxchg16b)
12251 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
12252 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
12253 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12254 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
12255 cast<AtomicSDNode>(Node)->getMemoryVT(),
12256 Node->getOperand(0),
12257 Node->getOperand(1), Node->getOperand(2),
12258 cast<AtomicSDNode>(Node)->getMemOperand(),
12259 cast<AtomicSDNode>(Node)->getOrdering(),
12260 cast<AtomicSDNode>(Node)->getSynchScope());
12261 return Swap.getValue(1);
12263 // Other atomic stores have a simple pattern.
12267 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
12268 EVT VT = Op.getNode()->getValueType(0);
12270 // Let legalize expand this if it isn't a legal type yet.
12271 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
12274 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
12277 bool ExtraOp = false;
12278 switch (Op.getOpcode()) {
12279 default: llvm_unreachable("Invalid code");
12280 case ISD::ADDC: Opc = X86ISD::ADD; break;
12281 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
12282 case ISD::SUBC: Opc = X86ISD::SUB; break;
12283 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
12287 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
12289 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
12290 Op.getOperand(1), Op.getOperand(2));
12293 SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
12294 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
12296 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
12297 // which returns the values in two XMM registers.
12298 DebugLoc dl = Op.getDebugLoc();
12299 SDValue Arg = Op.getOperand(0);
12300 EVT ArgVT = Arg.getValueType();
12301 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
12304 ArgListEntry Entry;
12308 Entry.isSExt = false;
12309 Entry.isZExt = false;
12310 Args.push_back(Entry);
12312 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
12313 // the small struct {f32, f32} is returned in (eax, edx). For f64,
12314 // the results are returned via SRet in memory.
12315 const char *LibcallName = (ArgVT == MVT::f64)
12316 ? "__sincos_stret" : "__sincosf_stret";
12317 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
12319 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
12321 CallLoweringInfo CLI(DAG.getEntryNode(), RetTy,
12322 false, false, false, false, 0,
12323 CallingConv::C, /*isTaillCall=*/false,
12324 /*doesNotRet=*/false, /*isReturnValueUsed*/true,
12325 Callee, Args, DAG, dl);
12326 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
12327 return CallResult.first;
12330 /// LowerOperation - Provide custom lowering hooks for some operations.
12332 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
12333 switch (Op.getOpcode()) {
12334 default: llvm_unreachable("Should not custom lower this!");
12335 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
12336 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, Subtarget, DAG);
12337 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
12338 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op, Subtarget, DAG);
12339 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
12340 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
12341 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
12342 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
12343 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
12344 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
12345 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
12346 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
12347 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
12348 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
12349 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
12350 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
12351 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
12352 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
12353 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
12354 case ISD::SHL_PARTS:
12355 case ISD::SRA_PARTS:
12356 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
12357 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
12358 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
12359 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
12360 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
12361 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
12362 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, DAG);
12363 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
12364 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
12365 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
12366 case ISD::FABS: return LowerFABS(Op, DAG);
12367 case ISD::FNEG: return LowerFNEG(Op, DAG);
12368 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
12369 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
12370 case ISD::SETCC: return LowerSETCC(Op, DAG);
12371 case ISD::SELECT: return LowerSELECT(Op, DAG);
12372 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
12373 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
12374 case ISD::VASTART: return LowerVASTART(Op, DAG);
12375 case ISD::VAARG: return LowerVAARG(Op, DAG);
12376 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
12377 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
12378 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
12379 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
12380 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
12381 case ISD::FRAME_TO_ARGS_OFFSET:
12382 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
12383 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
12384 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
12385 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
12386 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
12387 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
12388 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
12389 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
12390 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
12391 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
12392 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
12393 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
12396 case ISD::SHL: return LowerShift(Op, DAG);
12402 case ISD::UMULO: return LowerXALUO(Op, DAG);
12403 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
12404 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
12408 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
12409 case ISD::ADD: return LowerADD(Op, DAG);
12410 case ISD::SUB: return LowerSUB(Op, DAG);
12411 case ISD::SDIV: return LowerSDIV(Op, DAG);
12412 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
12416 static void ReplaceATOMIC_LOAD(SDNode *Node,
12417 SmallVectorImpl<SDValue> &Results,
12418 SelectionDAG &DAG) {
12419 DebugLoc dl = Node->getDebugLoc();
12420 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
12422 // Convert wide load -> cmpxchg8b/cmpxchg16b
12423 // FIXME: On 32-bit, load -> fild or movq would be more efficient
12424 // (The only way to get a 16-byte load is cmpxchg16b)
12425 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
12426 SDValue Zero = DAG.getConstant(0, VT);
12427 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
12428 Node->getOperand(0),
12429 Node->getOperand(1), Zero, Zero,
12430 cast<AtomicSDNode>(Node)->getMemOperand(),
12431 cast<AtomicSDNode>(Node)->getOrdering(),
12432 cast<AtomicSDNode>(Node)->getSynchScope());
12433 Results.push_back(Swap.getValue(0));
12434 Results.push_back(Swap.getValue(1));
12438 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
12439 SelectionDAG &DAG, unsigned NewOp) {
12440 DebugLoc dl = Node->getDebugLoc();
12441 assert (Node->getValueType(0) == MVT::i64 &&
12442 "Only know how to expand i64 atomics");
12444 SDValue Chain = Node->getOperand(0);
12445 SDValue In1 = Node->getOperand(1);
12446 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
12447 Node->getOperand(2), DAG.getIntPtrConstant(0));
12448 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
12449 Node->getOperand(2), DAG.getIntPtrConstant(1));
12450 SDValue Ops[] = { Chain, In1, In2L, In2H };
12451 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
12453 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
12454 cast<MemSDNode>(Node)->getMemOperand());
12455 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
12456 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
12457 Results.push_back(Result.getValue(2));
12460 /// ReplaceNodeResults - Replace a node with an illegal result type
12461 /// with a new node built out of custom code.
12462 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
12463 SmallVectorImpl<SDValue>&Results,
12464 SelectionDAG &DAG) const {
12465 DebugLoc dl = N->getDebugLoc();
12466 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12467 switch (N->getOpcode()) {
12469 llvm_unreachable("Do not know how to custom type legalize this operation!");
12470 case ISD::SIGN_EXTEND_INREG:
12475 // We don't want to expand or promote these.
12477 case ISD::FP_TO_SINT:
12478 case ISD::FP_TO_UINT: {
12479 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
12481 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
12484 std::pair<SDValue,SDValue> Vals =
12485 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
12486 SDValue FIST = Vals.first, StackSlot = Vals.second;
12487 if (FIST.getNode() != 0) {
12488 EVT VT = N->getValueType(0);
12489 // Return a load from the stack slot.
12490 if (StackSlot.getNode() != 0)
12491 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
12492 MachinePointerInfo(),
12493 false, false, false, 0));
12495 Results.push_back(FIST);
12499 case ISD::UINT_TO_FP: {
12500 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
12501 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
12502 N->getValueType(0) != MVT::v2f32)
12504 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
12506 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
12508 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
12509 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
12510 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
12511 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
12512 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
12513 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
12516 case ISD::FP_ROUND: {
12517 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
12519 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
12520 Results.push_back(V);
12523 case ISD::READCYCLECOUNTER: {
12524 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12525 SDValue TheChain = N->getOperand(0);
12526 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
12527 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
12529 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
12531 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
12532 SDValue Ops[] = { eax, edx };
12533 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
12534 Results.push_back(edx.getValue(1));
12537 case ISD::ATOMIC_CMP_SWAP: {
12538 EVT T = N->getValueType(0);
12539 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
12540 bool Regs64bit = T == MVT::i128;
12541 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
12542 SDValue cpInL, cpInH;
12543 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
12544 DAG.getConstant(0, HalfT));
12545 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
12546 DAG.getConstant(1, HalfT));
12547 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
12548 Regs64bit ? X86::RAX : X86::EAX,
12550 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
12551 Regs64bit ? X86::RDX : X86::EDX,
12552 cpInH, cpInL.getValue(1));
12553 SDValue swapInL, swapInH;
12554 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
12555 DAG.getConstant(0, HalfT));
12556 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
12557 DAG.getConstant(1, HalfT));
12558 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
12559 Regs64bit ? X86::RBX : X86::EBX,
12560 swapInL, cpInH.getValue(1));
12561 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
12562 Regs64bit ? X86::RCX : X86::ECX,
12563 swapInH, swapInL.getValue(1));
12564 SDValue Ops[] = { swapInH.getValue(0),
12566 swapInH.getValue(1) };
12567 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
12568 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
12569 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
12570 X86ISD::LCMPXCHG8_DAG;
12571 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
12573 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
12574 Regs64bit ? X86::RAX : X86::EAX,
12575 HalfT, Result.getValue(1));
12576 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
12577 Regs64bit ? X86::RDX : X86::EDX,
12578 HalfT, cpOutL.getValue(2));
12579 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
12580 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
12581 Results.push_back(cpOutH.getValue(1));
12584 case ISD::ATOMIC_LOAD_ADD:
12585 case ISD::ATOMIC_LOAD_AND:
12586 case ISD::ATOMIC_LOAD_NAND:
12587 case ISD::ATOMIC_LOAD_OR:
12588 case ISD::ATOMIC_LOAD_SUB:
12589 case ISD::ATOMIC_LOAD_XOR:
12590 case ISD::ATOMIC_LOAD_MAX:
12591 case ISD::ATOMIC_LOAD_MIN:
12592 case ISD::ATOMIC_LOAD_UMAX:
12593 case ISD::ATOMIC_LOAD_UMIN:
12594 case ISD::ATOMIC_SWAP: {
12596 switch (N->getOpcode()) {
12597 default: llvm_unreachable("Unexpected opcode");
12598 case ISD::ATOMIC_LOAD_ADD:
12599 Opc = X86ISD::ATOMADD64_DAG;
12601 case ISD::ATOMIC_LOAD_AND:
12602 Opc = X86ISD::ATOMAND64_DAG;
12604 case ISD::ATOMIC_LOAD_NAND:
12605 Opc = X86ISD::ATOMNAND64_DAG;
12607 case ISD::ATOMIC_LOAD_OR:
12608 Opc = X86ISD::ATOMOR64_DAG;
12610 case ISD::ATOMIC_LOAD_SUB:
12611 Opc = X86ISD::ATOMSUB64_DAG;
12613 case ISD::ATOMIC_LOAD_XOR:
12614 Opc = X86ISD::ATOMXOR64_DAG;
12616 case ISD::ATOMIC_LOAD_MAX:
12617 Opc = X86ISD::ATOMMAX64_DAG;
12619 case ISD::ATOMIC_LOAD_MIN:
12620 Opc = X86ISD::ATOMMIN64_DAG;
12622 case ISD::ATOMIC_LOAD_UMAX:
12623 Opc = X86ISD::ATOMUMAX64_DAG;
12625 case ISD::ATOMIC_LOAD_UMIN:
12626 Opc = X86ISD::ATOMUMIN64_DAG;
12628 case ISD::ATOMIC_SWAP:
12629 Opc = X86ISD::ATOMSWAP64_DAG;
12632 ReplaceATOMIC_BINARY_64(N, Results, DAG, Opc);
12635 case ISD::ATOMIC_LOAD:
12636 ReplaceATOMIC_LOAD(N, Results, DAG);
12640 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
12642 default: return NULL;
12643 case X86ISD::BSF: return "X86ISD::BSF";
12644 case X86ISD::BSR: return "X86ISD::BSR";
12645 case X86ISD::SHLD: return "X86ISD::SHLD";
12646 case X86ISD::SHRD: return "X86ISD::SHRD";
12647 case X86ISD::FAND: return "X86ISD::FAND";
12648 case X86ISD::FOR: return "X86ISD::FOR";
12649 case X86ISD::FXOR: return "X86ISD::FXOR";
12650 case X86ISD::FSRL: return "X86ISD::FSRL";
12651 case X86ISD::FILD: return "X86ISD::FILD";
12652 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
12653 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
12654 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
12655 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
12656 case X86ISD::FLD: return "X86ISD::FLD";
12657 case X86ISD::FST: return "X86ISD::FST";
12658 case X86ISD::CALL: return "X86ISD::CALL";
12659 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
12660 case X86ISD::BT: return "X86ISD::BT";
12661 case X86ISD::CMP: return "X86ISD::CMP";
12662 case X86ISD::COMI: return "X86ISD::COMI";
12663 case X86ISD::UCOMI: return "X86ISD::UCOMI";
12664 case X86ISD::SETCC: return "X86ISD::SETCC";
12665 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
12666 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
12667 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
12668 case X86ISD::CMOV: return "X86ISD::CMOV";
12669 case X86ISD::BRCOND: return "X86ISD::BRCOND";
12670 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
12671 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
12672 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
12673 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
12674 case X86ISD::Wrapper: return "X86ISD::Wrapper";
12675 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
12676 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
12677 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
12678 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
12679 case X86ISD::PINSRB: return "X86ISD::PINSRB";
12680 case X86ISD::PINSRW: return "X86ISD::PINSRW";
12681 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
12682 case X86ISD::ANDNP: return "X86ISD::ANDNP";
12683 case X86ISD::PSIGN: return "X86ISD::PSIGN";
12684 case X86ISD::BLENDV: return "X86ISD::BLENDV";
12685 case X86ISD::BLENDI: return "X86ISD::BLENDI";
12686 case X86ISD::SUBUS: return "X86ISD::SUBUS";
12687 case X86ISD::HADD: return "X86ISD::HADD";
12688 case X86ISD::HSUB: return "X86ISD::HSUB";
12689 case X86ISD::FHADD: return "X86ISD::FHADD";
12690 case X86ISD::FHSUB: return "X86ISD::FHSUB";
12691 case X86ISD::UMAX: return "X86ISD::UMAX";
12692 case X86ISD::UMIN: return "X86ISD::UMIN";
12693 case X86ISD::SMAX: return "X86ISD::SMAX";
12694 case X86ISD::SMIN: return "X86ISD::SMIN";
12695 case X86ISD::FMAX: return "X86ISD::FMAX";
12696 case X86ISD::FMIN: return "X86ISD::FMIN";
12697 case X86ISD::FMAXC: return "X86ISD::FMAXC";
12698 case X86ISD::FMINC: return "X86ISD::FMINC";
12699 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
12700 case X86ISD::FRCP: return "X86ISD::FRCP";
12701 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
12702 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
12703 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
12704 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
12705 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
12706 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
12707 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
12708 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
12709 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
12710 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
12711 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
12712 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
12713 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
12714 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
12715 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
12716 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
12717 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
12718 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
12719 case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
12720 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
12721 case X86ISD::VZEXT: return "X86ISD::VZEXT";
12722 case X86ISD::VSEXT: return "X86ISD::VSEXT";
12723 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
12724 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
12725 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
12726 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
12727 case X86ISD::VSHL: return "X86ISD::VSHL";
12728 case X86ISD::VSRL: return "X86ISD::VSRL";
12729 case X86ISD::VSRA: return "X86ISD::VSRA";
12730 case X86ISD::VSHLI: return "X86ISD::VSHLI";
12731 case X86ISD::VSRLI: return "X86ISD::VSRLI";
12732 case X86ISD::VSRAI: return "X86ISD::VSRAI";
12733 case X86ISD::CMPP: return "X86ISD::CMPP";
12734 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
12735 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
12736 case X86ISD::ADD: return "X86ISD::ADD";
12737 case X86ISD::SUB: return "X86ISD::SUB";
12738 case X86ISD::ADC: return "X86ISD::ADC";
12739 case X86ISD::SBB: return "X86ISD::SBB";
12740 case X86ISD::SMUL: return "X86ISD::SMUL";
12741 case X86ISD::UMUL: return "X86ISD::UMUL";
12742 case X86ISD::INC: return "X86ISD::INC";
12743 case X86ISD::DEC: return "X86ISD::DEC";
12744 case X86ISD::OR: return "X86ISD::OR";
12745 case X86ISD::XOR: return "X86ISD::XOR";
12746 case X86ISD::AND: return "X86ISD::AND";
12747 case X86ISD::BLSI: return "X86ISD::BLSI";
12748 case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
12749 case X86ISD::BLSR: return "X86ISD::BLSR";
12750 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
12751 case X86ISD::PTEST: return "X86ISD::PTEST";
12752 case X86ISD::TESTP: return "X86ISD::TESTP";
12753 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
12754 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
12755 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
12756 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
12757 case X86ISD::SHUFP: return "X86ISD::SHUFP";
12758 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
12759 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
12760 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
12761 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
12762 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
12763 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
12764 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
12765 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
12766 case X86ISD::MOVSD: return "X86ISD::MOVSD";
12767 case X86ISD::MOVSS: return "X86ISD::MOVSS";
12768 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
12769 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
12770 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
12771 case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
12772 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
12773 case X86ISD::VPERMV: return "X86ISD::VPERMV";
12774 case X86ISD::VPERMI: return "X86ISD::VPERMI";
12775 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
12776 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
12777 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
12778 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
12779 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
12780 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
12781 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
12782 case X86ISD::SAHF: return "X86ISD::SAHF";
12783 case X86ISD::RDRAND: return "X86ISD::RDRAND";
12784 case X86ISD::FMADD: return "X86ISD::FMADD";
12785 case X86ISD::FMSUB: return "X86ISD::FMSUB";
12786 case X86ISD::FNMADD: return "X86ISD::FNMADD";
12787 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
12788 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
12789 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
12790 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
12791 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
12792 case X86ISD::XTEST: return "X86ISD::XTEST";
12796 // isLegalAddressingMode - Return true if the addressing mode represented
12797 // by AM is legal for this target, for a load/store of the specified type.
12798 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
12800 // X86 supports extremely general addressing modes.
12801 CodeModel::Model M = getTargetMachine().getCodeModel();
12802 Reloc::Model R = getTargetMachine().getRelocationModel();
12804 // X86 allows a sign-extended 32-bit immediate field as a displacement.
12805 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
12810 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
12812 // If a reference to this global requires an extra load, we can't fold it.
12813 if (isGlobalStubReference(GVFlags))
12816 // If BaseGV requires a register for the PIC base, we cannot also have a
12817 // BaseReg specified.
12818 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
12821 // If lower 4G is not available, then we must use rip-relative addressing.
12822 if ((M != CodeModel::Small || R != Reloc::Static) &&
12823 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
12827 switch (AM.Scale) {
12833 // These scales always work.
12838 // These scales are formed with basereg+scalereg. Only accept if there is
12843 default: // Other stuff never works.
12850 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
12851 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
12853 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
12854 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
12855 return NumBits1 > NumBits2;
12858 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
12859 return isInt<32>(Imm);
12862 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
12863 // Can also use sub to handle negated immediates.
12864 return isInt<32>(Imm);
12867 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
12868 if (!VT1.isInteger() || !VT2.isInteger())
12870 unsigned NumBits1 = VT1.getSizeInBits();
12871 unsigned NumBits2 = VT2.getSizeInBits();
12872 return NumBits1 > NumBits2;
12875 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
12876 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
12877 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
12880 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
12881 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
12882 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
12885 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
12886 EVT VT1 = Val.getValueType();
12887 if (isZExtFree(VT1, VT2))
12890 if (Val.getOpcode() != ISD::LOAD)
12893 if (!VT1.isSimple() || !VT1.isInteger() ||
12894 !VT2.isSimple() || !VT2.isInteger())
12897 switch (VT1.getSimpleVT().SimpleTy) {
12902 // X86 has 8, 16, and 32-bit zero-extending loads.
12909 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
12910 // i16 instructions are longer (0x66 prefix) and potentially slower.
12911 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
12914 /// isShuffleMaskLegal - Targets can use this to indicate that they only
12915 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
12916 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
12917 /// are assumed to be legal.
12919 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
12921 // Very little shuffling can be done for 64-bit vectors right now.
12922 if (VT.getSizeInBits() == 64)
12925 // FIXME: pshufb, blends, shifts.
12926 return (VT.getVectorNumElements() == 2 ||
12927 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
12928 isMOVLMask(M, VT) ||
12929 isSHUFPMask(M, VT, Subtarget->hasFp256()) ||
12930 isPSHUFDMask(M, VT) ||
12931 isPSHUFHWMask(M, VT, Subtarget->hasInt256()) ||
12932 isPSHUFLWMask(M, VT, Subtarget->hasInt256()) ||
12933 isPALIGNRMask(M, VT, Subtarget) ||
12934 isUNPCKLMask(M, VT, Subtarget->hasInt256()) ||
12935 isUNPCKHMask(M, VT, Subtarget->hasInt256()) ||
12936 isUNPCKL_v_undef_Mask(M, VT, Subtarget->hasInt256()) ||
12937 isUNPCKH_v_undef_Mask(M, VT, Subtarget->hasInt256()));
12941 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
12943 unsigned NumElts = VT.getVectorNumElements();
12944 // FIXME: This collection of masks seems suspect.
12947 if (NumElts == 4 && VT.is128BitVector()) {
12948 return (isMOVLMask(Mask, VT) ||
12949 isCommutedMOVLMask(Mask, VT, true) ||
12950 isSHUFPMask(Mask, VT, Subtarget->hasFp256()) ||
12951 isSHUFPMask(Mask, VT, Subtarget->hasFp256(), /* Commuted */ true));
12956 //===----------------------------------------------------------------------===//
12957 // X86 Scheduler Hooks
12958 //===----------------------------------------------------------------------===//
12960 /// Utility function to emit xbegin specifying the start of an RTM region.
12961 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
12962 const TargetInstrInfo *TII) {
12963 DebugLoc DL = MI->getDebugLoc();
12965 const BasicBlock *BB = MBB->getBasicBlock();
12966 MachineFunction::iterator I = MBB;
12969 // For the v = xbegin(), we generate
12980 MachineBasicBlock *thisMBB = MBB;
12981 MachineFunction *MF = MBB->getParent();
12982 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
12983 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
12984 MF->insert(I, mainMBB);
12985 MF->insert(I, sinkMBB);
12987 // Transfer the remainder of BB and its successor edges to sinkMBB.
12988 sinkMBB->splice(sinkMBB->begin(), MBB,
12989 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
12990 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
12994 // # fallthrough to mainMBB
12995 // # abortion to sinkMBB
12996 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
12997 thisMBB->addSuccessor(mainMBB);
12998 thisMBB->addSuccessor(sinkMBB);
13002 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
13003 mainMBB->addSuccessor(sinkMBB);
13006 // EAX is live into the sinkMBB
13007 sinkMBB->addLiveIn(X86::EAX);
13008 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13009 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
13012 MI->eraseFromParent();
13016 // Get CMPXCHG opcode for the specified data type.
13017 static unsigned getCmpXChgOpcode(EVT VT) {
13018 switch (VT.getSimpleVT().SimpleTy) {
13019 case MVT::i8: return X86::LCMPXCHG8;
13020 case MVT::i16: return X86::LCMPXCHG16;
13021 case MVT::i32: return X86::LCMPXCHG32;
13022 case MVT::i64: return X86::LCMPXCHG64;
13026 llvm_unreachable("Invalid operand size!");
13029 // Get LOAD opcode for the specified data type.
13030 static unsigned getLoadOpcode(EVT VT) {
13031 switch (VT.getSimpleVT().SimpleTy) {
13032 case MVT::i8: return X86::MOV8rm;
13033 case MVT::i16: return X86::MOV16rm;
13034 case MVT::i32: return X86::MOV32rm;
13035 case MVT::i64: return X86::MOV64rm;
13039 llvm_unreachable("Invalid operand size!");
13042 // Get opcode of the non-atomic one from the specified atomic instruction.
13043 static unsigned getNonAtomicOpcode(unsigned Opc) {
13045 case X86::ATOMAND8: return X86::AND8rr;
13046 case X86::ATOMAND16: return X86::AND16rr;
13047 case X86::ATOMAND32: return X86::AND32rr;
13048 case X86::ATOMAND64: return X86::AND64rr;
13049 case X86::ATOMOR8: return X86::OR8rr;
13050 case X86::ATOMOR16: return X86::OR16rr;
13051 case X86::ATOMOR32: return X86::OR32rr;
13052 case X86::ATOMOR64: return X86::OR64rr;
13053 case X86::ATOMXOR8: return X86::XOR8rr;
13054 case X86::ATOMXOR16: return X86::XOR16rr;
13055 case X86::ATOMXOR32: return X86::XOR32rr;
13056 case X86::ATOMXOR64: return X86::XOR64rr;
13058 llvm_unreachable("Unhandled atomic-load-op opcode!");
13061 // Get opcode of the non-atomic one from the specified atomic instruction with
13063 static unsigned getNonAtomicOpcodeWithExtraOpc(unsigned Opc,
13064 unsigned &ExtraOpc) {
13066 case X86::ATOMNAND8: ExtraOpc = X86::NOT8r; return X86::AND8rr;
13067 case X86::ATOMNAND16: ExtraOpc = X86::NOT16r; return X86::AND16rr;
13068 case X86::ATOMNAND32: ExtraOpc = X86::NOT32r; return X86::AND32rr;
13069 case X86::ATOMNAND64: ExtraOpc = X86::NOT64r; return X86::AND64rr;
13070 case X86::ATOMMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVL32rr;
13071 case X86::ATOMMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVL16rr;
13072 case X86::ATOMMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVL32rr;
13073 case X86::ATOMMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVL64rr;
13074 case X86::ATOMMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVG32rr;
13075 case X86::ATOMMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVG16rr;
13076 case X86::ATOMMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVG32rr;
13077 case X86::ATOMMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVG64rr;
13078 case X86::ATOMUMAX8: ExtraOpc = X86::CMP8rr; return X86::CMOVB32rr;
13079 case X86::ATOMUMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVB16rr;
13080 case X86::ATOMUMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVB32rr;
13081 case X86::ATOMUMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVB64rr;
13082 case X86::ATOMUMIN8: ExtraOpc = X86::CMP8rr; return X86::CMOVA32rr;
13083 case X86::ATOMUMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVA16rr;
13084 case X86::ATOMUMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVA32rr;
13085 case X86::ATOMUMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVA64rr;
13087 llvm_unreachable("Unhandled atomic-load-op opcode!");
13090 // Get opcode of the non-atomic one from the specified atomic instruction for
13091 // 64-bit data type on 32-bit target.
13092 static unsigned getNonAtomic6432Opcode(unsigned Opc, unsigned &HiOpc) {
13094 case X86::ATOMAND6432: HiOpc = X86::AND32rr; return X86::AND32rr;
13095 case X86::ATOMOR6432: HiOpc = X86::OR32rr; return X86::OR32rr;
13096 case X86::ATOMXOR6432: HiOpc = X86::XOR32rr; return X86::XOR32rr;
13097 case X86::ATOMADD6432: HiOpc = X86::ADC32rr; return X86::ADD32rr;
13098 case X86::ATOMSUB6432: HiOpc = X86::SBB32rr; return X86::SUB32rr;
13099 case X86::ATOMSWAP6432: HiOpc = X86::MOV32rr; return X86::MOV32rr;
13100 case X86::ATOMMAX6432: HiOpc = X86::SETLr; return X86::SETLr;
13101 case X86::ATOMMIN6432: HiOpc = X86::SETGr; return X86::SETGr;
13102 case X86::ATOMUMAX6432: HiOpc = X86::SETBr; return X86::SETBr;
13103 case X86::ATOMUMIN6432: HiOpc = X86::SETAr; return X86::SETAr;
13105 llvm_unreachable("Unhandled atomic-load-op opcode!");
13108 // Get opcode of the non-atomic one from the specified atomic instruction for
13109 // 64-bit data type on 32-bit target with extra opcode.
13110 static unsigned getNonAtomic6432OpcodeWithExtraOpc(unsigned Opc,
13112 unsigned &ExtraOpc) {
13114 case X86::ATOMNAND6432:
13115 ExtraOpc = X86::NOT32r;
13116 HiOpc = X86::AND32rr;
13117 return X86::AND32rr;
13119 llvm_unreachable("Unhandled atomic-load-op opcode!");
13122 // Get pseudo CMOV opcode from the specified data type.
13123 static unsigned getPseudoCMOVOpc(EVT VT) {
13124 switch (VT.getSimpleVT().SimpleTy) {
13125 case MVT::i8: return X86::CMOV_GR8;
13126 case MVT::i16: return X86::CMOV_GR16;
13127 case MVT::i32: return X86::CMOV_GR32;
13131 llvm_unreachable("Unknown CMOV opcode!");
13134 // EmitAtomicLoadArith - emit the code sequence for pseudo atomic instructions.
13135 // They will be translated into a spin-loop or compare-exchange loop from
13138 // dst = atomic-fetch-op MI.addr, MI.val
13144 // t1 = LOAD MI.addr
13146 // t4 = phi(t1, t3 / loop)
13147 // t2 = OP MI.val, t4
13149 // LCMPXCHG [MI.addr], t2, [EAX is implicitly used & defined]
13155 MachineBasicBlock *
13156 X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
13157 MachineBasicBlock *MBB) const {
13158 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13159 DebugLoc DL = MI->getDebugLoc();
13161 MachineFunction *MF = MBB->getParent();
13162 MachineRegisterInfo &MRI = MF->getRegInfo();
13164 const BasicBlock *BB = MBB->getBasicBlock();
13165 MachineFunction::iterator I = MBB;
13168 assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
13169 "Unexpected number of operands");
13171 assert(MI->hasOneMemOperand() &&
13172 "Expected atomic-load-op to have one memoperand");
13174 // Memory Reference
13175 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13176 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13178 unsigned DstReg, SrcReg;
13179 unsigned MemOpndSlot;
13181 unsigned CurOp = 0;
13183 DstReg = MI->getOperand(CurOp++).getReg();
13184 MemOpndSlot = CurOp;
13185 CurOp += X86::AddrNumOperands;
13186 SrcReg = MI->getOperand(CurOp++).getReg();
13188 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
13189 MVT::SimpleValueType VT = *RC->vt_begin();
13190 unsigned t1 = MRI.createVirtualRegister(RC);
13191 unsigned t2 = MRI.createVirtualRegister(RC);
13192 unsigned t3 = MRI.createVirtualRegister(RC);
13193 unsigned t4 = MRI.createVirtualRegister(RC);
13194 unsigned PhyReg = getX86SubSuperRegister(X86::EAX, VT);
13196 unsigned LCMPXCHGOpc = getCmpXChgOpcode(VT);
13197 unsigned LOADOpc = getLoadOpcode(VT);
13199 // For the atomic load-arith operator, we generate
13202 // t1 = LOAD [MI.addr]
13204 // t4 = phi(t1 / thisMBB, t3 / mainMBB)
13205 // t1 = OP MI.val, EAX
13207 // LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
13213 MachineBasicBlock *thisMBB = MBB;
13214 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13215 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13216 MF->insert(I, mainMBB);
13217 MF->insert(I, sinkMBB);
13219 MachineInstrBuilder MIB;
13221 // Transfer the remainder of BB and its successor edges to sinkMBB.
13222 sinkMBB->splice(sinkMBB->begin(), MBB,
13223 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13224 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13227 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1);
13228 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13229 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13231 NewMO.setIsKill(false);
13232 MIB.addOperand(NewMO);
13234 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
13235 unsigned flags = (*MMOI)->getFlags();
13236 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
13237 MachineMemOperand *MMO =
13238 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
13239 (*MMOI)->getSize(),
13240 (*MMOI)->getBaseAlignment(),
13241 (*MMOI)->getTBAAInfo(),
13242 (*MMOI)->getRanges());
13243 MIB.addMemOperand(MMO);
13246 thisMBB->addSuccessor(mainMBB);
13249 MachineBasicBlock *origMainMBB = mainMBB;
13252 MachineInstr *Phi = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4)
13253 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
13255 unsigned Opc = MI->getOpcode();
13258 llvm_unreachable("Unhandled atomic-load-op opcode!");
13259 case X86::ATOMAND8:
13260 case X86::ATOMAND16:
13261 case X86::ATOMAND32:
13262 case X86::ATOMAND64:
13264 case X86::ATOMOR16:
13265 case X86::ATOMOR32:
13266 case X86::ATOMOR64:
13267 case X86::ATOMXOR8:
13268 case X86::ATOMXOR16:
13269 case X86::ATOMXOR32:
13270 case X86::ATOMXOR64: {
13271 unsigned ARITHOpc = getNonAtomicOpcode(Opc);
13272 BuildMI(mainMBB, DL, TII->get(ARITHOpc), t2).addReg(SrcReg)
13276 case X86::ATOMNAND8:
13277 case X86::ATOMNAND16:
13278 case X86::ATOMNAND32:
13279 case X86::ATOMNAND64: {
13280 unsigned Tmp = MRI.createVirtualRegister(RC);
13282 unsigned ANDOpc = getNonAtomicOpcodeWithExtraOpc(Opc, NOTOpc);
13283 BuildMI(mainMBB, DL, TII->get(ANDOpc), Tmp).addReg(SrcReg)
13285 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2).addReg(Tmp);
13288 case X86::ATOMMAX8:
13289 case X86::ATOMMAX16:
13290 case X86::ATOMMAX32:
13291 case X86::ATOMMAX64:
13292 case X86::ATOMMIN8:
13293 case X86::ATOMMIN16:
13294 case X86::ATOMMIN32:
13295 case X86::ATOMMIN64:
13296 case X86::ATOMUMAX8:
13297 case X86::ATOMUMAX16:
13298 case X86::ATOMUMAX32:
13299 case X86::ATOMUMAX64:
13300 case X86::ATOMUMIN8:
13301 case X86::ATOMUMIN16:
13302 case X86::ATOMUMIN32:
13303 case X86::ATOMUMIN64: {
13305 unsigned CMOVOpc = getNonAtomicOpcodeWithExtraOpc(Opc, CMPOpc);
13307 BuildMI(mainMBB, DL, TII->get(CMPOpc))
13311 if (Subtarget->hasCMov()) {
13312 if (VT != MVT::i8) {
13314 BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
13318 // Promote i8 to i32 to use CMOV32
13319 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
13320 const TargetRegisterClass *RC32 =
13321 TRI->getSubClassWithSubReg(getRegClassFor(MVT::i32), X86::sub_8bit);
13322 unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
13323 unsigned AccReg32 = MRI.createVirtualRegister(RC32);
13324 unsigned Tmp = MRI.createVirtualRegister(RC32);
13326 unsigned Undef = MRI.createVirtualRegister(RC32);
13327 BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
13329 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), SrcReg32)
13332 .addImm(X86::sub_8bit);
13333 BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
13336 .addImm(X86::sub_8bit);
13338 BuildMI(mainMBB, DL, TII->get(CMOVOpc), Tmp)
13342 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t2)
13343 .addReg(Tmp, 0, X86::sub_8bit);
13346 // Use pseudo select and lower them.
13347 assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
13348 "Invalid atomic-load-op transformation!");
13349 unsigned SelOpc = getPseudoCMOVOpc(VT);
13350 X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
13351 assert(CC != X86::COND_INVALID && "Invalid atomic-load-op transformation!");
13352 MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t2)
13353 .addReg(SrcReg).addReg(t4)
13355 mainMBB = EmitLoweredSelect(MIB, mainMBB);
13356 // Replace the original PHI node as mainMBB is changed after CMOV
13358 BuildMI(*origMainMBB, Phi, DL, TII->get(X86::PHI), t4)
13359 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
13360 Phi->eraseFromParent();
13366 // Copy PhyReg back from virtual register.
13367 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), PhyReg)
13370 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
13371 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13372 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13374 NewMO.setIsKill(false);
13375 MIB.addOperand(NewMO);
13378 MIB.setMemRefs(MMOBegin, MMOEnd);
13380 // Copy PhyReg back to virtual register.
13381 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3)
13384 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
13386 mainMBB->addSuccessor(origMainMBB);
13387 mainMBB->addSuccessor(sinkMBB);
13390 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13391 TII->get(TargetOpcode::COPY), DstReg)
13394 MI->eraseFromParent();
13398 // EmitAtomicLoadArith6432 - emit the code sequence for pseudo atomic
13399 // instructions. They will be translated into a spin-loop or compare-exchange
13403 // dst = atomic-fetch-op MI.addr, MI.val
13409 // t1L = LOAD [MI.addr + 0]
13410 // t1H = LOAD [MI.addr + 4]
13412 // t4L = phi(t1L, t3L / loop)
13413 // t4H = phi(t1H, t3H / loop)
13414 // t2L = OP MI.val.lo, t4L
13415 // t2H = OP MI.val.hi, t4H
13420 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
13428 MachineBasicBlock *
13429 X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
13430 MachineBasicBlock *MBB) const {
13431 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13432 DebugLoc DL = MI->getDebugLoc();
13434 MachineFunction *MF = MBB->getParent();
13435 MachineRegisterInfo &MRI = MF->getRegInfo();
13437 const BasicBlock *BB = MBB->getBasicBlock();
13438 MachineFunction::iterator I = MBB;
13441 assert(MI->getNumOperands() <= X86::AddrNumOperands + 7 &&
13442 "Unexpected number of operands");
13444 assert(MI->hasOneMemOperand() &&
13445 "Expected atomic-load-op32 to have one memoperand");
13447 // Memory Reference
13448 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13449 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13451 unsigned DstLoReg, DstHiReg;
13452 unsigned SrcLoReg, SrcHiReg;
13453 unsigned MemOpndSlot;
13455 unsigned CurOp = 0;
13457 DstLoReg = MI->getOperand(CurOp++).getReg();
13458 DstHiReg = MI->getOperand(CurOp++).getReg();
13459 MemOpndSlot = CurOp;
13460 CurOp += X86::AddrNumOperands;
13461 SrcLoReg = MI->getOperand(CurOp++).getReg();
13462 SrcHiReg = MI->getOperand(CurOp++).getReg();
13464 const TargetRegisterClass *RC = &X86::GR32RegClass;
13465 const TargetRegisterClass *RC8 = &X86::GR8RegClass;
13467 unsigned t1L = MRI.createVirtualRegister(RC);
13468 unsigned t1H = MRI.createVirtualRegister(RC);
13469 unsigned t2L = MRI.createVirtualRegister(RC);
13470 unsigned t2H = MRI.createVirtualRegister(RC);
13471 unsigned t3L = MRI.createVirtualRegister(RC);
13472 unsigned t3H = MRI.createVirtualRegister(RC);
13473 unsigned t4L = MRI.createVirtualRegister(RC);
13474 unsigned t4H = MRI.createVirtualRegister(RC);
13476 unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
13477 unsigned LOADOpc = X86::MOV32rm;
13479 // For the atomic load-arith operator, we generate
13482 // t1L = LOAD [MI.addr + 0]
13483 // t1H = LOAD [MI.addr + 4]
13485 // t4L = phi(t1L / thisMBB, t3L / mainMBB)
13486 // t4H = phi(t1H / thisMBB, t3H / mainMBB)
13487 // t2L = OP MI.val.lo, t4L
13488 // t2H = OP MI.val.hi, t4H
13491 // LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
13499 MachineBasicBlock *thisMBB = MBB;
13500 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
13501 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
13502 MF->insert(I, mainMBB);
13503 MF->insert(I, sinkMBB);
13505 MachineInstrBuilder MIB;
13507 // Transfer the remainder of BB and its successor edges to sinkMBB.
13508 sinkMBB->splice(sinkMBB->begin(), MBB,
13509 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
13510 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
13514 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1L);
13515 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13516 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13518 NewMO.setIsKill(false);
13519 MIB.addOperand(NewMO);
13521 for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
13522 unsigned flags = (*MMOI)->getFlags();
13523 flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
13524 MachineMemOperand *MMO =
13525 MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
13526 (*MMOI)->getSize(),
13527 (*MMOI)->getBaseAlignment(),
13528 (*MMOI)->getTBAAInfo(),
13529 (*MMOI)->getRanges());
13530 MIB.addMemOperand(MMO);
13532 MachineInstr *LowMI = MIB;
13535 MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1H);
13536 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13537 if (i == X86::AddrDisp) {
13538 MIB.addDisp(MI->getOperand(MemOpndSlot + i), 4); // 4 == sizeof(i32)
13540 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13542 NewMO.setIsKill(false);
13543 MIB.addOperand(NewMO);
13546 MIB.setMemRefs(LowMI->memoperands_begin(), LowMI->memoperands_end());
13548 thisMBB->addSuccessor(mainMBB);
13551 MachineBasicBlock *origMainMBB = mainMBB;
13554 MachineInstr *PhiL = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4L)
13555 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
13556 MachineInstr *PhiH = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4H)
13557 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
13559 unsigned Opc = MI->getOpcode();
13562 llvm_unreachable("Unhandled atomic-load-op6432 opcode!");
13563 case X86::ATOMAND6432:
13564 case X86::ATOMOR6432:
13565 case X86::ATOMXOR6432:
13566 case X86::ATOMADD6432:
13567 case X86::ATOMSUB6432: {
13569 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
13570 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(t4L)
13572 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(t4H)
13576 case X86::ATOMNAND6432: {
13577 unsigned HiOpc, NOTOpc;
13578 unsigned LoOpc = getNonAtomic6432OpcodeWithExtraOpc(Opc, HiOpc, NOTOpc);
13579 unsigned TmpL = MRI.createVirtualRegister(RC);
13580 unsigned TmpH = MRI.createVirtualRegister(RC);
13581 BuildMI(mainMBB, DL, TII->get(LoOpc), TmpL).addReg(SrcLoReg)
13583 BuildMI(mainMBB, DL, TII->get(HiOpc), TmpH).addReg(SrcHiReg)
13585 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2L).addReg(TmpL);
13586 BuildMI(mainMBB, DL, TII->get(NOTOpc), t2H).addReg(TmpH);
13589 case X86::ATOMMAX6432:
13590 case X86::ATOMMIN6432:
13591 case X86::ATOMUMAX6432:
13592 case X86::ATOMUMIN6432: {
13594 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
13595 unsigned cL = MRI.createVirtualRegister(RC8);
13596 unsigned cH = MRI.createVirtualRegister(RC8);
13597 unsigned cL32 = MRI.createVirtualRegister(RC);
13598 unsigned cH32 = MRI.createVirtualRegister(RC);
13599 unsigned cc = MRI.createVirtualRegister(RC);
13600 // cl := cmp src_lo, lo
13601 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
13602 .addReg(SrcLoReg).addReg(t4L);
13603 BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
13604 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
13605 // ch := cmp src_hi, hi
13606 BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
13607 .addReg(SrcHiReg).addReg(t4H);
13608 BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
13609 BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
13610 // cc := if (src_hi == hi) ? cl : ch;
13611 if (Subtarget->hasCMov()) {
13612 BuildMI(mainMBB, DL, TII->get(X86::CMOVE32rr), cc)
13613 .addReg(cH32).addReg(cL32);
13615 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), cc)
13616 .addReg(cH32).addReg(cL32)
13617 .addImm(X86::COND_E);
13618 mainMBB = EmitLoweredSelect(MIB, mainMBB);
13620 BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
13621 if (Subtarget->hasCMov()) {
13622 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2L)
13623 .addReg(SrcLoReg).addReg(t4L);
13624 BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2H)
13625 .addReg(SrcHiReg).addReg(t4H);
13627 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2L)
13628 .addReg(SrcLoReg).addReg(t4L)
13629 .addImm(X86::COND_NE);
13630 mainMBB = EmitLoweredSelect(MIB, mainMBB);
13631 // As the lowered CMOV won't clobber EFLAGS, we could reuse it for the
13632 // 2nd CMOV lowering.
13633 mainMBB->addLiveIn(X86::EFLAGS);
13634 MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2H)
13635 .addReg(SrcHiReg).addReg(t4H)
13636 .addImm(X86::COND_NE);
13637 mainMBB = EmitLoweredSelect(MIB, mainMBB);
13638 // Replace the original PHI node as mainMBB is changed after CMOV
13640 BuildMI(*origMainMBB, PhiL, DL, TII->get(X86::PHI), t4L)
13641 .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
13642 BuildMI(*origMainMBB, PhiH, DL, TII->get(X86::PHI), t4H)
13643 .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
13644 PhiL->eraseFromParent();
13645 PhiH->eraseFromParent();
13649 case X86::ATOMSWAP6432: {
13651 unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
13652 BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg);
13653 BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg);
13658 // Copy EDX:EAX back from HiReg:LoReg
13659 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(t4L);
13660 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(t4H);
13661 // Copy ECX:EBX from t1H:t1L
13662 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t2L);
13663 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t2H);
13665 MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
13666 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
13667 MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
13669 NewMO.setIsKill(false);
13670 MIB.addOperand(NewMO);
13672 MIB.setMemRefs(MMOBegin, MMOEnd);
13674 // Copy EDX:EAX back to t3H:t3L
13675 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3L).addReg(X86::EAX);
13676 BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3H).addReg(X86::EDX);
13678 BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
13680 mainMBB->addSuccessor(origMainMBB);
13681 mainMBB->addSuccessor(sinkMBB);
13684 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13685 TII->get(TargetOpcode::COPY), DstLoReg)
13687 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
13688 TII->get(TargetOpcode::COPY), DstHiReg)
13691 MI->eraseFromParent();
13695 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
13696 // or XMM0_V32I8 in AVX all of this code can be replaced with that
13697 // in the .td file.
13698 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
13699 const TargetInstrInfo *TII) {
13701 switch (MI->getOpcode()) {
13702 default: llvm_unreachable("illegal opcode!");
13703 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
13704 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
13705 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
13706 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
13707 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
13708 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
13709 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
13710 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
13713 DebugLoc dl = MI->getDebugLoc();
13714 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
13716 unsigned NumArgs = MI->getNumOperands();
13717 for (unsigned i = 1; i < NumArgs; ++i) {
13718 MachineOperand &Op = MI->getOperand(i);
13719 if (!(Op.isReg() && Op.isImplicit()))
13720 MIB.addOperand(Op);
13722 if (MI->hasOneMemOperand())
13723 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
13725 BuildMI(*BB, MI, dl,
13726 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
13727 .addReg(X86::XMM0);
13729 MI->eraseFromParent();
13733 // FIXME: Custom handling because TableGen doesn't support multiple implicit
13734 // defs in an instruction pattern
13735 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
13736 const TargetInstrInfo *TII) {
13738 switch (MI->getOpcode()) {
13739 default: llvm_unreachable("illegal opcode!");
13740 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
13741 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
13742 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
13743 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
13744 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
13745 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
13746 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
13747 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
13750 DebugLoc dl = MI->getDebugLoc();
13751 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
13753 unsigned NumArgs = MI->getNumOperands(); // remove the results
13754 for (unsigned i = 1; i < NumArgs; ++i) {
13755 MachineOperand &Op = MI->getOperand(i);
13756 if (!(Op.isReg() && Op.isImplicit()))
13757 MIB.addOperand(Op);
13759 if (MI->hasOneMemOperand())
13760 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
13762 BuildMI(*BB, MI, dl,
13763 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
13766 MI->eraseFromParent();
13770 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
13771 const TargetInstrInfo *TII,
13772 const X86Subtarget* Subtarget) {
13773 DebugLoc dl = MI->getDebugLoc();
13775 // Address into RAX/EAX, other two args into ECX, EDX.
13776 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
13777 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
13778 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
13779 for (int i = 0; i < X86::AddrNumOperands; ++i)
13780 MIB.addOperand(MI->getOperand(i));
13782 unsigned ValOps = X86::AddrNumOperands;
13783 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
13784 .addReg(MI->getOperand(ValOps).getReg());
13785 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
13786 .addReg(MI->getOperand(ValOps+1).getReg());
13788 // The instruction doesn't actually take any operands though.
13789 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
13791 MI->eraseFromParent(); // The pseudo is gone now.
13795 MachineBasicBlock *
13796 X86TargetLowering::EmitVAARG64WithCustomInserter(
13798 MachineBasicBlock *MBB) const {
13799 // Emit va_arg instruction on X86-64.
13801 // Operands to this pseudo-instruction:
13802 // 0 ) Output : destination address (reg)
13803 // 1-5) Input : va_list address (addr, i64mem)
13804 // 6 ) ArgSize : Size (in bytes) of vararg type
13805 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
13806 // 8 ) Align : Alignment of type
13807 // 9 ) EFLAGS (implicit-def)
13809 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
13810 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
13812 unsigned DestReg = MI->getOperand(0).getReg();
13813 MachineOperand &Base = MI->getOperand(1);
13814 MachineOperand &Scale = MI->getOperand(2);
13815 MachineOperand &Index = MI->getOperand(3);
13816 MachineOperand &Disp = MI->getOperand(4);
13817 MachineOperand &Segment = MI->getOperand(5);
13818 unsigned ArgSize = MI->getOperand(6).getImm();
13819 unsigned ArgMode = MI->getOperand(7).getImm();
13820 unsigned Align = MI->getOperand(8).getImm();
13822 // Memory Reference
13823 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
13824 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
13825 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
13827 // Machine Information
13828 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
13829 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
13830 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
13831 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
13832 DebugLoc DL = MI->getDebugLoc();
13834 // struct va_list {
13837 // i64 overflow_area (address)
13838 // i64 reg_save_area (address)
13840 // sizeof(va_list) = 24
13841 // alignment(va_list) = 8
13843 unsigned TotalNumIntRegs = 6;
13844 unsigned TotalNumXMMRegs = 8;
13845 bool UseGPOffset = (ArgMode == 1);
13846 bool UseFPOffset = (ArgMode == 2);
13847 unsigned MaxOffset = TotalNumIntRegs * 8 +
13848 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
13850 /* Align ArgSize to a multiple of 8 */
13851 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
13852 bool NeedsAlign = (Align > 8);
13854 MachineBasicBlock *thisMBB = MBB;
13855 MachineBasicBlock *overflowMBB;
13856 MachineBasicBlock *offsetMBB;
13857 MachineBasicBlock *endMBB;
13859 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
13860 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
13861 unsigned OffsetReg = 0;
13863 if (!UseGPOffset && !UseFPOffset) {
13864 // If we only pull from the overflow region, we don't create a branch.
13865 // We don't need to alter control flow.
13866 OffsetDestReg = 0; // unused
13867 OverflowDestReg = DestReg;
13870 overflowMBB = thisMBB;
13873 // First emit code to check if gp_offset (or fp_offset) is below the bound.
13874 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
13875 // If not, pull from overflow_area. (branch to overflowMBB)
13880 // offsetMBB overflowMBB
13885 // Registers for the PHI in endMBB
13886 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
13887 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
13889 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
13890 MachineFunction *MF = MBB->getParent();
13891 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13892 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13893 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
13895 MachineFunction::iterator MBBIter = MBB;
13898 // Insert the new basic blocks
13899 MF->insert(MBBIter, offsetMBB);
13900 MF->insert(MBBIter, overflowMBB);
13901 MF->insert(MBBIter, endMBB);
13903 // Transfer the remainder of MBB and its successor edges to endMBB.
13904 endMBB->splice(endMBB->begin(), thisMBB,
13905 llvm::next(MachineBasicBlock::iterator(MI)),
13907 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
13909 // Make offsetMBB and overflowMBB successors of thisMBB
13910 thisMBB->addSuccessor(offsetMBB);
13911 thisMBB->addSuccessor(overflowMBB);
13913 // endMBB is a successor of both offsetMBB and overflowMBB
13914 offsetMBB->addSuccessor(endMBB);
13915 overflowMBB->addSuccessor(endMBB);
13917 // Load the offset value into a register
13918 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
13919 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
13923 .addDisp(Disp, UseFPOffset ? 4 : 0)
13924 .addOperand(Segment)
13925 .setMemRefs(MMOBegin, MMOEnd);
13927 // Check if there is enough room left to pull this argument.
13928 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
13930 .addImm(MaxOffset + 8 - ArgSizeA8);
13932 // Branch to "overflowMBB" if offset >= max
13933 // Fall through to "offsetMBB" otherwise
13934 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
13935 .addMBB(overflowMBB);
13938 // In offsetMBB, emit code to use the reg_save_area.
13940 assert(OffsetReg != 0);
13942 // Read the reg_save_area address.
13943 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
13944 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
13949 .addOperand(Segment)
13950 .setMemRefs(MMOBegin, MMOEnd);
13952 // Zero-extend the offset
13953 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
13954 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
13957 .addImm(X86::sub_32bit);
13959 // Add the offset to the reg_save_area to get the final address.
13960 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
13961 .addReg(OffsetReg64)
13962 .addReg(RegSaveReg);
13964 // Compute the offset for the next argument
13965 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
13966 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
13968 .addImm(UseFPOffset ? 16 : 8);
13970 // Store it back into the va_list.
13971 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
13975 .addDisp(Disp, UseFPOffset ? 4 : 0)
13976 .addOperand(Segment)
13977 .addReg(NextOffsetReg)
13978 .setMemRefs(MMOBegin, MMOEnd);
13981 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
13986 // Emit code to use overflow area
13989 // Load the overflow_area address into a register.
13990 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
13991 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
13996 .addOperand(Segment)
13997 .setMemRefs(MMOBegin, MMOEnd);
13999 // If we need to align it, do so. Otherwise, just copy the address
14000 // to OverflowDestReg.
14002 // Align the overflow address
14003 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
14004 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
14006 // aligned_addr = (addr + (align-1)) & ~(align-1)
14007 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
14008 .addReg(OverflowAddrReg)
14011 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
14013 .addImm(~(uint64_t)(Align-1));
14015 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
14016 .addReg(OverflowAddrReg);
14019 // Compute the next overflow address after this argument.
14020 // (the overflow address should be kept 8-byte aligned)
14021 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
14022 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
14023 .addReg(OverflowDestReg)
14024 .addImm(ArgSizeA8);
14026 // Store the new overflow address.
14027 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
14032 .addOperand(Segment)
14033 .addReg(NextAddrReg)
14034 .setMemRefs(MMOBegin, MMOEnd);
14036 // If we branched, emit the PHI to the front of endMBB.
14038 BuildMI(*endMBB, endMBB->begin(), DL,
14039 TII->get(X86::PHI), DestReg)
14040 .addReg(OffsetDestReg).addMBB(offsetMBB)
14041 .addReg(OverflowDestReg).addMBB(overflowMBB);
14044 // Erase the pseudo instruction
14045 MI->eraseFromParent();
14050 MachineBasicBlock *
14051 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
14053 MachineBasicBlock *MBB) const {
14054 // Emit code to save XMM registers to the stack. The ABI says that the
14055 // number of registers to save is given in %al, so it's theoretically
14056 // possible to do an indirect jump trick to avoid saving all of them,
14057 // however this code takes a simpler approach and just executes all
14058 // of the stores if %al is non-zero. It's less code, and it's probably
14059 // easier on the hardware branch predictor, and stores aren't all that
14060 // expensive anyway.
14062 // Create the new basic blocks. One block contains all the XMM stores,
14063 // and one block is the final destination regardless of whether any
14064 // stores were performed.
14065 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
14066 MachineFunction *F = MBB->getParent();
14067 MachineFunction::iterator MBBIter = MBB;
14069 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
14070 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
14071 F->insert(MBBIter, XMMSaveMBB);
14072 F->insert(MBBIter, EndMBB);
14074 // Transfer the remainder of MBB and its successor edges to EndMBB.
14075 EndMBB->splice(EndMBB->begin(), MBB,
14076 llvm::next(MachineBasicBlock::iterator(MI)),
14078 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
14080 // The original block will now fall through to the XMM save block.
14081 MBB->addSuccessor(XMMSaveMBB);
14082 // The XMMSaveMBB will fall through to the end block.
14083 XMMSaveMBB->addSuccessor(EndMBB);
14085 // Now add the instructions.
14086 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14087 DebugLoc DL = MI->getDebugLoc();
14089 unsigned CountReg = MI->getOperand(0).getReg();
14090 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
14091 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
14093 if (!Subtarget->isTargetWin64()) {
14094 // If %al is 0, branch around the XMM save block.
14095 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
14096 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
14097 MBB->addSuccessor(EndMBB);
14100 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
14101 // In the XMM save block, save all the XMM argument registers.
14102 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
14103 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
14104 MachineMemOperand *MMO =
14105 F->getMachineMemOperand(
14106 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
14107 MachineMemOperand::MOStore,
14108 /*Size=*/16, /*Align=*/16);
14109 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
14110 .addFrameIndex(RegSaveFrameIndex)
14111 .addImm(/*Scale=*/1)
14112 .addReg(/*IndexReg=*/0)
14113 .addImm(/*Disp=*/Offset)
14114 .addReg(/*Segment=*/0)
14115 .addReg(MI->getOperand(i).getReg())
14116 .addMemOperand(MMO);
14119 MI->eraseFromParent(); // The pseudo instruction is gone now.
14124 // The EFLAGS operand of SelectItr might be missing a kill marker
14125 // because there were multiple uses of EFLAGS, and ISel didn't know
14126 // which to mark. Figure out whether SelectItr should have had a
14127 // kill marker, and set it if it should. Returns the correct kill
14129 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
14130 MachineBasicBlock* BB,
14131 const TargetRegisterInfo* TRI) {
14132 // Scan forward through BB for a use/def of EFLAGS.
14133 MachineBasicBlock::iterator miI(llvm::next(SelectItr));
14134 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
14135 const MachineInstr& mi = *miI;
14136 if (mi.readsRegister(X86::EFLAGS))
14138 if (mi.definesRegister(X86::EFLAGS))
14139 break; // Should have kill-flag - update below.
14142 // If we hit the end of the block, check whether EFLAGS is live into a
14144 if (miI == BB->end()) {
14145 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
14146 sEnd = BB->succ_end();
14147 sItr != sEnd; ++sItr) {
14148 MachineBasicBlock* succ = *sItr;
14149 if (succ->isLiveIn(X86::EFLAGS))
14154 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
14155 // out. SelectMI should have a kill flag on EFLAGS.
14156 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
14160 MachineBasicBlock *
14161 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
14162 MachineBasicBlock *BB) const {
14163 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14164 DebugLoc DL = MI->getDebugLoc();
14166 // To "insert" a SELECT_CC instruction, we actually have to insert the
14167 // diamond control-flow pattern. The incoming instruction knows the
14168 // destination vreg to set, the condition code register to branch on, the
14169 // true/false values to select between, and a branch opcode to use.
14170 const BasicBlock *LLVM_BB = BB->getBasicBlock();
14171 MachineFunction::iterator It = BB;
14177 // cmpTY ccX, r1, r2
14179 // fallthrough --> copy0MBB
14180 MachineBasicBlock *thisMBB = BB;
14181 MachineFunction *F = BB->getParent();
14182 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
14183 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
14184 F->insert(It, copy0MBB);
14185 F->insert(It, sinkMBB);
14187 // If the EFLAGS register isn't dead in the terminator, then claim that it's
14188 // live into the sink and copy blocks.
14189 const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
14190 if (!MI->killsRegister(X86::EFLAGS) &&
14191 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
14192 copy0MBB->addLiveIn(X86::EFLAGS);
14193 sinkMBB->addLiveIn(X86::EFLAGS);
14196 // Transfer the remainder of BB and its successor edges to sinkMBB.
14197 sinkMBB->splice(sinkMBB->begin(), BB,
14198 llvm::next(MachineBasicBlock::iterator(MI)),
14200 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
14202 // Add the true and fallthrough blocks as its successors.
14203 BB->addSuccessor(copy0MBB);
14204 BB->addSuccessor(sinkMBB);
14206 // Create the conditional branch instruction.
14208 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
14209 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
14212 // %FalseValue = ...
14213 // # fallthrough to sinkMBB
14214 copy0MBB->addSuccessor(sinkMBB);
14217 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
14219 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14220 TII->get(X86::PHI), MI->getOperand(0).getReg())
14221 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
14222 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
14224 MI->eraseFromParent(); // The pseudo instruction is gone now.
14228 MachineBasicBlock *
14229 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
14230 bool Is64Bit) const {
14231 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14232 DebugLoc DL = MI->getDebugLoc();
14233 MachineFunction *MF = BB->getParent();
14234 const BasicBlock *LLVM_BB = BB->getBasicBlock();
14236 assert(getTargetMachine().Options.EnableSegmentedStacks);
14238 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
14239 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
14242 // ... [Till the alloca]
14243 // If stacklet is not large enough, jump to mallocMBB
14246 // Allocate by subtracting from RSP
14247 // Jump to continueMBB
14250 // Allocate by call to runtime
14254 // [rest of original BB]
14257 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14258 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14259 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
14261 MachineRegisterInfo &MRI = MF->getRegInfo();
14262 const TargetRegisterClass *AddrRegClass =
14263 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
14265 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
14266 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
14267 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
14268 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
14269 sizeVReg = MI->getOperand(1).getReg(),
14270 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
14272 MachineFunction::iterator MBBIter = BB;
14275 MF->insert(MBBIter, bumpMBB);
14276 MF->insert(MBBIter, mallocMBB);
14277 MF->insert(MBBIter, continueMBB);
14279 continueMBB->splice(continueMBB->begin(), BB, llvm::next
14280 (MachineBasicBlock::iterator(MI)), BB->end());
14281 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
14283 // Add code to the main basic block to check if the stack limit has been hit,
14284 // and if so, jump to mallocMBB otherwise to bumpMBB.
14285 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
14286 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
14287 .addReg(tmpSPVReg).addReg(sizeVReg);
14288 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
14289 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
14290 .addReg(SPLimitVReg);
14291 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
14293 // bumpMBB simply decreases the stack pointer, since we know the current
14294 // stacklet has enough space.
14295 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
14296 .addReg(SPLimitVReg);
14297 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
14298 .addReg(SPLimitVReg);
14299 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
14301 // Calls into a routine in libgcc to allocate more space from the heap.
14302 const uint32_t *RegMask =
14303 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
14305 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
14307 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
14308 .addExternalSymbol("__morestack_allocate_stack_space")
14309 .addRegMask(RegMask)
14310 .addReg(X86::RDI, RegState::Implicit)
14311 .addReg(X86::RAX, RegState::ImplicitDefine);
14313 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
14315 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
14316 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
14317 .addExternalSymbol("__morestack_allocate_stack_space")
14318 .addRegMask(RegMask)
14319 .addReg(X86::EAX, RegState::ImplicitDefine);
14323 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
14326 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
14327 .addReg(Is64Bit ? X86::RAX : X86::EAX);
14328 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
14330 // Set up the CFG correctly.
14331 BB->addSuccessor(bumpMBB);
14332 BB->addSuccessor(mallocMBB);
14333 mallocMBB->addSuccessor(continueMBB);
14334 bumpMBB->addSuccessor(continueMBB);
14336 // Take care of the PHI nodes.
14337 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
14338 MI->getOperand(0).getReg())
14339 .addReg(mallocPtrVReg).addMBB(mallocMBB)
14340 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
14342 // Delete the original pseudo instruction.
14343 MI->eraseFromParent();
14346 return continueMBB;
14349 MachineBasicBlock *
14350 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
14351 MachineBasicBlock *BB) const {
14352 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14353 DebugLoc DL = MI->getDebugLoc();
14355 assert(!Subtarget->isTargetEnvMacho());
14357 // The lowering is pretty easy: we're just emitting the call to _alloca. The
14358 // non-trivial part is impdef of ESP.
14360 if (Subtarget->isTargetWin64()) {
14361 if (Subtarget->isTargetCygMing()) {
14362 // ___chkstk(Mingw64):
14363 // Clobbers R10, R11, RAX and EFLAGS.
14365 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
14366 .addExternalSymbol("___chkstk")
14367 .addReg(X86::RAX, RegState::Implicit)
14368 .addReg(X86::RSP, RegState::Implicit)
14369 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
14370 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
14371 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14373 // __chkstk(MSVCRT): does not update stack pointer.
14374 // Clobbers R10, R11 and EFLAGS.
14375 // FIXME: RAX(allocated size) might be reused and not killed.
14376 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
14377 .addExternalSymbol("__chkstk")
14378 .addReg(X86::RAX, RegState::Implicit)
14379 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14380 // RAX has the offset to subtracted from RSP.
14381 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
14386 const char *StackProbeSymbol =
14387 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
14389 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
14390 .addExternalSymbol(StackProbeSymbol)
14391 .addReg(X86::EAX, RegState::Implicit)
14392 .addReg(X86::ESP, RegState::Implicit)
14393 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
14394 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
14395 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
14398 MI->eraseFromParent(); // The pseudo instruction is gone now.
14402 MachineBasicBlock *
14403 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
14404 MachineBasicBlock *BB) const {
14405 // This is pretty easy. We're taking the value that we received from
14406 // our load from the relocation, sticking it in either RDI (x86-64)
14407 // or EAX and doing an indirect call. The return value will then
14408 // be in the normal return register.
14409 const X86InstrInfo *TII
14410 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
14411 DebugLoc DL = MI->getDebugLoc();
14412 MachineFunction *F = BB->getParent();
14414 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
14415 assert(MI->getOperand(3).isGlobal() && "This should be a global");
14417 // Get a register mask for the lowered call.
14418 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
14419 // proper register mask.
14420 const uint32_t *RegMask =
14421 getTargetMachine().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
14422 if (Subtarget->is64Bit()) {
14423 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14424 TII->get(X86::MOV64rm), X86::RDI)
14426 .addImm(0).addReg(0)
14427 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14428 MI->getOperand(3).getTargetFlags())
14430 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
14431 addDirectMem(MIB, X86::RDI);
14432 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
14433 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
14434 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14435 TII->get(X86::MOV32rm), X86::EAX)
14437 .addImm(0).addReg(0)
14438 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14439 MI->getOperand(3).getTargetFlags())
14441 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
14442 addDirectMem(MIB, X86::EAX);
14443 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
14445 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
14446 TII->get(X86::MOV32rm), X86::EAX)
14447 .addReg(TII->getGlobalBaseReg(F))
14448 .addImm(0).addReg(0)
14449 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
14450 MI->getOperand(3).getTargetFlags())
14452 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
14453 addDirectMem(MIB, X86::EAX);
14454 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
14457 MI->eraseFromParent(); // The pseudo instruction is gone now.
14461 MachineBasicBlock *
14462 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
14463 MachineBasicBlock *MBB) const {
14464 DebugLoc DL = MI->getDebugLoc();
14465 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14467 MachineFunction *MF = MBB->getParent();
14468 MachineRegisterInfo &MRI = MF->getRegInfo();
14470 const BasicBlock *BB = MBB->getBasicBlock();
14471 MachineFunction::iterator I = MBB;
14474 // Memory Reference
14475 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14476 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14479 unsigned MemOpndSlot = 0;
14481 unsigned CurOp = 0;
14483 DstReg = MI->getOperand(CurOp++).getReg();
14484 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
14485 assert(RC->hasType(MVT::i32) && "Invalid destination!");
14486 unsigned mainDstReg = MRI.createVirtualRegister(RC);
14487 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
14489 MemOpndSlot = CurOp;
14491 MVT PVT = getPointerTy();
14492 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
14493 "Invalid Pointer Size!");
14495 // For v = setjmp(buf), we generate
14498 // buf[LabelOffset] = restoreMBB
14499 // SjLjSetup restoreMBB
14505 // v = phi(main, restore)
14510 MachineBasicBlock *thisMBB = MBB;
14511 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
14512 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
14513 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
14514 MF->insert(I, mainMBB);
14515 MF->insert(I, sinkMBB);
14516 MF->push_back(restoreMBB);
14518 MachineInstrBuilder MIB;
14520 // Transfer the remainder of BB and its successor edges to sinkMBB.
14521 sinkMBB->splice(sinkMBB->begin(), MBB,
14522 llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
14523 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
14526 unsigned PtrStoreOpc = 0;
14527 unsigned LabelReg = 0;
14528 const int64_t LabelOffset = 1 * PVT.getStoreSize();
14529 Reloc::Model RM = getTargetMachine().getRelocationModel();
14530 bool UseImmLabel = (getTargetMachine().getCodeModel() == CodeModel::Small) &&
14531 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
14533 // Prepare IP either in reg or imm.
14534 if (!UseImmLabel) {
14535 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
14536 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
14537 LabelReg = MRI.createVirtualRegister(PtrRC);
14538 if (Subtarget->is64Bit()) {
14539 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
14543 .addMBB(restoreMBB)
14546 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
14547 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
14548 .addReg(XII->getGlobalBaseReg(MF))
14551 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
14555 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
14557 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
14558 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14559 if (i == X86::AddrDisp)
14560 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
14562 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
14565 MIB.addReg(LabelReg);
14567 MIB.addMBB(restoreMBB);
14568 MIB.setMemRefs(MMOBegin, MMOEnd);
14570 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
14571 .addMBB(restoreMBB);
14572 MIB.addRegMask(RegInfo->getNoPreservedMask());
14573 thisMBB->addSuccessor(mainMBB);
14574 thisMBB->addSuccessor(restoreMBB);
14578 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
14579 mainMBB->addSuccessor(sinkMBB);
14582 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
14583 TII->get(X86::PHI), DstReg)
14584 .addReg(mainDstReg).addMBB(mainMBB)
14585 .addReg(restoreDstReg).addMBB(restoreMBB);
14588 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
14589 BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
14590 restoreMBB->addSuccessor(sinkMBB);
14592 MI->eraseFromParent();
14596 MachineBasicBlock *
14597 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
14598 MachineBasicBlock *MBB) const {
14599 DebugLoc DL = MI->getDebugLoc();
14600 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14602 MachineFunction *MF = MBB->getParent();
14603 MachineRegisterInfo &MRI = MF->getRegInfo();
14605 // Memory Reference
14606 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
14607 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
14609 MVT PVT = getPointerTy();
14610 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
14611 "Invalid Pointer Size!");
14613 const TargetRegisterClass *RC =
14614 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
14615 unsigned Tmp = MRI.createVirtualRegister(RC);
14616 // Since FP is only updated here but NOT referenced, it's treated as GPR.
14617 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
14618 unsigned SP = RegInfo->getStackRegister();
14620 MachineInstrBuilder MIB;
14622 const int64_t LabelOffset = 1 * PVT.getStoreSize();
14623 const int64_t SPOffset = 2 * PVT.getStoreSize();
14625 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
14626 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
14629 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
14630 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
14631 MIB.addOperand(MI->getOperand(i));
14632 MIB.setMemRefs(MMOBegin, MMOEnd);
14634 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
14635 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14636 if (i == X86::AddrDisp)
14637 MIB.addDisp(MI->getOperand(i), LabelOffset);
14639 MIB.addOperand(MI->getOperand(i));
14641 MIB.setMemRefs(MMOBegin, MMOEnd);
14643 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
14644 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
14645 if (i == X86::AddrDisp)
14646 MIB.addDisp(MI->getOperand(i), SPOffset);
14648 MIB.addOperand(MI->getOperand(i));
14650 MIB.setMemRefs(MMOBegin, MMOEnd);
14652 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
14654 MI->eraseFromParent();
14658 MachineBasicBlock *
14659 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
14660 MachineBasicBlock *BB) const {
14661 switch (MI->getOpcode()) {
14662 default: llvm_unreachable("Unexpected instr type to insert");
14663 case X86::TAILJMPd64:
14664 case X86::TAILJMPr64:
14665 case X86::TAILJMPm64:
14666 llvm_unreachable("TAILJMP64 would not be touched here.");
14667 case X86::TCRETURNdi64:
14668 case X86::TCRETURNri64:
14669 case X86::TCRETURNmi64:
14671 case X86::WIN_ALLOCA:
14672 return EmitLoweredWinAlloca(MI, BB);
14673 case X86::SEG_ALLOCA_32:
14674 return EmitLoweredSegAlloca(MI, BB, false);
14675 case X86::SEG_ALLOCA_64:
14676 return EmitLoweredSegAlloca(MI, BB, true);
14677 case X86::TLSCall_32:
14678 case X86::TLSCall_64:
14679 return EmitLoweredTLSCall(MI, BB);
14680 case X86::CMOV_GR8:
14681 case X86::CMOV_FR32:
14682 case X86::CMOV_FR64:
14683 case X86::CMOV_V4F32:
14684 case X86::CMOV_V2F64:
14685 case X86::CMOV_V2I64:
14686 case X86::CMOV_V8F32:
14687 case X86::CMOV_V4F64:
14688 case X86::CMOV_V4I64:
14689 case X86::CMOV_GR16:
14690 case X86::CMOV_GR32:
14691 case X86::CMOV_RFP32:
14692 case X86::CMOV_RFP64:
14693 case X86::CMOV_RFP80:
14694 return EmitLoweredSelect(MI, BB);
14696 case X86::FP32_TO_INT16_IN_MEM:
14697 case X86::FP32_TO_INT32_IN_MEM:
14698 case X86::FP32_TO_INT64_IN_MEM:
14699 case X86::FP64_TO_INT16_IN_MEM:
14700 case X86::FP64_TO_INT32_IN_MEM:
14701 case X86::FP64_TO_INT64_IN_MEM:
14702 case X86::FP80_TO_INT16_IN_MEM:
14703 case X86::FP80_TO_INT32_IN_MEM:
14704 case X86::FP80_TO_INT64_IN_MEM: {
14705 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
14706 DebugLoc DL = MI->getDebugLoc();
14708 // Change the floating point control register to use "round towards zero"
14709 // mode when truncating to an integer value.
14710 MachineFunction *F = BB->getParent();
14711 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
14712 addFrameReference(BuildMI(*BB, MI, DL,
14713 TII->get(X86::FNSTCW16m)), CWFrameIdx);
14715 // Load the old value of the high byte of the control word...
14717 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
14718 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
14721 // Set the high part to be round to zero...
14722 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
14725 // Reload the modified control word now...
14726 addFrameReference(BuildMI(*BB, MI, DL,
14727 TII->get(X86::FLDCW16m)), CWFrameIdx);
14729 // Restore the memory image of control word to original value
14730 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
14733 // Get the X86 opcode to use.
14735 switch (MI->getOpcode()) {
14736 default: llvm_unreachable("illegal opcode!");
14737 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
14738 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
14739 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
14740 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
14741 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
14742 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
14743 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
14744 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
14745 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
14749 MachineOperand &Op = MI->getOperand(0);
14751 AM.BaseType = X86AddressMode::RegBase;
14752 AM.Base.Reg = Op.getReg();
14754 AM.BaseType = X86AddressMode::FrameIndexBase;
14755 AM.Base.FrameIndex = Op.getIndex();
14757 Op = MI->getOperand(1);
14759 AM.Scale = Op.getImm();
14760 Op = MI->getOperand(2);
14762 AM.IndexReg = Op.getImm();
14763 Op = MI->getOperand(3);
14764 if (Op.isGlobal()) {
14765 AM.GV = Op.getGlobal();
14767 AM.Disp = Op.getImm();
14769 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
14770 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
14772 // Reload the original control word now.
14773 addFrameReference(BuildMI(*BB, MI, DL,
14774 TII->get(X86::FLDCW16m)), CWFrameIdx);
14776 MI->eraseFromParent(); // The pseudo instruction is gone now.
14779 // String/text processing lowering.
14780 case X86::PCMPISTRM128REG:
14781 case X86::VPCMPISTRM128REG:
14782 case X86::PCMPISTRM128MEM:
14783 case X86::VPCMPISTRM128MEM:
14784 case X86::PCMPESTRM128REG:
14785 case X86::VPCMPESTRM128REG:
14786 case X86::PCMPESTRM128MEM:
14787 case X86::VPCMPESTRM128MEM:
14788 assert(Subtarget->hasSSE42() &&
14789 "Target must have SSE4.2 or AVX features enabled");
14790 return EmitPCMPSTRM(MI, BB, getTargetMachine().getInstrInfo());
14792 // String/text processing lowering.
14793 case X86::PCMPISTRIREG:
14794 case X86::VPCMPISTRIREG:
14795 case X86::PCMPISTRIMEM:
14796 case X86::VPCMPISTRIMEM:
14797 case X86::PCMPESTRIREG:
14798 case X86::VPCMPESTRIREG:
14799 case X86::PCMPESTRIMEM:
14800 case X86::VPCMPESTRIMEM:
14801 assert(Subtarget->hasSSE42() &&
14802 "Target must have SSE4.2 or AVX features enabled");
14803 return EmitPCMPSTRI(MI, BB, getTargetMachine().getInstrInfo());
14805 // Thread synchronization.
14807 return EmitMonitor(MI, BB, getTargetMachine().getInstrInfo(), Subtarget);
14811 return EmitXBegin(MI, BB, getTargetMachine().getInstrInfo());
14813 // Atomic Lowering.
14814 case X86::ATOMAND8:
14815 case X86::ATOMAND16:
14816 case X86::ATOMAND32:
14817 case X86::ATOMAND64:
14820 case X86::ATOMOR16:
14821 case X86::ATOMOR32:
14822 case X86::ATOMOR64:
14824 case X86::ATOMXOR16:
14825 case X86::ATOMXOR8:
14826 case X86::ATOMXOR32:
14827 case X86::ATOMXOR64:
14829 case X86::ATOMNAND8:
14830 case X86::ATOMNAND16:
14831 case X86::ATOMNAND32:
14832 case X86::ATOMNAND64:
14834 case X86::ATOMMAX8:
14835 case X86::ATOMMAX16:
14836 case X86::ATOMMAX32:
14837 case X86::ATOMMAX64:
14839 case X86::ATOMMIN8:
14840 case X86::ATOMMIN16:
14841 case X86::ATOMMIN32:
14842 case X86::ATOMMIN64:
14844 case X86::ATOMUMAX8:
14845 case X86::ATOMUMAX16:
14846 case X86::ATOMUMAX32:
14847 case X86::ATOMUMAX64:
14849 case X86::ATOMUMIN8:
14850 case X86::ATOMUMIN16:
14851 case X86::ATOMUMIN32:
14852 case X86::ATOMUMIN64:
14853 return EmitAtomicLoadArith(MI, BB);
14855 // This group does 64-bit operations on a 32-bit host.
14856 case X86::ATOMAND6432:
14857 case X86::ATOMOR6432:
14858 case X86::ATOMXOR6432:
14859 case X86::ATOMNAND6432:
14860 case X86::ATOMADD6432:
14861 case X86::ATOMSUB6432:
14862 case X86::ATOMMAX6432:
14863 case X86::ATOMMIN6432:
14864 case X86::ATOMUMAX6432:
14865 case X86::ATOMUMIN6432:
14866 case X86::ATOMSWAP6432:
14867 return EmitAtomicLoadArith6432(MI, BB);
14869 case X86::VASTART_SAVE_XMM_REGS:
14870 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
14872 case X86::VAARG_64:
14873 return EmitVAARG64WithCustomInserter(MI, BB);
14875 case X86::EH_SjLj_SetJmp32:
14876 case X86::EH_SjLj_SetJmp64:
14877 return emitEHSjLjSetJmp(MI, BB);
14879 case X86::EH_SjLj_LongJmp32:
14880 case X86::EH_SjLj_LongJmp64:
14881 return emitEHSjLjLongJmp(MI, BB);
14885 //===----------------------------------------------------------------------===//
14886 // X86 Optimization Hooks
14887 //===----------------------------------------------------------------------===//
14889 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
14892 const SelectionDAG &DAG,
14893 unsigned Depth) const {
14894 unsigned BitWidth = KnownZero.getBitWidth();
14895 unsigned Opc = Op.getOpcode();
14896 assert((Opc >= ISD::BUILTIN_OP_END ||
14897 Opc == ISD::INTRINSIC_WO_CHAIN ||
14898 Opc == ISD::INTRINSIC_W_CHAIN ||
14899 Opc == ISD::INTRINSIC_VOID) &&
14900 "Should use MaskedValueIsZero if you don't know whether Op"
14901 " is a target node!");
14903 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
14917 // These nodes' second result is a boolean.
14918 if (Op.getResNo() == 0)
14921 case X86ISD::SETCC:
14922 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
14924 case ISD::INTRINSIC_WO_CHAIN: {
14925 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14926 unsigned NumLoBits = 0;
14929 case Intrinsic::x86_sse_movmsk_ps:
14930 case Intrinsic::x86_avx_movmsk_ps_256:
14931 case Intrinsic::x86_sse2_movmsk_pd:
14932 case Intrinsic::x86_avx_movmsk_pd_256:
14933 case Intrinsic::x86_mmx_pmovmskb:
14934 case Intrinsic::x86_sse2_pmovmskb_128:
14935 case Intrinsic::x86_avx2_pmovmskb: {
14936 // High bits of movmskp{s|d}, pmovmskb are known zero.
14938 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14939 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
14940 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
14941 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
14942 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
14943 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
14944 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
14945 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
14947 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
14956 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
14957 unsigned Depth) const {
14958 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
14959 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
14960 return Op.getValueType().getScalarType().getSizeInBits();
14966 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
14967 /// node is a GlobalAddress + offset.
14968 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
14969 const GlobalValue* &GA,
14970 int64_t &Offset) const {
14971 if (N->getOpcode() == X86ISD::Wrapper) {
14972 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
14973 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
14974 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
14978 return TargetLowering::isGAPlusOffset(N, GA, Offset);
14981 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
14982 /// same as extracting the high 128-bit part of 256-bit vector and then
14983 /// inserting the result into the low part of a new 256-bit vector
14984 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
14985 EVT VT = SVOp->getValueType(0);
14986 unsigned NumElems = VT.getVectorNumElements();
14988 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
14989 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
14990 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
14991 SVOp->getMaskElt(j) >= 0)
14997 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
14998 /// same as extracting the low 128-bit part of 256-bit vector and then
14999 /// inserting the result into the high part of a new 256-bit vector
15000 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
15001 EVT VT = SVOp->getValueType(0);
15002 unsigned NumElems = VT.getVectorNumElements();
15004 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15005 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
15006 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
15007 SVOp->getMaskElt(j) >= 0)
15013 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
15014 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
15015 TargetLowering::DAGCombinerInfo &DCI,
15016 const X86Subtarget* Subtarget) {
15017 DebugLoc dl = N->getDebugLoc();
15018 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
15019 SDValue V1 = SVOp->getOperand(0);
15020 SDValue V2 = SVOp->getOperand(1);
15021 EVT VT = SVOp->getValueType(0);
15022 unsigned NumElems = VT.getVectorNumElements();
15024 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
15025 V2.getOpcode() == ISD::CONCAT_VECTORS) {
15029 // V UNDEF BUILD_VECTOR UNDEF
15031 // CONCAT_VECTOR CONCAT_VECTOR
15034 // RESULT: V + zero extended
15036 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
15037 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
15038 V1.getOperand(1).getOpcode() != ISD::UNDEF)
15041 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
15044 // To match the shuffle mask, the first half of the mask should
15045 // be exactly the first vector, and all the rest a splat with the
15046 // first element of the second one.
15047 for (unsigned i = 0; i != NumElems/2; ++i)
15048 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
15049 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
15052 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
15053 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
15054 if (Ld->hasNUsesOfValue(1, 0)) {
15055 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
15056 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
15058 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2,
15060 Ld->getPointerInfo(),
15061 Ld->getAlignment(),
15062 false/*isVolatile*/, true/*ReadMem*/,
15063 false/*WriteMem*/);
15065 // Make sure the newly-created LOAD is in the same position as Ld in
15066 // terms of dependency. We create a TokenFactor for Ld and ResNode,
15067 // and update uses of Ld's output chain to use the TokenFactor.
15068 if (Ld->hasAnyUseOfValue(1)) {
15069 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
15070 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
15071 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
15072 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
15073 SDValue(ResNode.getNode(), 1));
15076 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
15080 // Emit a zeroed vector and insert the desired subvector on its
15082 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15083 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
15084 return DCI.CombineTo(N, InsV);
15087 //===--------------------------------------------------------------------===//
15088 // Combine some shuffles into subvector extracts and inserts:
15091 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
15092 if (isShuffleHigh128VectorInsertLow(SVOp)) {
15093 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
15094 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
15095 return DCI.CombineTo(N, InsV);
15098 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
15099 if (isShuffleLow128VectorInsertHigh(SVOp)) {
15100 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
15101 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
15102 return DCI.CombineTo(N, InsV);
15108 /// PerformShuffleCombine - Performs several different shuffle combines.
15109 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
15110 TargetLowering::DAGCombinerInfo &DCI,
15111 const X86Subtarget *Subtarget) {
15112 DebugLoc dl = N->getDebugLoc();
15113 EVT VT = N->getValueType(0);
15115 // Don't create instructions with illegal types after legalize types has run.
15116 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15117 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
15120 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
15121 if (Subtarget->hasFp256() && VT.is256BitVector() &&
15122 N->getOpcode() == ISD::VECTOR_SHUFFLE)
15123 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
15125 // Only handle 128 wide vector from here on.
15126 if (!VT.is128BitVector())
15129 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
15130 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
15131 // consecutive, non-overlapping, and in the right order.
15132 SmallVector<SDValue, 16> Elts;
15133 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
15134 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
15136 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
15139 /// PerformTruncateCombine - Converts truncate operation to
15140 /// a sequence of vector shuffle operations.
15141 /// It is possible when we truncate 256-bit vector to 128-bit vector
15142 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
15143 TargetLowering::DAGCombinerInfo &DCI,
15144 const X86Subtarget *Subtarget) {
15148 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
15149 /// specific shuffle of a load can be folded into a single element load.
15150 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
15151 /// shuffles have been customed lowered so we need to handle those here.
15152 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
15153 TargetLowering::DAGCombinerInfo &DCI) {
15154 if (DCI.isBeforeLegalizeOps())
15157 SDValue InVec = N->getOperand(0);
15158 SDValue EltNo = N->getOperand(1);
15160 if (!isa<ConstantSDNode>(EltNo))
15163 EVT VT = InVec.getValueType();
15165 bool HasShuffleIntoBitcast = false;
15166 if (InVec.getOpcode() == ISD::BITCAST) {
15167 // Don't duplicate a load with other uses.
15168 if (!InVec.hasOneUse())
15170 EVT BCVT = InVec.getOperand(0).getValueType();
15171 if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
15173 InVec = InVec.getOperand(0);
15174 HasShuffleIntoBitcast = true;
15177 if (!isTargetShuffle(InVec.getOpcode()))
15180 // Don't duplicate a load with other uses.
15181 if (!InVec.hasOneUse())
15184 SmallVector<int, 16> ShuffleMask;
15186 if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
15190 // Select the input vector, guarding against out of range extract vector.
15191 unsigned NumElems = VT.getVectorNumElements();
15192 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
15193 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
15194 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
15195 : InVec.getOperand(1);
15197 // If inputs to shuffle are the same for both ops, then allow 2 uses
15198 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
15200 if (LdNode.getOpcode() == ISD::BITCAST) {
15201 // Don't duplicate a load with other uses.
15202 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
15205 AllowedUses = 1; // only allow 1 load use if we have a bitcast
15206 LdNode = LdNode.getOperand(0);
15209 if (!ISD::isNormalLoad(LdNode.getNode()))
15212 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
15214 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
15217 if (HasShuffleIntoBitcast) {
15218 // If there's a bitcast before the shuffle, check if the load type and
15219 // alignment is valid.
15220 unsigned Align = LN0->getAlignment();
15221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15222 unsigned NewAlign = TLI.getDataLayout()->
15223 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
15225 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
15229 // All checks match so transform back to vector_shuffle so that DAG combiner
15230 // can finish the job
15231 DebugLoc dl = N->getDebugLoc();
15233 // Create shuffle node taking into account the case that its a unary shuffle
15234 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
15235 Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
15236 InVec.getOperand(0), Shuffle,
15238 Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
15239 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
15243 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
15244 /// generation and convert it from being a bunch of shuffles and extracts
15245 /// to a simple store and scalar loads to extract the elements.
15246 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
15247 TargetLowering::DAGCombinerInfo &DCI) {
15248 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
15249 if (NewOp.getNode())
15252 SDValue InputVector = N->getOperand(0);
15253 // Detect whether we are trying to convert from mmx to i32 and the bitcast
15254 // from mmx to v2i32 has a single usage.
15255 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
15256 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
15257 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
15258 return DAG.getNode(X86ISD::MMX_MOVD2W, InputVector.getDebugLoc(),
15259 N->getValueType(0),
15260 InputVector.getNode()->getOperand(0));
15262 // Only operate on vectors of 4 elements, where the alternative shuffling
15263 // gets to be more expensive.
15264 if (InputVector.getValueType() != MVT::v4i32)
15267 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
15268 // single use which is a sign-extend or zero-extend, and all elements are
15270 SmallVector<SDNode *, 4> Uses;
15271 unsigned ExtractedElements = 0;
15272 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
15273 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
15274 if (UI.getUse().getResNo() != InputVector.getResNo())
15277 SDNode *Extract = *UI;
15278 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
15281 if (Extract->getValueType(0) != MVT::i32)
15283 if (!Extract->hasOneUse())
15285 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
15286 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
15288 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
15291 // Record which element was extracted.
15292 ExtractedElements |=
15293 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
15295 Uses.push_back(Extract);
15298 // If not all the elements were used, this may not be worthwhile.
15299 if (ExtractedElements != 15)
15302 // Ok, we've now decided to do the transformation.
15303 DebugLoc dl = InputVector.getDebugLoc();
15305 // Store the value to a temporary stack slot.
15306 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
15307 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
15308 MachinePointerInfo(), false, false, 0);
15310 // Replace each use (extract) with a load of the appropriate element.
15311 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
15312 UE = Uses.end(); UI != UE; ++UI) {
15313 SDNode *Extract = *UI;
15315 // cOMpute the element's address.
15316 SDValue Idx = Extract->getOperand(1);
15318 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
15319 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
15320 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15321 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
15323 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
15324 StackPtr, OffsetVal);
15326 // Load the scalar.
15327 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
15328 ScalarAddr, MachinePointerInfo(),
15329 false, false, false, 0);
15331 // Replace the exact with the load.
15332 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
15335 // The replacement was made in place; don't return anything.
15339 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
15340 static unsigned matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS,
15341 SDValue RHS, SelectionDAG &DAG,
15342 const X86Subtarget *Subtarget) {
15343 if (!VT.isVector())
15346 switch (VT.getSimpleVT().SimpleTy) {
15351 if (!Subtarget->hasAVX2())
15356 if (!Subtarget->hasSSE2())
15360 // SSE2 has only a small subset of the operations.
15361 bool hasUnsigned = Subtarget->hasSSE41() ||
15362 (Subtarget->hasSSE2() && VT == MVT::v16i8);
15363 bool hasSigned = Subtarget->hasSSE41() ||
15364 (Subtarget->hasSSE2() && VT == MVT::v8i16);
15366 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15368 // Check for x CC y ? x : y.
15369 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
15370 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
15375 return hasUnsigned ? X86ISD::UMIN : 0;
15378 return hasUnsigned ? X86ISD::UMAX : 0;
15381 return hasSigned ? X86ISD::SMIN : 0;
15384 return hasSigned ? X86ISD::SMAX : 0;
15386 // Check for x CC y ? y : x -- a min/max with reversed arms.
15387 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
15388 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
15393 return hasUnsigned ? X86ISD::UMAX : 0;
15396 return hasUnsigned ? X86ISD::UMIN : 0;
15399 return hasSigned ? X86ISD::SMAX : 0;
15402 return hasSigned ? X86ISD::SMIN : 0;
15409 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
15411 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
15412 TargetLowering::DAGCombinerInfo &DCI,
15413 const X86Subtarget *Subtarget) {
15414 DebugLoc DL = N->getDebugLoc();
15415 SDValue Cond = N->getOperand(0);
15416 // Get the LHS/RHS of the select.
15417 SDValue LHS = N->getOperand(1);
15418 SDValue RHS = N->getOperand(2);
15419 EVT VT = LHS.getValueType();
15421 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
15422 // instructions match the semantics of the common C idiom x<y?x:y but not
15423 // x<=y?x:y, because of how they handle negative zero (which can be
15424 // ignored in unsafe-math mode).
15425 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
15426 VT != MVT::f80 && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
15427 (Subtarget->hasSSE2() ||
15428 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
15429 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15431 unsigned Opcode = 0;
15432 // Check for x CC y ? x : y.
15433 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
15434 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
15438 // Converting this to a min would handle NaNs incorrectly, and swapping
15439 // the operands would cause it to handle comparisons between positive
15440 // and negative zero incorrectly.
15441 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
15442 if (!DAG.getTarget().Options.UnsafeFPMath &&
15443 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
15445 std::swap(LHS, RHS);
15447 Opcode = X86ISD::FMIN;
15450 // Converting this to a min would handle comparisons between positive
15451 // and negative zero incorrectly.
15452 if (!DAG.getTarget().Options.UnsafeFPMath &&
15453 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
15455 Opcode = X86ISD::FMIN;
15458 // Converting this to a min would handle both negative zeros and NaNs
15459 // incorrectly, but we can swap the operands to fix both.
15460 std::swap(LHS, RHS);
15464 Opcode = X86ISD::FMIN;
15468 // Converting this to a max would handle comparisons between positive
15469 // and negative zero incorrectly.
15470 if (!DAG.getTarget().Options.UnsafeFPMath &&
15471 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
15473 Opcode = X86ISD::FMAX;
15476 // Converting this to a max would handle NaNs incorrectly, and swapping
15477 // the operands would cause it to handle comparisons between positive
15478 // and negative zero incorrectly.
15479 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
15480 if (!DAG.getTarget().Options.UnsafeFPMath &&
15481 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
15483 std::swap(LHS, RHS);
15485 Opcode = X86ISD::FMAX;
15488 // Converting this to a max would handle both negative zeros and NaNs
15489 // incorrectly, but we can swap the operands to fix both.
15490 std::swap(LHS, RHS);
15494 Opcode = X86ISD::FMAX;
15497 // Check for x CC y ? y : x -- a min/max with reversed arms.
15498 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
15499 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
15503 // Converting this to a min would handle comparisons between positive
15504 // and negative zero incorrectly, and swapping the operands would
15505 // cause it to handle NaNs incorrectly.
15506 if (!DAG.getTarget().Options.UnsafeFPMath &&
15507 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
15508 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
15510 std::swap(LHS, RHS);
15512 Opcode = X86ISD::FMIN;
15515 // Converting this to a min would handle NaNs incorrectly.
15516 if (!DAG.getTarget().Options.UnsafeFPMath &&
15517 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
15519 Opcode = X86ISD::FMIN;
15522 // Converting this to a min would handle both negative zeros and NaNs
15523 // incorrectly, but we can swap the operands to fix both.
15524 std::swap(LHS, RHS);
15528 Opcode = X86ISD::FMIN;
15532 // Converting this to a max would handle NaNs incorrectly.
15533 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
15535 Opcode = X86ISD::FMAX;
15538 // Converting this to a max would handle comparisons between positive
15539 // and negative zero incorrectly, and swapping the operands would
15540 // cause it to handle NaNs incorrectly.
15541 if (!DAG.getTarget().Options.UnsafeFPMath &&
15542 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
15543 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
15545 std::swap(LHS, RHS);
15547 Opcode = X86ISD::FMAX;
15550 // Converting this to a max would handle both negative zeros and NaNs
15551 // incorrectly, but we can swap the operands to fix both.
15552 std::swap(LHS, RHS);
15556 Opcode = X86ISD::FMAX;
15562 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
15565 // If this is a select between two integer constants, try to do some
15567 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
15568 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
15569 // Don't do this for crazy integer types.
15570 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
15571 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
15572 // so that TrueC (the true value) is larger than FalseC.
15573 bool NeedsCondInvert = false;
15575 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
15576 // Efficiently invertible.
15577 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
15578 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
15579 isa<ConstantSDNode>(Cond.getOperand(1))))) {
15580 NeedsCondInvert = true;
15581 std::swap(TrueC, FalseC);
15584 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
15585 if (FalseC->getAPIntValue() == 0 &&
15586 TrueC->getAPIntValue().isPowerOf2()) {
15587 if (NeedsCondInvert) // Invert the condition if needed.
15588 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
15589 DAG.getConstant(1, Cond.getValueType()));
15591 // Zero extend the condition if needed.
15592 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
15594 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
15595 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
15596 DAG.getConstant(ShAmt, MVT::i8));
15599 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
15600 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
15601 if (NeedsCondInvert) // Invert the condition if needed.
15602 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
15603 DAG.getConstant(1, Cond.getValueType()));
15605 // Zero extend the condition if needed.
15606 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
15607 FalseC->getValueType(0), Cond);
15608 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15609 SDValue(FalseC, 0));
15612 // Optimize cases that will turn into an LEA instruction. This requires
15613 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
15614 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
15615 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
15616 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
15618 bool isFastMultiplier = false;
15620 switch ((unsigned char)Diff) {
15622 case 1: // result = add base, cond
15623 case 2: // result = lea base( , cond*2)
15624 case 3: // result = lea base(cond, cond*2)
15625 case 4: // result = lea base( , cond*4)
15626 case 5: // result = lea base(cond, cond*4)
15627 case 8: // result = lea base( , cond*8)
15628 case 9: // result = lea base(cond, cond*8)
15629 isFastMultiplier = true;
15634 if (isFastMultiplier) {
15635 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
15636 if (NeedsCondInvert) // Invert the condition if needed.
15637 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
15638 DAG.getConstant(1, Cond.getValueType()));
15640 // Zero extend the condition if needed.
15641 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
15643 // Scale the condition by the difference.
15645 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
15646 DAG.getConstant(Diff, Cond.getValueType()));
15648 // Add the base if non-zero.
15649 if (FalseC->getAPIntValue() != 0)
15650 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15651 SDValue(FalseC, 0));
15658 // Canonicalize max and min:
15659 // (x > y) ? x : y -> (x >= y) ? x : y
15660 // (x < y) ? x : y -> (x <= y) ? x : y
15661 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
15662 // the need for an extra compare
15663 // against zero. e.g.
15664 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
15666 // testl %edi, %edi
15668 // cmovgl %edi, %eax
15672 // cmovsl %eax, %edi
15673 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
15674 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
15675 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
15676 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15681 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
15682 Cond = DAG.getSetCC(Cond.getDebugLoc(), Cond.getValueType(),
15683 Cond.getOperand(0), Cond.getOperand(1), NewCC);
15684 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
15689 // Match VSELECTs into subs with unsigned saturation.
15690 if (!DCI.isBeforeLegalize() &&
15691 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
15692 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
15693 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
15694 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
15695 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
15697 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
15698 // left side invert the predicate to simplify logic below.
15700 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
15702 CC = ISD::getSetCCInverse(CC, true);
15703 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
15707 if (Other.getNode() && Other->getNumOperands() == 2 &&
15708 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
15709 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
15710 SDValue CondRHS = Cond->getOperand(1);
15712 // Look for a general sub with unsigned saturation first.
15713 // x >= y ? x-y : 0 --> subus x, y
15714 // x > y ? x-y : 0 --> subus x, y
15715 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
15716 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
15717 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
15719 // If the RHS is a constant we have to reverse the const canonicalization.
15720 // x > C-1 ? x+-C : 0 --> subus x, C
15721 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
15722 isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
15723 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
15724 if (CondRHS.getConstantOperandVal(0) == -A-1)
15725 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
15726 DAG.getConstant(-A, VT));
15729 // Another special case: If C was a sign bit, the sub has been
15730 // canonicalized into a xor.
15731 // FIXME: Would it be better to use ComputeMaskedBits to determine whether
15732 // it's safe to decanonicalize the xor?
15733 // x s< 0 ? x^C : 0 --> subus x, C
15734 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
15735 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
15736 isSplatVector(OpRHS.getNode())) {
15737 APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
15739 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
15744 // Try to match a min/max vector operation.
15745 if (!DCI.isBeforeLegalize() &&
15746 N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC)
15747 if (unsigned Op = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget))
15748 return DAG.getNode(Op, DL, N->getValueType(0), LHS, RHS);
15750 // If we know that this node is legal then we know that it is going to be
15751 // matched by one of the SSE/AVX BLEND instructions. These instructions only
15752 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
15753 // to simplify previous instructions.
15754 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15755 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
15756 !DCI.isBeforeLegalize() && TLI.isOperationLegal(ISD::VSELECT, VT)) {
15757 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
15759 // Don't optimize vector selects that map to mask-registers.
15763 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
15764 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
15766 APInt KnownZero, KnownOne;
15767 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
15768 DCI.isBeforeLegalizeOps());
15769 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
15770 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne, TLO))
15771 DCI.CommitTargetLoweringOpt(TLO);
15777 // Check whether a boolean test is testing a boolean value generated by
15778 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
15781 // Simplify the following patterns:
15782 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
15783 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
15784 // to (Op EFLAGS Cond)
15786 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
15787 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
15788 // to (Op EFLAGS !Cond)
15790 // where Op could be BRCOND or CMOV.
15792 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
15793 // Quit if not CMP and SUB with its value result used.
15794 if (Cmp.getOpcode() != X86ISD::CMP &&
15795 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
15798 // Quit if not used as a boolean value.
15799 if (CC != X86::COND_E && CC != X86::COND_NE)
15802 // Check CMP operands. One of them should be 0 or 1 and the other should be
15803 // an SetCC or extended from it.
15804 SDValue Op1 = Cmp.getOperand(0);
15805 SDValue Op2 = Cmp.getOperand(1);
15808 const ConstantSDNode* C = 0;
15809 bool needOppositeCond = (CC == X86::COND_E);
15811 if ((C = dyn_cast<ConstantSDNode>(Op1)))
15813 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
15815 else // Quit if all operands are not constants.
15818 if (C->getZExtValue() == 1)
15819 needOppositeCond = !needOppositeCond;
15820 else if (C->getZExtValue() != 0)
15821 // Quit if the constant is neither 0 or 1.
15824 // Skip 'zext' node.
15825 if (SetCC.getOpcode() == ISD::ZERO_EXTEND)
15826 SetCC = SetCC.getOperand(0);
15828 switch (SetCC.getOpcode()) {
15829 case X86ISD::SETCC:
15830 // Set the condition code or opposite one if necessary.
15831 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
15832 if (needOppositeCond)
15833 CC = X86::GetOppositeBranchCondition(CC);
15834 return SetCC.getOperand(1);
15835 case X86ISD::CMOV: {
15836 // Check whether false/true value has canonical one, i.e. 0 or 1.
15837 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
15838 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
15839 // Quit if true value is not a constant.
15842 // Quit if false value is not a constant.
15844 // A special case for rdrand, where 0 is set if false cond is found.
15845 SDValue Op = SetCC.getOperand(0);
15846 if (Op.getOpcode() != X86ISD::RDRAND)
15849 // Quit if false value is not the constant 0 or 1.
15850 bool FValIsFalse = true;
15851 if (FVal && FVal->getZExtValue() != 0) {
15852 if (FVal->getZExtValue() != 1)
15854 // If FVal is 1, opposite cond is needed.
15855 needOppositeCond = !needOppositeCond;
15856 FValIsFalse = false;
15858 // Quit if TVal is not the constant opposite of FVal.
15859 if (FValIsFalse && TVal->getZExtValue() != 1)
15861 if (!FValIsFalse && TVal->getZExtValue() != 0)
15863 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
15864 if (needOppositeCond)
15865 CC = X86::GetOppositeBranchCondition(CC);
15866 return SetCC.getOperand(3);
15873 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
15874 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
15875 TargetLowering::DAGCombinerInfo &DCI,
15876 const X86Subtarget *Subtarget) {
15877 DebugLoc DL = N->getDebugLoc();
15879 // If the flag operand isn't dead, don't touch this CMOV.
15880 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
15883 SDValue FalseOp = N->getOperand(0);
15884 SDValue TrueOp = N->getOperand(1);
15885 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
15886 SDValue Cond = N->getOperand(3);
15888 if (CC == X86::COND_E || CC == X86::COND_NE) {
15889 switch (Cond.getOpcode()) {
15893 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
15894 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
15895 return (CC == X86::COND_E) ? FalseOp : TrueOp;
15901 Flags = checkBoolTestSetCCCombine(Cond, CC);
15902 if (Flags.getNode() &&
15903 // Extra check as FCMOV only supports a subset of X86 cond.
15904 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
15905 SDValue Ops[] = { FalseOp, TrueOp,
15906 DAG.getConstant(CC, MVT::i8), Flags };
15907 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(),
15908 Ops, array_lengthof(Ops));
15911 // If this is a select between two integer constants, try to do some
15912 // optimizations. Note that the operands are ordered the opposite of SELECT
15914 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
15915 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
15916 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
15917 // larger than FalseC (the false value).
15918 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
15919 CC = X86::GetOppositeBranchCondition(CC);
15920 std::swap(TrueC, FalseC);
15921 std::swap(TrueOp, FalseOp);
15924 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
15925 // This is efficient for any integer data type (including i8/i16) and
15927 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
15928 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15929 DAG.getConstant(CC, MVT::i8), Cond);
15931 // Zero extend the condition if needed.
15932 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
15934 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
15935 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
15936 DAG.getConstant(ShAmt, MVT::i8));
15937 if (N->getNumValues() == 2) // Dead flag value?
15938 return DCI.CombineTo(N, Cond, SDValue());
15942 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
15943 // for any integer data type, including i8/i16.
15944 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
15945 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15946 DAG.getConstant(CC, MVT::i8), Cond);
15948 // Zero extend the condition if needed.
15949 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
15950 FalseC->getValueType(0), Cond);
15951 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15952 SDValue(FalseC, 0));
15954 if (N->getNumValues() == 2) // Dead flag value?
15955 return DCI.CombineTo(N, Cond, SDValue());
15959 // Optimize cases that will turn into an LEA instruction. This requires
15960 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
15961 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
15962 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
15963 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
15965 bool isFastMultiplier = false;
15967 switch ((unsigned char)Diff) {
15969 case 1: // result = add base, cond
15970 case 2: // result = lea base( , cond*2)
15971 case 3: // result = lea base(cond, cond*2)
15972 case 4: // result = lea base( , cond*4)
15973 case 5: // result = lea base(cond, cond*4)
15974 case 8: // result = lea base( , cond*8)
15975 case 9: // result = lea base(cond, cond*8)
15976 isFastMultiplier = true;
15981 if (isFastMultiplier) {
15982 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
15983 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
15984 DAG.getConstant(CC, MVT::i8), Cond);
15985 // Zero extend the condition if needed.
15986 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
15988 // Scale the condition by the difference.
15990 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
15991 DAG.getConstant(Diff, Cond.getValueType()));
15993 // Add the base if non-zero.
15994 if (FalseC->getAPIntValue() != 0)
15995 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
15996 SDValue(FalseC, 0));
15997 if (N->getNumValues() == 2) // Dead flag value?
15998 return DCI.CombineTo(N, Cond, SDValue());
16005 // Handle these cases:
16006 // (select (x != c), e, c) -> select (x != c), e, x),
16007 // (select (x == c), c, e) -> select (x == c), x, e)
16008 // where the c is an integer constant, and the "select" is the combination
16009 // of CMOV and CMP.
16011 // The rationale for this change is that the conditional-move from a constant
16012 // needs two instructions, however, conditional-move from a register needs
16013 // only one instruction.
16015 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
16016 // some instruction-combining opportunities. This opt needs to be
16017 // postponed as late as possible.
16019 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
16020 // the DCI.xxxx conditions are provided to postpone the optimization as
16021 // late as possible.
16023 ConstantSDNode *CmpAgainst = 0;
16024 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
16025 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
16026 !isa<ConstantSDNode>(Cond.getOperand(0))) {
16028 if (CC == X86::COND_NE &&
16029 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
16030 CC = X86::GetOppositeBranchCondition(CC);
16031 std::swap(TrueOp, FalseOp);
16034 if (CC == X86::COND_E &&
16035 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
16036 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
16037 DAG.getConstant(CC, MVT::i8), Cond };
16038 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops,
16039 array_lengthof(Ops));
16047 /// PerformMulCombine - Optimize a single multiply with constant into two
16048 /// in order to implement it with two cheaper instructions, e.g.
16049 /// LEA + SHL, LEA + LEA.
16050 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
16051 TargetLowering::DAGCombinerInfo &DCI) {
16052 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
16055 EVT VT = N->getValueType(0);
16056 if (VT != MVT::i64)
16059 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
16062 uint64_t MulAmt = C->getZExtValue();
16063 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
16066 uint64_t MulAmt1 = 0;
16067 uint64_t MulAmt2 = 0;
16068 if ((MulAmt % 9) == 0) {
16070 MulAmt2 = MulAmt / 9;
16071 } else if ((MulAmt % 5) == 0) {
16073 MulAmt2 = MulAmt / 5;
16074 } else if ((MulAmt % 3) == 0) {
16076 MulAmt2 = MulAmt / 3;
16079 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
16080 DebugLoc DL = N->getDebugLoc();
16082 if (isPowerOf2_64(MulAmt2) &&
16083 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
16084 // If second multiplifer is pow2, issue it first. We want the multiply by
16085 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
16087 std::swap(MulAmt1, MulAmt2);
16090 if (isPowerOf2_64(MulAmt1))
16091 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
16092 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
16094 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
16095 DAG.getConstant(MulAmt1, VT));
16097 if (isPowerOf2_64(MulAmt2))
16098 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
16099 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
16101 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
16102 DAG.getConstant(MulAmt2, VT));
16104 // Do not add new nodes to DAG combiner worklist.
16105 DCI.CombineTo(N, NewMul, false);
16110 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
16111 SDValue N0 = N->getOperand(0);
16112 SDValue N1 = N->getOperand(1);
16113 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
16114 EVT VT = N0.getValueType();
16116 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
16117 // since the result of setcc_c is all zero's or all ones.
16118 if (VT.isInteger() && !VT.isVector() &&
16119 N1C && N0.getOpcode() == ISD::AND &&
16120 N0.getOperand(1).getOpcode() == ISD::Constant) {
16121 SDValue N00 = N0.getOperand(0);
16122 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
16123 ((N00.getOpcode() == ISD::ANY_EXTEND ||
16124 N00.getOpcode() == ISD::ZERO_EXTEND) &&
16125 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
16126 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
16127 APInt ShAmt = N1C->getAPIntValue();
16128 Mask = Mask.shl(ShAmt);
16130 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
16131 N00, DAG.getConstant(Mask, VT));
16135 // Hardware support for vector shifts is sparse which makes us scalarize the
16136 // vector operations in many cases. Also, on sandybridge ADD is faster than
16138 // (shl V, 1) -> add V,V
16139 if (isSplatVector(N1.getNode())) {
16140 assert(N0.getValueType().isVector() && "Invalid vector shift type");
16141 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1->getOperand(0));
16142 // We shift all of the values by one. In many cases we do not have
16143 // hardware support for this operation. This is better expressed as an ADD
16145 if (N1C && (1 == N1C->getZExtValue())) {
16146 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N0);
16153 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
16155 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
16156 TargetLowering::DAGCombinerInfo &DCI,
16157 const X86Subtarget *Subtarget) {
16158 if (N->getOpcode() == ISD::SHL) {
16159 SDValue V = PerformSHLCombine(N, DAG);
16160 if (V.getNode()) return V;
16166 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
16167 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
16168 // and friends. Likewise for OR -> CMPNEQSS.
16169 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
16170 TargetLowering::DAGCombinerInfo &DCI,
16171 const X86Subtarget *Subtarget) {
16174 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
16175 // we're requiring SSE2 for both.
16176 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
16177 SDValue N0 = N->getOperand(0);
16178 SDValue N1 = N->getOperand(1);
16179 SDValue CMP0 = N0->getOperand(1);
16180 SDValue CMP1 = N1->getOperand(1);
16181 DebugLoc DL = N->getDebugLoc();
16183 // The SETCCs should both refer to the same CMP.
16184 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
16187 SDValue CMP00 = CMP0->getOperand(0);
16188 SDValue CMP01 = CMP0->getOperand(1);
16189 EVT VT = CMP00.getValueType();
16191 if (VT == MVT::f32 || VT == MVT::f64) {
16192 bool ExpectingFlags = false;
16193 // Check for any users that want flags:
16194 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
16195 !ExpectingFlags && UI != UE; ++UI)
16196 switch (UI->getOpcode()) {
16201 ExpectingFlags = true;
16203 case ISD::CopyToReg:
16204 case ISD::SIGN_EXTEND:
16205 case ISD::ZERO_EXTEND:
16206 case ISD::ANY_EXTEND:
16210 if (!ExpectingFlags) {
16211 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
16212 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
16214 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
16215 X86::CondCode tmp = cc0;
16220 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
16221 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
16222 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
16223 X86ISD::NodeType NTOperator = is64BitFP ?
16224 X86ISD::FSETCCsd : X86ISD::FSETCCss;
16225 // FIXME: need symbolic constants for these magic numbers.
16226 // See X86ATTInstPrinter.cpp:printSSECC().
16227 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
16228 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
16229 DAG.getConstant(x86cc, MVT::i8));
16230 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
16232 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
16233 DAG.getConstant(1, MVT::i32));
16234 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
16235 return OneBitOfTruth;
16243 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
16244 /// so it can be folded inside ANDNP.
16245 static bool CanFoldXORWithAllOnes(const SDNode *N) {
16246 EVT VT = N->getValueType(0);
16248 // Match direct AllOnes for 128 and 256-bit vectors
16249 if (ISD::isBuildVectorAllOnes(N))
16252 // Look through a bit convert.
16253 if (N->getOpcode() == ISD::BITCAST)
16254 N = N->getOperand(0).getNode();
16256 // Sometimes the operand may come from a insert_subvector building a 256-bit
16258 if (VT.is256BitVector() &&
16259 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
16260 SDValue V1 = N->getOperand(0);
16261 SDValue V2 = N->getOperand(1);
16263 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
16264 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
16265 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
16266 ISD::isBuildVectorAllOnes(V2.getNode()))
16273 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
16274 // register. In most cases we actually compare or select YMM-sized registers
16275 // and mixing the two types creates horrible code. This method optimizes
16276 // some of the transition sequences.
16277 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
16278 TargetLowering::DAGCombinerInfo &DCI,
16279 const X86Subtarget *Subtarget) {
16280 EVT VT = N->getValueType(0);
16281 if (!VT.is256BitVector())
16284 assert((N->getOpcode() == ISD::ANY_EXTEND ||
16285 N->getOpcode() == ISD::ZERO_EXTEND ||
16286 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
16288 SDValue Narrow = N->getOperand(0);
16289 EVT NarrowVT = Narrow->getValueType(0);
16290 if (!NarrowVT.is128BitVector())
16293 if (Narrow->getOpcode() != ISD::XOR &&
16294 Narrow->getOpcode() != ISD::AND &&
16295 Narrow->getOpcode() != ISD::OR)
16298 SDValue N0 = Narrow->getOperand(0);
16299 SDValue N1 = Narrow->getOperand(1);
16300 DebugLoc DL = Narrow->getDebugLoc();
16302 // The Left side has to be a trunc.
16303 if (N0.getOpcode() != ISD::TRUNCATE)
16306 // The type of the truncated inputs.
16307 EVT WideVT = N0->getOperand(0)->getValueType(0);
16311 // The right side has to be a 'trunc' or a constant vector.
16312 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
16313 bool RHSConst = (isSplatVector(N1.getNode()) &&
16314 isa<ConstantSDNode>(N1->getOperand(0)));
16315 if (!RHSTrunc && !RHSConst)
16318 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16320 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
16323 // Set N0 and N1 to hold the inputs to the new wide operation.
16324 N0 = N0->getOperand(0);
16326 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
16327 N1->getOperand(0));
16328 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
16329 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, &C[0], C.size());
16330 } else if (RHSTrunc) {
16331 N1 = N1->getOperand(0);
16334 // Generate the wide operation.
16335 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
16336 unsigned Opcode = N->getOpcode();
16338 case ISD::ANY_EXTEND:
16340 case ISD::ZERO_EXTEND: {
16341 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
16342 APInt Mask = APInt::getAllOnesValue(InBits);
16343 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
16344 return DAG.getNode(ISD::AND, DL, VT,
16345 Op, DAG.getConstant(Mask, VT));
16347 case ISD::SIGN_EXTEND:
16348 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
16349 Op, DAG.getValueType(NarrowVT));
16351 llvm_unreachable("Unexpected opcode");
16355 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
16356 TargetLowering::DAGCombinerInfo &DCI,
16357 const X86Subtarget *Subtarget) {
16358 EVT VT = N->getValueType(0);
16359 if (DCI.isBeforeLegalizeOps())
16362 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
16366 // Create BLSI, and BLSR instructions
16367 // BLSI is X & (-X)
16368 // BLSR is X & (X-1)
16369 if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
16370 SDValue N0 = N->getOperand(0);
16371 SDValue N1 = N->getOperand(1);
16372 DebugLoc DL = N->getDebugLoc();
16374 // Check LHS for neg
16375 if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
16376 isZero(N0.getOperand(0)))
16377 return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
16379 // Check RHS for neg
16380 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
16381 isZero(N1.getOperand(0)))
16382 return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
16384 // Check LHS for X-1
16385 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
16386 isAllOnes(N0.getOperand(1)))
16387 return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
16389 // Check RHS for X-1
16390 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
16391 isAllOnes(N1.getOperand(1)))
16392 return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
16397 // Want to form ANDNP nodes:
16398 // 1) In the hopes of then easily combining them with OR and AND nodes
16399 // to form PBLEND/PSIGN.
16400 // 2) To match ANDN packed intrinsics
16401 if (VT != MVT::v2i64 && VT != MVT::v4i64)
16404 SDValue N0 = N->getOperand(0);
16405 SDValue N1 = N->getOperand(1);
16406 DebugLoc DL = N->getDebugLoc();
16408 // Check LHS for vnot
16409 if (N0.getOpcode() == ISD::XOR &&
16410 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
16411 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
16412 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
16414 // Check RHS for vnot
16415 if (N1.getOpcode() == ISD::XOR &&
16416 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
16417 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
16418 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
16423 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
16424 TargetLowering::DAGCombinerInfo &DCI,
16425 const X86Subtarget *Subtarget) {
16426 EVT VT = N->getValueType(0);
16427 if (DCI.isBeforeLegalizeOps())
16430 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
16434 SDValue N0 = N->getOperand(0);
16435 SDValue N1 = N->getOperand(1);
16437 // look for psign/blend
16438 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
16439 if (!Subtarget->hasSSSE3() ||
16440 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
16443 // Canonicalize pandn to RHS
16444 if (N0.getOpcode() == X86ISD::ANDNP)
16446 // or (and (m, y), (pandn m, x))
16447 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
16448 SDValue Mask = N1.getOperand(0);
16449 SDValue X = N1.getOperand(1);
16451 if (N0.getOperand(0) == Mask)
16452 Y = N0.getOperand(1);
16453 if (N0.getOperand(1) == Mask)
16454 Y = N0.getOperand(0);
16456 // Check to see if the mask appeared in both the AND and ANDNP and
16460 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
16461 // Look through mask bitcast.
16462 if (Mask.getOpcode() == ISD::BITCAST)
16463 Mask = Mask.getOperand(0);
16464 if (X.getOpcode() == ISD::BITCAST)
16465 X = X.getOperand(0);
16466 if (Y.getOpcode() == ISD::BITCAST)
16467 Y = Y.getOperand(0);
16469 EVT MaskVT = Mask.getValueType();
16471 // Validate that the Mask operand is a vector sra node.
16472 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
16473 // there is no psrai.b
16474 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
16475 unsigned SraAmt = ~0;
16476 if (Mask.getOpcode() == ISD::SRA) {
16477 SDValue Amt = Mask.getOperand(1);
16478 if (isSplatVector(Amt.getNode())) {
16479 SDValue SclrAmt = Amt->getOperand(0);
16480 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt))
16481 SraAmt = C->getZExtValue();
16483 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
16484 SDValue SraC = Mask.getOperand(1);
16485 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
16487 if ((SraAmt + 1) != EltBits)
16490 DebugLoc DL = N->getDebugLoc();
16492 // Now we know we at least have a plendvb with the mask val. See if
16493 // we can form a psignb/w/d.
16494 // psign = x.type == y.type == mask.type && y = sub(0, x);
16495 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
16496 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
16497 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
16498 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
16499 "Unsupported VT for PSIGN");
16500 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
16501 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
16503 // PBLENDVB only available on SSE 4.1
16504 if (!Subtarget->hasSSE41())
16507 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
16509 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
16510 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
16511 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
16512 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
16513 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
16517 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
16520 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
16521 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
16523 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
16525 if (!N0.hasOneUse() || !N1.hasOneUse())
16528 SDValue ShAmt0 = N0.getOperand(1);
16529 if (ShAmt0.getValueType() != MVT::i8)
16531 SDValue ShAmt1 = N1.getOperand(1);
16532 if (ShAmt1.getValueType() != MVT::i8)
16534 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
16535 ShAmt0 = ShAmt0.getOperand(0);
16536 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
16537 ShAmt1 = ShAmt1.getOperand(0);
16539 DebugLoc DL = N->getDebugLoc();
16540 unsigned Opc = X86ISD::SHLD;
16541 SDValue Op0 = N0.getOperand(0);
16542 SDValue Op1 = N1.getOperand(0);
16543 if (ShAmt0.getOpcode() == ISD::SUB) {
16544 Opc = X86ISD::SHRD;
16545 std::swap(Op0, Op1);
16546 std::swap(ShAmt0, ShAmt1);
16549 unsigned Bits = VT.getSizeInBits();
16550 if (ShAmt1.getOpcode() == ISD::SUB) {
16551 SDValue Sum = ShAmt1.getOperand(0);
16552 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
16553 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
16554 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
16555 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
16556 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
16557 return DAG.getNode(Opc, DL, VT,
16559 DAG.getNode(ISD::TRUNCATE, DL,
16562 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
16563 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
16565 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
16566 return DAG.getNode(Opc, DL, VT,
16567 N0.getOperand(0), N1.getOperand(0),
16568 DAG.getNode(ISD::TRUNCATE, DL,
16575 // Generate NEG and CMOV for integer abs.
16576 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
16577 EVT VT = N->getValueType(0);
16579 // Since X86 does not have CMOV for 8-bit integer, we don't convert
16580 // 8-bit integer abs to NEG and CMOV.
16581 if (VT.isInteger() && VT.getSizeInBits() == 8)
16584 SDValue N0 = N->getOperand(0);
16585 SDValue N1 = N->getOperand(1);
16586 DebugLoc DL = N->getDebugLoc();
16588 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
16589 // and change it to SUB and CMOV.
16590 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
16591 N0.getOpcode() == ISD::ADD &&
16592 N0.getOperand(1) == N1 &&
16593 N1.getOpcode() == ISD::SRA &&
16594 N1.getOperand(0) == N0.getOperand(0))
16595 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
16596 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
16597 // Generate SUB & CMOV.
16598 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
16599 DAG.getConstant(0, VT), N0.getOperand(0));
16601 SDValue Ops[] = { N0.getOperand(0), Neg,
16602 DAG.getConstant(X86::COND_GE, MVT::i8),
16603 SDValue(Neg.getNode(), 1) };
16604 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue),
16605 Ops, array_lengthof(Ops));
16610 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
16611 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
16612 TargetLowering::DAGCombinerInfo &DCI,
16613 const X86Subtarget *Subtarget) {
16614 EVT VT = N->getValueType(0);
16615 if (DCI.isBeforeLegalizeOps())
16618 if (Subtarget->hasCMov()) {
16619 SDValue RV = performIntegerAbsCombine(N, DAG);
16624 // Try forming BMI if it is available.
16625 if (!Subtarget->hasBMI())
16628 if (VT != MVT::i32 && VT != MVT::i64)
16631 assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
16633 // Create BLSMSK instructions by finding X ^ (X-1)
16634 SDValue N0 = N->getOperand(0);
16635 SDValue N1 = N->getOperand(1);
16636 DebugLoc DL = N->getDebugLoc();
16638 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
16639 isAllOnes(N0.getOperand(1)))
16640 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
16642 if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
16643 isAllOnes(N1.getOperand(1)))
16644 return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
16649 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
16650 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
16651 TargetLowering::DAGCombinerInfo &DCI,
16652 const X86Subtarget *Subtarget) {
16653 LoadSDNode *Ld = cast<LoadSDNode>(N);
16654 EVT RegVT = Ld->getValueType(0);
16655 EVT MemVT = Ld->getMemoryVT();
16656 DebugLoc dl = Ld->getDebugLoc();
16657 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16658 unsigned RegSz = RegVT.getSizeInBits();
16660 // On Sandybridge unaligned 256bit loads are inefficient.
16661 ISD::LoadExtType Ext = Ld->getExtensionType();
16662 unsigned Alignment = Ld->getAlignment();
16663 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
16664 if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
16665 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
16666 unsigned NumElems = RegVT.getVectorNumElements();
16670 SDValue Ptr = Ld->getBasePtr();
16671 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
16673 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
16675 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
16676 Ld->getPointerInfo(), Ld->isVolatile(),
16677 Ld->isNonTemporal(), Ld->isInvariant(),
16679 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16680 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
16681 Ld->getPointerInfo(), Ld->isVolatile(),
16682 Ld->isNonTemporal(), Ld->isInvariant(),
16683 std::min(16U, Alignment));
16684 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
16686 Load2.getValue(1));
16688 SDValue NewVec = DAG.getUNDEF(RegVT);
16689 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
16690 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
16691 return DCI.CombineTo(N, NewVec, TF, true);
16694 // If this is a vector EXT Load then attempt to optimize it using a
16695 // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
16696 // expansion is still better than scalar code.
16697 // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
16698 // emit a shuffle and a arithmetic shift.
16699 // TODO: It is possible to support ZExt by zeroing the undef values
16700 // during the shuffle phase or after the shuffle.
16701 if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
16702 (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
16703 assert(MemVT != RegVT && "Cannot extend to the same type");
16704 assert(MemVT.isVector() && "Must load a vector from memory");
16706 unsigned NumElems = RegVT.getVectorNumElements();
16707 unsigned MemSz = MemVT.getSizeInBits();
16708 assert(RegSz > MemSz && "Register size must be greater than the mem size");
16710 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
16713 // All sizes must be a power of two.
16714 if (!isPowerOf2_32(RegSz * MemSz * NumElems))
16717 // Attempt to load the original value using scalar loads.
16718 // Find the largest scalar type that divides the total loaded size.
16719 MVT SclrLoadTy = MVT::i8;
16720 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
16721 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
16722 MVT Tp = (MVT::SimpleValueType)tp;
16723 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
16728 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
16729 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
16731 SclrLoadTy = MVT::f64;
16733 // Calculate the number of scalar loads that we need to perform
16734 // in order to load our vector from memory.
16735 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
16736 if (Ext == ISD::SEXTLOAD && NumLoads > 1)
16739 unsigned loadRegZize = RegSz;
16740 if (Ext == ISD::SEXTLOAD && RegSz == 256)
16743 // Represent our vector as a sequence of elements which are the
16744 // largest scalar that we can load.
16745 EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
16746 loadRegZize/SclrLoadTy.getSizeInBits());
16748 // Represent the data using the same element type that is stored in
16749 // memory. In practice, we ''widen'' MemVT.
16751 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
16752 loadRegZize/MemVT.getScalarType().getSizeInBits());
16754 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
16755 "Invalid vector type");
16757 // We can't shuffle using an illegal type.
16758 if (!TLI.isTypeLegal(WideVecVT))
16761 SmallVector<SDValue, 8> Chains;
16762 SDValue Ptr = Ld->getBasePtr();
16763 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
16764 TLI.getPointerTy());
16765 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
16767 for (unsigned i = 0; i < NumLoads; ++i) {
16768 // Perform a single load.
16769 SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
16770 Ptr, Ld->getPointerInfo(),
16771 Ld->isVolatile(), Ld->isNonTemporal(),
16772 Ld->isInvariant(), Ld->getAlignment());
16773 Chains.push_back(ScalarLoad.getValue(1));
16774 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
16775 // another round of DAGCombining.
16777 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
16779 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
16780 ScalarLoad, DAG.getIntPtrConstant(i));
16782 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16785 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
16788 // Bitcast the loaded value to a vector of the original element type, in
16789 // the size of the target vector type.
16790 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
16791 unsigned SizeRatio = RegSz/MemSz;
16793 if (Ext == ISD::SEXTLOAD) {
16794 // If we have SSE4.1 we can directly emit a VSEXT node.
16795 if (Subtarget->hasSSE41()) {
16796 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
16797 return DCI.CombineTo(N, Sext, TF, true);
16800 // Otherwise we'll shuffle the small elements in the high bits of the
16801 // larger type and perform an arithmetic shift. If the shift is not legal
16802 // it's better to scalarize.
16803 if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
16806 // Redistribute the loaded elements into the different locations.
16807 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
16808 for (unsigned i = 0; i != NumElems; ++i)
16809 ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
16811 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
16812 DAG.getUNDEF(WideVecVT),
16815 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16817 // Build the arithmetic shift.
16818 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
16819 MemVT.getVectorElementType().getSizeInBits();
16820 Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
16821 DAG.getConstant(Amt, RegVT));
16823 return DCI.CombineTo(N, Shuff, TF, true);
16826 // Redistribute the loaded elements into the different locations.
16827 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
16828 for (unsigned i = 0; i != NumElems; ++i)
16829 ShuffleVec[i*SizeRatio] = i;
16831 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
16832 DAG.getUNDEF(WideVecVT),
16835 // Bitcast to the requested type.
16836 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16837 // Replace the original load with the new sequence
16838 // and return the new chain.
16839 return DCI.CombineTo(N, Shuff, TF, true);
16845 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
16846 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
16847 const X86Subtarget *Subtarget) {
16848 StoreSDNode *St = cast<StoreSDNode>(N);
16849 EVT VT = St->getValue().getValueType();
16850 EVT StVT = St->getMemoryVT();
16851 DebugLoc dl = St->getDebugLoc();
16852 SDValue StoredVal = St->getOperand(1);
16853 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16855 // If we are saving a concatenation of two XMM registers, perform two stores.
16856 // On Sandy Bridge, 256-bit memory operations are executed by two
16857 // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
16858 // memory operation.
16859 unsigned Alignment = St->getAlignment();
16860 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
16861 if (VT.is256BitVector() && !Subtarget->hasInt256() &&
16862 StVT == VT && !IsAligned) {
16863 unsigned NumElems = VT.getVectorNumElements();
16867 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
16868 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
16870 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
16871 SDValue Ptr0 = St->getBasePtr();
16872 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
16874 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
16875 St->getPointerInfo(), St->isVolatile(),
16876 St->isNonTemporal(), Alignment);
16877 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
16878 St->getPointerInfo(), St->isVolatile(),
16879 St->isNonTemporal(),
16880 std::min(16U, Alignment));
16881 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
16884 // Optimize trunc store (of multiple scalars) to shuffle and store.
16885 // First, pack all of the elements in one place. Next, store to memory
16886 // in fewer chunks.
16887 if (St->isTruncatingStore() && VT.isVector()) {
16888 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16889 unsigned NumElems = VT.getVectorNumElements();
16890 assert(StVT != VT && "Cannot truncate to the same type");
16891 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
16892 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
16894 // From, To sizes and ElemCount must be pow of two
16895 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
16896 // We are going to use the original vector elt for storing.
16897 // Accumulated smaller vector elements must be a multiple of the store size.
16898 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
16900 unsigned SizeRatio = FromSz / ToSz;
16902 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
16904 // Create a type on which we perform the shuffle
16905 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
16906 StVT.getScalarType(), NumElems*SizeRatio);
16908 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
16910 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
16911 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
16912 for (unsigned i = 0; i != NumElems; ++i)
16913 ShuffleVec[i] = i * SizeRatio;
16915 // Can't shuffle using an illegal type.
16916 if (!TLI.isTypeLegal(WideVecVT))
16919 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
16920 DAG.getUNDEF(WideVecVT),
16922 // At this point all of the data is stored at the bottom of the
16923 // register. We now need to save it to mem.
16925 // Find the largest store unit
16926 MVT StoreType = MVT::i8;
16927 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
16928 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
16929 MVT Tp = (MVT::SimpleValueType)tp;
16930 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
16934 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
16935 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
16936 (64 <= NumElems * ToSz))
16937 StoreType = MVT::f64;
16939 // Bitcast the original vector into a vector of store-size units
16940 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
16941 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
16942 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
16943 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
16944 SmallVector<SDValue, 8> Chains;
16945 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
16946 TLI.getPointerTy());
16947 SDValue Ptr = St->getBasePtr();
16949 // Perform one or more big stores into memory.
16950 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
16951 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
16952 StoreType, ShuffWide,
16953 DAG.getIntPtrConstant(i));
16954 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
16955 St->getPointerInfo(), St->isVolatile(),
16956 St->isNonTemporal(), St->getAlignment());
16957 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16958 Chains.push_back(Ch);
16961 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
16965 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
16966 // the FP state in cases where an emms may be missing.
16967 // A preferable solution to the general problem is to figure out the right
16968 // places to insert EMMS. This qualifies as a quick hack.
16970 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
16971 if (VT.getSizeInBits() != 64)
16974 const Function *F = DAG.getMachineFunction().getFunction();
16975 bool NoImplicitFloatOps = F->getAttributes().
16976 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
16977 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
16978 && Subtarget->hasSSE2();
16979 if ((VT.isVector() ||
16980 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
16981 isa<LoadSDNode>(St->getValue()) &&
16982 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
16983 St->getChain().hasOneUse() && !St->isVolatile()) {
16984 SDNode* LdVal = St->getValue().getNode();
16985 LoadSDNode *Ld = 0;
16986 int TokenFactorIndex = -1;
16987 SmallVector<SDValue, 8> Ops;
16988 SDNode* ChainVal = St->getChain().getNode();
16989 // Must be a store of a load. We currently handle two cases: the load
16990 // is a direct child, and it's under an intervening TokenFactor. It is
16991 // possible to dig deeper under nested TokenFactors.
16992 if (ChainVal == LdVal)
16993 Ld = cast<LoadSDNode>(St->getChain());
16994 else if (St->getValue().hasOneUse() &&
16995 ChainVal->getOpcode() == ISD::TokenFactor) {
16996 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
16997 if (ChainVal->getOperand(i).getNode() == LdVal) {
16998 TokenFactorIndex = i;
16999 Ld = cast<LoadSDNode>(St->getValue());
17001 Ops.push_back(ChainVal->getOperand(i));
17005 if (!Ld || !ISD::isNormalLoad(Ld))
17008 // If this is not the MMX case, i.e. we are just turning i64 load/store
17009 // into f64 load/store, avoid the transformation if there are multiple
17010 // uses of the loaded value.
17011 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
17014 DebugLoc LdDL = Ld->getDebugLoc();
17015 DebugLoc StDL = N->getDebugLoc();
17016 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
17017 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
17019 if (Subtarget->is64Bit() || F64IsLegal) {
17020 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
17021 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
17022 Ld->getPointerInfo(), Ld->isVolatile(),
17023 Ld->isNonTemporal(), Ld->isInvariant(),
17024 Ld->getAlignment());
17025 SDValue NewChain = NewLd.getValue(1);
17026 if (TokenFactorIndex != -1) {
17027 Ops.push_back(NewChain);
17028 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17031 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
17032 St->getPointerInfo(),
17033 St->isVolatile(), St->isNonTemporal(),
17034 St->getAlignment());
17037 // Otherwise, lower to two pairs of 32-bit loads / stores.
17038 SDValue LoAddr = Ld->getBasePtr();
17039 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
17040 DAG.getConstant(4, MVT::i32));
17042 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
17043 Ld->getPointerInfo(),
17044 Ld->isVolatile(), Ld->isNonTemporal(),
17045 Ld->isInvariant(), Ld->getAlignment());
17046 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
17047 Ld->getPointerInfo().getWithOffset(4),
17048 Ld->isVolatile(), Ld->isNonTemporal(),
17050 MinAlign(Ld->getAlignment(), 4));
17052 SDValue NewChain = LoLd.getValue(1);
17053 if (TokenFactorIndex != -1) {
17054 Ops.push_back(LoLd);
17055 Ops.push_back(HiLd);
17056 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
17060 LoAddr = St->getBasePtr();
17061 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
17062 DAG.getConstant(4, MVT::i32));
17064 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
17065 St->getPointerInfo(),
17066 St->isVolatile(), St->isNonTemporal(),
17067 St->getAlignment());
17068 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
17069 St->getPointerInfo().getWithOffset(4),
17071 St->isNonTemporal(),
17072 MinAlign(St->getAlignment(), 4));
17073 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
17078 /// isHorizontalBinOp - Return 'true' if this vector operation is "horizontal"
17079 /// and return the operands for the horizontal operation in LHS and RHS. A
17080 /// horizontal operation performs the binary operation on successive elements
17081 /// of its first operand, then on successive elements of its second operand,
17082 /// returning the resulting values in a vector. For example, if
17083 /// A = < float a0, float a1, float a2, float a3 >
17085 /// B = < float b0, float b1, float b2, float b3 >
17086 /// then the result of doing a horizontal operation on A and B is
17087 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
17088 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
17089 /// A horizontal-op B, for some already available A and B, and if so then LHS is
17090 /// set to A, RHS to B, and the routine returns 'true'.
17091 /// Note that the binary operation should have the property that if one of the
17092 /// operands is UNDEF then the result is UNDEF.
17093 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
17094 // Look for the following pattern: if
17095 // A = < float a0, float a1, float a2, float a3 >
17096 // B = < float b0, float b1, float b2, float b3 >
17098 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
17099 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
17100 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
17101 // which is A horizontal-op B.
17103 // At least one of the operands should be a vector shuffle.
17104 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
17105 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
17108 EVT VT = LHS.getValueType();
17110 assert((VT.is128BitVector() || VT.is256BitVector()) &&
17111 "Unsupported vector type for horizontal add/sub");
17113 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
17114 // operate independently on 128-bit lanes.
17115 unsigned NumElts = VT.getVectorNumElements();
17116 unsigned NumLanes = VT.getSizeInBits()/128;
17117 unsigned NumLaneElts = NumElts / NumLanes;
17118 assert((NumLaneElts % 2 == 0) &&
17119 "Vector type should have an even number of elements in each lane");
17120 unsigned HalfLaneElts = NumLaneElts/2;
17122 // View LHS in the form
17123 // LHS = VECTOR_SHUFFLE A, B, LMask
17124 // If LHS is not a shuffle then pretend it is the shuffle
17125 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
17126 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
17129 SmallVector<int, 16> LMask(NumElts);
17130 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
17131 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
17132 A = LHS.getOperand(0);
17133 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
17134 B = LHS.getOperand(1);
17135 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
17136 std::copy(Mask.begin(), Mask.end(), LMask.begin());
17138 if (LHS.getOpcode() != ISD::UNDEF)
17140 for (unsigned i = 0; i != NumElts; ++i)
17144 // Likewise, view RHS in the form
17145 // RHS = VECTOR_SHUFFLE C, D, RMask
17147 SmallVector<int, 16> RMask(NumElts);
17148 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
17149 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
17150 C = RHS.getOperand(0);
17151 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
17152 D = RHS.getOperand(1);
17153 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
17154 std::copy(Mask.begin(), Mask.end(), RMask.begin());
17156 if (RHS.getOpcode() != ISD::UNDEF)
17158 for (unsigned i = 0; i != NumElts; ++i)
17162 // Check that the shuffles are both shuffling the same vectors.
17163 if (!(A == C && B == D) && !(A == D && B == C))
17166 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
17167 if (!A.getNode() && !B.getNode())
17170 // If A and B occur in reverse order in RHS, then "swap" them (which means
17171 // rewriting the mask).
17173 CommuteVectorShuffleMask(RMask, NumElts);
17175 // At this point LHS and RHS are equivalent to
17176 // LHS = VECTOR_SHUFFLE A, B, LMask
17177 // RHS = VECTOR_SHUFFLE A, B, RMask
17178 // Check that the masks correspond to performing a horizontal operation.
17179 for (unsigned i = 0; i != NumElts; ++i) {
17180 int LIdx = LMask[i], RIdx = RMask[i];
17182 // Ignore any UNDEF components.
17183 if (LIdx < 0 || RIdx < 0 ||
17184 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
17185 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
17188 // Check that successive elements are being operated on. If not, this is
17189 // not a horizontal operation.
17190 unsigned Src = (i/HalfLaneElts) % 2; // each lane is split between srcs
17191 unsigned LaneStart = (i/NumLaneElts) * NumLaneElts;
17192 int Index = 2*(i%HalfLaneElts) + NumElts*Src + LaneStart;
17193 if (!(LIdx == Index && RIdx == Index + 1) &&
17194 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
17198 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
17199 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
17203 /// PerformFADDCombine - Do target-specific dag combines on floating point adds.
17204 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
17205 const X86Subtarget *Subtarget) {
17206 EVT VT = N->getValueType(0);
17207 SDValue LHS = N->getOperand(0);
17208 SDValue RHS = N->getOperand(1);
17210 // Try to synthesize horizontal adds from adds of shuffles.
17211 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
17212 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
17213 isHorizontalBinOp(LHS, RHS, true))
17214 return DAG.getNode(X86ISD::FHADD, N->getDebugLoc(), VT, LHS, RHS);
17218 /// PerformFSUBCombine - Do target-specific dag combines on floating point subs.
17219 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
17220 const X86Subtarget *Subtarget) {
17221 EVT VT = N->getValueType(0);
17222 SDValue LHS = N->getOperand(0);
17223 SDValue RHS = N->getOperand(1);
17225 // Try to synthesize horizontal subs from subs of shuffles.
17226 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
17227 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
17228 isHorizontalBinOp(LHS, RHS, false))
17229 return DAG.getNode(X86ISD::FHSUB, N->getDebugLoc(), VT, LHS, RHS);
17233 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
17234 /// X86ISD::FXOR nodes.
17235 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
17236 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
17237 // F[X]OR(0.0, x) -> x
17238 // F[X]OR(x, 0.0) -> x
17239 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
17240 if (C->getValueAPF().isPosZero())
17241 return N->getOperand(1);
17242 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
17243 if (C->getValueAPF().isPosZero())
17244 return N->getOperand(0);
17248 /// PerformFMinFMaxCombine - Do target-specific dag combines on X86ISD::FMIN and
17249 /// X86ISD::FMAX nodes.
17250 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
17251 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
17253 // Only perform optimizations if UnsafeMath is used.
17254 if (!DAG.getTarget().Options.UnsafeFPMath)
17257 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
17258 // into FMINC and FMAXC, which are Commutative operations.
17259 unsigned NewOp = 0;
17260 switch (N->getOpcode()) {
17261 default: llvm_unreachable("unknown opcode");
17262 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
17263 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
17266 return DAG.getNode(NewOp, N->getDebugLoc(), N->getValueType(0),
17267 N->getOperand(0), N->getOperand(1));
17270 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
17271 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
17272 // FAND(0.0, x) -> 0.0
17273 // FAND(x, 0.0) -> 0.0
17274 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
17275 if (C->getValueAPF().isPosZero())
17276 return N->getOperand(0);
17277 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
17278 if (C->getValueAPF().isPosZero())
17279 return N->getOperand(1);
17283 static SDValue PerformBTCombine(SDNode *N,
17285 TargetLowering::DAGCombinerInfo &DCI) {
17286 // BT ignores high bits in the bit index operand.
17287 SDValue Op1 = N->getOperand(1);
17288 if (Op1.hasOneUse()) {
17289 unsigned BitWidth = Op1.getValueSizeInBits();
17290 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
17291 APInt KnownZero, KnownOne;
17292 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
17293 !DCI.isBeforeLegalizeOps());
17294 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17295 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
17296 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
17297 DCI.CommitTargetLoweringOpt(TLO);
17302 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
17303 SDValue Op = N->getOperand(0);
17304 if (Op.getOpcode() == ISD::BITCAST)
17305 Op = Op.getOperand(0);
17306 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
17307 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
17308 VT.getVectorElementType().getSizeInBits() ==
17309 OpVT.getVectorElementType().getSizeInBits()) {
17310 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
17315 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
17316 const X86Subtarget *Subtarget) {
17317 EVT VT = N->getValueType(0);
17318 if (!VT.isVector())
17321 SDValue N0 = N->getOperand(0);
17322 SDValue N1 = N->getOperand(1);
17323 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
17324 DebugLoc dl = N->getDebugLoc();
17326 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
17327 // both SSE and AVX2 since there is no sign-extended shift right
17328 // operation on a vector with 64-bit elements.
17329 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
17330 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
17331 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
17332 N0.getOpcode() == ISD::SIGN_EXTEND)) {
17333 SDValue N00 = N0.getOperand(0);
17335 // EXTLOAD has a better solution on AVX2,
17336 // it may be replaced with X86ISD::VSEXT node.
17337 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
17338 if (!ISD::isNormalLoad(N00.getNode()))
17341 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
17342 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
17344 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
17350 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
17351 TargetLowering::DAGCombinerInfo &DCI,
17352 const X86Subtarget *Subtarget) {
17353 if (!DCI.isBeforeLegalizeOps())
17356 if (!Subtarget->hasFp256())
17359 EVT VT = N->getValueType(0);
17360 if (VT.isVector() && VT.getSizeInBits() == 256) {
17361 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
17369 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
17370 const X86Subtarget* Subtarget) {
17371 DebugLoc dl = N->getDebugLoc();
17372 EVT VT = N->getValueType(0);
17374 // Let legalize expand this if it isn't a legal type yet.
17375 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
17378 EVT ScalarVT = VT.getScalarType();
17379 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
17380 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
17383 SDValue A = N->getOperand(0);
17384 SDValue B = N->getOperand(1);
17385 SDValue C = N->getOperand(2);
17387 bool NegA = (A.getOpcode() == ISD::FNEG);
17388 bool NegB = (B.getOpcode() == ISD::FNEG);
17389 bool NegC = (C.getOpcode() == ISD::FNEG);
17391 // Negative multiplication when NegA xor NegB
17392 bool NegMul = (NegA != NegB);
17394 A = A.getOperand(0);
17396 B = B.getOperand(0);
17398 C = C.getOperand(0);
17402 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
17404 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
17406 return DAG.getNode(Opcode, dl, VT, A, B, C);
17409 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
17410 TargetLowering::DAGCombinerInfo &DCI,
17411 const X86Subtarget *Subtarget) {
17412 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
17413 // (and (i32 x86isd::setcc_carry), 1)
17414 // This eliminates the zext. This transformation is necessary because
17415 // ISD::SETCC is always legalized to i8.
17416 DebugLoc dl = N->getDebugLoc();
17417 SDValue N0 = N->getOperand(0);
17418 EVT VT = N->getValueType(0);
17420 if (N0.getOpcode() == ISD::AND &&
17422 N0.getOperand(0).hasOneUse()) {
17423 SDValue N00 = N0.getOperand(0);
17424 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
17425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
17426 if (!C || C->getZExtValue() != 1)
17428 return DAG.getNode(ISD::AND, dl, VT,
17429 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
17430 N00.getOperand(0), N00.getOperand(1)),
17431 DAG.getConstant(1, VT));
17435 if (VT.is256BitVector()) {
17436 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
17444 // Optimize x == -y --> x+y == 0
17445 // x != -y --> x+y != 0
17446 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
17447 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
17448 SDValue LHS = N->getOperand(0);
17449 SDValue RHS = N->getOperand(1);
17451 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
17452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
17453 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
17454 SDValue addV = DAG.getNode(ISD::ADD, N->getDebugLoc(),
17455 LHS.getValueType(), RHS, LHS.getOperand(1));
17456 return DAG.getSetCC(N->getDebugLoc(), N->getValueType(0),
17457 addV, DAG.getConstant(0, addV.getValueType()), CC);
17459 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
17460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
17461 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
17462 SDValue addV = DAG.getNode(ISD::ADD, N->getDebugLoc(),
17463 RHS.getValueType(), LHS, RHS.getOperand(1));
17464 return DAG.getSetCC(N->getDebugLoc(), N->getValueType(0),
17465 addV, DAG.getConstant(0, addV.getValueType()), CC);
17470 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
17471 // as "sbb reg,reg", since it can be extended without zext and produces
17472 // an all-ones bit which is more useful than 0/1 in some cases.
17473 static SDValue MaterializeSETB(DebugLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
17474 return DAG.getNode(ISD::AND, DL, MVT::i8,
17475 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
17476 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
17477 DAG.getConstant(1, MVT::i8));
17480 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
17481 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
17482 TargetLowering::DAGCombinerInfo &DCI,
17483 const X86Subtarget *Subtarget) {
17484 DebugLoc DL = N->getDebugLoc();
17485 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
17486 SDValue EFLAGS = N->getOperand(1);
17488 if (CC == X86::COND_A) {
17489 // Try to convert COND_A into COND_B in an attempt to facilitate
17490 // materializing "setb reg".
17492 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
17493 // cannot take an immediate as its first operand.
17495 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
17496 EFLAGS.getValueType().isInteger() &&
17497 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
17498 SDValue NewSub = DAG.getNode(X86ISD::SUB, EFLAGS.getDebugLoc(),
17499 EFLAGS.getNode()->getVTList(),
17500 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
17501 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
17502 return MaterializeSETB(DL, NewEFLAGS, DAG);
17506 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
17507 // a zext and produces an all-ones bit which is more useful than 0/1 in some
17509 if (CC == X86::COND_B)
17510 return MaterializeSETB(DL, EFLAGS, DAG);
17514 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
17515 if (Flags.getNode()) {
17516 SDValue Cond = DAG.getConstant(CC, MVT::i8);
17517 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
17523 // Optimize branch condition evaluation.
17525 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
17526 TargetLowering::DAGCombinerInfo &DCI,
17527 const X86Subtarget *Subtarget) {
17528 DebugLoc DL = N->getDebugLoc();
17529 SDValue Chain = N->getOperand(0);
17530 SDValue Dest = N->getOperand(1);
17531 SDValue EFLAGS = N->getOperand(3);
17532 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
17536 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
17537 if (Flags.getNode()) {
17538 SDValue Cond = DAG.getConstant(CC, MVT::i8);
17539 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
17546 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
17547 const X86TargetLowering *XTLI) {
17548 SDValue Op0 = N->getOperand(0);
17549 EVT InVT = Op0->getValueType(0);
17551 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
17552 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
17553 DebugLoc dl = N->getDebugLoc();
17554 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
17555 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
17556 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
17559 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
17560 // a 32-bit target where SSE doesn't support i64->FP operations.
17561 if (Op0.getOpcode() == ISD::LOAD) {
17562 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
17563 EVT VT = Ld->getValueType(0);
17564 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
17565 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
17566 !XTLI->getSubtarget()->is64Bit() &&
17567 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
17568 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
17569 Ld->getChain(), Op0, DAG);
17570 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
17577 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
17578 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
17579 X86TargetLowering::DAGCombinerInfo &DCI) {
17580 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
17581 // the result is either zero or one (depending on the input carry bit).
17582 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
17583 if (X86::isZeroNode(N->getOperand(0)) &&
17584 X86::isZeroNode(N->getOperand(1)) &&
17585 // We don't have a good way to replace an EFLAGS use, so only do this when
17587 SDValue(N, 1).use_empty()) {
17588 DebugLoc DL = N->getDebugLoc();
17589 EVT VT = N->getValueType(0);
17590 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
17591 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
17592 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
17593 DAG.getConstant(X86::COND_B,MVT::i8),
17595 DAG.getConstant(1, VT));
17596 return DCI.CombineTo(N, Res1, CarryOut);
17602 // fold (add Y, (sete X, 0)) -> adc 0, Y
17603 // (add Y, (setne X, 0)) -> sbb -1, Y
17604 // (sub (sete X, 0), Y) -> sbb 0, Y
17605 // (sub (setne X, 0), Y) -> adc -1, Y
17606 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
17607 DebugLoc DL = N->getDebugLoc();
17609 // Look through ZExts.
17610 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
17611 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
17614 SDValue SetCC = Ext.getOperand(0);
17615 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
17618 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
17619 if (CC != X86::COND_E && CC != X86::COND_NE)
17622 SDValue Cmp = SetCC.getOperand(1);
17623 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
17624 !X86::isZeroNode(Cmp.getOperand(1)) ||
17625 !Cmp.getOperand(0).getValueType().isInteger())
17628 SDValue CmpOp0 = Cmp.getOperand(0);
17629 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
17630 DAG.getConstant(1, CmpOp0.getValueType()));
17632 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
17633 if (CC == X86::COND_NE)
17634 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
17635 DL, OtherVal.getValueType(), OtherVal,
17636 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
17637 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
17638 DL, OtherVal.getValueType(), OtherVal,
17639 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
17642 /// PerformADDCombine - Do target-specific dag combines on integer adds.
17643 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
17644 const X86Subtarget *Subtarget) {
17645 EVT VT = N->getValueType(0);
17646 SDValue Op0 = N->getOperand(0);
17647 SDValue Op1 = N->getOperand(1);
17649 // Try to synthesize horizontal adds from adds of shuffles.
17650 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
17651 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
17652 isHorizontalBinOp(Op0, Op1, true))
17653 return DAG.getNode(X86ISD::HADD, N->getDebugLoc(), VT, Op0, Op1);
17655 return OptimizeConditionalInDecrement(N, DAG);
17658 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
17659 const X86Subtarget *Subtarget) {
17660 SDValue Op0 = N->getOperand(0);
17661 SDValue Op1 = N->getOperand(1);
17663 // X86 can't encode an immediate LHS of a sub. See if we can push the
17664 // negation into a preceding instruction.
17665 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
17666 // If the RHS of the sub is a XOR with one use and a constant, invert the
17667 // immediate. Then add one to the LHS of the sub so we can turn
17668 // X-Y -> X+~Y+1, saving one register.
17669 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
17670 isa<ConstantSDNode>(Op1.getOperand(1))) {
17671 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
17672 EVT VT = Op0.getValueType();
17673 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
17675 DAG.getConstant(~XorC, VT));
17676 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
17677 DAG.getConstant(C->getAPIntValue()+1, VT));
17681 // Try to synthesize horizontal adds from adds of shuffles.
17682 EVT VT = N->getValueType(0);
17683 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
17684 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
17685 isHorizontalBinOp(Op0, Op1, true))
17686 return DAG.getNode(X86ISD::HSUB, N->getDebugLoc(), VT, Op0, Op1);
17688 return OptimizeConditionalInDecrement(N, DAG);
17691 /// performVZEXTCombine - Performs build vector combines
17692 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
17693 TargetLowering::DAGCombinerInfo &DCI,
17694 const X86Subtarget *Subtarget) {
17695 // (vzext (bitcast (vzext (x)) -> (vzext x)
17696 SDValue In = N->getOperand(0);
17697 while (In.getOpcode() == ISD::BITCAST)
17698 In = In.getOperand(0);
17700 if (In.getOpcode() != X86ISD::VZEXT)
17703 return DAG.getNode(X86ISD::VZEXT, N->getDebugLoc(), N->getValueType(0),
17707 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
17708 DAGCombinerInfo &DCI) const {
17709 SelectionDAG &DAG = DCI.DAG;
17710 switch (N->getOpcode()) {
17712 case ISD::EXTRACT_VECTOR_ELT:
17713 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
17715 case ISD::SELECT: return PerformSELECTCombine(N, DAG, DCI, Subtarget);
17716 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
17717 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
17718 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
17719 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
17720 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
17723 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
17724 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
17725 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
17726 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
17727 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
17728 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
17729 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
17730 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
17731 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
17733 case X86ISD::FOR: return PerformFORCombine(N, DAG);
17735 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
17736 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
17737 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
17738 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
17739 case ISD::ANY_EXTEND:
17740 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
17741 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
17742 case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
17743 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
17744 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
17745 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
17746 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
17747 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
17748 case X86ISD::SHUFP: // Handle all target specific shuffles
17749 case X86ISD::PALIGNR:
17750 case X86ISD::UNPCKH:
17751 case X86ISD::UNPCKL:
17752 case X86ISD::MOVHLPS:
17753 case X86ISD::MOVLHPS:
17754 case X86ISD::PSHUFD:
17755 case X86ISD::PSHUFHW:
17756 case X86ISD::PSHUFLW:
17757 case X86ISD::MOVSS:
17758 case X86ISD::MOVSD:
17759 case X86ISD::VPERMILP:
17760 case X86ISD::VPERM2X128:
17761 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
17762 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
17768 /// isTypeDesirableForOp - Return true if the target has native support for
17769 /// the specified value type and it is 'desirable' to use the type for the
17770 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
17771 /// instruction encodings are longer and some i16 instructions are slow.
17772 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
17773 if (!isTypeLegal(VT))
17775 if (VT != MVT::i16)
17782 case ISD::SIGN_EXTEND:
17783 case ISD::ZERO_EXTEND:
17784 case ISD::ANY_EXTEND:
17797 /// IsDesirableToPromoteOp - This method query the target whether it is
17798 /// beneficial for dag combiner to promote the specified node. If true, it
17799 /// should return the desired promotion type by reference.
17800 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
17801 EVT VT = Op.getValueType();
17802 if (VT != MVT::i16)
17805 bool Promote = false;
17806 bool Commute = false;
17807 switch (Op.getOpcode()) {
17810 LoadSDNode *LD = cast<LoadSDNode>(Op);
17811 // If the non-extending load has a single use and it's not live out, then it
17812 // might be folded.
17813 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
17814 Op.hasOneUse()*/) {
17815 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
17816 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
17817 // The only case where we'd want to promote LOAD (rather then it being
17818 // promoted as an operand is when it's only use is liveout.
17819 if (UI->getOpcode() != ISD::CopyToReg)
17826 case ISD::SIGN_EXTEND:
17827 case ISD::ZERO_EXTEND:
17828 case ISD::ANY_EXTEND:
17833 SDValue N0 = Op.getOperand(0);
17834 // Look out for (store (shl (load), x)).
17835 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
17848 SDValue N0 = Op.getOperand(0);
17849 SDValue N1 = Op.getOperand(1);
17850 if (!Commute && MayFoldLoad(N1))
17852 // Avoid disabling potential load folding opportunities.
17853 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
17855 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
17865 //===----------------------------------------------------------------------===//
17866 // X86 Inline Assembly Support
17867 //===----------------------------------------------------------------------===//
17870 // Helper to match a string separated by whitespace.
17871 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
17872 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
17874 for (unsigned i = 0, e = args.size(); i != e; ++i) {
17875 StringRef piece(*args[i]);
17876 if (!s.startswith(piece)) // Check if the piece matches.
17879 s = s.substr(piece.size());
17880 StringRef::size_type pos = s.find_first_not_of(" \t");
17881 if (pos == 0) // We matched a prefix.
17889 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
17892 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
17893 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
17895 std::string AsmStr = IA->getAsmString();
17897 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
17898 if (!Ty || Ty->getBitWidth() % 16 != 0)
17901 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
17902 SmallVector<StringRef, 4> AsmPieces;
17903 SplitString(AsmStr, AsmPieces, ";\n");
17905 switch (AsmPieces.size()) {
17906 default: return false;
17908 // FIXME: this should verify that we are targeting a 486 or better. If not,
17909 // we will turn this bswap into something that will be lowered to logical
17910 // ops instead of emitting the bswap asm. For now, we don't support 486 or
17911 // lower so don't worry about this.
17913 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
17914 matchAsm(AsmPieces[0], "bswapl", "$0") ||
17915 matchAsm(AsmPieces[0], "bswapq", "$0") ||
17916 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
17917 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
17918 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
17919 // No need to check constraints, nothing other than the equivalent of
17920 // "=r,0" would be valid here.
17921 return IntrinsicLowering::LowerToByteSwap(CI);
17924 // rorw $$8, ${0:w} --> llvm.bswap.i16
17925 if (CI->getType()->isIntegerTy(16) &&
17926 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
17927 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
17928 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
17930 const std::string &ConstraintsStr = IA->getConstraintString();
17931 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
17932 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
17933 if (AsmPieces.size() == 4 &&
17934 AsmPieces[0] == "~{cc}" &&
17935 AsmPieces[1] == "~{dirflag}" &&
17936 AsmPieces[2] == "~{flags}" &&
17937 AsmPieces[3] == "~{fpsr}")
17938 return IntrinsicLowering::LowerToByteSwap(CI);
17942 if (CI->getType()->isIntegerTy(32) &&
17943 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
17944 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
17945 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
17946 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
17948 const std::string &ConstraintsStr = IA->getConstraintString();
17949 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
17950 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
17951 if (AsmPieces.size() == 4 &&
17952 AsmPieces[0] == "~{cc}" &&
17953 AsmPieces[1] == "~{dirflag}" &&
17954 AsmPieces[2] == "~{flags}" &&
17955 AsmPieces[3] == "~{fpsr}")
17956 return IntrinsicLowering::LowerToByteSwap(CI);
17959 if (CI->getType()->isIntegerTy(64)) {
17960 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
17961 if (Constraints.size() >= 2 &&
17962 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
17963 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
17964 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
17965 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
17966 matchAsm(AsmPieces[1], "bswap", "%edx") &&
17967 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
17968 return IntrinsicLowering::LowerToByteSwap(CI);
17976 /// getConstraintType - Given a constraint letter, return the type of
17977 /// constraint it is for this target.
17978 X86TargetLowering::ConstraintType
17979 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
17980 if (Constraint.size() == 1) {
17981 switch (Constraint[0]) {
17992 return C_RegisterClass;
18016 return TargetLowering::getConstraintType(Constraint);
18019 /// Examine constraint type and operand type and determine a weight value.
18020 /// This object must already have been set up with the operand type
18021 /// and the current alternative constraint selected.
18022 TargetLowering::ConstraintWeight
18023 X86TargetLowering::getSingleConstraintMatchWeight(
18024 AsmOperandInfo &info, const char *constraint) const {
18025 ConstraintWeight weight = CW_Invalid;
18026 Value *CallOperandVal = info.CallOperandVal;
18027 // If we don't have a value, we can't do a match,
18028 // but allow it at the lowest weight.
18029 if (CallOperandVal == NULL)
18031 Type *type = CallOperandVal->getType();
18032 // Look at the constraint type.
18033 switch (*constraint) {
18035 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
18046 if (CallOperandVal->getType()->isIntegerTy())
18047 weight = CW_SpecificReg;
18052 if (type->isFloatingPointTy())
18053 weight = CW_SpecificReg;
18056 if (type->isX86_MMXTy() && Subtarget->hasMMX())
18057 weight = CW_SpecificReg;
18061 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
18062 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
18063 weight = CW_Register;
18066 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
18067 if (C->getZExtValue() <= 31)
18068 weight = CW_Constant;
18072 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18073 if (C->getZExtValue() <= 63)
18074 weight = CW_Constant;
18078 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18079 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
18080 weight = CW_Constant;
18084 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18085 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
18086 weight = CW_Constant;
18090 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18091 if (C->getZExtValue() <= 3)
18092 weight = CW_Constant;
18096 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18097 if (C->getZExtValue() <= 0xff)
18098 weight = CW_Constant;
18103 if (dyn_cast<ConstantFP>(CallOperandVal)) {
18104 weight = CW_Constant;
18108 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18109 if ((C->getSExtValue() >= -0x80000000LL) &&
18110 (C->getSExtValue() <= 0x7fffffffLL))
18111 weight = CW_Constant;
18115 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
18116 if (C->getZExtValue() <= 0xffffffff)
18117 weight = CW_Constant;
18124 /// LowerXConstraint - try to replace an X constraint, which matches anything,
18125 /// with another that has more specific requirements based on the type of the
18126 /// corresponding operand.
18127 const char *X86TargetLowering::
18128 LowerXConstraint(EVT ConstraintVT) const {
18129 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
18130 // 'f' like normal targets.
18131 if (ConstraintVT.isFloatingPoint()) {
18132 if (Subtarget->hasSSE2())
18134 if (Subtarget->hasSSE1())
18138 return TargetLowering::LowerXConstraint(ConstraintVT);
18141 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
18142 /// vector. If it is invalid, don't add anything to Ops.
18143 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
18144 std::string &Constraint,
18145 std::vector<SDValue>&Ops,
18146 SelectionDAG &DAG) const {
18147 SDValue Result(0, 0);
18149 // Only support length 1 constraints for now.
18150 if (Constraint.length() > 1) return;
18152 char ConstraintLetter = Constraint[0];
18153 switch (ConstraintLetter) {
18156 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18157 if (C->getZExtValue() <= 31) {
18158 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18164 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18165 if (C->getZExtValue() <= 63) {
18166 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18172 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18173 if (isInt<8>(C->getSExtValue())) {
18174 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18180 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18181 if (C->getZExtValue() <= 255) {
18182 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18188 // 32-bit signed value
18189 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18190 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
18191 C->getSExtValue())) {
18192 // Widen to 64 bits here to get it sign extended.
18193 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
18196 // FIXME gcc accepts some relocatable values here too, but only in certain
18197 // memory models; it's complicated.
18202 // 32-bit unsigned value
18203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
18204 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
18205 C->getZExtValue())) {
18206 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
18210 // FIXME gcc accepts some relocatable values here too, but only in certain
18211 // memory models; it's complicated.
18215 // Literal immediates are always ok.
18216 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
18217 // Widen to 64 bits here to get it sign extended.
18218 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
18222 // In any sort of PIC mode addresses need to be computed at runtime by
18223 // adding in a register or some sort of table lookup. These can't
18224 // be used as immediates.
18225 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
18228 // If we are in non-pic codegen mode, we allow the address of a global (with
18229 // an optional displacement) to be used with 'i'.
18230 GlobalAddressSDNode *GA = 0;
18231 int64_t Offset = 0;
18233 // Match either (GA), (GA+C), (GA+C1+C2), etc.
18235 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
18236 Offset += GA->getOffset();
18238 } else if (Op.getOpcode() == ISD::ADD) {
18239 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
18240 Offset += C->getZExtValue();
18241 Op = Op.getOperand(0);
18244 } else if (Op.getOpcode() == ISD::SUB) {
18245 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
18246 Offset += -C->getZExtValue();
18247 Op = Op.getOperand(0);
18252 // Otherwise, this isn't something we can handle, reject it.
18256 const GlobalValue *GV = GA->getGlobal();
18257 // If we require an extra load to get this address, as in PIC mode, we
18258 // can't accept it.
18259 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
18260 getTargetMachine())))
18263 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
18264 GA->getValueType(0), Offset);
18269 if (Result.getNode()) {
18270 Ops.push_back(Result);
18273 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
18276 std::pair<unsigned, const TargetRegisterClass*>
18277 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
18279 // First, see if this is a constraint that directly corresponds to an LLVM
18281 if (Constraint.size() == 1) {
18282 // GCC Constraint Letters
18283 switch (Constraint[0]) {
18285 // TODO: Slight differences here in allocation order and leaving
18286 // RIP in the class. Do they matter any more here than they do
18287 // in the normal allocation?
18288 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
18289 if (Subtarget->is64Bit()) {
18290 if (VT == MVT::i32 || VT == MVT::f32)
18291 return std::make_pair(0U, &X86::GR32RegClass);
18292 if (VT == MVT::i16)
18293 return std::make_pair(0U, &X86::GR16RegClass);
18294 if (VT == MVT::i8 || VT == MVT::i1)
18295 return std::make_pair(0U, &X86::GR8RegClass);
18296 if (VT == MVT::i64 || VT == MVT::f64)
18297 return std::make_pair(0U, &X86::GR64RegClass);
18300 // 32-bit fallthrough
18301 case 'Q': // Q_REGS
18302 if (VT == MVT::i32 || VT == MVT::f32)
18303 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
18304 if (VT == MVT::i16)
18305 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
18306 if (VT == MVT::i8 || VT == MVT::i1)
18307 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
18308 if (VT == MVT::i64)
18309 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
18311 case 'r': // GENERAL_REGS
18312 case 'l': // INDEX_REGS
18313 if (VT == MVT::i8 || VT == MVT::i1)
18314 return std::make_pair(0U, &X86::GR8RegClass);
18315 if (VT == MVT::i16)
18316 return std::make_pair(0U, &X86::GR16RegClass);
18317 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
18318 return std::make_pair(0U, &X86::GR32RegClass);
18319 return std::make_pair(0U, &X86::GR64RegClass);
18320 case 'R': // LEGACY_REGS
18321 if (VT == MVT::i8 || VT == MVT::i1)
18322 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
18323 if (VT == MVT::i16)
18324 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
18325 if (VT == MVT::i32 || !Subtarget->is64Bit())
18326 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
18327 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
18328 case 'f': // FP Stack registers.
18329 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
18330 // value to the correct fpstack register class.
18331 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
18332 return std::make_pair(0U, &X86::RFP32RegClass);
18333 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
18334 return std::make_pair(0U, &X86::RFP64RegClass);
18335 return std::make_pair(0U, &X86::RFP80RegClass);
18336 case 'y': // MMX_REGS if MMX allowed.
18337 if (!Subtarget->hasMMX()) break;
18338 return std::make_pair(0U, &X86::VR64RegClass);
18339 case 'Y': // SSE_REGS if SSE2 allowed
18340 if (!Subtarget->hasSSE2()) break;
18342 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
18343 if (!Subtarget->hasSSE1()) break;
18345 switch (VT.getSimpleVT().SimpleTy) {
18347 // Scalar SSE types.
18350 return std::make_pair(0U, &X86::FR32RegClass);
18353 return std::make_pair(0U, &X86::FR64RegClass);
18361 return std::make_pair(0U, &X86::VR128RegClass);
18369 return std::make_pair(0U, &X86::VR256RegClass);
18375 // Use the default implementation in TargetLowering to convert the register
18376 // constraint into a member of a register class.
18377 std::pair<unsigned, const TargetRegisterClass*> Res;
18378 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
18380 // Not found as a standard register?
18381 if (Res.second == 0) {
18382 // Map st(0) -> st(7) -> ST0
18383 if (Constraint.size() == 7 && Constraint[0] == '{' &&
18384 tolower(Constraint[1]) == 's' &&
18385 tolower(Constraint[2]) == 't' &&
18386 Constraint[3] == '(' &&
18387 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
18388 Constraint[5] == ')' &&
18389 Constraint[6] == '}') {
18391 Res.first = X86::ST0+Constraint[4]-'0';
18392 Res.second = &X86::RFP80RegClass;
18396 // GCC allows "st(0)" to be called just plain "st".
18397 if (StringRef("{st}").equals_lower(Constraint)) {
18398 Res.first = X86::ST0;
18399 Res.second = &X86::RFP80RegClass;
18404 if (StringRef("{flags}").equals_lower(Constraint)) {
18405 Res.first = X86::EFLAGS;
18406 Res.second = &X86::CCRRegClass;
18410 // 'A' means EAX + EDX.
18411 if (Constraint == "A") {
18412 Res.first = X86::EAX;
18413 Res.second = &X86::GR32_ADRegClass;
18419 // Otherwise, check to see if this is a register class of the wrong value
18420 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
18421 // turn into {ax},{dx}.
18422 if (Res.second->hasType(VT))
18423 return Res; // Correct type already, nothing to do.
18425 // All of the single-register GCC register classes map their values onto
18426 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
18427 // really want an 8-bit or 32-bit register, map to the appropriate register
18428 // class and return the appropriate register.
18429 if (Res.second == &X86::GR16RegClass) {
18430 if (VT == MVT::i8 || VT == MVT::i1) {
18431 unsigned DestReg = 0;
18432 switch (Res.first) {
18434 case X86::AX: DestReg = X86::AL; break;
18435 case X86::DX: DestReg = X86::DL; break;
18436 case X86::CX: DestReg = X86::CL; break;
18437 case X86::BX: DestReg = X86::BL; break;
18440 Res.first = DestReg;
18441 Res.second = &X86::GR8RegClass;
18443 } else if (VT == MVT::i32 || VT == MVT::f32) {
18444 unsigned DestReg = 0;
18445 switch (Res.first) {
18447 case X86::AX: DestReg = X86::EAX; break;
18448 case X86::DX: DestReg = X86::EDX; break;
18449 case X86::CX: DestReg = X86::ECX; break;
18450 case X86::BX: DestReg = X86::EBX; break;
18451 case X86::SI: DestReg = X86::ESI; break;
18452 case X86::DI: DestReg = X86::EDI; break;
18453 case X86::BP: DestReg = X86::EBP; break;
18454 case X86::SP: DestReg = X86::ESP; break;
18457 Res.first = DestReg;
18458 Res.second = &X86::GR32RegClass;
18460 } else if (VT == MVT::i64 || VT == MVT::f64) {
18461 unsigned DestReg = 0;
18462 switch (Res.first) {
18464 case X86::AX: DestReg = X86::RAX; break;
18465 case X86::DX: DestReg = X86::RDX; break;
18466 case X86::CX: DestReg = X86::RCX; break;
18467 case X86::BX: DestReg = X86::RBX; break;
18468 case X86::SI: DestReg = X86::RSI; break;
18469 case X86::DI: DestReg = X86::RDI; break;
18470 case X86::BP: DestReg = X86::RBP; break;
18471 case X86::SP: DestReg = X86::RSP; break;
18474 Res.first = DestReg;
18475 Res.second = &X86::GR64RegClass;
18478 } else if (Res.second == &X86::FR32RegClass ||
18479 Res.second == &X86::FR64RegClass ||
18480 Res.second == &X86::VR128RegClass) {
18481 // Handle references to XMM physical registers that got mapped into the
18482 // wrong class. This can happen with constraints like {xmm0} where the
18483 // target independent register mapper will just pick the first match it can
18484 // find, ignoring the required type.
18486 if (VT == MVT::f32 || VT == MVT::i32)
18487 Res.second = &X86::FR32RegClass;
18488 else if (VT == MVT::f64 || VT == MVT::i64)
18489 Res.second = &X86::FR64RegClass;
18490 else if (X86::VR128RegClass.hasType(VT))
18491 Res.second = &X86::VR128RegClass;
18492 else if (X86::VR256RegClass.hasType(VT))
18493 Res.second = &X86::VR256RegClass;